Hey. Oh, let's go.
Ladies and gentlemen, please welcome Lisa Spelman.
Let's go. Hey, oh, let's go. The moment is better.
All right. Thank you. Welcome, everyone. It is great to see you here, and thank you so much for coming and joining us. It's always fun when we get to bring our nearest and dearest into the halls of Intel and give them a little bit of a view of what we're working on and what we're doing. I think today is a unique opportunity for us and hopefully for you, too, to have an opportunity to sit together, share some time, and really dive in a little bit deeper into the strategies and the business and what we're driving across the innovation spectrum in the data-centric era. We're going to have a lot of our leaders across the business unit share with you their thoughts and views on all of what our markets are doing and how they're being shaped and what the upcoming trends are.
We're going to have a chance to show you how we're servicing our customers across the globe. We're going to start off the day, I'm showing up the morning schedule for you, with an overview of the entire business from Navin Shenoy, then we're going to follow on again, hearing from leaders across enterprise and cloud and our comm service providers around what's going from core data centers through the network out to the edge across cloud, just the transformation that's happening everywhere. In the afternoon, we're going to come back together, we'll have two separate tracks running. We're going to first focus in a little bit more on the silicon foundation that's driving data-centric computing, then we're going to cut across the rest of the product portfolio.
You'll get updates on our connectivity business, on our FPGA business, see how these are all coming together into a solutions profile. For our technical press, we have a tech track session that will also be running concurrently, that'll be a good chance for those in the audience to really dig a little bit deeper in on some of the themes and trends that we've talked about throughout the rest of the day. We'll close with a reception at the end. I want to just thank you again for joining us, without further ado, welcome Navin Shenoy to the stage to get us kicked off. There you go.
Okay. Thank you. Good morning, welcome to the Data-Centric Innovation Summit. At Intel, we have long been the foundation for the world's most important innovations and advancements, we've always been about two primary things. We've been about innovation, we've been about scale. As we make this shift to a data-centric company, we see even more opportunity ahead. Together with my colleagues that Lisa mentioned this morning, we will lay out our strategy and plans for the future. We'll cover a wide range of topics, from AI to the enterprise to the network to 5G to the cloud. We'll cover a broad set of products from our microprocessor portfolio to FPGAs to ASICs to connectivity to Optane persistent memory. I'll give you an update on our roadmap, I'll give you an update on how we continue to win. A lot of ground to cover.
Let's get started. It won't be any surprise to you that we believe that data is defining the future of our industry and defining the future of Intel. Five years ago, roughly one-third of our revenue was data-centric. Today, nearly half of our revenue is data-centric and growing at double digits. The data-centric trend is well documented. Many have been writing about the data deluge or how data is the new oil or how data is driving power. Some are even writing headlines and asking if data is the new God. I'm not sure I'd go that far, but if you dig a little deeper beneath these headlines, there's many reasons to be optimistic about the long-term trends for our industry and for Intel. I think one of the most stunning statistics is that 90% of the world's data has been created in the last two years.
Even more stunning, perhaps, is that only about 1% of that data is being utilized to create any sort of meaningful business value. Tons of room for our industry to grow. We've seen this exponential growth in data come at the same time that we've seen dramatic reductions in the cost of computing, in the cost of storage, and the increase in compute performance. When you combine these trends with the evolution and innovations and breakthroughs we've had in software and algorithmic development, it's not that hard to believe why we believe that we're entering this golden age of data. We believe that Intel is uniquely positioned to enable this insatiable appetite the world will have to process, analyze, store, and move data to drive business value.
This data-centric transformation, if you will, when I talk to customers in industry after industry, in economy and geography after geography, is quite profound, and we see it happening across every industry. Let's just take one example, the example I want to use is the transportation industry. This is an industry that clearly is ripe for disruption. Automobiles have extremely low utilization when you consider how much they cost. Of course, they're the cause of accidents and injuries and lost lives. We've been busy trying to think about how can we solve some of these problems with technology. Intel's Mobileye Group recently announced our first autonomous test fleet, and it's now on the road in Silicon Valley here and in Jerusalem, in Israel. I had an opportunity to drive, or be driven in one of these autonomous cars recently.
I think it's a great example when you think about it, of an end-to-end compute problem that exercises all of the assets of Intel. The car will be the greatest data collector in the world. Autonomous vehicles will run on data just like today's vehicles run on gasoline. I wanted to just give you a sense of the technology involved in making an autonomous drive happen. This is a map of the building that we're sitting in here in the campus at Intel. I just wanted to give you a simple example of how a simple drive around this campus, which we're doing now with our autonomous test fleet, actually happens. If you wanted to just drive in the circle that I'm drawing here, the first thing you need is a high-definition map. The car cannot drive autonomously without being able to see where it's going.
To create that map, of course, the vehicle needs to be able to see. Similar to how humans drive, you need to have a view of the landscape that you're driving in, the operational domain that you're participating in. The cameras on the car will be the source of the information to analyze the environment in real time. Now, the pilot cars that we're driving around, the test fleet cars, are generating four terabytes of data per hour. It's collecting sensing and policy data, and that data is then being sent to the cloud, where it's combined with data from other cars, and we're training then the driving and perception models. That's generating a huge increase in the amount of compute resources in the cloud.
once those models are then deployed, you have to blend the intelligence in the cloud with the intelligence and the vision streams in the car to enable essentially a continuously updating real-time high-definition map. The car itself will need 10 times the intelligence that it has today, and the cloud and the edge will then need to handle the delivery of that data back to the car in a highly efficient way. By the end of 2018, Intel, our Mobileye Group, will have 2 million consumer vehicles on the road, updating and developing what we call crowdsourced maps. That is what's going to help enable this to happen at scale.
These maps will enable safe operation of the vehicle, as the number of vehicles scale, we believe that more and more edge compute will need to be deployed so that the delivery of the maps back to the car can be done in an efficient way. Remember, the car itself can only store locally a certain amount of data, 20 or 30 or 40 kilometers of range. As the car is driving down the highway, it will need to be continually updated, and that latency of going back to a faraway data center over time will not be the way things are done. In autonomous driving, it's clear to us that this is not about a single product. This is going to take an industrial approach, an industry of computer vision, of mapping, of AI, of the cloud, of edge, of networking, and much more.
This combination of the in-vehicle compute, the edge compute, and the cloud compute, really requires this tight integration between the computing resources, all of those different areas. An end-to-end platform, if you will. Autonomous driving is a simple example of this data-centric transformation we're seeing in industry after industry. It's just one example of why we believe that we see a larger and ever-increasing, expanding total available market in the future. A year ago, we shared this view of the total available market that we thought Intel would pursue. From the cloud to the edge, we think we are incredibly well-positioned to move and store and process more data faster. This view shows $160 billion of silicon TAM that we were going after in 2021. Fast-forward one year to today, and our views have become even more optimistic.
Our updated view of the total available market that we're pursuing is $200 billion in 2022. That is the data-centric view of the TAM. If you added the client view of the TAM and the mobile view of the TAM, that number would be north of $300 billion. It's not lost on me that after 50 years as a company, this represents the biggest opportunity in the company's history. The TAM itself is growing at nearly double digits, and when we think about this TAM, we start with the mindset that we only have about 20% of this market today. Our ambitions are to grow faster than this TAM and to grow our share of an ever-growing market.
Over the course of today with my colleagues, we will sort of expound on why we believe this TAM will grow and of course, our strategy to drive our growth in an ever-increasing TAM. Let me spend a few minutes and just talk about three of the growth drivers underlying and underpinning this newly expanded view of our market: the cloud, the network, and AI. We've talked a lot about the cloud in the past. It is a massive disruption in the way that compute is delivered and consumed. It provided an easy on-demand access to data center resources that wasn't available before. Our business in the public cloud continues to grow. It is our largest single data-centric business, and in Q2 of this year, it grew for us at 41% year-over-year.
We're seeing that architecture from the public cloud increasingly permeate other parts of the business. It's moving into the enterprise, the cloud architecture approach. It's moving into the network. For us, we don't think about the public cloud growth as a zero-sum game. Two-thirds of our cloud business is TAM expansive. New consumer services, for example, Facebook, Twitter, online video streaming, online gaming, these are all new expansive opportunities for us. In the enterprise, new use cases like AI are coming to be in the public cloud that simply wouldn't have happened in a traditional on-premise enterprise setup. We view that as expansive. Only one-third of our cloud business is really offload from the traditional enterprise to the public cloud.
In addition, this diversity of needs of our public cloud companies and customers has led to a deeper, intimate relationship between Intel and each of our customers. Over time, we've further customized our products for their needs. Five years ago, 18% of our cloud service provider CPU volume was custom in nature. Last year, half of our volume shipping into the CSPs was custom in nature, customized products, highly customized products for their needs. You're going to hear a lot more about the public cloud, the dynamics, our strategy, how we intend to grow, and our customer impact from Raejeanne a little bit later this morning. The second big sort of TAM driver is the network, and I'm going to include 5G and the edge when I talk about this trend. We see a profound shift happening in the network.
You might call it the cloudification of the network. This is a $24 billion silicon TAM that we're pursuing. We have about 20% of that market today, and this is a market in transformation. Essentially moving off of proprietary equipment to standard high-volume servers. Much of that technology that the cloud service providers have deployed and adopted, the communication service providers now are looking to benefit from those same economics. All of the technologies we've developed in the data center, in the cloud, on the right-hand side of this page, we are now applying increasingly through the core of the network, the edge of the network, the access of the network.
Success here requires really a top-to-bottom approach, a top-to-bottom portfolio, low power envelope products, medium power envelope products, high performance products, workload optimized instruction set optimizations, which we've been driving in the network space for a decade now. Then, of course, software and ecosystem enabling to bring this all to life. I wanted to just show you one example of one of the things that we've been working on. This is a reference design. This is a 4G or 5G wireless baseband reference design. You could use it for edge compute as well. Essentially, it's a rugged server platform, and it's optimized for space. As you can see, it's optimized for power, it's optimized for cooling, it's built in a way that it can be used in an outdoor environment.
This reference design includes inside of it a Xeon D processor, an Intel network interface card, an Intel SSD. As the world shifts from 4G to 5G, you can imagine a day where every cell phone tower that we drive by has one of these servers, or hopefully many of these servers on it. One example of one way to think about this is there's $1.5 million wireless base stations sold every year. Again, as the world moves from 4G to 5G, we see that part of the market becoming more and more compute intense. I see just a massive opportunity for Intel here. It's perhaps one of the most attractive opportunities for Intel in that the secular trend is moving in our direction.
You're going to hear a lot more from Dan Rodriguez on our strategy in this market, how we've invested to win in this market, and how our customers are responding to that. The third data center trend driving TAM is AI, and this trend has been well documented, and we've all talked a lot about this. It's finally reaching, I believe, an inflection point. It's moving from a technology that was sort of in the domain of the scientific elite into moving into broad deployment used by businesses around the world. This is a $2.5 billion silicon TAM market today, growing at double digits to $10 billion in 2022. We're investing heavily in this portfolio. It's our single biggest incremental area of investment. We're building purpose-built products for AI. We're also infusing artificial intelligence into all of our existing products.
You're going to hear a lot more from Naveen Rao on this after I speak, then I'll touch a little bit more on AI in a few minutes. Our strategy, very simply, is to drive a new era of data center technology. Very simply, data-centric infrastructure of the future will need to move data faster, remove the bottlenecks that we see inside of the data center as traffic continues to grow. We'll need to store more data with the ability to quickly access that data to deliver rapid and real-time insight. Of course, we'll need to process everything through a combination of general purpose computing, but also workload accelerated computing products. I'm going to dig into this in a bit more detail, starting first with moving data faster.
In the last few years, we've seen an explosion in network traffic in the data center, this chart plots that traffic growth over a multiple year period. To me, what's interesting here is the orange portion of this graph. The orange portion of this graph is the data traffic within the data center, the traffic moving data from rack to rack, right? As this traffic explodes, as this growth just continues to grow, this east-west traffic in the data center, connectivity is increasingly becoming the bottleneck for our customers to fully utilize the compute that resides within the racks that they've set in the data center. You've seen Intel increase our investments in this area to help solve this problem. We have a full portfolio of connectivity products going after a $4 billion TAM today.
That TAM is growing to $11 billion in 2022, a 25% compound annual growth rate. We've had great success in high-performance fabric products with our Omni-Path portfolio. Our Ethernet 10 Gig NIC portfolio of products has a number 1 market segment share position. We're announcing today that we are expanding that portfolio into an emerging category that the industry refers to as Smart NICs, essentially adding intelligence to the Ethernet network interface card, to drive infrastructure offload. Many of our cloud service provider customers are looking at ways to drive more infrastructure offload so that they can free up compute resources inside of their infrastructure. Our Smart NICs will combine the assets we have in our FPGA portfolio with our Ethernet portfolio, you're going to hear more from Alexis on how we intend to do that. The biggest opportunity for us in connectivity comes in optical connectivity.
Optical connectivity is the idea that we can take electrons and convert them to photons and send them down these glass tubes at the speed of light and move data around rack to rack at very, very fast speeds. Intel Silicon Photonics is something that we've been at for a number of years. We have a unique advantage there in that we are able to integrate the laser in silicon. Our roadmap over time will deliver the lowest cost per bit, the highest bandwidth density, the lowest power per bit, we've made tremendous progress on this business in the last 18 months or so since we entered it. You're going to hear a lot more from Alexis on this as well in terms of our impact and our progress with top cloud service provider customers, OEMs and more. Let me shift to moving data.
Data center infrastructure must, as we talked about at the beginning, store massive amounts of data. Equally importantly, we have to find a way to deliver real-time insight from that data. We've invested heavily in this part of the portfolio as well. If you look at the memory storage hierarchy in a typical data center, you see that over a number of years, these layers have developed. This hot layer, this warm layer, and this cold layer of data access. In a world of ever-increasing amounts of data, this model is completely insufficient. The traditional DRAM layer, for example, is costly. It offers limited capacity for large data sets. While the software developer community has been very creative in trying to engineer around memory, it's not sufficient for the types of workload footprint that we see our customers demanding in the future.
At the same time, that warm layer and that cold layer has many limitations. Limitations in the speed of data transfer, the latency of data transfer, inherently, hard drives and even tape drives still used in data centers, inherently limit the efficiency in expanded views of how data is evolving. We've been hard at work at trying to attack some of these gaps that we see in the memory and storage hierarchy. Intel's 3D NAND portfolio is something we've been at for a while. Most recently, we've reached a new threshold in terms of cost, by adding QLC NAND capability, four-bit-per-cell capability. We recently announced the world's first PCIe-based data center QLC NAND solution. Just yesterday at the Flash Memory Summit, we announced the world's first QLC NAND consumer hard drive solution. That's sort of at the bottom layer of this graph.
In the middle part, in the storage layer, we've introduced Optane SSDs, and those have been out in the market for a few quarters now, and those products offer dramatically lower latency than traditional SSDs, 40x lower latency. Much higher endurance than a typical NAND SSD, and it's helping us break this storage bottleneck that we see. Finally, at the top layer of this memory storage hierarchy is something we're super excited about, and that is Optane persistent memory, an entirely new category of capability for data center system architecture. This is going to enable new workloads that weren't possible before, and it's going to solve many of these bottlenecks that we talked about in the DRAM tier. I want to expound on that particular part of this diagram and our plans with Optane persistent memory. Market opportunity, first and foremost, is huge.
I don't think it's an understatement to say that Optane persistent memory can completely transform the memory storage hierarchy that I showed you on that previous page. It's a $10 billion market. That's a subset of the existing data center memory market in 2022. It's a great example of something that Intel is uniquely positioned to deliver. We've been at this for about a decade now. We had to invent the Optane media. That was a breakthrough in material science and fundamental technology. We had to revamp the integrated memory controller on the microprocessor. We had to develop the software. We had to enable a big software and hardware ecosystem. The net result of all that is some of the numbers that you see on this page. In performance, we're able to deliver an 8x performance gain in certain use cases like Spark SQL over DRAM-only systems.
In other areas where there's limited system memory resources, we're able to enable 9x the read transactions and 11x the number of users per system. In addition to those metrics, Optane persistent memory delivers what the name implies, persistence, the ability to keep the data intact when you lose the power. What that does for our data center customers is quite profound. In high availability systems, you now have the ability to go from minutes of restart time, reboot time to seconds. You'll see some demonstrations of that a little bit later. You're able to go from 3 nines of availability to 5 nines of availability. That is huge for a mission-critical system in the data center. We're super excited about Optane persistent memory. We've made a ton of progress.
Rather than me sort of talk about this endlessly, I thought it'd be interesting for you to hear from a customer that has been with us on this multi-year journey. I'd like you to join me in welcoming Bart Sano, the VP of Platforms at Google. How you doing?
Doing very well. Thank you for having me.
Good to see you, Bart. Good to see you. Bart, our relationship goes back decades. We've been innovating together for a very long time.
Yes.
You and I have gotten to know each other quite a bit over the last few years.
Indeed.
In the last two years, our collaboration has sort of really oriented around Google Cloud in addition to everything else that we've done together.
Right.
I think one of the things that we set out to do is to help your customers adopt new technology faster.
Yes.
For example, Google Cloud was the first provider to deploy Skylake in the market and ramp it to very high volume, and, of course, we're working together on Optane persistent memory. Maybe you could just tell this audience a little bit about how we've been collaborating.
It'll be a pleasure. Thank you again for having us. I've only been at Google for about 12 years, but during that period of time, I can honestly say that we've had a very collaborative engineering engagement with Intel between Google and Data Center Innovation. As was mentioned, our joint early ship program with the Xeon Skylake was really an important one and a good example. That was when we both worked together to bring it to the cloud first and to Google first, that it really demonstrated how you can transform the cloud and the benefits to the end user customers. It did it in two different ways. The first one is that, it really lowered the barrier to get it out there.
Instead of waiting for sort of the nominal timelines that you have for a product release cycle or for data center refresh cycles that people have to go through, we were able to give it to the customers at scale and as soon as possible. The second benefit, obviously, is that the higher performance and the greater capability of that device was demonstrated in some of the performance on the, for example, on the AVX instruction set that showed somewhere between 20%-40% improvement and even 100% when the customers optimized it. The end result of all of that is at the end of last year, it had a lot of pull. We had deployed it quickly to majority of our regions. We had thousands of customers and tens of millions of hours working on the product. It was an extremely important launch.
I do want to recognize that that was part of why we value our partnership. A week or so ago at Google Cloud Next, we recognized Intel as our Infrastructure Partner of the Year. That's a really important thing for us because we value Intel as our partner.
Well, thanks, Bart. We appreciate that, and that recognition is great. I asked your team earlier if anyone had won that award two years in a row. They said no. We're going to try to make that happen again.
Very good.
We're not stopping at the innovations that you talked about around the processor.
No.
We're pushing the envelope, we're expanding into this Optane persistent memory area.
That's right.
To really deal with these memory bound workloads that we see.
That's right.
Can you share a little bit about how you see your customers benefiting from that?
Well, sure. Thank you. Yeah. This is very exciting for us because it's different capabilities as you're talking about, They're capabilities for our customers that are in a very demanding workload. Google Cloud strives to bring first to market all of these innovative technologies from Intel and to bring it out as soon as possible. In that vein, we've announced a partnership not only with the key partnership with Intel, but also with SAP to basically enable SAP HANA workloads running on GCP VMs powered by the Optane persistent memory, but also coupled with that, the next generation processor technology, the Xeon processor technology that's known as Cascade Lake. We see this as another big improvement in two ways.
One, as you mentioned, it's a persistent memory that the D in the DRAM is a problem that in order to scale something, you have to put a lot of DRAM and unfortunately, when you have to either reboot or launch new software updates, you have to bring down the server, and it takes a lot of time to bring that up. Because it's persistent, you don't have to do that. That's a huge thing. The larger the data sets. The more insight our customers are gaining, the more learning they're gaining, and that versus the alternatives with DRAM. Those are two really big, important capabilities of scale and availability, as you say. We're seeing a lot of pull and a lot of interest, similarly, as we saw with the Skylake interest. We're very excited about this.
That's great. Thanks, Bart. This is exciting, and we're really happy about this, but I know that Google has been anxiously awaiting-
Yeah
the first production units.
We're always anxious.
I'm happy to report that our first production units of Optane persistent memory shipped out of our factory yesterday.
Wow.
They're on their way.
Wow
to you.
I have my car here too, if you need to.
Bart, I know Google is anxiously awaiting this, but you personally.
Yes
are anxiously
Yes
awaiting this. I have actually in my hands here
Wow
the first production module of Optane
Excellent
persistent memory, and I'm going to hand it to you.
All righty.
You may have to rip it out of my hand. I'm going to hand it to you now.
Oh, thank you very much.
There you go, buddy.
Thank you.
Thank you so much for your partnership.
Pleasure.
Yeah.
I can't wait to get this into the data centers to torture it and to qualify it, I mean, for our customers.
We can't wait for the first check.
All right.
Thank you very much. See you, Bart. We're thrilled about our partnership with Google, and we're also excited about the broad industry support we have on Optane persistent memory. You can see every company you could imagine in the cloud service provider space, in our OEM customer base, our ISVs, are all working hard on getting ready for Optane persistent memory. I just wanted to take a minute and congratulate our engineers who have worked tirelessly on this product, some of whom have been on this for a decade, some of whom, for the last year, have been working 365 days a year, and in the last two months, probably pretty close to 24 by seven. I wanted to say thank you to them. I also want to thank, of course, all of our partners here on this page.
We're just scratching the surface of what's possible with Optane persistent memory. You're going to hear a lot more from Alper and from us over time as we expand and ramp this product. Finally, let's talk about our strategy to process everything. We all know that workload-optimized processing is going to come in all sizes and all shapes, and we've been investing in decades in this broad portfolio of products for this segment of the market. You're going to hear more from Dan McNamara on our plans in FPGAs. You're going to hear more from Naveen on our plans in network neural processors. I want to spend our last few minutes together this morning on the microprocessor.
I want to start by maybe letting you know that in addition to this year being our 50th anniversary as a company, it's also our 20th anniversary for the Xeon processor family. In those 20 years, Xeon has become the industry-leading processor for data centers everywhere. We have consistently delivered advancements year in and year out for our customers. In those 20 years, we've shipped over 220 million Xeons and delivered over $130 billion in revenue from this portfolio of products. It's one of the primary reasons that we have been so successful in the data center over time. The reason for that success, I would argue, is that we have gotten deeper and deeper with our customers to understand their needs and tailored Xeon to the workloads that they care about. You can see that over time.
In 2000, we added support for greater than two-socket support for scale-up workloads. In 2005, we added virtualization support, which we continue to improve. In 2010, we added AES-NI for network encryption. In 2011, we delivered our first custom cloud SKU, Quick Sync Video for video encode and decode for the visual cloud, DDIO for storage acceleration. In 2015, we added TSX for database acceleration. In 2017, we added integrated Quick Assist Technology for crypto and compression acceleration. Of course, AVX-512 for HPC and AI acceleration. This know-how we've developed over a long period of time gives us confidence that we can continue to tailor Xeon for the broad, diverse set of workloads that we all know our customers demand. It also gives us confidence in our ability to continue to deliver leadership for Xeon in the future.
A year ago, we delivered the biggest advancement for Xeon in the past decade. That was Xeon Scalable. It's doing very well in the market. You heard from Bart earlier. It represents our largest early ship program. That was the program that we went under and worked on with Google. It was the fastest Xeon ramp in our history to the first 1 million units. It is 50% of our volume today in Xeon. Despite that milestone, it has significant room to grow. We're only halfway done. In Q2, we shipped over 2 million Xeon Scalables. So far, through the first four weeks of this quarter, we have shipped already 1 million Xeon Scalables. The rate of adoption is accelerating. I get asked all the time, how do we stack up against our competition?
This chart shows you a small sample of the broad set of workloads upon which Xeon Scalable has absolute leadership. 48% better per core performance for network workloads, 1.7x better L3 forwarding for HPC workloads such as LINPACK, 3x better performance. For databases, 1.85x better performance. Memory optimized caching type of workloads, 1.45x better performance. This level of performance advantage is much more and much greater than a process node of leadership. In addition to the performance leadership that you see on this chart, we're offering the ultimate flexibility through that 20 years of innovation. One to greater than eight sockets of support, 60 SKUs on Xeon Scalable that covers an unparalleled range of frequencies, price points, power envelopes. This is what allows us to compete up and down the stack.
One of the things that we're particularly proud of on Xeon Scalable is how we've been optimizing it for one of the most important workloads, artificial intelligence. From the launch of our Haswell-based Xeon processor in 2014 to the launch of Xeon Scalable in July of last year, we improved our AI inference performance by 277 times, our AI training performance by 240 times, some of which was enabled by the AVX-512 capability, some of which was enabled by the software optimizations that we've driven. Since that point in July of last year, you can see here we've further improved the performance on Xeon Scalable. This time entirely through the software optimizations that we've driven. 5.4x improvement since we launched Xeon Scalable in July. This is inference performance that we're measuring here on Caffe ResNet-50. That level of performance has been well-received by our customers.
We work closely with many of our cloud customers, our enterprise customers, you see some examples of things that our customers have announced. For example, we work closely with Amazon to optimize their ML frameworks on their highly customized Xeon processor and delivered a 7x performance improvement for them on Xeon Scalable. At our AI developer conference a few months ago, Facebook explained how their workloads are increasingly diverse and broad, they needed a platform with ultimate flexibility, they've chosen to run their inference workloads on Xeon. You can see a broad set of other companies and customers that are now using Xeon Scalable for their AI workloads. I'm very happy with that progress that we've made, I'm even more happy to report that this step function increase in performance has led to a meaningful business impact for us.
In 2017, we recorded $1 billion in revenue from customers that are running AI on Xeon processors in the data center. It's the first time we've disclosed that. It's a meaningful impact, the growth rate that we see in that AI adoption of Xeon is quite stunning, we expect that to continue as we look forward. The obvious question is what's next? Our next generation Xeon, Bart alluded to it, is codenamed Cascade Lake. It's on track to ship at the end of 2018. It has many new enhancements, a new integrated memory controller. That's what enables the support for Optane persistent memory. It adds hardware enhanced security features for capabilities for mitigation, such as the Spectre and Meltdown vulnerabilities. It has higher frequencies, cache optimizations, new instructions, continued performance leadership.
As I mentioned earlier, AI is a domain in which we've been pushing the envelope, sort of reinventing Xeon for AI. Today, we are announcing that we're adding a new AI extension to Xeon that we call Intel Deep Learning Boost, that will come with Cascade Lake. Intel DL Boost extends the Intel AVX-512. It adds a new vector neural network instruction, this instruction essentially can handle int8 convolutions with fewer instructions, thereby speeding up the performance. This shows you that we expect to deliver an 11x improvement from the Skylake launch in July to the time we introduce Cascade Lake for inference performance, given that new capability.
Rather than just talk about it, I thought it'd be interesting to ask John to join me on stage, and he's going to show you the world's first public demonstration of Cascade Lake running Intel DL Boost. Come on up here, John.
Thanks, Pat.
Good to see you. What do we have?
Today we're going to show a performance demonstration on the two servers behind me. The server on the left is our Intel Xeon Scalable processor that we launched just about a year ago. The server on the right is our next generation Intel Xeon Scalable processor, codenamed Cascade Lake, featuring that Intel Deep Learning Boost technology that you just mentioned a moment ago. Let's launch a workload and see how it runs. We're running ResNet-50, which is a popular AI workload for image classification, and we're going to switch over to a screen and show the performance results from the two servers. If we'll switch the screen. There we go. We're showing about 11x speedup on the server with Intel DL Boost technology.
Image classification is used in many different types of applications in the cloud and in the enterprise, a speed up like this really delivers a better TCO for our customers and delivers a better experience for the end users of these applications.
What's behind that performance benefit? How did that come about?
Really this is mainly being driven by Intel Deep Learning Boost technology. Deep Learning Boost technology was designed to accelerate AI workloads, all of those instructions, as you talked about before, can be completed in one instruction.
Cool. That's awesome, John. Thank you very much. Appreciate it.
Thanks, Shane.
We're excited about Cascade Lake. It's on track. We're going to ship it at the end of this year. Now, all that great innovation is fantastic, but we need to do more than that. Our customers tell us consistently that one of the greatest challenges they face is the cost and time it takes them to deploy complex solutions. We spent some time thinking about how could we help them with that. Last year with the launch of Xeon Scalable, we also launched Intel Select Solutions. This is a program, a deep sort of technology-focused program, where we verify configurations with workload-optimized performance. We engineer the solutions. We validate those solutions. We test those solutions with our customers at the hardware level, the software level, and essentially, we try to give them the easy button.
We make it easier for them to deploy these complex solutions so they can accelerate their time to market. Since the launch of Xeon Scalable, we've worked with over 30 industry leaders here to bring solutions for analytics, for hybrid cloud, for HPC, for network infrastructure, and today, we're announcing an expansion of that portfolio. I'm excited about this. We have a new program and Select solutions around AI for BigDL using Apache Spark. We have a new program, Intel Select Solutions, around blockchain for security kinds of workloads. Finally, we have a new, SAP HANA certified appliance Intel Select Solutions, and this one is interesting. This one we've been at for 10 years with SAP, to work on the optimizations of Optane persistent memory with Xeon, with the software, so that we could accelerate time to production for their and our mutual end customers.
Intel Select, absolutely critical to us to help our customers to deploy all that great innovation that we talked about a bit earlier. All of that innovation and all of that solution optimization I've described so far, it wouldn't be possible without a solid foundation, a solid technological underpinning. In my 23 years at Intel, it's always been true that our engineers needed to take a broad approach to apply all of the assets of the company to solve the technical and customer problems that we see. They've harnessed all of these assets year after year after year. Transistor and packaging technology, of course, is fundamental, foundational, but there's much more than that.
Architectural innovation that we're driving, such as the AI extensions that we talked about here this morning, our memory investments that we've made, such as the Optane persistent memory technology, investments in interconnects, IO, how do we make it easier to move data faster, such as the investments we're making in silicon photonics. Of course, the expanded view we have in responding quickly to an ever-increasing, changing security landscape, and then, of course, software and solutions investments. All of that needs to come together and come to life in the hearts and minds of our engineers. I thought you guys might be interested to hear from one of our newest engineers. He gets embarrassed when he hears this term, but he's been referred to as a rockstar engineer.
I'd like to welcome Jim Keller to the stage, our Senior Vice President of the Silicon Engineering. Hey, Jim.
By the way, I'm not embarrassed by that. I really use it with my kids. I have two girls. They're 13 and 14. Well, 12 and 14, but soon. That was pretty cool to be cool with teenagers.
Jim.
Super happy about that, keep it up. Really good with that.
Jim, you and I have talked quite a bit. It's been three months.
We've talked a lot about kind of how we can expand the boundaries and solve problems that we've never solved before.
Maybe share two things with the audience. One, why did you come to Intel? Two, what have your sort of initial observations been?
It's an amazing opportunity for me to join Intel, and it's partly the scale and the excellence in engineering and to be able to contribute to that team. It's really kind of wild. I'm a little humbled by the whole thing. There's so many engineers here. I've spent the last couple of months, I've met so many people, and so many great people and enthusiastic. A guy called me yesterday, he wanted to meet because he felt like we weren't changing the world fast enough. He was a little frustrated about the roadmap and getting the next technology out, that's such a great problem to have, it's super exciting. What I really like is the diversity of all the technology. We have clients and servers, processors, graphics, AI, memory, IO.
Software, the scale of the problem set is really interesting, the engineers that can meet that challenge is really interesting. So far so good.
What are you working on right now? What's your kind of intense focus right now, here now?
Technology is so interesting because it's an endless train, right? We got today's problems, the next year's, the future stuff. Obviously, I came in and started working on getting We have some great new products in 14 and 10 nanometer, focusing on that and really getting up to speed with the fab guys about where we are in detail on 14 and 10. The technology is so good, getting it out is really complicated and hard. Focused on performance, on yield, on the analysis that it takes to get the products out. I'm pretty hands-on that kind of stuff.
You sort of brought this sort of outside perspective to the table a little bit, right? I think you're starting to help solve some problems, that is pretty interesting.
Yeah. I guess you're kind of saying I've been around a little bit. Yeah. Let me say it this way.
He said that. I didn't say that.
One thing I know is every company has unique culture, right? I'm not only a technology nerd, I'm kind of a culture nerd. For me, I'm kind of a new guy, but understanding how Intel works and what's great about it, and then being able to contribute to that, like for me, that's really exciting. We have really interesting technologists and a really interesting way of doing things and getting up to speed on all that is a monumental task, by the way, and then contributing to it in a positive way, that's my goal.
Yeah. Maybe just lastly, talk a little bit about the long term. What are some of the things that you're excited about? What are some of the things that you're working on for the long term?
Okay. The industry, I always like Ray Kurzweil's line, which is, "Future accelerates." Like, the next 25 years of progress is going to be bigger than the last 25. You kind of think there's certain inflection points like the Internet, mobile computing, PCs, the AI revolution is really big. How that's impacting the data center, the client, graphics, like you name it's showing up everywhere. In terms of the future kind of thinking, like Raja and I are talking a lot about how do we have a coherent view of client and server of CPU, GPU, AI acceleration, and how do we build the memory system that meets the bandwidth and demands of that, and do it in a way that meets the software. Like, it's not just one kind of software solution anymore. Software runs on everything and in very different ways.
The dimension of that is really exciting and thinking about it coherently across the range of products. We have like the half a watt to megawatt problem, and that's a pretty broad range. CPU, GPU, AI acceleration, that's a broad range. The memory system is a really broad range. To put that together in a plan that you can say, "That makes a lot of sense and that meets customer needs," yeah, it's pretty exciting.
Cool.
Super fun.
Well, Jim, it's great to have you at Intel. We'll look forward to the impact you're going to make.
All right.
I know you're not waiting for the long term to make an impact.
No.
You're working on it right now.
Every day. Yeah.
You actually need to get back to work, so I'm going to let you go.
I will.
Good to see you, Jim.
Hey, thanks a lot.
Thank you. Thank you. Good to see you. Okay. Before I end, I wanted to just give you a sneak peek into what we have planned in the future on the road map. We talked about Cascade Lake. We will deliver leadership performance on Cascade Lake, the Optane persistent memory, the continued AI workload acceleration. This is going to be a very attractive product, and I expect this to ramp very fast in 2019. We're confident that we will maintain our performance leadership in 2019 on the back of this road map that I'm showing you today. At our earnings call, looking out a little bit further, we told you that we'd have a 2020, 10 nanometer data center product. It's called Ice Lake. Our product will be out in the market in that time frame.
Today, what I'm sharing that is new is that we have created a flexible, feature-rich platform that allows our customers to select the right CPU for their workloads that will support both a new 14 nanometer-based CPU called Cooper Lake, and the 10 nanometer Ice Lake product as a fast follow-on. Cooper Lake will be available towards the end of 2019. It's going to generate and deliver a significantly better generation on generation performance improvement. We're going to continue to expand and extend the 14 nanometer generation on performance. We're making process improvements, we're adding architectural advancements, and we'll continue to push on the software front as well. One new feature that we will add, that we are adding into Cooper Lake is another new AI extension under the DL Boost family. It's called bfloat16. It's a new numeric format. It's principally used for training kinds of workloads.
We're aggressively standardizing on bfloat16 and infusing it into all of our products in Xeon, in our network neural processor family, and so on. You can expect us to build on that 20-year heritage that we talked about earlier and drive an aggressive push over the course of the second half of this year through 2019 and into 2020, and maintain our leadership position in the data center. We've covered a lot of ground this morning, and I just wanted to leave you with three final thoughts. First, we're in a new era of data-centric computing, the cloud, 5G, the edge, AI, really is driving a profound shift in the way we think about the market. Massive amounts of largely untapped data. The data-centric opportunity that we see is huge. Is larger than we've ever had in the history of our company, $200 billion TAM by 2022.
It's growing at near double digits, and our ambitions are to grow faster than the market. Third, we have an unparalleled array of assets that you've heard about from me this morning, and you'll hear about through the rest of today. An ecosystem that spans that entire data-centric product portfolio, and I think probably most importantly, I hope you get this sense from us today, is that we are hungry to get after this market in 2018, 2019 and beyond. Thank you very much. This may be the first time in my life I've done this, but I'm going to introduce another Naveen to come up after me, Naveen Rao, to talk about AI. Please welcome Naveen. Naveen.
Welcome everyone, and good morning. Lucky you get two Naveens for the price of one. This has caused a lot of confusion around the company, by the way. It's not something either of us have ever really had to deal with, so it's something we're all getting used to. Yeah. Today we're going to cover a couple of topics, largely some updates on our AI strategy. I shared a lot about this at Intel AI DevCon in May, our first ever Intel AI DevCon, which I think was very successful, and we'll go through some of the numbers there. We're continually evolving our strategy and adding new features and products and clarifications as we go forward. I'm going to talk through that a little bit. How our complete AI portfolio powers AI applications end to end, and we'll show you some proof points there.
Continued improvements with our tools, including Xeon, but also our dedicated silicon and our software. How we're delivering that software and interacting with the ecosystem, to ease the development life cycle. We really want to make AI more broadly applicable and usable by a greater number of developers. It was alluded to in Navin's talk that AI has really been in the hands of an elite few, and like any other new technology, as the tools develop and mature, we'll see more broad adoption. In Navin's talk, we talked about the size of the market here. Our view is that it's about a $2.5 billion market encompassing both inference and training. Today, there's a very strict dichotomy between those things. We do believe there's an opportunity for about $10 billion at a 30% CAGR, in the 2022 timeframe.
There's going to be a portion of that that's strictly training and a portion of that that's strictly inference. However, there are new paradigms emerging, and it's actually the line between these things is blurred, and you can actually see the line blurred. We expect that reinforcement learning is going to start coming on the scenes in a strong way, combined with simulation capabilities, transfer learning, and sort of hybrid models. We see that there's a future where learning will be distributed, endpoint, edge to cloud. These lines become a lot more blurred going forward, but we do know there's a very big opportunity out there. AI is becoming the major workload going forward. I'm a little bit biased, but it's my view that computing is AI, right?
We're just seeing that trend evolve and continue. We've gotten better at it. We've taken more inspiration from neuroscience and actually start developing the underpinnings of intelligence, I think it's a very exciting time for a technologist right now. Computing is in this new architectural phase, which just doesn't happen very often, right? The internet was one of them, where we saw the network become a very important part of computing. Now we're seeing the core of computing changing. This is extremely exciting for me, and Intel is very well positioned to lead this future of AI, as we have a broad portfolio of computing products that address different segments and new capabilities coming online. As I mentioned, AI is evolving. We started five years ago with proof of concepts.
There was a paper published by Andrew Ng while working with Google Brain, where they used 16,000 CPUs running for, I think, three weeks, to find cat faces in YouTube videos. Of course, that sounds funny now, but back then, that was extremely hard. Doing that reliably was not so simple. That was a breakthrough. Today, we're actually taking those concepts and applying them to real-world problems. I'm going to show you a demo in a little bit, with one of our customers, Novartis, where they're actually using these techniques to automate processes that decrease the time and cost of developing drugs. In addition, we've done a lot of work in the industry on natural language processing. We have capabilities now to do full conversational translation, which 10 years ago, if anyone used a speech-to-text tool, it was pretty miserable.
Today, it's actually pretty darn good, and it's really useful, and that's really due to innovations on these techniques and computing. Intel is driving this field forward as well through our open source efforts. We have the NLP Architect, released by Intel AI Lab, on GitHub in the open source. We're seeing adoption of all these things happening at a pretty fair clip. What's interesting to me is that AI is really becoming a general purpose paradigm, right? We've unlocked some fundamental foundational building blocks of finding value in data, which if you think about what a brain does, that's really what it does. We have this one to many relationship between these primitives of computing that we're seeing emerge and being able to solve problems in multiple domains. We're also seeing the application space for AI expand.
Applications are now bridging from all the way from the endpoint to the data center. One of our close partners, AWS, Amazon Web Services, has released something called SageMaker and Greengrass. We launched a device called DeepLens with them last year. DeepLens is a camera with Intel computing inside of it, and there's a software infrastructure to write applications in the cloud and deploy them at the edge. We're starting to see this sort of spanning of an application between edge and data center, and I believe that's going to continue in full force as we see the IoT space grow and we see AI capabilities being infused all across the entire spectrum. This combination of dense compute capabilities and general purpose compute is how we're going to get there.
We still have to use the fundamental container technologies that we use in the data center, being able to move those flexibly to different devices is how we're going to be able to do this at scale and manage the complexity. We have a comprehensive AI portfolio that allows us to actually take those capabilities that were designed for high scale in the data center and move them to various places where the application demands. One thing I want to get across here is that the application space is extremely diverse. It's not so simple to think we can build one computing architecture, and solve all of these problems. In terms of products, we're spanning from milliwatts all the way to hundreds of watts, as Jim alluded to. One architecture to do that is simply going to break down.
You're going to have things that are optimized for one set of applications or one set of power constraints that won't apply in other places. Intel is very well positioned here because we play in all of these different segments, we have technologies there today, and we're investing heavily in software that binds them and makes them work together and be able to migrate between these different kinds of technologies. Where the compute resides is really down to the constraints of the application itself. For instance, in the data center, we care about flexibility. If you're a data scientist and you're iterating on different neural networks to solve your problem, you need that kind of flexibility. You need the tools, you need the tool chain, all of that working for you.
Whereas at the edge, like in a drone for instance, you might care much more about power constraints, and you'll do whatever it takes from the software side to make that work, because your application depends on it. It's very different constraints, very different requirements from these two different domains. I like to think of this as Xeon Plus. General purpose is the foundation. It's where we build the foundational technology to scale and the foundational capabilities for AI. We can prototype algorithms quickly on general purpose computing. We actually add the additional specialized hardware to optimize that, especially when we talk about scale. We embrace this heterogeneous world that's emerging, especially when we talk about our high scale customers. We want to make sure that we have consistency in our IP portfolio across these things.
If there's a capability that's at the endpoint that is something that's relevant in our data center, we're going to use the same IP and be able to leverage the software investment that went into that IP from the original application at the edge back in the data center. We have a very rich IP portfolio emerging across all of these things, and we're sharing it quite openly inside of Intel to build the best products. This whole thing is really working. As Navin said, we've hit a $1 billion business for Intel today. This is really no surprise because, again, my view is AI computing is AI. Intel is computing. It's no surprise that once AI became a thing, that Intel benefited greatly from it. We're seeing that across the board, CSPs, cloud service providers, and enterprise and government.
By the way, this $1 billion actually doesn't include FPGA or IoT. It's only Xeon in the data center. There's a lot more here. The whole world is exploding, the application space is exploding. This is just a few names here all the way from the edge endpoint to the data center that are part of this journey with us. I'm going to take a step back and talk about the life cycle here. At Intel, we're really looking at building the best tools to improve the reach of AI. I look at the internet and say, like, the tools were very impoverished in the early '90s, but the capabilities were there. It's only a few companies could actually access those technologies. You fast-forward to the mid-2000s, and we saw a great proliferation of these things because the tools became much more mature.
10 people could write an application that could scale to hundreds of millions of people. That phase hasn't happened yet in AI. It's still in the realm of experts. I was asked what ballgame we're in or what inning we're in for AI, if we're going to think of it as a ballgame. I'm thinking it's like top of the second, right? This is still very early stage. When we look at the total development cycle for actually deploying an AI solution in an application, aggregating data is where you start. There's a development cycle where you're trying to make sense of that data and drive some new experience or new capability, you actually got to deploy that. This is actually a realistic timeline we see. Let's blow this part of it up where we're talking about the development cycle.
Even within this, labeling data and just dealing with that is actually the vast majority. About 30% of this is training, which is model parameter fitting. GPUs have very early lead here due to their parallel computing capabilities. AI is much, much more than that. That's one small piece of it. The rest of this is actually running on Xeon today, and we are going after building the best training and inference solutions going forward. Overall, we care about making our customers shrink the development time of AI and getting it to market faster. Today, AI is powered by data. I think tomorrow it'll be powered by data plus simulation. Intel platforms have been a de facto standard for both of those things. Data is managed on Intel. We move data around, we process it, we extract meaning from it on CPUs even today.
Naveen mentioned, only less than 1% of data is actually being used today to drive business forward. We want to increase that number. We want to make it easier and make it faster. We want to empower more of the world to use AI to derive value from data. This is really brought to life by data scientists today. We're building the tools to make their life easier. We mentioned much more about our roadmap on Xeon going forward, and I just want to make it clear again, I said this message when I was at AIDC that let's bust a couple of myths here. First, that GPUs are the only thing out there for AI. The reality is almost all inference in the world runs on Xeon today.
That the performance gap between general purpose computing and specific kinds of computing, like GPU, is some enormous gap, like 100X. It's more like 3X. That's okay because general purpose computing has a scale that a specific solution can't really achieve. Everything has its place. Once AI starts making its way into general purpose computing, we've achieved a scale with this new technology that we simply couldn't before. It's incumbent upon us to continually evolve our platform to make it the best it can be for AI, as well as everything else that Xeon does today. Xeon wasn't well optimized from a software perspective two years ago, but just from the launch in July of 2017, we've increased the performance and inference by 5.5X, and even in training by 1.4X.
I think Naveen alluded to a 200+X number on both of these things since two years ago. That was in the Haswell generation. We've added things like vector and matrix multiplication and SIMD instructions in the Skylake launch to continue this gen on gen improvement. Then, of course, in Cascade Lake, we are adding DL Boost, this family of new capabilities. We showed you the 11X improvement that we're already seeing. We're going to continue this process. Bfloat was mentioned as something that we are standardizing around. I'm just going to give you a little bit of a primer on what that is. Scientific computing emerged in the last 20 years or so as a way to simulate the real world. We can simulate the weather or earthquakes or what have you. We need very high precision of numbers.
The models there are built out of physics, and they tend to be continuous numbers. We need 64 bits of precision. We call that double precision floating point. That was the de facto standard for scientific computing. Most general purpose computing used 32 bits, 32-bit floating point. AI is actually quite an interesting use case because it's actually very tolerant of lower precisions. Why do we care? Lower precisions allow us to put more parallelism on a chip. We can have more of these operations happening at the same time at a lower power and do more, and without even degrading the algorithmic performance. What I mean by that is the classification rate or how well it's translating speech-to-text, something like that. You're not degrading that while also achieving higher levels of parallelism at a lower power.
We've invested a lot of time and energy in making lower precision models work, and you're seeing us standardize now on bfloat, which was originally published by Google, across our entire line. Like I said, this is going to be coming into Cooper Lake going forward, and we'll continue the innovation cycle on subsequent generations after that. We also talked about at AI DevCon, we're looking at the inference market because inference today runs on Xeon, but we want to make sure that we continue to defend that. The neural network processor family is how we're approaching this from an AI specific perspective. We're building a variant for the inference market, which we haven't disclosed yet, but I just wanted to let you know about the existence of it.
We're also talking about the NNP-L, or for learning, which will be coming out in the end of 2019. We continue to invest in this to push the boundaries of AI. As I said, I think computing is evolving. This is the tip of the spear. The idea is that we bring new capabilities here, things like bfloat, which are going to be present in our first commercial NNP, as well as new capabilities in terms of scaling. Chip-to-chip interconnect, very low latency, high bandwidth. This is the kind of stuff that you can push the capabilities from an algorithmic standpoint. It actually changes fundamentally what you can compute because now certain bottlenecks have gone away. That's hard. That's how innovation has to happen.
You start with one place, the innovators on the algorithm side will take it and build wonderful things that you never even thought of. I think we want to be part of that entire cycle of innovation. What we find is that we've iterated with our customers on our first rev, which we call Lake Crest. We got some really good results, and it validated that they want maximum performance. They don't really care about theoretical teraops. That doesn't really matter. What really matters is what can I really shove through the system and effectively get my model trained? That's all they really care about. At very high scale, like when you're deploying 100,000 or more compute nodes, you actually care a lot about efficiency and you care about even 5% increases in performance.
Those matter because at that scale, it has a huge monetary component to it. We're leading the charge in this new category of compute. We're working with our customers directly, I'm very excited that we're going to have this coming out in late 2019. Software is essential. Some people have even said that AI is software. I would agree with that to a certain extent, except for the fact that if you don't have hardware that can run it'll take forever to actually train a model. 50% of AIPG, which is the Artificial Intelligence Products Group at Intel, is actually dedicated directly to software. We're interacting with the open source community, which I'll talk about in just a moment, quite a lot there and putting a lot of new tools into the industry.
As a tool stack emerges and matures, people move up the stack with more abstraction, and we can address a larger audience. Today, data scientists and AI researchers are a few thousand around the world. We want to get to enabling the true enterprise developer, that 25, 35 million population, where we can actually see massive scale for adoption and proliferation of these technologies. Supporting the full stack and recognizing a different audience is supported by each one of these is extremely important. At the very lowest level, we see the foundational libraries. Things like MKL-DNN. This is a set of linear algebra kind of libraries that underlie neural networks. You may have heard of cuDNN from NVIDIA. It's a very similar level of abstraction. The reality is today that no one really codes there.
These capabilities are sort of encapsulated in the libraries, which continue to evolve, the hardware vendors are the ones who know how to do that stuff the best. It's not really a lock-in there. The entry point for most developers in this space is actually higher up on the stack. The libraries and deep learning frameworks are really where data scientists interact. They take things like TensorFlow, they just want to download it and get the best experience out of the box. We're working with our partners at Google to make sure that happens. We're continually investing in the foundational layers to get the best performance out of our hardware, we want to make sure that it's upstreamed into the open-source community so they get access to it. That's a continual process that we're working on.
That's going to continue as we have our new hardware platforms out there. One thing I want everyone to take note of here is that this whole stack is emerging almost in real time. I don't think TensorFlow was even announced until December or November of 2015. It sort of becomes a de facto standard in people's minds, but this is very new. This whole stack is just continually evolving. I've never seen anything so quick to change in my career. Above this, we're actually driving a higher level of abstraction to make applications easier to develop for edge devices like Movidius or FPGA. Toolkits like OpenVINO take some of the complexity out of taking a neural network and compressing it and changing the quantization and really just making it optimized for that edge platform where power matters quite a lot.
Same thing with the Intel Movidius Neural Compute SDK. We're building tools for different personas of developers that make their life easier. Again, higher abstraction, more accessibility to developers. We want to proliferate these technologies to the greatest possible population of developers out there. Another part of this strategy is a project we call nGraph. I've shown you that there's a high diversity of hardware solutions and architectures to solve different problems. That all sounds great, you're like, "Well, how do I use these things together?" We're looking at a way to boil down the primitives, the building blocks of neural networks, and actually standardize that, then come up with a way to translate those to optimize libraries or optimize code for various hardware platforms.
If we look at Xeon today, we have nGraph connected from TensorFlow or MXNet all the way to Xeon. We're seeing actually very good performance. In fact, sometimes even better than directly connecting these things. By doing it this way, we just make it easier to see the performance benefits of the low-level software efforts we're putting in and make that more accessible to developers that are iterating and changing the topologies on the neural network side. Instead of having to optimize each one of those topologies on each piece of hardware that you're working on, we can leverage investments being made on each piece of hardware for a number of different topologies. We're also working closely with the ONNX consortium, of which we're a part, where we're defining a set of APIs, a standard API for inference.
We can take a set of inference calls and actually run it through nGraph and make it optimized for whatever hardware platform. You're going to hear more about that in Arjun Bansal's talk in the technical track. Let's get to some customer use cases. I talked about earlier that we're working with Novartis on drug discovery. One thing that was really challenging here is that they want us to analyze images that are much larger than we typically see in some of these toy examples that demonstrate the capabilities of AI. ImageNet, which you might have heard of, has really driven the field forward in a lot of ways, but they're tiny images, 224 by 224. You look at these things, you can't even make out details on them because we don't deal with such small images anymore in the real world.
The ones here are actually not even that high resolution, but they're much higher resolution than ImageNet. This actually presents a lot of challenges for things like GPUs. CPUs actually turn out to be a very good solution here because this moves into the regime of what we call IO bound. It's not compute bound. You can throw more compute at the problem, but if you can't get the data to the compute, it really doesn't matter. We worked with Novartis to optimize TensorFlow running on Xeon Gold CPUs and using much larger memory footprints. Before we started, it took about 11 hours running on a previous generation of Xeon. Let me show you a quick demo of how we scaled this up. We went from optimized software on one Xeon Gold, and then we actually scaled it to eight.
We're able to achieve about six and a half, seven x speed up from their data sets and their specific solution running on eight CPUs. I think this is a great testament of the capabilities that we have in terms of optimization, as well as just what Xeon can do when you apply it to this problem. It has the right IO and balance between compute to actually solve large-scale problems like this. We actually reduced the training time down from 11 hours from start to about 30 minutes. You can imagine if you're a data scientist working on this problem, that's a gift from heaven, right? You have a timeline, you're trying to iterate, and you're trying to get the best performance out there, and if it takes you a day each time, you're going to make very slow progress.
This, you can do it multiple iterations in an afternoon. Here we talk a little bit about our technologies. We worked with Google. We're working continuously with Google to upstream our latest innovations into TensorFlow. Those come from MKL-DNN, which is our primary vehicle to deliver those optimizations. In this particular case, we actually worked with Novartis to build out an infrastructure with Omni-Path Architecture. This is an example of this holistic approach that Naveen alluded to, where we're really thinking about the whole data center, how data moves, how it's processed, and how we can optimize every part of that stack. Next, let's look at scale-out inference. You may have heard of a company called Taboola. They're the world's largest content recommendation engine, serving over 360 billion recommendations to over 1 billion unique visitors monthly.
They actually evaluated both CPUs and GPUs. CPUs actually provided the best scale for memory-intensive inference workloads across seven data centers around the world while maintaining their efficient operational cost. You may have heard us talk about TCO, total cost of ownership. This is what it is all about. Inference tends to be embedded into a workflow with many other things going on. You are serving websites, you are serving REST API requests, and you are also doing inference. It is actually quite complicated to manage if you have a set up infrastructure just for inference. Xeon actually gets a lot of traction because of this total cost of ownership benefit. In using Intel's Xeon Scalable processors, Taboola was able to increase throughput by about 2.5x over the baseline working with us. This is an example of how performance for actual workloads really matters.
I want to talk about the AI Builders Program a little bit. We launched this back in May, and we have seen an incredible response. We have 95+ members in just three months. We are addressing multiple segments, full verticals, horizontal technologies, and cross-vertical providers. This is something where we are seeing the proliferation of AI into many different layers and capabilities. I am actually extremely happy to see this kind of adoption because this is exactly why we designed this program, was to accelerate adoption into the customer's hands and make it easier for customers to access and leverage our technology. Engaging with the developers is an extremely important part of the puzzle here.
You can build the best hardware out there, if no one knows how to program it really does not matter. We have been very active in the open source space, putting out things like NLP Architect, which I mentioned earlier. RL Coach, it is a reinforcement learning paradigm that allows you to build these kind of solutions. Distiller, which can optimize neural networks for sparsity and other kinds of low-precision arithmetic. We have seen just from zero stars on GitHub to three months out, we see a pretty quick ramp. This just shows me that there is a really big appetite for these kind of solutions. We also launched the AI Academy and AI DevCloud, and to date, we have trained 110,000 developers, 150,000 using online resources each month, and 800 shared projects. Some of the shared projects are from Carnegie Mellon, Shanghai Jiao Tong University, and IIT Kharagpur.
They have presented papers and actually built course material around this. Again, we want to make sure that the next generation of developers actually understands how to use our tools. Back in May, we had our first ever AI developers conference. A lot more demand than we even expected. We had over 950 attendees, 50+ sessions, and 50% of that content was actually led by customers. Many of the workshops and talks were standing room only. This is the first of many. In fact, just 12 hours ago in India, we had our first international AI DevCon, we are going to be continuing this in Europe and China next. I will quickly summarize here. Intel Xeon is a real business for us today. It is $1 billion in AI specifically, again, not even including the IoT side of things.
This is having a real economic impact upon this company. We're delivering the best tools and software to simplify that development of AI. We are a tools company. We want to build that foundational layer that enables all other applications. We're investing in cutting-edge silicon to drive the next phase of AI. Before I conclude, I want to share an exciting example of how all this comes together in the real world, software, hardware, tools, and something that we're actually going to see on the big screen on Friday. Ziva Dynamics is a company that's actually utilizing Xeon and AI capabilities running on Xeon to improve graphics rendering for movies. This character is called Zeke. There he goes.
It's a pretty cool example, very realistic, and they actually build it up from fundamental physics of body tissues and skeletal relationships, which I think is really neat. There's so much you can do with calculation because we can characterize these things at a pretty high level. At some point, it actually becomes difficult to make it perfect. There's always something missing. If you've seen rendering over the last 20 years, you've seen things where it's like it looks a little plastic-like. It doesn't look quite right. That's really what Ziva has brought with their new pipeline, the set of tools that they bring to industry. Let's move on to the next. Okay. They've really tried to bring this next level of realism into rendering.
On the left side there, you see sort of the old way of doing it, from the fundamental physics of the body movements. On the right side, it's actually applying subtle tweaks, but basically taking what real tissues do and training an AI algorithm to actually map from that input on the left to something more on the output on the right. If you look at the hindquarters there, you'll see how the skin folds. These are things that are very difficult to model, but we can use AI to actually build something that's very close to reality. I think this is a great example of how we can use AI in the real world in a very visual format. Go to the next one. How does this translate to the big screen?
Let's take a look at a virtual character created by Scanline VFX for upcoming Warner Bros. movie called "The Meg," which will be released on August 10th, this Friday.
The Meg," rated PG-13. Experience it in IMAX August 10th.
Hope you all go catch that one. All right. Now I'd like to introduce Raejeanne Skillern, Vice President and General Manager of the Cloud Service Providers Group, to talk about cloud in driving transformation and growth. Thank you.
Thanks. Amazing. I'm really thrilled to be with you today, and I can clearly say that out of 110,000 people at Intel, I believe I have the best job. I think in 30 minutes from now, or maybe 27, we'll see how it goes, I think you'll agree with me. Why? Because I have spent the last 10 years of my 24-year career at Intel focused on growing the public cloud and building deep collaborative partnerships with cloud service providers around the entire world. Things were different 10 years ago when I started, very different. We were at 35% market segment share, and the cloud was nascent. We have learned a lot, and we have worked very hard over the last decade to get to the position we are in today.
All of those learnings are foundational to the strategy that we are taking forward to continue servicing this growth. A couple of things that I have learned on the way are absolutely critical. One, one size does not fit all. Customization is critical, and I know that Naveen mentioned this this morning. Two, technologies need to provide more than just better performance efficiency. They need to help the cloud service providers differentiate themselves in a competitive market. Three, long-term relationships are hard. It takes years to prove that you can design, deliver, support this type of infrastructure at scale across millions of CPUs. With that learning, we have had an incredible 2018 so far, and the past few years have been pretty fantastic as well, which I'll share with you in a minute.
I want to share my vision and strategy for why the cloud segment is going to continue to be the largest revenue contributor to the Data Center Group. 10 years in, where are we today, for me, at least? We're still at the beginning. Cloud is everywhere. It touches everything we do at home and at work. We're still at the beginning of untapping the possibilities of what it can do. Cloud has created a massive disruption in how compute resources are delivered and consumed. Cloud is also a new way of architecting a data center that delivers extreme performance at scale efficiently. Cloud enables millions and billions of connected devices, which drives not only new service opportunities, but generates massive amounts of data that we can use to drive businesses. Therefore, cloud is the foundation of the digital transformation.
How we shop, how we receive medical treatment, how we consume our media, and how we work, none of this would be possible without the power, performance, efficiency, and flexibility of the cloud today. You will see that reflected in our growth numbers. We have grown over 2x over the last three years, which is a 30% CAGR from 2014 to 2017. Through this growth, we have gone from being 23% of the data center business to 43% of the business, becoming the largest segment revenue contributor. With the phenomenal growth that we've had already in 2018, I think the question that might be on your mind is: Will this growth continue? Maybe not at 43%, although I do strive for that every single day, but it will continue to be a high-growth segment due to the many contributors that are fueling this growth.
As Naveen said earlier, we know that this growth today is two-thirds TAM expansive, and only one-third is that conversion of enterprise workloads. We also know that the growth isn't just coming from the big getting bigger. It's coming from the next wave cloud service providers as well. Over the last few years, they have grown at an equal clip year-over-year revenue growth to the large ones. We also know that the growth is fueled, whether you're consumer or business, whether you're small or large, we know that there is a wave of digital services that is driving this industry. Today, only 13%, if you look at the top three infrastructure providers in the world, only 13% of their revenue comes from public cloud or infrastructure PaaS and SaaS, while 70% of their revenue is coming from digital services. What are these digital services? You all know e-commerce.
E-commerce, by 2021, is expected to grow 21% to be a nearly $5 trillion industry of goods and services sold over the internet. Today, only 18% of retail transactions happen over the internet, so you can see we have a long way to grow and a lot of opportunity. Of course, when we connect on the internet, there's digital advertising. That's growing at 15% to be a $400 billion market, driven by all of us engaging and getting online through online news, media, and gaming. My favorite is digital media and video. It's the smallest segment, but it's growing the fastest at 25% year-over-year to be a $120 billion industry by 2021. That's just driven by, once again, all the new content that is being created and delivered, so we can watch these anytime we want, on the device we want, and any program we want.
We have that flexibility. Of course, what is also growing at an equal pace is the public cloud and infrastructure PaaS and SaaS. It is going to be, by 2021, a $300 billion market. That is a 22% growth rate because of its benefiting from the explosion of digitally native organizations and all of the artificial intelligence that is needed to service both their enterprise customers as well as their entire consumer base that they are servicing. With all of this growth, Intel is uniquely positioned to win. I have seen the diversity in CSP requirements from their workloads, to their service types, to their scale. It demands a very focused and flexible design methodology to truly understand their needs. This has given us the opportunity to build these deep partnerships with the global cloud service providers to really uncover innovation and differentiation.
The three things that I think are really unmatched capabilities for Intel that set us apart in the industry, I am going to hit on a few of these. One, we start with designing tailored and custom products across the entire platform. We are known for compute, but we do this across compute, network, and storage. What also makes us unique is we partner alongside our CSPs with hardware and software engineering resources to ensure that every ounce of performance, every feature and capability that is in that architecture, is exposed upward to their applications and optimized uniquely for their workloads. Third, we invest. We invest in large-scale go-to-market programs, co-marketing programs, and sales programs so that we are out accelerating the demand for cloud services that are built on the Intel architecture and accelerating this TAM expansive growth for the market.
I want to go a little bit deeper into each one of these, and I want to start with how we deliver our products. One of the things that makes us unique is that in addition to delivering our standard roadmap that you know very well, we offer the ability to optimize our SKUs, specifically for a unique workload and instance type. Navin kind of showed our history of Xeon. We actually did our first optimization work with Google in 2008, but in 2012, we released the first ever optimized Intel Xeon processor SKU on Sandy Bridge. Now, six generations later, Navin showed five Xeons, I added sixth, because I add Xeon D, which our Xeon D line was actually created out of a custom joint innovation with Facebook.
Six generations later, 50% of our volume is optimized, spanning over 10 different cloud service providers and nearly 30 different SKUs, in addition to the SKUs that you saw that cover our standard roadmap. These are off-roadmap SKUs because they are customized uniquely for the service provider and they provide differentiation, so it is unique to them, and they want to keep it off-roadmap. One example of this is based on work we have done with Amazon. They have made some recent announcements that have highlighted the value that an optimized SKU provides for their instances. I have a quote. Our collaboration with Amazon spans a decade. We work together not only to optimize their infrastructure for the best price performance, but also for their market differentiation.
In addition to the M5 and C5 that they launched on the Intel Xeon Scalable processor, they recently announced the R5 as well as the Z1d EC2 instance. This instance is based on a customized high-frequency Intel Xeon Scalable processor that is uniquely optimized to create this instance for relational databases and electronic design automation, or EDA. This is just the beginning. We take this further. We have taken this and built the capability to deliver semi-custom products as well. We can integrate the IP from our customers, we can mix and match our Intel IP, and we can integrate other third-party IP. We have packaging technologies, and we can hone a product, a semi-custom product now, not just tweaking knobs, specifically for our customer base.
We can scale it up when they need it, we can scale it down when there are things that they just don't need in that CPU and don't want to pay for. What also makes us unique is we can take it one step further. We deliver fully custom ASICs. We have the ability to deliver custom silicon when our customers' workloads demand. You take that ability, this broad productization spectrum, and combine it with the fact that we can apply it uniquely across the entirety of the platform, compute, memory, and storage, connectivity. In my segment, there's a ton of Ethernet, high-speed Ethernet, as well as silicon photonics and accelerators, whether it's our FPGAs or the purpose-built accelerators that we've been talking about that we're building as well.
We can work across all of these technologies to optimize specifically for our customers. I want to share an example from Toutiao. You might not know of Toutiao. You may have heard of ByteDance, which is what they're known more in the U.S. They're significantly a venture capital-funded company. They are one of the fastest-growing machine learning and AI-based apps in China. They have over 200 million monthly users on their app, and these users are estimated to spend 74 minutes a day on the app. I mean, that's a lot of time, right? It's a lot of data. They are building their next-generation infrastructure, not only using our Intel Xeon Scalable processors, but our Intel SSDs and our networking products. Because of the balanced platform-level design approach we took, we were able to prove out the value of that platform.
I'm going to let Toutiao tell you in their words about the engagement.
Toutiao is a data mining-based recommendation engine product. Machine learning and artificial intelligence are the cornerstones of Toutiao's personalized recommendation engine. Toutiao and Intel continue to innovate in data-centric computing. Toutiao next generation data-driven infrastructure is based on Intel Xeon Scalable processor and uses innovative platform design philosophy, achieving balanced system design with Intel high-speed networking and SSD storage to unleash the huge potential of today's data center performance, providing our users with the best user experience.
If you don't even recognize ByteDance, if you have children and you know the app Musical.ly, you know Toutiao. That is one of their apps. It is also kind of sometimes strange for a silicon company to be talking about software, yet we've done that a lot today, and that's because we know software is the key to unlocking the full potential of any infrastructure and hardware. It's also, of course, the key to creating new services that our cloud service providers put out into the market. This might surprise you, but I have just shy of 200 engineers, hardware and software engineers. Intel has thousands. These 200 engineers are actively, directly engaged with cloud service providers. They don't do the generalized support. They go out to our customers on their premise, and they do hands-on, side-by-side engineering with the global cloud service provider base.
In 2018, this team of engineers is already working on 150 cloud service provider projects this year. We keep them busy. We call them the ninjas because literally, we can pick them up and drop them on a customer premises, and sometimes they'll stay for a month, sometimes they'll stay for two days. They are there and connected to solve the problems and opportunities. One of these projects is with Oath. Oath is Verizon's digital content subsidiary, which includes the former Yahoo. We've been working closely with them, not just on optimization of their cloud infrastructure, but also infrastructure modernization. This is critical because there is a lot of times where the cloud service provider's infrastructure, their entire software stack, may be at a point that it can't take advantage or easily move to the next generation technology.
When we get in and help with infrastructure modernization, we are touching everything, compilers, orchestration, telemetry, their software stack, and of course, the hardware. For Oath, we did it all within their OpenStack environment. This work has led to a custom CPU definition, which I will let Hugo from Oath tell you about.
Oath Inc. is the merger of AOL and Yahoo Inc. under Verizon Communications Inc. as a big media and technology company. We have spent a lot of time researching technology coming out, determine what is applicable for our situation, moving from an old way of run data center to cloud, specific data center, to give the technology to the user much faster. We have been working very closely with Intel on both modernization and optimization. Cloud optimization for us is key. Intel has been a great partner in this area because they are touching everything from compilers, software, hardware, as well as understanding how data centers are built and designed. Intel have been helping us with porting our OpenStack version into a mainstream version so we can benefit from all other releases coming out in the future.
On top of that, with Intel, we have been able to work very closely to come up with a customized CPU. We can reconfigure it on the fly to fit applications. To me, optimization is all about making the application better so we can deploy application very quickly, save time keeping the cost low, and also focusing on good performance for the end user.
They have been a great partner to us for sure, but I want to share a few more examples because this is so important to the way that we work with our customers. When we can ensure that their applications and services are exposed to all of the libraries and instructions and performance that we have in our architecture, they get value and they can do more with their infrastructure. Another example is when we worked with Tencent's WeChat team to optimize their voice recognition algorithm on the latest Intel Xeon processor. How many people know WeChat? Have you heard of it yet? Good, because if you haven't, you will. WeChat has over 1 billion monthly active users. My 78-year-old dad, who lives in Oregon, uses it every day because he has friends around the world, and that is his app of choice.
That is a technology breakthrough, really great just right there. WeChat is an application and platform that allows other partner apps to run on top of it. You can make a voice or video call, hail a cab, or if you do not want to catch a cab, you can make a plane or train ticket. You can make your dinner reservation or order takeout and do many other things within this app and you will never have to leave the app. That is a lot of people, 1 billion users, and a lot of data. As I mentioned earlier about digital services, WeChat is part of the 82% of Tencent's revenue that comes from their digital services. They also have gaming and are a global leader in gaming, which you will hear more about from Lynn Comp later.
Today, only about 2% of their services revenue comes from cloud, public cloud, but it is rapidly growing as well. With all of this data, their AI performance is critical. Working side by side with Tencent, our engineering teams increased AI performance by 16x for inference. That means much faster voice-to-text translation, and think about how important that is, especially as technologies like WeChat move into the autonomous cars and move into cars for use. It is going to give a much safer as well as a much more pleasant experience for their users. Another optimization we have been doing is around deep learning training and inference with Amazon. I am going to share another training example. Amazon launched their deep learning AMIs that come with TensorFlow 1.6.
This TensorFlow 1.6 is built with and built on our Intel Advanced Vector Instructions, as well as the math kernel library for deep neural networks that Naveen Rao just talked about, to specifically optimize it for training on Intel architecture. With these optimizations, you can train a ResNet-50 benchmark with a synthetic ImageNet data set on Amazon C5, their high-performance computing instance, and you are going to get 7.4x faster speedup than using stock TensorFlow binaries. This is the power of that joint engineering and innovation. We can unlock 7.4x better performance by working together. Two more examples because these are just to me incredible numbers, and you can instantly see the value that we provide, not only to our cloud service providers, but this value extends to their users and consumers of their service.
Python, a very common high-level programming language, we engineered alongside Google to deliver a 21 to 23x performance speed-up on scikit-learn running on a 96 virtual CPU GCP instance based on the Intel Xeon processor. 21 to 23x speed-up through joint engineering. I just threw in one more. It kind of plays off where Naveen Rao was just, last week there was an article on enterprisetech.com. Alan Poor, who is the VP and GM of Engineering for RenderMan products at Pixar. In this quote, I love this quote. He says, "For over two years now, Intel engineers have been helping us modernize a very old, a very cranky code base. We parallelized and vectorized the software to substantially improve performance." Then he indicated that through this work, the performance that they are seeing speed up is 2x-4x in their environment.
If you all were wondering why you liked Coco so much, it is because the code base was not cranky anymore, and it was running a lot more sufficiently. If you remember, I said there are two things. There is performance optimization and modernization, but there is also services differentiation. This is something I believe passionately in. When we can help cloud service providers either create new service offerings or differentiate their service provider offerings in the market, we create value for them and for their consumers. We have been working with Microsoft, and hopefully you have heard about this before, to develop the Azure Confidential Computing services that take advantage of the Intel Software Guard instructions or Intel SGX.
As you can see from this quote, the collaboration is built on a differentiated technology, Intel SGX, and it has now resulted in a service offering that ensures data is in a secure enclave while being processed, and it extends those protections to encrypt the data when it is at rest or in transit. If you can imagine, when you're moving enterprise critical workloads into the public cloud, this level of security is critical. This is a highly differentiated service for Microsoft, and we're really pleased that they've chosen us to co-engineer and partner with them. I could give you many examples, but I'm not going to. They are just one example, we have worked together this year, just this year, with CSPs around the world of all sizes, and we have helped deliver 50 new cloud services built on and optimized for the Intel architecture.
Now that these services are out in the market, I want to talk about the third piece of our strategy, because once again, this makes us different in our capabilities that we bring to this market. We work with hundreds of cloud service providers via our Cloud Insider program. The Cloud Insider program, you can get technical training and support, industry insights, marketing and sales resources all in one place to help cloud service providers build industry-leading infrastructure. Today, we have greater than 650 members in the program enrolled, I remember about five years ago when we launched our first marketing campaign and this program. We launched it with Amazon, the big thing was we realized, well, wait a minute. If we brand Intel in physical compute and physical storage, cloud virtual compute and storage would also make sense.
People care about what's inside their compute and storage and systems. Amazon agreed, they joined us, they put the Intel Xeon Powered by logo on their instances, for the first time, showed everybody exactly which CPUs were running their Intel-based instances. That has just progressed to this program. I'm going to touch on a couple of elements of just how far this program has come. First, we invest. For 2018, we are going to invest about $100 million in co-marketing alongside and with our cloud service providers to grow the market and create pull for their services that are built on IA. You can see, obviously, this strategy is working. It's contributing to the phenomenal growth numbers that I showed you earlier, driving that TAM expansive growth.
In addition to just helping them sell the services that they have today, we're helping create the pipeline for future services. This year, we launched the Intel Advanced Technology Sandbox. I'll call it ATS for short. It is an easy way to get early access to pre-launch hardware for all of our cloud service providers who are in the program. This will include Cascade Lake and the Intel Optane persistent memory that Naveen talked about that's going to launch later this year because customers want to come in and use the sandbox to start their use case development. We've already tested 25 different workloads in the ATS since it's been launched, we have a huge pipeline ready for our cloud service providers to come in and continue to test on our new technologies. Unlike others, to complement this program, we have dedicated support.
Our customers can reach Intel engineers, Intel architects. They can reach me anytime they need. They know who to call and pick up a phone because we have those relationships, not just with the big cloud service providers, but with many cloud service providers around the world. They know who Intel is, we're on their premises, and they know who to call when they need help or want to innovate. For us, this is a win-win investment. When the cloud service providers grow faster, the market grows faster, and of course, our revenue grows faster. I'll give you a data point to think about. I've been tracking our engagements and our investments, for the next wave CSPs that we work with, when we make these investments across engineering and marketing and support, we see about a 30% revenue uplift from those engagements. Why?
We modernize the infrastructure and optimize, which enables selling up the CPU stack, helping them transition to the next generation technology more quickly than they would have, and we create pull for our adjacencies because we can show them if you design at the system level, you're going to get a balanced platform and better TCO. We prove this out with them. We don't just give them a marketing foil. We have proved out this value with them side by side with their engineers so that they know that when they buy these systems, they're getting the best. We are proud. I'm so proud of this one that even though Bart mentioned it before, I just had to pull up because it's just such a nice picture. There's Mark Hocking, he runs our Google worldwide account team, and Diane and her staff.
We were just recently, just a couple of weeks ago, it was the day before Google Cloud Next, at the Google Cloud Partner Summit. We received the Innovative Solution and Infrastructure Award, and we are going to win it two years in a row. I've already set my sights on it. This award, it's not just that it was an award and an opportunity to showcase our collaboration, but it really represents the massive body of work. We've been collaborating for over 15 years. In 2017, with the services launch, this massive body of work that we did, Intel, alongside Google with our alliance, to deliver and support their growth in the industry. I could talk for hours on this subject, and I do, but it's time for me to summarize. I hope you see why I do love my job. It is different every single day.
The opportunities are endless, we are just at the beginning of growth. I've talked about many things in my half an hour, but there's three things I'd like you to remember. One, TAM expansive growth continues. Two, if you want to get to the heart of real innovation and differentiation for cloud service providers, it takes deep collaborative partnerships. Three, Intel has unmatched capabilities in the industry to serve these cloud service providers going forward. We provide the broadest set of silicon productization. I remember I mentioned from standard roadmap to optimized, to semi-custom, to fully custom silicon and ASICs. We do it and are the only one in the industry that can apply that across compute, network storage, memory, and acceleration. We wrap it around it and couple it with software engineering investment, hands-on co-engineering to unlock value and drive and create their differentiation.
We back it up. We back it up with robust marketing programs and sales programs. We stand by our customers, and we go out and sell the value of services that are better because they're built on the Intel architecture. I want to thank you for the time. I also get to tell you that we are going to take a 15-minute break, and there will be refreshments served in the lobby. Thank you for your time, and have a nice cup of coffee. Thanks.
Let's do a little bit of a historical walk. I think it's necessary just for us to calibrate ourselves and get a view of the future versus trying to look at this industry by looking at your rearview mirror. The realities are the realities. Over 2014 through 2017, IDC predicted, and we saw a 6% decline in on-premise infrastructure spend. Our own business declined at about a negative 4% CAGR during that same period. In particular, units as a measure of consumption and Xeon SP or Xeon server units declined 12% in that period. The industry sentiment matched that reality. People wrote, some of you may recognize these things, that indeed, CIOs were increasingly stalling in their plans to spend on infrastructure. On-premise hardware spends and projected spends were declining.
Perhaps most importantly, there was almost a thought starting to creep in to the industry that this would be a systemic decline. There were questions like the Forbes quote on, "Are corporate data centers obsolete in the cloud era?" Was this the beginning of the end? We looked at the factors that obviously dictated the results that we saw. There was macroeconomic and somewhat geopolitical instability that caused GDP, especially in 2015, to hit a low. It was hard to transform your business, invest in the future when there was such large macroeconomic uncertainty around demand and instability driven by a number of political factors. The second was the cloud was here, and it confused many, many in the industry as to what its long-term use would be.
Would everything go to the public cloud driven by the massive scale and TCO advantages, the ability to test in dev, the agility that the public cloud offered? Would the end state be that everything would move there, obsoleting the need for corporate data centers? IT infrastructure declined, spending declined. The clip at which the past had refreshed and put in on-premise infrastructure in the data center stalled, and that was clearly a decline in the trend of increased spends before that. At this point in time, it's probably interesting to say I took on this role in 2015. Don't blame me for these trends. I didn't realize at that point in time, to be honest, that we were in the middle of a transformation. That transformation wasn't just our business transformation, but that transformation was a transformation primarily first in our customers.
The big enterprises of the world needed to serve their customers differently with new products and services, new ways to do things, and new levels of efficiencies and capabilities that they hadn't used in the past. Indeed, our customers were transforming. Our business had to transform to enable those customers to meet their aspirations. If you looked at it, again, in this period, if you talk to the hundreds of CIOs that I had the pleasure of talking to, it was indeed a different conversation. It was always a conversation more so not about the next product or the TCO benefits. It was always a conversation around strategy. What's our cloud strategy? What's our data strategy? What's our security strategy? Where do we place workloads? Indeed, even what's technology possibilities that could inform those strategies?
In this era of looking at the transformation while you're in it, we came up with some convictions. We believed through these hundreds of conversations and watching what technology could also do at the same time, that this business transformation was inevitable. As Naveen said earlier, automobile companies would transform themselves to being mobility services companies. Companies like John Deere would stop being farm equipment companies and would become AI-driven intelligent agriculture services. The list goes on and on in almost every vertical, and we'll talk about some of those with specific examples today. That business transformation had a very big impact on what we do in our business, or what we do in our business had a very big impact on that business transformation. Fundamentally, it was the era of IT is it. IT no longer supported the business. IT was the business in many ways.
It came home to us in three particular aspects. One is legacy infrastructure. In 2015, the install base was the average age of a server was about 5.3 years and rapidly growing. That legacy infrastructure would age faster. What do I mean by that? It would mean it would show its age faster as new capabilities were needed, as new levels of TCO had to be achieved, and efficiencies. The penalty for not going to the latest generation of technology was just higher. The second is, this was a very important part of our conviction, that enterprises would embrace clouds, but they would go hybrid. That the world would be one in which they had intelligent placement of applications and services in the public cloud and on premise, and they would adopt private clouds to bring the cloud economics even to the on-premise portion of their infrastructure.
The third was that AI would drive phenomenal growth as it drove a complete transformation in how businesses and enterprises did what they did and what they offered to customers. Their products and services would be AI inspired, and the way they operated their businesses would be AI advanced. With this as the conviction, as increasing clarity and focus, we then invested in our strategies. We had fundamentally three different places we focused. Instead of trying to return the industry, a daunting task, to the heydays it was, we invested to accelerate it to where our customers want it to be. There were three particular areas of investment. One was accelerating private hybrid cloud growth. Again, this was fundamental to infrastructure being more capable and more TCO friendly. We worked extensively with the leaders.
We worked with VMware on vSAN, vCloud Foundation, and actually created Intel Select Solutions based on optimizations for Xeon. We worked with Azure Stack. We worked with Microsoft to make Azure Stack, the on-premise portion of their cloud offering, feature consistent and run seamlessly between Azure the cloud and the on-premise private cloud called Azure Stack. Looking forward to an era where you would actually deliver and develop natively to the cloud, and using Kubernetes as a multi-cloud, not just a hybrid cloud, but a multi-cloud portability mechanism for the industry, all with the goal of easing the adoption and making the private cloud a reality to move forward in this transformation.
The second was an area we'll talk again in detail as we go forward, is expanding analytics, which was already in use with older generation technologies like rule-based systems, but expanding analytics to a broader portion of an enterprise's operations and offerings, and deepening it with AI. That is, making it more capable and offering new things with advances in artificial intelligence. Again, working with many that are centerpieces in the enterprise IT ecosystem landscape to drive this forward. Working with Cloudera, with SAP, with Spark, and most importantly, perhaps not just optimizing what we do, our silicon platforms, but most importantly, optimizing it for capabilities that were yet still to come and just in their nascent early days. The third thing we focused on was taking, in a period of rapid change and adoption confusion and concerns, accelerating time to value through Intel Select Solutions.
Navin talked about this extensively before me. The whole goal was to take what was the art of now clearly possible and make it deployable and consumable in the shortest amount of time. The whole point of workload optimized offerings, the whole point about efficiently managed and deployed infrastructure and scalable recipes. We went from reference architectures that we worked on for decades with the industry, with our supply chain partners, our OEM friends, and advanced that to include ISVs and SIs who had the deployment practices and the mechanisms to go make a real application solution for an end customer happen. We'll talk about that a little bit as well as we go forward. Our belief and our conviction as we saw this transition, this transformation come through, really define our investment priorities, and those investment priorities are working.
The three areas that we have seen tremendous growth, response, and an aspiration are still in their early stages. First is private cloud growth. Through the efforts of mobilizing the private cloud solution supply chain, if you will, what we've seen is a doubling of the adoption of private cloud on-premise in the last five years. Right? In 2013, 6% of on-premise infrastructure was in the form of a private cloud. In 2018, that's gone to 12%, and this is just the beginning. Part of that story is the infrastructure transformation and modernization that our customers, enterprises are doing in order to stay abreast and lead into this era of transformation.
The knowledge of what technologies can do, as well as working very closely with them on workload placement and workload mobility across public cloud and all of the advantages and great value it brings, as Raejeanne described, and what to keep on-premise, grow on-premise, how to place workflows where you would do test and dev on the public cloud and operationalize on-premise. That became a center point discussion and result in a far more clear picture for my CIO friends to go then have a targeted investment that would drive infrastructure upgrades. This is an IDC report, again. 80% of companies are reporting that they are actively working on workload placement, and some of that requires repatriation from public cloud to private cloud, even as they build out new services on the public cloud.
Contrast that to the sentiment 3 years ago, where many said a cloud strategy might mean I'm taking everything to the public cloud. The third area that our focus to drive the penetration of analytics and infuse it with the next generation of AI is also working. We see it in our own results. Between 2014 and 2016, the growth rate of CPUs on-premise that ran AI and analytics workloads was about 6%. That doubled to almost 12%. That will double based on what we see between 2017 and 2021. Again, based on the great value that we will deliver through the product lines and the platforms, but most importantly, driven by the need of our customers that matches that value. Results speak louder than any words.
If you look at the last 3 years' performance of the business, again, as it started, as it reflected the beginning of this transformation, we were down, the business was down 4%. The first half of this year, the business is up 6%. Perhaps most importantly, as I said, in the last look back, we had units going down, declining. In the first half of this year, we are seeing growth of units. In fact, Xeon server, Xeon SP, is a 1.5% unit growth in the first half relative to the year before. Again, based on all of the increasing demands of workloads, the value we've delivered through the CPU generations in Xeon in the marketplace. The entire goal to set up an infrastructure that can be systematically upgraded and expanded as these workloads explode.
In a nutshell, we see the business having stabilized. Obviously, that's both a financial statement, but perhaps more importantly, it's a statement around what we now see as much more stable trends, underlying trends in the business that's going to drive future investment in our business results. The period of confusion is largely over. The period of clarity, focus, and execution is here. Let's talk about what that future brings. As we look into the future, the Intel Xeon processor remains the heartbeat of the enterprise. It is the biggest, especially our recent, as we say, the Skylake processor, is the workhorse of this industry. It is the biggest platform innovation in a decade. 65% performance gains across the broadest range of workloads. Again, it is an enterprise, this doesn't do one thing.
And what it really values is that tremendous performance uplift across the broadest set of things it has to deal with. For some applications, like high-performance computing applications based on features like AVX2, the gain is almost 120%. If you're looking increasingly to deploy infrastructure in a virtualized environment, in a software-defined environment, leadership virtualization performance is key. 47% gen-on-gen virtualization performance or applications deployed in a virtual machine and 35% increase in VM density. A huge TCO driver for an enterprise IT CIO. Thirdly, it's legacy virtualization friendly. You can add multiple generations as you move through your upgrade cycles and expand and extend your data center and simplified manageability up and down the stack with consistency of features, of critical enterprise features like RAS. The result is, it's been the fastest ramp of a Xeon processor in the enterprise since Ivy Bridge.
As Naveen alluded to this morning, year to date, we're north of 50% of processors shipped to the enterprise being the Intel Xeon Scalable processor, AKA Skylake. Perhaps more importantly, it is the highest mix of top-end SKUs in a sustained manner. What does that mean? It means it's delivering value to the emerging and soon-to-be-established workloads in the data center that require high performance, require large memory capacities. Well matched the needs of the golden age of data. It's being widely adopted, it's generating great value, and we are getting paid for that value. It would be remiss on my part if I just said it was all about the processor. While the processor is the heartbeat, there's a body around it called the platform. It's tempting, and many people ask me, these are platform things you just plug into the Xeon, right?
If you look at something like Omni-Path Fabric, that's not only a critical part of our high-performance compute platform architecture and differentiation, but it's even now being used in emerging scale-out workloads like AI training, like Naveen Rao mentioned. If you look at our SSDs, they offer for the enterprise, in particular, a level of reliable latency reduction and bandwidth improvement along with capacity per dollar that is unforeseen or unmatched today. Dan will talk about FPGAs. This set of platform capabilities, co-designed with the CPU, is a unique platform advantage and a unique value delivered to our enterprise customers. Talking about unique value, let's talk about revolutionary capabilities. Much has been said this morning already about Intel Data Center Optane Persistent Memory.
Most people just have tempted to call it non-volatile, but if you're really kind of a follower in the space of the details, AKA, a memory nerd, you will understand that this is the true first persistent memory since the days of core memory, and yes, unfortunately, I do remember core memories. What it has done is it's fundamentally rearchitected the memory storage hierarchy. One of those examples is what I show on the right, SAP, a longtime partner that we worked with for many years now on Optane. A huge part of SAP's business, of course, and HANA, the value of HANA, is having a database up and running so you can run analytics. Unfortunately, with growing data sizes or dataset sizes, the cost of actually having to reload a database after some downtime event becomes disruptive to the workflow.
You could be in hours, several hours to reload a database. The value of Optane in this space is a revolution in how that workflow and its TCO is managed. Using Optane, we saw a 12.5x improvement in faster startup times, going from about 50 minutes on a 6 terabyte data set using DRAM and SSD, to about 4 minutes. Why is this revolutionary? Because it allows them to be consistent with their SLAs with far greater accuracy, thereby increasing the value delivered in those SLAs. It also allows them to have higher throughput in a multi-SLA environment. All in all, this is what they are clamoring to bring to market, and we're looking forward with excitement to unleashing this capability. If you look at it one step further, it's not just about up times.
What it does is actually through this re-architecture of memory and storage, where you no longer look at Optane as storage, but it's a memory tier, like Naveen mentioned, it allows you to do now things in memory real time. Real-time analytics, without having to take your database down as it's getting updated, becomes a possibility, and it becomes a value adder, and really an enabler of the value of analytics in the enterprise. The storage re-architecture also allows a very significant increase in capacity per socket because of the re-architecture and how much memory you can now put in the various tiers of that chart that Naveen demonstrated. First time, 3 terabyte total memory capacity per socket.
This is going to be a revolution for the golden age of data as more and more single dataset sizes versus distributed across an overall distributed system becomes the key to delivering performance and doing work that couldn't have been done before. The last thing I will mention, which has been mentioned before, but should be underscored, this does not happen without co-design and a deep partnership. This is not about putting yet another component next to a high volume standard CPU. This comes about re-architecting portions of the CPU, re-architecting portions of the memory device itself, writing the software to be aware of both, and sometimes even changing the software to accommodate the characteristics of that CPU memory pipeline is absolutely essential. That is what we've been able to do because of our decades-long partnership with companies like SAP in the ecosystem.
The next is the deepening of enterprise. Don't worry, I will not go through the whole architecture diagram on the left. That could get us through lunch. The key point is the enterprise's requirements are unique. The enterprise has an established workflow. It has a set of constraints, operational constraints, processes, policies, and commitments to multiple portions of the enterprise, whether it's an internal reporting, to a customer decision or to new products and services through marketing. It has to do that over a consistent data architecture and incorporate new capabilities on that consistent architecture without disrupting any of those operational flows. It isn't just about buying a DGX box and putting it on the side of the data center. We've got that feedback consistently from many customers, Mastercard, others in the banking industry, in the drug industry.
It's quite clear, for AI to be successful in the enterprise, you must run it on the infrastructure that runs the rest of the enterprise, especially drive value from an overall TCO for the workflow. UnionPay is an example, one of many, where they have a very important task of fraud detection as their user base grows and the penalties for actually not being able to deal with fraud is in the $ tens of billions this year for them. They looked at artificial intelligence techniques, particularly deep learning, to go look at solving this fraud detection problem. They actually did try GPGPUs. What they found was not only the performance, but the ability to integrate the solution into their operational workflow was best done on Xeon. First Broadwell and now rapidly expanding to Skylake.
A 60% increase in coverage, while 20% increase in accuracy amounts to $ billions of value created. This is a trendsetter, it's just the beginning for how the enterprises, through our work in the platforms, in the software ecosystem, and the tools we provide, bring this together into that necessary one single operational workflow. While we talk about all these innovations, we obviously are driving them hard into the ecosystem because the currency of business is time to value. Complexity and delays dilute that currency. We've invested in 13 solutions to date in just over a year in areas of infrastructure and private cloud, in analytics, artificial intelligence, and even high-performance computing. This is essential to drive the rate and pace of innovation at customer speed in the future. This is the way business will be done. This is the way winners will play.
Talking about the future I want to talk about one particular thing that's a testament to almost all of the things that makes Intel who we are for our customers. In 2015, I had the personal honor of being invited to the White House as the nation and President Obama handed out an executive order to create a National Strategic Computing Initiative. It was a bold journey to have a system in 2021 that would converge classical modeling and simulation, data analytics, and AI, and solve problems that have never been solved before. For instance, I'll give you just one example that won me over in two minutes, or actually 20 seconds. Today, cancer is treated by a cocktail of drugs. There's about 200 combinations. There are 200 cocktails, 200 by 200. It's a combinatorial nightmare.
Unfortunately, by the time you use computation to figure out what the right cocktail is for your pathology, the disease has won. Using AI techniques, we are able to look at all of the combinations and things from the past and learn and reduce that time to weeks, but most importantly, give a chance for the disease to lose. That is transformational. There are many such applications in which battery life, through the creation of new materials, learned from data which you can't really simulate or model. You can obsolete the term battery life. It could literally go on for years and outlive the device itself. This is the scope of changes that's possible when you bring all of the power of high-performance infrastructure and algorithmic advances like AI to the marketplace. The U.S. government and us are partnering to do this.
The Department of Energy, in particular, is partnering to do this, to take what is today's best capability. It's about 180 petaflop system, and take it to exaflop. That's a 5x almost improvement in overall performance for just two times more the power. Serve and create with us a software ecosystem that can make use of this tremendous scale of capability, and then drive it further down into the commercial marketplace. When I talked to the Assistant Secretary of Energy at the Department of Energy, the one statement was, "We are relying on Intel to come through with what it's done in the past so well." This brings it right back into our backyard. New microarchitecture, advanced interconnect, novel memory storage hierarchy, and high-performance converged software. This is what we are best at doing, solving customer problems, solving nations' problems, solving humanity's problems through technology and partnerships.
In summary, I said the third thing I would leave you with is why I feel good. Winning with three unmatched assets, zero distance from our customers, breakthrough innovations, and unmatched partner scale, built and honed over three to four decades of being in this business. We will win with unmatched capabilities, unrivaled scale, and a scope in an infinitely growing TAM that gives us the opportunity to bring the best of what we do to solve our customers' problems. The future of the enterprise, which was the title of my presentation, is bright. We are winning, and we are investing both aggressively and smartly to drive the transformation and make our customers the winners, and us come to the party along with them. Thank you.
I'd like to invite my friend and colleague, Dan Rodriguez, who if we were in the Air Force and had call signs, would be called Spiky.
Thank you so much. Thanks for that colorful introduction, Raj. I really appreciate it. Anyhow, you heard from Naveen and other DCG staff members really talk about our strategy and our vision in this data-centric world. What I'm going to do over the next 30 minutes is talk about how we're driving network transformation. Now, global IP traffic that flows across a network is expected to increase roughly 24% through 2021, according to IHS. If you think about that global traffic and you think about video as a simple use case, way back in 2012, believe it or not, we only watched roughly four minutes of online video a day from our mobile handset.
Into 2018, that number is expected to increase to 30 minutes. I don't know about you, but I certainly know a lot of people that watch way more than 30 minutes of video a day, and there's probably some of you in the audience. Not only do service providers face that challenge, they also face the opportunity and really the promise of 5G. With 5G, you're going to see a ton of new use cases, new services, and really new experiences. That's going to cause the network to become much more intelligent, flexible, as well as cloud-ready. As you would not want your 5G network to prioritize a smart parking meter over maybe a smart health monitor or perhaps one of your financial transactions. Because of this, the network needs to become something different.
It's no longer sufficient to roll out a series of fixed function appliances or single-purpose appliances. Rather, we must move the network to a series of general purpose servers using cloud technologies such as virtualization to drive the right scale, flexibility, as well as agility. Intel is partnering with our customers to deliver upon this network transformation vision, and we are winning in the marketplace today. When we talk about network transformation, our focus is really across all three networks: comms service provider networks, cloud networks, as well as enterprise networks. To better imagine this, let's take this simple example of when you download a YouTube video. The first thing you do is you connect to a cell tower, and at the bottom of that sits the wireless base station. Next, you do several hops across the edge network. You eventually find yourself at the core.
At that point, you are authenticated as a user, checked against your subscriber plan. You go to the internet backbone, eventually the cloud data center through several routers and switches, and then your video file is served up. Then it is streamed in the exact same way, but in the opposite direction, until you can watch your video. That is the opportunity that we're talking about today, and that network transformation opportunity represents a $24 billion network compute TAM by 2022. Our strategy, it was really built in collaboration with our customers. It was built with the comms service providers, the ISVs, as well as the equipment manufacturers. The first part of our strategy, it is fairly simple, it is again, to move the market from a series of these fixed function or single-purpose appliances to general purpose servers.
If you think about your PC just for a quick second, you would not dream of running things like PowerPoint, Excel, and email on three separate machines, as that would be completely absurd. If you look at the way the networks are rolled out today, that is exactly what they look like. Our job one is to invest in our architectures, to innovate in our CPUs, our IO, our memory, our FPGAs to ensure that we can run all four major workloads across Intel architecture-based platforms, being application, control, packet, as well as signal. Once you have this general purpose machine throughout the overall network, you have the opportunity to apply cloud technologies like virtualization. To drive virtualization into the marketplace and into the network, we partnered with the industry, and we created an ETSI project known as NFV.
What that means is network functions virtualization. The goal of that is to deliver virtualization on a network in both an open and a standard way. Since that time, we've seen a slew of network deployments. The third part of our strategy is to do this with the overall industry. To partner with industry standards bodies, partner with open source software projects, also partner deeply with the customers, again, those service providers, those equipment manufacturers, as well as the ISVs. Just to give you a quick example, signal processing. One common workload is media, and we partnered with the service providers to showcase how you can deliver broadcast quality video on a Xeon-based processor.
Furthermore, if you think about some of the new innovations we've been driving, we have just released our Intel Core processor with Radeon graphics, and we're working with the ecosystem to enable cloud gaming services using that platform. Overall, our strategy is simple but effective. It is to converge all network workloads onto Intel architecture. It is to virtualize them, to provide flexibility, then deeply partner with the industry. You can see that our strategy overall has delivered results when you look at these financials. The way we look at our business is really in two different views. The first view is the view that you see on your left. This is what we call our horizontal networking view. Again, think back, I talked about how we're going to transform all networks.
This includes the revenue we're generating from comms service providers, cloud service providers, as well as enterprise-based networks. In this view, over the past three years, we have delivered revenue growth of about 40%, and that has outpaced the market by roughly 5x and fueled our MSS from 8% to 19%. Obviously, there's still some room to grow in this market. The second view is the view that you see on the right, this is our comms service provider vertical view. We pay particular attention to comms service providers because we believe this is the area where we're going to see the greatest amount of network disruption and where network transformation will firmly take hold.
When at this view, we're showing our revenue CAGRs that is inclusive of the revenue we drive with service providers for their networking, for their IT data center spends, as well as comms service providers' public cloud hosting business. With this, in the last few years, network transformation has fueled the growth to over 22%, 31% in the first half of 2018 alone. This comms service provider business represents roughly 21% of all DCG revenue. We, with our strategy and the market dynamics at play here, are very confident that we're going to continue to deliver very strong financial results and growth for the corporation. We just talked about the numbers a little bit, and you saw the strong results. One of the things that really fuels our overall business is adoption of virtualization technology.
If you look at these two numbers I'm sharing, it's really at a high level, you see two markets. One is the fixed function or physical appliance market, the other one is the virtualized server market. You can see that both are growing. The good news here really is that Intel is actually winning in both market segments. Furthermore, in the market segment that is growing fastest, we are uniquely positioned and have a distinct competitive advantage, as we have been delivering virtualized solutions to the market for well over the decade and have deep partnerships with ISVs to ensure those solutions are running efficiently on top of Intel architecture.
In addition to this, we have a host of all sorts of innovative technologies to ensure that when you run your workload in a virtual machine, you can do it efficiently on top of our platforms. Because of this, nearly all of the network virtualized deployments to date are running based on Intel architecture. To continue to deliver on upon our network transformation strategy, we of course must invest in our silicon capabilities. Of course, our CPUs as well as our platform technologies. The CPUs, they are the cornerstone of our business. For networking, we have a very strong roadmap. We offer top-to-bottom scalability across multiple design points for our customers: performance, power, as well as price.
With that scalability as well as our architectural consistency across those platforms, our customers can seamlessly scale their software, reducing the amount of R&D they have to spend to get a product to market, as well as reducing their overall time to market. When you look at the top part of our stack, it is of course our Intel Xeon Scalable processor, which is our flagship product and delivers very strong performance as well as scale for networking workloads. One example of a workload of course is packet processing. Through the innovations we've made in CPU, IO, memory, cache, Ethernet, as well as FPGAs, we can deliver approximately 600 gigabits of packet processing throughput on a dual socket platform. This is up over 1.7x compared to our prior generation and over 40x over the last decade. Furthermore, we also have other products.
We have the Xeon D, which offers very strong performance per watt and performance density. You saw earlier today, Naveen showcase his ruggedized design for wireless base stations, which is a great example of where you can utilize Xeon D. Furthermore, we also have our Atom class products that deliver scalability into both power and price-sensitive segments. In addition to these CPUs, we have a host of platform technologies that really allow us to complete the overall platform. We have our Ethernet solutions that not only provide the right level of connectivity, but they also ensure that the packets are passed to Intel architecture in an efficient manner so we can drive very strong packet processing. It includes new technologies such as dynamic device personalization technology.
Furthermore, we have our FPGAs that allow our customers to add all sorts of unique algorithms right into hardware on their platform to deliver overall system speed up. We also have the new Optane DC Persistent Memory and Optane DC SSD solutions for networking our customers as well as a host of data center customers. Finally, we also have Intel Silicon Photonics. Intel Silicon Photonics will become a great asset in the era of 5G, as it will enable you to connect up remote radio heads to virtualize and centralize base stations, sometimes up to 10 kilometers apart, but do it at the speeds demanded from a 5G network. Our CPUs and platforms offer not only the right performance, but also the right scalability to help drive this network transformation.
Of course, it takes much more than just building better silicon products to really drive the sort of transformation that we're talking about today. As such, we invest heavily in software and we partner with ISVs out there, as well as many open source projects to drive innovations and ensure our platforms can be consumed very easily, but also in a standard fashion. For example, we collaborate with the Linux Foundation, and we collaborate on projects ranging from OPNFV, which stands for Open NFV, so we want to make sure that NFV is done in an open and a standard way. We also collaborate with something called ONAP, which allows service providers to place, manage, and orchestrate their workloads across the entire network. Finally, we also contribute to something called DPDK, or the Data Plane Development Kit.
This is a set of libraries that sit on top of Linux. The goal of these libraries is to ensure that you can move packets very efficiently. This is something that Intel originally open sourced way back in 2010, and now it has become a de facto standard in the industry for running packet processing on general purpose servers. Beyond software, we know it takes an entire ecosystem to drive this type of transformation. Thus, we have created a vast ecosystem program with over 300 members in it to drive network transformation. Those 300 members have stood up over 80 unique solutions, proof of concepts to support proof of concepts, trials, as well as deployments. Finally, you heard Naveen earlier and Raj talk about the Intel Select Solutions. We are going to build on that as well.
For this, we have 2 classes of solutions. We have 1 class of solution that is known as our Network Function Virtualization Infrastructure solution, or NFVI for short. What it does is it allows our customers to converge multiple virtual network functions onto a 1 platform to drive this transformation. This solution, of course, is made available through our partners. The second solution that we have is something called an ISS solution for universal CP. What this is this is a platform that enables service providers to deliver managed enterprise services such as software-defined WAN onto an enterprise prem, basically onto a virtualized server. This not only provides great opportunities for the service providers to drive increased revenue for them, but also provides great flexibility and agility for that enterprise.
I am confident that through these strategies, the software, the ecosystem, and these Select Solutions, that we can comfortably accelerate network transformation. With that, I'd like to shift gears a bit and talk about a future opportunity, which is edge computing. To kickstart that, let's play a short video from our customers.
With the advance of 5G and advanced machine communication use cases, we are seeing much more ultra low latency requirements coming from the cloud-based networks, as well as high bandwidth requirements. This is the inception of Edge Cloud. To deploy data centers at the edge of the network, Nokia has designed the AirFrame open edge cloud infrastructure solution. We are not pushing all the applications to the edge of the network, though. We keep centralized the ones which we can for efficiency and scale and distribute the ones that we must. Through our partnership with Intel, we have been optimizing our data centers for optimal footprint, power consumption, bandwidth, and latency. Over time, we believe that there will be as much compute at the edge of the network as there are in central data centers.
When you put together innovations on the chipset side, along with the requirements needed from the telco side, a lot of things can happen. The next generation of innovations from an applications and network services perspective from all kinds of developers will be enabled. We've been working with Intel on Xeon and the next generation of it, because when you get closer to the edge of the network, the ability to scale and also be power efficient is super important in that case.
All right, you just heard a video from a couple of our customers. Let's just talk about the edge a bit more. When you think about the services and use cases of today, most of them are really delivered through the cloud, with the network as really just a simple transport mechanism. That is all going to change in the later stages of LTE as well as 5G. We expect a diverse set of use cases, again, to flow across those networks. It could range anything from media, which could include cloud gaming to video streaming. It could include industrial automation as well as smart retail. With those diverse use cases come diverse requirements. Some customers may, or use cases may need high bandwidth, others low latency. Some may need both.
Some may have a greater need for privacy, others may need to store data locally. To deal with this, comm service providers are not only going to build up the capacity of their networks, they're also going to place a greater amount of compute at the network's edge, much closer to the user. In addition to this, they're also going to utilize artificial intelligence and analytics throughout the overall network to not only automate the network, but also ensure that they're delivering the right quality of experience for unique services. This edge of the future, it will include both on-prem and off-prem equipment. Take the example of a smart retail store. You're going to have all sorts of sensors in that store. They're going to throw off terabytes of data a day.
There's going to be a smart IoT edge node to ingest that data and make smart decisions locally. Furthermore, you could see tons of different devices, robots, as well as control systems in an industrial factory floor. If you go off-prem for a moment and you think about the service provider network, you can think about a base station. That base station located roughly a mile from the user will be transformed. It will be virtualized and will have greater amounts of compute for intelligence. Finally, one step back behind the base station sits a node called the central office, and it is a great example of an edge node and is a great example of a node that can be transformed. Let's talk a little bit more about that.
Here you see a map of the U.S., the light dots represent where a service provider would typically have central offices. The dark blue really represents open spaces where you may have less population density, of course you have less central offices. You can see the different building types. You could have a small central office in rural America, a mid-size building, perhaps in suburbia, a big building in a big city like Manhattan. In terms of the types of compute you would see, you could see anywhere from a single system to a rack of servers to racks of servers. In terms of the size of the opportunity, there's 20,000 COs alone in the U.S. and over 70,000 in China. The great thing about this is the industry recognizes opportunity.
They recognize the opportunity to transform and modernize this, they've created a concept called Next Generation CO. You can think about this Next Generation CO, as really a mini data center at the edge of the network that is fiber rich. It is a mini edge data center that's capable of converging multiple workloads, both mobile, residential, as well as enterprise. Furthermore, it'd be capable of supporting up to 80,000 subscribers compared to a typical CO of 5,000 today. Finally, you can think about it as a local edge data center that has much smaller and power and area footprints as compared to a traditional or hyperscale data center, but still offers the same sort of scale, agility, and flexibility.
Service providers, they not only view the network edge as a place where you're going to run these network functions, but also a strategic location in their network to deploy IoT, 5G, immersive media, and much more. If you go back to the roadmap discussion that we had earlier, you can imagine that a Xeon Scalable processor could be utilized in that big city like Manhattan to deliver the right capacity. That Xeon D processor could be delivered in a suburban central office location. Again, because of that architectural consistency, service providers can easily place and manage those workloads. Furthermore, because of the seamless scalability, they can reduce their overall time to market as well as R&D. To better imagine what this next generation CO looks like, let's watch a short video, which is really a demo of this concept.
By transforming NGCO into a mini data center, comm service providers can support a variety of mobile and fixed networks on general purpose servers rather than dedicated fixed function appliances. This next generation CO becomes a pooled resource for virtual network functions and for services to be deployed by the service providers behind the enterprise, residential, mobile, and cable access points. Multiple virtual functions can run on a single processor or can scale up to multiple blades or racks when demand requires it. This provides incredible flexibility on how you provision your network and brings the same economics that drove cloud computing to the network edge. A legacy static configuration would inefficiently provision individual appliances for each network function without sharing of underlying resources or future-proofing. To better understand the efficiency gained, let's look at an enterprise, residential, and mobile network over 24 hours in a fixed mobile convergence scenario.
Aside from a few night owls, the demand is low across all the networks in the early morning. The platforms supporting the various network functions are in a low-power state. As morning approaches, the mobile network ramps for a short period, but as people get into the office, the enterprise network demand ramps significantly. Additional platforms are brought online to support the enterprise provider edge, which is now supported across three platforms. This flexibility means more efficient reuse of deployed hardware. High enterprise demand continues during the workday, but as we move into the late afternoon, the demand reduces as people return home. Now the NGCO reallocates the platforms to support a greater residential BNG throughput as the demand on the enterprise provider edge falls off.
This pooled resource approach replaces a number of fixed function appliances in this location and delivers the efficiency and flexibility we now require. We can also take a look at a network use case where a video-on-demand service is streaming from a number of local edge NGCOs and from a central location. Network analytics identifies that it is now more efficient to host the streaming service locally. The local NGCO has enough processing capacity. The video streaming service is then provisioned on the local NGCO. The connected subscribers are supported locally. Bringing the service closer to the subscribers improves their experience and also reduces the load on the backhaul network, allowing service providers to more efficiently manage their traffic. The transformation of the edge will provide necessary compute power, agility, security, and speed to support these edge services at the right quality of experience and the right cost.
You just watched this video, so hopefully, you now understand both the opportunity as well as the value of a Next Generation Central Office. We of course, we cannot talk about network transformation without talking about 5G. 5G, it will be a historical inflection point in the industry where it's going to deliver massive amount of compute power, seamless connectivity, really to every person, device, and thing. If you think about the prior standards, 2G was just about voice. 3G introduced mobile data. 4G was about faster data rates. 5G, yes, it will be about faster data rates and more capacity on the network. It is really so much more than that. It will provide comm service providers the opportunity to deliver all sorts of unique and tailored services to many different vertical markets, some which I've talked about today, industrial, retail, et cetera.
Because of these vast array of use cases, we believe that network transformation is truly required and foundational to deliver the 5G experience. For this, Intel is uniquely positioned in the industry. We have assets in the devices or the things. We have assets, of course, in the network, and we've been driving the enterprise and the cloud for a number of years. Because of this, we are a great partner for the industry to collaborate with. With that, we have driven with our partners over 25 5G trials. Due to those trials, plus the momentum that we have on LTE networks, the design wins that we've won in 5G, Intel is tracking to the leadership position in 5G wireless infrastructure. To better illustrate how network transformation is truly required and foundational for 5G, let me share a quick example or two.
We collaborated earlier with China Unicom, Nokia, and Tencent in Mercedes-Benz Arena. What we did was, we enabled an amazing experience for concertgoers. Essentially, even if you were in the worst seat in the house, you could view the concert as if you were in the best seat in the house by simply opting in on your mobile device. You were then treated to view the concert from multiple real-time video angles throughout the overall arena. Furthermore, you had the opportunity to enjoy social media and do real-time video chatting with a celebrity right in the stadium. All of this was made possible because you put Xeon-based servers at the network's edge and very close to these users. What that did is it enabled this video service that I'm talking about to even be possible.
It cut the latencies from 30 seconds down to half a second. Furthermore, because you weren't shipping all that data back to the cloud, it reduced the cost of that backhaul by 80%. That was just one use case, a smart stadium use case. You can imagine in the 5G era how utilizing 5G compute can be applied to all sorts of latency-sensitive applications. Again, think about that factory floor and think about those latency-sensitive applications. You've got a ton of devices out there, a ton of robots, as well as a ton of control systems, some of which are using compute vision, which is a resource-intensive workload. In addition to this, we also, with our partners, really drove the largest 5G trial of them all, which was, of course, the Winter Olympics.
In that trial, we stood up a network with over 3,800 terabytes of capacity, over 22 5G links, and over 10 sites. The experiences at the Olympics were built on top of a virtualized infrastructure running on Intel. We believe the flexibility and scale that network transformation is truly required and essential for 5G. In essence, to really deliver the promise of 5G and all these experiences that we've talked about, you must transform the network. In summary, we have a massive opportunity in front of us. It's a $24 billion network compute TAM by 2022. This is an opportunity that Intel is uniquely positioned in, where we have a distinct competitive advantage. The strategy that we have developed, created, and executed, it has driven financial results, driving a revenue CAGR of over 40% over the last few years.
Furthermore, if you think about the last two topics, the edge topic, as well as the 5G discussion we just had, those are not only great growth opportunities for Intel, but they're also accelerants for the underlying network transformation strategy that we have. Finally, we will continue to innovate on our technologies and we'll continue to partner with our customers to drive wins in the business, and we will continue to win in this marketplace with network transformation as a backdrop and continue to deliver significant financial results for the corporation. With that, I'd like to thank you and then welcome Navin back on stage for a little bit of question and answer. Here we go.
Okay. I think we've got mic runners here. We'll just start, in the audience. Go ahead, Trey.
Thanks, Navin. It's Ross Seymore from Deutsche Bank. Lots of great presentations about the breadth of your DCG capabilities today, a lot of concern on Wall Street about the lithography aspect that you mentioned. Talk a little bit about what your customers' reaction is to the elongating of the 14 nanometer node, pivoting a little bit off that, what's your view on the competitive intensity you're going to feel as we look into 2019 versus 2018?
Sure. I don't talk to customers about nanometers. At the end of the day, what they care about is delivered system-level performance, and they care about it on the workloads that matter to them. As you've probably heard and may have taken away from this morning, this breadth of need to optimize for workload after workload after workload comes through a multitude of factors. Microarchitectural innovation, software investments, and broadening, quite frankly, our portfolio beyond the microprocessor to new domains. The Optane persistent memory example is a very good case of that. To solve in-memory transactional database workload problems, we couldn't do that with just the microprocessor.
The conversation we've been having with our customers as we've been working through the roadmaps is really about, hey, our job is to deliver you a consistent level of improvements in performance year after year after year, just like we've done for 20 years. The roadmap we laid out to them several quarters ago, and we told you about today, does that. Look, I think in terms of competitive intensity, we walked into this year expecting a competitive year. We expected the dynamics to be competitive. It's not new to us. We're used to competing. You've seen that the team has done a very admirable job in outperforming really what we had expected as we came into this year. Three quarters of 20% year-on-year growth, was not what we had expected. That's a function of really three things.
It's a function of the TAM environment being good. That's a function of the products that we've put forward with Xeon Scalable and the broad array of our adjacent products, and that's a function of our expansion into new domains like AI, and we talked a little bit about that today. We're not confused about competitive intensity. We're not scared by competitive intensity. Our roadmap and the products that we're putting forward, gives us confidence that we're going to continue to win. I'll take one from this side.
Thank you. Navin Ambrish from BMO. We were talking about it a little bit during the break. I wanted to get back to the 10 nanometer and microarchitectures.
You were explaining that, I hope I get the nomenclature right, Ice Lake will be a new microarchitecture. Along with a node shrink, you're having a microarchitecture change as well. If you could please help us understand the cadence that you've had in the past when you do that, also when you shift from client to server this time versus what has happened in the past. Just kind of related to what Ross was asking, concern in the investment community is that, what are the performance gains, you highlighted some of that, it's not just about a node. You have Intel DL Boost, I forget some of the features, including persistent memory. What are the performance gains for most workloads that you will have between now and 10-
with the two microarchitectures that are coming in between? Thank you.
Let me see if I can remember all of that. First of all, I'd tell you that over the last several generations, we've been driving microarchitectural improvements within the node as well as when we transition to a new node. For example, if I go back to Broadwell, the way I think about Broadwell is we added new microarchitectural enhancements to that, even though it was the first product on the 14 nanometer node. The notion that we can't push and drive microarchitectural improvements on node transitions is really not a fact, and we've been able to do that over the last several generations. I don't think we've really gotten into too much detail about the microarchitectural enhancements on Ice Lake on 10 nanometer. We'll, as we get closer, talk a lot about that. It's a great product. It's going to deliver amazing performance improvements.
Winding back from that to our 14 nanometer portfolio, Skylake was a big microarchitectural bump. That's what led to us calling it the biggest advancement in the last decade, and the performance advantages that I highlighted when I spoke this morning. Our follow-on to that, Cascade Lake, adds new capabilities. They don't always have to be at the microprocessor level. Adding a new memory controller that enables us to deliver Optane persistent memory at the platform level delivers massive performance improvement. We talked about some of the examples today, 8x on certain memory intense database workloads. Quite frankly, we don't fully know yet how our customers and our partners are going to innovate around things like Optane persistent memory. It is a breakthrough, kind of revolutionary, once in a decade kind of change. I don't expect it to be a niche thing.
I expect it to be broadly used as customers figure out how to take more and more advantage of it. You asked about the client and data center dynamics. Historically, there was a year, a year and a half lag between our client introduction of products on a new node and our data center products on a new node. That gap and time is going to reduce. We talked in our earnings call about having client products on 10-nanometer on shelf for holiday of 2019. We will have our data center products in market in much less than the typical lag of about a year, with the Ice Lake product you referred to. We'll come back to this side.
Thank you. Navin, this is Srini Pajjuri from Macquarie.
Oh, there you are. Yeah.
Yeah, I'm here.
Sorry. These lights.
Yeah, I have, again, a competitive landscape question. I guess one of the concerns that we all have is that in the last five years, you had very little competition, but historically, at one point, AMD did have 20%-plus market share. Can you compare and contrast what's different about this time versus the last time AMD had 20%, 25% market share? Within the subsegments, enterprise, cloud, and comms, where do you think or where do you see most competitive intensity, or where do you expect that?
Sure. I think the biggest difference in the way we see the world now is the TAM we're going after is much larger. Five years ago, you could have said, "Hey, we're defined by the PC and server market as a microprocessor supplier." That's not the way we see the opportunity we're going after now. The $200 billion TAM that we outlined today is much bigger than we've ever gone after before. Our mindset, again, is that we have 20% of that market. I think it's important to recognize that the investments we're making to expand the portfolio we have is crucial to the way we think about the way we're competing and the way we're addressing our customers. We don't talk to them about just microprocessors.
We talk to them about silicon photonics, and we talk to them about SmartNICs, and we talk to them about Optane persistent memory, and we talk to them about FPGAs, and we talk to them about custom ASICs. That breadth of portfolio we have and our ability to stitch those capabilities together to deliver higher levels of performance is a big time difference between the way perhaps we were viewed five years ago to the way that at least we think about ourselves now. Competitive intensity, I would argue, we don't think of ourselves, and I've been here 23 years, I have in my brain paranoia. I don't think about periods of time where things are less competitive and when things are more competitive. We're always paranoid about anybody that's out in the market.
It's not as if all of a sudden we're waking up going, "Oh, wow, the competitive intensity has now increased dramatically." Right? We've been nervous about other suppliers in the market for many years. Between the segments, I really can't give you a perspective that the competitive intensity will differ. The competitors might differ a little bit, right? The way we look at the network infrastructure market that you just heard about from Dan, we'll have different competitors in that market than we do perhaps in the cloud and in the enterprise. Take one from this side.
Yeah, thank you. Aaron Rakers at Wells. I'm not going to ask about 10 nanometer. I'll skip the discussion a little bit. Clearly, the message here is your platform vision and everything that you're doing around the CPU. I'm curious of, number one, if you could help us understand when Optane becomes material. How much of revenue are you doing today within the non-volatile piece of your business, and when do you expect that to material accelerate? I'm curious, there wasn't a whole lot of discussion around FPGAs. I'm curious of how you see that fitting in your strategy vis-à-vis competing, it sounds like much more aggressively against, say, NVIDIA in AI instances. Thank you.
Yeah. On the second question, is Dan here?
Yeah. He's doing his afternoon thing.
Okay. Oh, there's Dan. Dan's going to talk this afternoon about FPGA, so that's one of the reasons why, but I'll let him answer briefly, so you can get a perspective on that. Sorry, the first part of your question was?
Optane.
Oh, Optane. The TAM we're going after is huge, and I'm going to break Optane into two segments. There's the SSD storage portion of Optane. That is reported in our NSG segment, and that business has been ramping. We believe that it's differentiated and will drive superior gross margins to the standard NAND business over time. The Optane persistent memory business is new, right? We just literally shipped our first revenue units yesterday. It's a $10 billion market as a subset of the overall DRAM market. I'm super optimistic about it. When does it become material? We'll talk about that as we ramp, but we have zero of it now. It's a $10 billion market. You can see what's happened in the DRAM market, the intensity our customers have in terms of consumption of DRAM content.
You can imagine that there is a high degree of interest from our customers in making this a real large-scale business over the next several years. Dan, do you want to answer on FPGAs?
Yes. Thank you for the question on FPGAs. It's not just an NVIDIA story for FPGAs. FPGAs are in the data center for a number of different reasons that I'm going to go through in the afternoon. If you think about FPGAs, it's really more of an inference play than a training play, and I think Naveen talked about the broad portfolio for AI across training and inference. I'll talk a little bit about the features for FPGAs. Very low latency, and the efficiency and the flexibility to react to all these emerging topologies is really the value. We're doing extremely well. I'm going to show an example later with Project Brainwave, which is their inference solution based on Intel FPGAs. We're also playing very strongly in the infrastructure portion of the data center.
Without stealing the thunder of later, we're seeing we've grown 140% in the first half of the year in FPGAs in the data center. Very good growth, very good traction, and we expect that to continue. Again, it's not just a head-to-head with NVIDIA. The last piece I would say is when you look at the two areas where FPGAs play in the data center, FPGAs can do both sort of what I would call lookaside acceleration and infrastructure. GPUs are only a lookaside play. I'll give you an example later of why that's so beneficial from an FPGA standpoint. Thank you.
All right. Thanks. Oh, there's one there.
Well, I think we're going to call it there so we can get everybody to lunch. For those on the webcast, thanks for joining. We're going to be back on the webcast at 1:30 P.M. For those here, we're going to walk to lunch, and we'll help you get there. Right out back.
Thank you, guys.
Welcome back, everyone, thank you for the great conversations over lunch. Obviously, they were very good because we had a hard time getting everyone out of there, I hope that everyone grabbed their desserts to go. That was a quick little hit there. This afternoon, we're shifting gears. We have spent the morning going over the business and the opportunities that we see walking through the TAM and our opportunity to address it. Now we're going to go in a little bit deeper into some of the portfolio elements and the silicon foundation that all of that TAM addressability is built on. We're going to kick off right now with an overview of our CPU architecture, and we're going to have Sailesh Kottapalli right here.
He's going to come up, he's our Intel fellow and our lead data center CPU architect, his team is the world's most advanced data center architecture team on the planet. I work very closely with Sailesh all the time. Our teams work together, he has a ton of great insight and customer insight, I think probably more of you have an opportunity to hear me speak than you'd actually have heard from Sailesh. I'm delighted to have a chance for you guys to all hear his insights and what he and the team are working on and how they think through these data-centric problems. Without further ado, I'll hand it over to Sailesh.
Thanks, Lisa. All right, there you go. All right.
Good afternoon. I think through the talks at the beginning of the day, you must have seen all the industry standard benchmarks, some of them that either we have published in our programs or our customers have published. I'm actually not going to talk about any industry standard benchmarks because fundamentally, when we build the products, there is much more we look at to make sure how we deliver performance in a data center across real customer workloads, which are running at scale, using large data sets, which usage models which are specific to the particular customers, and deployment models which change from data center to data center.
There are a number of factors that are very critical to deliver performance across that kind of a constraint, I'm going to talk about a few of the aspects that are critical, that I found to be very critical and vital. In order to actually set this up, what I'm going to do is actually introduce a couple of terms which actually people don't see generally when they're looking at a data center performance. The one thing that's very critical is what we call per-core performance. That is essentially the performance that a particular application instance sees when it's running on a core. Right? That is the performance that it actually realizes. That's what we call per-core performance.
The other one is the throughput performance, which a lot of people get, which is really the cumulative performance across either a processor or the entire server node. That's what we see as throughput performance. On top of that, I'm actually going to introduce a couple of qualifiers for these, which are actually very critical in how that plays out in the Data Center. The notion of a minimum per-core performance. When you're actually running a particular application, what is very critical is that all of these applications have certain performance expectations, some response time expectations, latency expectations, and so on. The minimum per-core performance defines essentially what is the level of performance that you need in each of the core when the entire server or the processor is running to be able to meet the SLA requirements or the latency requirements.
Anything less, you would actually not meet that requirement, right? With that, the qualifier that I'll introduce on the throughput is what we call effective throughput, which is the amount of cumulative performance that you can deliver when you have a service running with multiple requests, and each of those requests actually meeting the minimum SLA requirement or meeting the minimum per-core performance. What happens is that you can actually drive a higher throughput, but if essentially none of the cores actually meet that minimum requirement, the effective throughput at that point goes down to zero. You'll have to make sure that you maintain that per-core performance. Okay? In general, when we look at Data Center products, you really have to make sure that you're maintaining the balance of both of these things.
One of the things that we internally spend a lot of time in making sure people understand, why is per-core performance important? When actually you see industry standard benchmarks, the performance numbers that you see, they're actually what we say the maximum throughput that a particular processor can deliver. That actually has no bearing on what is the performance that every thread, every core was delivering. I'm going to talk a little bit about why that is critical. I gave one example of a response time. For example, you're doing a search on a web. If the search takes longer, that changes the response time, and people who provide search services actually monetize the response time.
It's the same thing as you're actually doing a trade on the Internet, and if it takes longer to trade, it's typically hard to get value out of it because anything that process faster. That's the response time. That's a very critical factor, and it's important to meet that response time. The second aspect is the notion of elastic compute. The infrastructure that has the ability to address all the different per-core or minimum SLA requirements of any workload. If it can address the entire spectrum, you have the best elasticity in your compute, so you can run any workload on that. If you have a lower per-core performance, the kind of workloads that you can run on it become limited. Right? You don't get the elasticity because you might have to deploy a different infrastructure to deal with something else. It gives you the highest elasticity.
The other thing that we have realized with the experience of deploying our processors in large data centers is that actually the performance at scale depends inherently on the performance of the core. If you're trying to actually scale your data center to a really large deployment, you need to make sure that every server is providing a good performance and every core is actually providing a good per-core performance in order to get scale and being able to increase scale. I think in general, a lot of people understand the Amdahl's law, which is essentially, if you're running a parallel program, it will always be limited by the serial portions at times. How fast you can actually run through the serial portion defines the performance across that entire parallel.
That serial portion, essentially, how fast it runs comes down to what kind of per-core performance you can provide. If you're in an enterprise kind of a setup, you're running a lot of ISV software, and all of these have software licensing constraints, and most of them are generally per-core licensing. The amount of performance you can get per license, again, depends on the per-core performance. Even if you go look at the public cloud providers, when they actually publish, "Hey, here's the expectation of performance in our public cloud," effectively what they're publishing is the performance of a virtual instance, and that actually is a per-core performance. That's not a throughput. Throughput is what they use to understand how to get the highest performance per TCO of the infrastructure, but the actual customer that they publish is really the per-core performance.
In general, this becomes a really critical factor. It's important to make sure that you deliver the right per-core performance and then establish throughput across that. When we are building products, that's a big consideration that in our architecture as to how we deliver both per-core performance and throughput. There are a number of things that we do to actually make sure that we get good per-core performance. I'm going to walk through some of those factors. The first one is basically you need to make sure that the actual CPU core performs really well. We actually build a number of things into our cores to actually deliver high performance. There are a number of factors that I talk about here.
The first one that I want to touch upon is the microarchitecture, which actually delivers what we in the technology space call it the instructions per cycle that represents how many instructions you can actually just process in parallel inside the core. There is a fair amount of microarchitecture tuning that we do to make sure that we can actually process three, four instructions per cycle and beyond. One thing I'd like to point out here is that, in general, if you look at the industry-standard benchmarks, none of them fully reflect the kind of complexity that's actually going on with the real data center workloads. For example, for most real-world workloads, you actually see that their code footprint that they're actually operating on is massive compared to any industry-standard benchmarks.
In order to optimize your core to actually deliver on workloads that have a large code footprint is a separate optimization point than what you would have to do for just industry-standard benchmarks. Those are the kinds of things we look at. We profile what's going on across all of these data centers, identify the microarchitecture, and then we drive the IPC improvements. The other factor is design. A lot of times people forget that. We have world-class design of our cores, which a lot of people don't understand. We can achieve frequencies of five gigahertz. By the way, IPC frequency, these are all performance terms which are exchangeable. If you have a really high-performance core, high-frequency core, that actually delivers performance. We try to make sure that we can deliver the frequency at low voltage with low power and so on.
The other thing is, as we look at all these workloads, what we try to identify is the type of processing that's going on, and we look at our instruction set architecture and identify, are there new instructions? Would it actually perform the same task with lot fewer instruction than what we have already built? Like, we talk about AVX. That's one of the example. A lot of people actually think, for example, the AVX, the vector, people associate it with floating point. We actually support integer vectors. One of the things that we were able to achieve with AVX and the kinds of instruction semantics we support with AVX is a massive speed-up with in-memory databases. All of the folks driving columnar in-memory databases were actually able to use the AVX capabilities to get multi-fold speed-up on our Xeon processors.
These are all the things that we do across the entire spectrum. Of course, there is a process that gives us speed up on the frequency on top of this. That's what we try to do to make sure that we have the best core performance. When it gets to a data center processor, that is the starting point. The other thing that's extremely important is really how well when you actually pack the cores, the performance across all those cores scales.
We internally, when we build a product, we look at a number of concepts of how we want to build a product, and I can tell you that you can easily build a product that has a scaling efficiency of 60%, which means you add 2x the number of cores, you get 1.6 of performance, throughput performance, or you can build a different kind of a product, which actually gives you 1.9x for 2x increase in cores. There is a very significant difference in how you actually scale the performance across the cores that you have on the processor. There are a number of factors that actually go into how you actually deliver that optimal performance.
The way you actually connect all these cores, how you actually manage the data movement, the movement to the caches, and so on, across all these cores is extremely important. What we did with our mesh architecture was the evolution of what we had done over the three or four generations. We had moved to a mesh architecture, which actually is really scalable architecture, and essentially across a really large core count, it allows all of them to be driving their traffic in a sustained manner without actually getting in the way of each other. That is extremely critical, right? The other thing is the cache hierarchy plays a very critical role in all of this. It is very effective in making sure that you can actually increase your effective latency that your core sees, and as the latency increases, the core is sitting idle.
If you can actually service it from the cache, you can actually get the core running faster with high utilization. The cache hierarchy that you build is very critical. We made changes in our cache hierarchy that actually drive very high performance with a large cache near the core and then a shared cache that can actually share contents across all the cores that we have. Right. How we manage the entire power of the SoC is another big factor that defines how you scale with power and so on.
Essentially, how optimally you can manage. You can assume a data center processor running a mix of workloads across all the cores, and you need to be able to identify what kind of workloads they are running at what level of utilization, and are you able to move the power from a core that is actually running at low utilization to a core that actually wants to run really fast with a lot of vectors and so on. How you optimally manage that delivers the highest throughput. The multi-core scaling or scaling within the processor, it depends on the actual SoC architecture. That is a big factor. There can be a huge difference between 1.5x kind of a delta. There is a significant. That is an important piece. The other piece is, I think when we look at the processor, we don't stop at just one processor.
We build processors which actually can scale one, two, four, and eight in a glueless fashion. What we try to do is make sure that as you actually bring in two sockets, we actually have a really optimal scaling. A four socket, same thing. That optimal scaling actually extends beyond that. You can think of as we add more cores in a server node, we actually maintain that scaling efficiency. Right? That is the critical piece of how you actually architect it to maintain that scaling efficiency, and that gives you throughput scaling across a really large core count as and when you need it. Right?
There are a number of things that we do in our architecture to make sure that the way we process the data across sockets and the traffic, the way we manage it, the coherence processing, actually gives us the most optimal scaling. That's actually one of the things that people don't talk is the way or the strength of that optimization is really why two-socket has been standard because people have not seen or going from one socket to two socket, the performance falls down, they would have gone to one socket. We actually have built an architecture that actually scales two socket. The same thing scales up to four and eight sockets. Depending on the solutions, we can actually build core count scalability all the way up to a very large number within a particular server node inside a server box. Right?
Underneath all of this is really the performance of the memory, how well you optimize that, and essentially when you have a server box with a large amount of workloads running on top of it, how you deliver low latency and effective bandwidth is a very critical aspect. We've been studying that along with our customers. With our customers on essentially the traffic patterns, what kind of traffic patterns that usually show up in a data center and try to optimize our memory subsystems to deliver the performance across this. All of these factors are super critical to ensure that you have really good per core performance.
That is, any one workload or any one thread running on it actually sees the optimal performance, and we have the right scalability to actually give optimal performance or linear scaling across cores in a processor and then across processors in a server node. Right? That's an important piece to keep in mind, how we expect to keep delivering leadership in both per core performance and throughput with the optimal per core performance and scaling architecture that we have so we can continue to actually drive that. Okay. That's one of the factors.
Let me actually touch a little bit on a second theme that actually shows up in pretty much every conversation that we have with our customers, especially all the large data center customers we have across different segments, whether that's cloud, comms provider, enterprise, and so on, is really when you're building a data center, a large data center, a big factor is what is the utilization you can get. If you're really setting up a large warehouse scale data center, you want to make sure that you can actually run a ton of workloads on that at scale, right? You can actually drive up the utilization.
There are a number of factors that are critical in achieving really high utilization in the data center. These are things that you actually have to understand by profiling the data center, understanding what are the bottlenecks to identify what are the architecture constructs that are needed to deliver this. Then start building into the product. We've been doing this for about four or five generations to understand what is needed at scale. We have a number of capabilities. We have internalized what is needed. We've been driving optimizations through that. I'll talk through a few of those. Let me first talk about the notion of consolidation. I think that's fairly standard. I think in enterprise, it started with if you want higher efficiency, you actually start orchestrating more workloads onto the same server. The process is the same.
What changes is that each customer's usage model, deployment models might be different. We've understood, and we've driven features in there that helps us drive efficiency in being able to consolidate or orchestrate more. There are a few things on that. Virtualization is a technology that's been there. We've supported it for a while. We continue to actually drive capabilities in there which are critical in a large-scale data center. Like you need to be able to live migrate a running work from one system to another, either because of a failure or because of load balancing and something, as opposed to stopping it and restarting it. Right. The size of the VM that you deploy. Right. The VM density that you can, and the density, meaning something that you can launch without actually seeing performance variabilities and across a large VM that you deploy.
Those are all really critical. The other portion that's really critical for consolidation when you're actually trying to drive a lot of workloads on the same server is you need to have profiling mechanisms that can actually tell you how your server is responding, and you can orchestrate it accordingly. We have capabilities to actually monitor the cache utilization, memory bandwidth utilization, and so on, that the orchestrator can look at and actually move workloads around. The other piece of consolidation is really the availability. When you're actually running a lot of workloads, you want to make sure that from an error perspective, there is right isolation between all the different workloads in a virtual environment.
We have technologies like Intel Run Sure Technology that actually limits the exposure, limits the error radius and so on, and actually makes sure that the server keeps running, which again, adds to your efficiency. The second piece of the same thing on utilization is around the performance consistency. When you are actually trying to run many workloads on the same server, one of the key thing that is critical is to make sure that you can actually provide what we call performance isolation in the sense that one workload is not adversely affecting another workload. There are things that we do in our architecture, specifically under the whole Resource Director Technology or RDT technology, that helps us limit the impacts.
Once you have that, what you're able to do is you can be more aggressive in how you orchestrate, because you know that there are technologies in the process that will actually reduce the amount of interference between the workloads. There are capabilities that we have driven in that helps people drive the efficiency. With the resource director kind of technologies that we drove in, some of our data center customers actually saw high single-digit improvement in utilization just from that one capability. The other aspect, when you're actually running the workloads in a data center and you have a large-scale service that you're providing, the jitter or the variation of performance of that service is a very critical factor.
What happens is that when you actually provide a service you have your average response time and you have your, what we call latency at tail, which means the outliers. A lot of work is done by the data center providers to make sure your variance is reduced, and they actually have to trade off the performance of the average in order to keep that variance in control. Right? We have built-in technologies that help actually reduce that variation so that you actually don't have to drop the amount of consolidation that you do. Right? The last piece that I'm going to touch upon in this utilization is really the data center efficiency. When you actually run a large data center, there is a fair amount of overheads that you have to manage.
This is really where you have services and things running on the data center, running on the CPU that is not part of the application. Right? That's not cycles available to the application because you have to just manage the system. That's either movement of data, compressing data, securing data, everything else in the data center, all of that actually has a certain amount of overhead on the data center. We have technologies that we've driven that actually helps reduce the amount of overheads that you have, either offloading that to capabilities like our QuickAssist Technology, where if you have to do compression of some data, you can actually do that outside of the core, so the core is available for actually processing. The ISA-L constructs that we have built in.
We've actually built in a specific ISA-L that can actually speed up crypto compression and so on. Right? We also have technologies like Data Direct I/O, or DDIO is what we call CBDMA. These are all technologies that can actually efficiently move data and keep it such that you reduce the overhead and keep the core from having to get involved to manage all of this. This actually allows more time to be available for applications. If you reduce your overhead by 10%, that is 10% time available for applications, which is 10% performance. None of the industry standard benchmarks can actually measure overheads in a data center. Right? These are all the things that we do to make sure that you have very high utilization.
A lot of these are capabilities that require us to work closely with the data center, understanding their usage models and their deployment models. Once we create the technologies, it actually takes us about 2 to 3 generations for them to deploy it at scale. Once they start using these capabilities, we find out that there are some things that, based on their usage model, we have to adjust and so on. These capabilities take time to actually be deployed, but once they are deployed, they deliver very significant value. Okay. I'm going to talk about a third factor here, which is the paradigm shift that's going on. Right? I think pretty much every talk today you've seen AI is a critical paradigm shift that's going on, and I think both Naveen and Raj talked about it.
There are a number of large-scale workloads in a data center that will start incorporating AI in some portion of that entire workload. The way we handle AI or process the AI portion of that workload is going to be critical to the performance that we deliver at the data center. We have actually driven a number of architecture changes to actually make sure that we can actually keep improving the performance of AI, and along with that, the overall workload performance. I'm going to talk a little bit about what are the kinds of things we've done in our process and continue to do to actually drive AI performance. The first one is the vector operations and the load operation. I think Naveen talked about the upcoming ISA that you will see in our process. That's a critical piece.
Those kinds of constructs allow us to deliver a very high compute capability in our process. The other thing that generally doesn't get talked about is the value of caches. One of the things that we see is that the AI workloads benefit significantly from the caches, the large caches, because there are workload footprint that actually can fit inside the cache. They fit really well. With the large L2 cache and the shared LLC caches for a lot of the AI workloads, what we are able to do is actually satisfy all the capacity and bandwidth needs without having to go to memory. What that does is it actually improves the efficiency and the utilization of our cores and actually reduces also the power-- or improves the power efficiency.
The other thing that we see is that all of these workloads have very unique data movement, data transformation that's involved. When people are doing all the linear algebra, the matrix operations, they have to really transform data as you're going through the workload. We've been studying those, we've been driving changes into our architecture to actually operate on all of these in an optimal manner. The other piece that I'm going to touch upon a little bit is really about the optimized software. I think in general, a lot of people have seen the performance of Xeon with AI a few years ago based on the out-of-box workload, which is basically a written application which just runs out of the box. What we see is that a lot of those applications are really not tuned for the capabilities that we have.
That's why when Naveen talked about 2 orders of magnitude improvement, 200 plus, that actually tells you the capabilities that we have that are untapped. A lot of these optimizations have been about optimizing that, and we've continued to add more capabilities. I think we've talked about the Intel DL Boost, a series of ISA enhancements that you see, there's some that we have disclosed and some that we have not disclosed. Those are all the things that we do. We expect to actually keep improving the performance of AI workloads as more and more workloads employ AI in their workload. This one is something I'm really excited about. I think we've talked about it quite a bit, I'll try to briefly summarize.
The way I look at it from an Optane perspective is I believe Optane has the capability of doing to the memory tier what SSDs did to the storage tier, which is they improved the responsiveness and the utilization of the data center immensely when SSDs were added. I expect the Optane to do similar. Again, this is really about the Optane as a memory, as a persistent memory. In order to bring Optane as a persistent memory into our processor, we actually had to completely transform the system architecture, the memory system architecture, to be able to allow what we call a load-store interface for an application to directly interface with persistent memory. Typically, applications go through a kernel to actually get to that, and there's this fair amount of overhead involved with that, which all boils down to how utilized your server is.
With this technology, we expect for the applications to be able to directly utilize. We had to actually go through our entire processor, all the way from core, uncore, all the pieces of our architecture, to make sure that we drove the changes that allow us to bring in a persistent NVM, high-performing persistent NVM into the memory tier. I think we've seen some of the performance numbers. Alper is going to come on. He's going to share some more of the performance numbers to actually highlight why we feel really excited about the technology that's probably going to change the computing. We'll make it more into a memory-centric computing. With that, I'm going to close.
I think the important thing to keep in mind is that industry-standard benchmarks are really important, there are a number of other things that a processor has to do to actually deliver high value in a large-scale data center. I've highlighted four themes that are really critical to actually have good performance and throughput leadership, we expect to continue to deliver that. You need to have technologies that can help a data center achieve high utilization. You need to be prepared for AI as more and more workloads are going to employ AI. The memory innovation, I think the capacity of the workloads, the data that they're dealing with as they grow, this is going to be revolutionary in terms of being able to bring in all that data in a memory tier as opposed to in storage.
All of these are going to be really critical, and these are all the technologies that we work on. Okay. With that, my timer shows 2 minutes for Q&A.
All right. Maybe we have time for one question.
The very last thing you mentioned, does that mean that any application that utilizes Optane needs to be somewhat rewritten to make that possible?
Yeah. The persistent aspect of it, in order for the application to benefit, it will have to realize that there is a capability for it to directly access the Optane in a persistent manner, as opposed to going through a kernel for some kind of a storage assist, right, or a file access kind of a thing. Yes.
The capability doesn't need rewriting to take advantage of the platform. I was just making sure that there was a clarification that applications can function on the platform without optimizations, but to take full advantage of the capabilities, additional work can be necessary.
Yeah. In general, everything that we design is done in a way that all applications will run out of the box, but we offer capabilities that when people optimize for that, they can get much more benefit out of it. The same thing works with vector extension and what we are seeing with AI. You can actually run something that doesn't use that, but when you optimize it, you actually see massive improvements. Okay. All right. Thanks. With that, I want to introduce Alexis, Ms. Connectivity (Chicatza).
Wow. I've never been called Ms. Connectivity. I'm really excited to share with you a tremendous growth story for the data center today, which is indeed our network connectivity growth story. As Intel transitions from being a CPU-centric to a data-centric company, moving and connecting and communicating data becomes ever more critical. We have a very strong position in the market today in our connectivity and networking portfolios. I'll share with you our strategy and what we're doing to continue to grow that leadership position. First, before we start that, I want to step back and have everybody think about what it means to be connected everywhere in a globally connected world, and what the impact that has on the needs for connectivity driving our customers, and our own product development.
Data is being generated absolutely everywhere, from sensors distributed in smart cities, to autonomous driving, to mobile. Videos being downloaded and uploaded at unprecedented rates. Our insatiable need to generate, analyze, store, create data is really creating a tremendous opportunity for us, because all this data needs to be moved. In 2016, the data center global IP traffic was 6.8 zettabytes per year. By 2021, that's going to triple. Today, as compute is now distributed, as you've heard throughout the day, from edge to core, our vision is a globally connected and globally distributed compute system. This is truly at a global scale. We were just talking over lunch about the cloud service providers deploying global networks and laying down undersea cables, and that gets me really excited because then it really truly is availability zones across the globe.
Just two weeks ago, Google had announced their fifth sub-sea cable project connecting Virginia to France. Facebook, Microsoft, Amazon, everybody has undersea cable projects in one form or another. We're hearing, actually, from all of our Cloud Service Provider customers that a major part of their innovation today is actually a focus on connecting data and enabling distributed compute at scale. As IBM's VP and GM, John Considine said, "Anyone who spends time in infrastructure knows what the real weakness in all clouds is, networking. How to get your components to talk together, to deliver performance, capability. It's about connectivity." Network connectivity is actually central to improving their overall services, performance, quality, and most importantly, availability to their customers and end users. In short, connectivity can either constrain or enhance performance. Our group here, the connectivity group, is working to address these problems across the board.
We're innovating, and in some cases, we're disrupting to define connectivity solutions that will drive performance at scale. Ultimately, we're connecting a global network connected with optical links, with smart network edges, Smart NICs, at the data centers and at the edge. Because this is so critical for our customers, it's representing an opportunity that's both large today and growing. Today, the opportunity for connectivity in the data center is $4 billion. In 2022, that'll be growing at a 25% CAGR to $11 billion. We project that the Ethernet NICs and the HPC fabrics will grow at about a 12% CAGR, whereas the silicon photonics growth CAGR is truly astounding, at 36%. Our business will look very different in 2022. We're going to utilize our strength in networking, silicon, combined with our photonics capabilities to capture this growth opportunity. We're uniquely positioned to deliver value to our customers through end-to-end solutions.
Let's talk about the role of connectivity in data centers today. For those of you who may not be familiar with connectivity in a modern data center, we'll take a closer look, both at the diagram behind me, but also talk about the performance that it requires. On the surface, the role of connectivity is pretty simple, very straightforward. All of our servers, accelerators, and storage elements need to be connected and communicate both with each other and to the outside world, out to the wide area network. As you dig into the details, though, you'll find it's an incredibly complex challenge to connect all this infrastructure to deliver the high performance and a consistent customer experience at cloud scale.
I think Sailesh was talking about this a little bit earlier, that no matter where your applications are being run in one of these massive data centers, the end user has to be billed the same amount. They have to have the exact same performance expectation. A modern hyperscale data center, being multiple football fields in size, has to optimize this across hundreds of thousands of servers and petabytes of storage. What I'm showing behind me here is a diagram of only 16 racks of servers. These are servers stacked up to the top. On each server, there's a NIC, a network interface card, that transports the data from the server up to the top-of-rack switch. The top-of-rack switch has hundreds of ports and spans out to connect to layers of switch infrastructure, all interconnected by fiber optic cables. Those yellow lines are what that shows.
The diagram here only shows a few hundred servers, a couple dozen switches, and represents a few thousand cables connecting it. In order to imagine a hyperscale data center, you'd have to imagine something over 500 times this size. Connectivity's not just about providing physical connections. It's now the network is more becoming integral to optimizing compute performance at scale. Our customers are seeking to virtualize their infrastructure, to disaggregate compute and storage, to deploy new compute models, such as containers and function as a service, all with increasing levels of security and resilience. To meet these hyperscale needs, it requires significant innovation such as we're making in our silicon photonics modules, with our network interface cards and the Smart NICs that we'll be talking about later in this talk. As well, ultimately, we have to optimize it as a fabric, an end-to-end system, like our Omni-Path fabric.
As you can see from the images on the left, we provide a wide range of products and technologies that address every layer of this network. This isn't just within the data center. This is truly at a global scale. Remember that warehouse scale compute on the bottom left-hand side. Warehouse-scale or hyperscale data centers, they're continuing to grow. A typical hyperscale data center today can be limited by the power distribution that's able to be brought to the site. Could be rated as a 65 MW data center, as an example. The infrastructure within these data centers, as they've grown, as the number of servers have grown 10x, the amount of switches required to connect it has grown 15x, and the amount of connectivity or fiber optics that's required has grown 30x. The network has grown super linearly with the amount of servers that are deployed.
In some cases, the spend on the network rivals that of server-based compute. Think about it. Every single server needs to be connected with every other server with 25, 50, or 100 gigabit per second links. This is all in order to maintain consistent performance, regardless of where data is stored or where compute is performed. You can never have the network be the bottleneck to be able to access compute or utilize compute. We need unconstrained connectivity. These data centers are then connected among themselves to form regions, and the regions are connected globally. As an example, Microsoft has, as of two weeks ago, they had 54 Azure regions available in 140 countries and having up to 1.6 petabytes of bandwidth each. To put it in context, a petabyte is 1 million gigabytes.
If you wanted to think about this in a more tangible way, if one byte were a grain of sand, that would be just a little bit shy of 10 billion pounds of sand. It's an awful lot of bandwidth that's required. We provide a wide range of interconnect solutions that we'll go over now, and we'll talk about where we're innovating to meet this explosion of data and bandwidth demand. What's happening in the data center? There are three key trends that are really defining network technologies and the evolution of network technologies today. The first is that our customers are always looking for ways to improve their compute performance. Earlier today, Naveen Rao talked about the need for purpose-built compute for application acceleration. Increasingly, our customers are focused on purpose-built infrastructure acceleration, so purpose-built accelerators to improve their network infrastructure performance.
Our strategy that you'll hear about soon, is to provide intelligent acceleration for infrastructure workloads. The NIC is actually becoming a critical innovation point in the network and in the data center as a whole. We're using FPGA technology and SoCs that complement the compute horsepower of the Xeon Scalable platform. FPGA-based accelerators provide the programmability and flexibility to accelerate emerging and evolving workloads, especially for both applications and infrastructure. The evolving part is very important because as people have said, we don't know what the future of compute will be. We don't know AI. We're at the infancy. The flexibility provided by FPGAs in the system is really important. The second key trend is that the cloud needs a pace of innovation to meet their scale-out needs. Large cloud providers are constantly seeking to improve efficiency and scale, and they're driving scale-out and disaggregated architectures.
To do that, they really are starting to co-innovate with the fabric and the network. In fact, we do a huge amount of connectivity collaboration with our key customers. As the cloud service providers are seeking to achieve 100% utilization of their compute resources, I'll just reiterate that unconstrained connectivity is required. Raejeanne actually had a nice video earlier this morning from Toutiao, where they really talked about a balanced design platform, including the high-speed networking element to achieve their data center performance. There's other areas where we really need to innovate as well. As the data rates and bandwidth increase, there's also an increasing amount of power that's being consumed from the network itself. Remember, these hyperscale data centers can be defined and limited by the power available to the site.
Our customers have told us that when they upgrade their fabric from 100G fabrics to 400G fabrics, they expect the power consumption from the network itself to double. This is really important. As the fabric grows more and more important to the overarching compute system, we need to innovate across every vector we can to enable the continued scale. The third key trend is that critical workloads are really evolving the network. I talked in the first trend about how custom-built compute is needed for infrastructure acceleration. The essence here is that you have to customize your network for the workloads if you want to really get the maximum performance. That's the fundamental basis of how HPC supercomputing clusters were designed.
We're seeing a true evolution in how hyperscale data centers are thinking about the architecture of their network to optimize the compute performance at the same time as they're optimizing the network connectivity. The key performance vectors are, of course, among the same ones that Sailesh was mentioning on a microscale: data bandwidth, latency, availability, and flow management across the network as a whole. As we also looked at earlier on the Novartis AI demo, the network fabric was actually critical to improving the training time in this particular example by 20X. Remember, this example demo that Naveen Rao went over, it was based on eight CPU-based servers, our Intel Omni-Path fabric, and optimized TensorFlow. Let me talk a little bit about the portfolio that we have that's addressing the connectivity needs of our customers. The first Is our Intel Ethernet products.
They provide a comprehensive level of interoperability through a range of media types with scalable and flexible IO. We've been in this market for over 30 years and have been continuously innovating to unleash the network performance. Our silicon photonics portfolio is really well-positioned, in fact, to disrupt the optical technology market, combining the innovation of our hybrid silicon laser with Intel's large-scale silicon manufacturing capability. We're leading the journey to integrated optics. The third that we'll talk about, it will be Omni-Path Architecture, which has been winning in the marketplace since it was first deployed about two years ago. It's the next generation of HPC fabrics, and it really spans the gamut from silicon to network adapter cards, to edge and director class switches, to a full software fabric. Full optimized software solutions. The Omni-Path fabric is truly changing fabric economics.
As I said, we've got a leadership position in the network. We have a number 1 market share position in server class Ethernet NICs. We co-defined the original Ethernet specification. We have a proven track record over 30 years of innovation. Intel Ethernet's actually the foundation of many networks across the world, from cloud service providers to Fortune 1,000 enterprises. Our foundational NIC product line provides Ethernet connectivity plus advanced functionality for fast packet processing and network virtualization. Complementing our Ethernet NIC portfolio is our Programmable Acceleration Card line. This is an FPGA product from the Programmable Solutions Group. This is a fully programmable PCIe card with Intel's leading FPGAs. PACs enable our customers to accelerate a wide range of applications from big data analytics to network applications. These are fully programmable and enable our customers to integrate either their own IP or third-party IP.
Today, I'm really excited to announce that we're extending our leadership into a new and innovative family of Smart NICs, specifically targeted at network and infrastructure acceleration. Smart NICs are the next step in optimizing compute performance by distributing intelligence across the platform. They provide additional levels of programmability and acceleration for critical infrastructure workloads. Altogether, we've got a leading portfolio of intelligent and configurable networking products. Let me tell you a little bit about the Smart NIC family. Our Intel Smart NIC family provides improved total cost of ownership and system performance for comms and cloud service providers. As the server network bandwidth increases to 25 gig and beyond, infrastructure workloads such as network, security, and storage begin to consume a significant number of compute cycles and cores on the platform itself. Smart NICs enable us to build a platform to run those most efficiently.
With Smart NICs, we're taking what we do best on compute on the Xeon platform when we're moving that and extending our leadership position into the NIC by incorporating processing and flexible logic elements into our products to optimize and balance across the platform itself. As you can see in the diagram behind me, the infrastructure acceleration implemented in the Smart NIC enables our customers to actually shift key infrastructure workloads from the host CPU, allowing them to run their host more efficiently. Through balancing the Xeon and the Smart NIC, we're enabling our customers to achieve the highest return on investment, at the same time improving overall compute utilization. This enables our customers to offer more services and support more workloads on their platform, taking advantage of the higher bandwidth and lower latency of our Smart NIC product line.
Comms service providers could take advantage of the reduced jitter to improve their NFV platform by accelerating the lower levels of the networking stack in the SmartNIC. Today, we're announcing our very first product in this family, Cascade Glacier. Cascade Glacier is the world's first SmartNIC to support hardware-based virtual host interfaces, which enables seamless live migration. I think Sailesh set this up pretty nicely in his last talk. Seamless live migration can be used to move VMs between legacy and new-built SmartNIC-enabled servers without any downtime and without changing how the servers behave in the system. By migrating their existing customers' VMs to a SmartNIC-enabled server, the customer now has access to a much higher performance network interface without the need to modify the software drivers. There are no changes needed to the customer's VMs.
In addition, since the VMs are now able to be controlled in hardware, the customers could take advantage of the improved security and quality of service that's afforded in a hardware-based system. This SmartNIC product line maximizes our core efficiency associated with the vSwitch by interacting with the VMs in their native driver. This is because it's a fully programmable SmartNIC, which enables our packet processing pipeline to be tuned to the customer's vSwitch. When we migrate this ensures there's no loss in fidelity between how the software-based vSwitch is implemented today and our hardware-based solution. Finally, the Cascade Glacier product line offers fully validated solution stacks to our customers. It's a plug-and-play NIC capability to afford our end users the ability to utilize all the features that we're sharing here.
Today, we're sampling our 2 by 25 Gig solution to leading customers, and we'll be in production by Q1 of 2019. Intel's future Glacier products will move closer to full hypervisor offload, and will also be providing higher bandwidth up to 100 Gig. I'll now change gears and talk a little bit about our Silicon Photonics platform. We've been talking about network connectivity from the perspective of unleashing performance in cloud data centers. One of the fastest-growing requirements in the data center is the need for cross-sectional bandwidth across the data center. 75% of all traffic flowing today, if not more, is east-west, connecting server to server to storage. In order to enable this, a massive amount of optical connections are needed to provide cost-efficient solutions that span the distance, at the density and the scale required. This is where Intel Silicon Photonics comes in.
What is Silicon Photonics? Silicon Photonics is essentially, at its core, the addition of light to Intel's world-class silicon manufacturing platform. Silicon can't produce light. We've spent the past number of years inventing a disruptive technology and process capability to add light to our platform. The output of this has been a technology that's critical to the future of data center networks, the hybrid silicon laser. We effectively, as Naveen talked about this morning, we take pure electrical signals on silicon and convert them to light on chip, enabling us to transmit at vast distances and bandwidth off chip across hyperscale data centers. This is the fundamental building block of an optical transceiver. We see a very large opportunity to disrupt the traditional fiber optics market using this technology.
Our silicon photonics replaces complex assemblies of discrete optical components that are today manufactured with disparate material systems, with fully integrated, self-contained optical circuits manufactured and tested in silicon at silicon wafer scale. In that sense, our silicon photonics is combining the bandwidth and reach of optical networking with the scale and manufacturing technology of CMOS integrated circuits. We're the only vertically integrated silicon photonics company. We have an in-house fab. We have our own silicon optical technology. We're the only people to be able to create light on silicon, and introduce the hybrid laser. We've got an in-house IC design team, and we have full-scale optical transceiver manufacturing backend. It takes innovation, collaboration, and technologies like silicon photonics to play a big role in the evolving data center needs. Our partners recognize this very well.
Let's hear from Jay Parikh, the Vice President of Infrastructure at Facebook, on the importance of the network and silicon photonics to Facebook's business.
Facebook is in the business of connecting people. The number of people using the service around the globe has grown from 150 million to over 2 billion. Connecting people at scale is a very large challenge. The amount of data is mind-boggling, network speed is very critical for us. Dealing with the speed of transmission, the number of terabytes we have to transport and move from machine to machine across the globe, that is one of our prime challenges in the networking space. Intel Silicon Photonics is one of the many bets that we've made that keep us on the leading edge of performance. Facebook has been working with Intel for several years now to make Intel Silicon Photonics a widely deployable technology that helps us solve the problem of optical transmission in a way never before done.
We've recently deployed it, and we're now rolling it out at scale. The partnership between Intel and Facebook is very special. It's founded on trust and transparency, deep technical collaboration between the two companies, and a deep focus on operational excellence. At hyperscale, you've got to innovate and be at the leading edge of the performance curve. Partnering with companies like Intel keeps us at the highest performance at the earliest time possible.
So as (TS) just mentioned, Intel Silicon Photonics is currently being scaled out in Facebook's data centers today. We are also shipping to other large cloud service providers to provide high-speed switch-to-switch links. Moving forward, silicon photonics is widely seen as the only technology offering the bandwidth density, cost, manufacturing scale to meet the future data center needs. We are continuing to gain significant momentum in the market since we launched our very first product in August of 2016. Since then, we have ramped our award-winning 100 Gig product lines with several hyperscale customers. We have been expanding beyond the data center as well, and are currently sampling our extended temperature 5G wireless radio front haul network device or module, as announced at Mobile World Congress.
If you remember from this morning when Naveen and Dan Rodriguez were up here and we had every radio base station can be connected from the radio to the front haul, by silicon photonics links. In addition, in the Optical Fiber Communication Conference, OFC, in Q1 of this year, we demonstrated our first 400 gigabit per second transceiver. Intel Silicon Photonics is in high volume production today and is poised for rapid and continuous growth. As I said, we are not stopping at 100 Gig. We have already demonstrated 400 Gig, and we are providing additional 400 Gig product samples later in 2018. We are actively engaged in early deployment customers to ensure a successful ramp. We expect to see initial 400 Gig ramp in 2019. In the future, silicon photonics will scale well beyond 400 Gig.
We will see the evolution from pluggable modules to onboard optics, to co-package silicon photonics with our networking logic silicon such as switches, FPGAs themselves. We are really excited about the future possibilities here. The possibilities are truly immense. When you think about the future of silicon photonics, that is truly optical IO. With our unique invention and our differentiated technology of the hybrid silicon laser, and our in-house manufacturing, we are the only company who could lay down dense arrays of 100 Gig links right next to each other, providing the bandwidth needed to get on and off of dense, high-bandwidth silicon. We believe this type of integration is ultimately the only technology. This type of integration can only be achieved with silicon photonics, which ultimately will enable the very low cost and high volume optical interconnects that will one day go to the servers themselves. All right.
Not only do we have a tremendous product line today, the market opportunity, we shared that a little bit earlier, is pretty large. The data center connectivity market for optical modules is projected to grow to $7 billion by 2022. And what is super exciting is the ability to grow beyond that into 5G, DCI, metro links, and LiDAR for automotive or autonomous driving sensor applications. We are just at the beginning of a renaissance where the network is critical to deliver efficient computing at scale. I am now going to share a little bit about our Omni-Path market success. And one of the best part about Omni-Path is that it really brings together everything that I have talked about so far, from the high-performance NIC and networking capability, to fully optical lengths, to the best example that we have of co-optimizing fabrics for specific workloads to run at scale.
Omni-Path has been in the market for about two years. We continue to see growing success from our first-generation fabric. As a testament to our success, Omni-Path has an incredibly strong presence on the TOP500 supercomputer list, which tracks all of the worldwide HPC clusters. The highest performing clusters today are based on 100 Gig fabrics, and Omni-Path is the fabric for 50% of those fabrics that are on the top 100. We also have a record number of systems on the list for ourselves. We're up to 40 systems that are deployed on the TOP500 that utilize Omni-Path. We've increased the number of flops by 36% since November 2017. Omni-Path is also the testament to its power efficiency. Omni-Path is the fabric interconnect for 27% of the top 100 green systems. Omni-Path success crosses multiple segments and verticals.
The chart here behind me represents just a sampling of the hundreds of customers where Omni-Path is already deployed. While a bulk of the deployments are for HPC workloads, many customers are using Omni-Path for a combination of uses, including AI. For example, Texas Advanced Computing Center offers cloud-based services to schedule and remotely run a variety of workloads from genomics to deep learning. Our Intel Omni-Path has a very large and growing ecosystem consisting of hardware vendors, open source software, and commercially available applications. What it truly does, and what I want to emphasize here, is that it really is the best example of how you can utilize the performance of a fabric to bring out the total system compute capability. In conclusion, and to wrap up, we're in a new era of data center connectivity.
Connectivity is critical to unleash the power of compute at scale. It's a huge and significant growth opportunity for Intel and for DCG. With a projected data center TAM of $11 billion by 2022. We're leveraging our technology foundation that's built on 30 years of continuous innovation, and we're collaborating closely with our customers, as you heard from Facebook, to deliver leading end-to-end solutions. Our Smart NICs that we announced today are adding intelligent accelerators to improve the platform performance and compute efficiency. Our silicon photonics are improving power density, power efficiency, and scale for hyperscale data centers. We're exceptionally well positioned to capture the growing data center connectivity networking opportunity through providing end-to-end solutions. With that, I believe we have at least a minute or two for questions.
Harlan Sur with J.P. Morgan. Thanks for the presentation. It's actually pretty timely because we are hearing more and more about the usage of Smart NICs with some of your big cloud customers.
The architectures that we're seeing, though, are SoC-based, multi-core Arm from companies like Broadcom, Mellanox, Cavium, to custom ASIC implementations that are also using multi-core Arm architectures. Can you help me compare or contrast an FPGA-based Smart NIC versus the conventional architectures that are out there today?
Sure. We're at the very beginning of the deployment of Smart NICs in infrastructure for infrastructure acceleration. I think we have technologies that are based both on FPGAs and, as I said, on our SoC technology. I think we've got the most flexible platform and the capability that can service any of our customers' needs. One of the things that cloud customers are really liking about the FPGA-based approach is that the workloads are continually evolving, and we have the ability to be flexible, and actually we don't have a three-year development timeline or two to three-year development timeline in silicon. We have an approach where we're going to deploy and we're co-optimizing with our customers in the FPGA, and that we can then harden into an ASIC or an SoC technology. They like the ability to co-innovate together.
They like the flexibility of being able to work across different workloads. Then, as I said earlier, we have the ability to implement hardware-based virtual host interfaces. We can do this by having the FPGA in the NIC itself. There are a lot of things that we're able to offer the world's first seamless live migration capability that nobody else is able to offer today. All right. With that I'd like to introduce you to Alper. There you go.
Good afternoon. This is actually a really happy day for me. You heard earlier that after about a decade of hard work, we had our first revenue shipments on our Optane DC Persistent Memory product. It's a huge achievement for the entire team. That really makes me super happy. On top of that, what also made me really happy all day today is from Naveen to Sailesh, all of my colleagues actually spent quite a bit of time sharing their excitement with you on how strongly they felt about the innovation that went into the product, how a differentiated solution and a great opportunity for their business this product is going to give. That embracing from the entire team, the enthusiasm, the sharing of the happiness is really awesome. I thank them all.
The downside is I have very little else to add to say today, they all pretty much said everything to be said. We'll see how it goes. All right. All right. The theme today you've seen all day long. We talked about these mountains of data, the explosion of data, and a great story that Lisa Spelman shared with me was how she used to spend hours with the Thomas Guide that her parents gave her before taking road trips and study and sort of figure out the route she was going to take, and that made me remember those good old days as well. Of course, that's all long gone. Then came, of course, MapQuest.
We could go to our computers and just pretty much type in where we want to go, and you get your printouts and just follow your directions with the GPS and mobile technologies. We started using Google Maps, and finally, with Waze, we are getting crowdsourced real-time traffic data and find our ways home. Essentially, if you think about it, within a fairly short period of time, we went from consuming no digital data to consuming significant amount of data to producing significant amount of data. This whole change is contributing across the world to these mountains of data, explosion of data. We talked about today over and over how organizations are viewing data more and more as an asset and extracting business value, driving technological breakthroughs, even social changes.
I don't want to go into this again, but there is a corollary to which my topic, memory and storage, actually really is related to here, and that one is the rising temperature of data. What I mean by that is, and Naveen talked about this a little bit that he said, we think about hot data, we think about warm data, we think about cold data. Hot data is very close to the CPU. It gets processed all the time, and warm data is also part of the working set, and we tend to access it. The trend that we are seeing is that the data temperature is rising. Organizations, when they are looking at the data and trying to extract value out of that, they're looking at larger and larger data sets.
What I used to just archive and forget my transactions from two years ago now are pretty relevant because I want to look at them, correlate them to what's going on today. Can I extract any more value out of this? More and more data is moving close to the CPU. This trend certainly has been quite beneficial to the storage and memory industries. When we're looking in, specifically in this case, just the memory industry alone, and we're looking at DRAM in the data center, we have observed doubling of the TAM between 2016-2017 almost, looking at the TAM growing to about $28 billion in 2022. This is sort of the opportunity size we are looking at. Against this backdrop, we are going to be introducing our Optane data center persistent memory product.
Obviously the opportunity is huge, more importantly, we're bringing to the table here a very differentiated, disruptive technology. The theme of my talk is disrupting memory storage hierarchy. We'll talk about this famous pyramid. You've seen this many times before. Before I get into this, what I like to do is, with your permission, jump on and start a demo. I do have a demo. Unfortunately, I have only 30 minutes, now 25. This thing runs quite a bit long. I'm going to start it. We'll come back to that in a minute. If you can switch the video, let's see if I can do this. All right. What I have here is two side-by-side systems, real-time running, and both of them have an in-memory database running. Actually, they haven't started yet.
One of the systems is a traditional DRAM and SSD-based system. The other one has our persistent memories in it. What we're going to do is restart these databases. I'm simulating essentially a situation where I've taken my server offline, I've done some servicing. It could be a kernel patch, it could be a security patch, a software upgrade. Whatever I needed to do, I've done it, and now I'm going to bring the system back online. We'll see how long it will take, and we'll get back to that. I'm starting this DRAM system now. Oh, here we go. Yes. All right. It started. We're going to time it. We'll get back to that. Let's move on. All right. Back to our pyramid. This is essentially the hierarchy of the memory and storage representing how data flows in a computer architecture.
We talked about DRAM being at the top of the hierarchy. It is very fast but quite expensive. The capacities are highly limited. Down in the storage hierarchy, we're seeing SSDs, and in the slower and cold, high capacity tier, we're seeing the hard disk drives and tapes. We always talk about it in this context, and we came to think of the memory and storage as a continuum, as if they were one and the same continuum space. I will like to offer a different view and talk about really how different beasts these memory and storage paradigms are. When you look at it from the perspective of software, hopefully, you'll see and agree with me as well. When we look at how memory really operates or an application accesses memory, all it does is issue a load or store instruction in the CPU.
Without any interruption, it will hit the DRAM through the very low latency and high bandwidth DDR bus, and data comes back. The whole process takes about 100 nanoseconds or so. This is how we operate the memory tier. However, when the application has to access storage, maybe because it ran out of memory and the data resides in the storage, it has to bring it, or maybe the application decided that the data needed to be persisted. It needs to be stored for future use. It needs to be saved because you can't afford to lose it. The path changes dramatically. You actually, at that point, what the application has to do is go to the kernel, the operating system kernel, execute a routine that takes it to the IO devices, read or write the data, and come back.
Now what the CPU does is run essentially bunch of extra code. While it's doing that, we are doing a context switch, lose all the contents of our caches, and come back with the data, and that's going to with a fast SSD, it might take you 100 microseconds, and if you're going to a hard disk drive, maybe milliseconds. Essentially, the latency difference between the memory and storage, we are looking at about 1,000x difference. As you can see, from the perspective of the application, memory and storage are very, very different. This is a big challenge for application developers. They have to constantly think about, hey, where do I put my data? How do I treat that? Should I be persisting it? What happens if I miss the data? How do I go get it if I don't have enough memory?
That's actually a significant efficiency gap that the developers are experiencing as well. Now that you've seen, we're introducing the persistent memory. It's not just a product, but we're introducing a totally new hierarchy, a new class of product. The persistent memory actually behaves like memory in that you access it directly through the load store instructions. It behaves like DRAM in that case. Unlike DRAM, it offers you large capacity points. Its price is lower, and very much unlike DRAM, it's persistent. It does not lose the contents of data once the power goes away. Now if you can think about this, we are offering you a new category that represents the best of the both worlds. You get something like DRAM that is persistent, and this changes the entire paradigm of how people compute and rewrite the applications.
We're very excited about this. We have a lot of ideas about, okay, here is the new applications, here's where we're going to see. A lot of our at least my excitement is about what are the things that we're not thinking about? Because every time I talk to one of our customers, they come back with ideas that I had never thought about. "Oh, we're going to use it like this. We're going to do this." I'm, "Oh, wow, great." The potential is immense, we're really, really excited about the potentials that we're going to realize through this new category of product. All right. Now let's talk about the product itself. I am holding in my hand a 512 gigabyte memory module. This is world's densest, highest capacity memory module.
One of the most important tenets of our product is that it offers big and affordable memory. The product will come in 128, 256, and 512 gigabytes. Compare that to DRAM, the sweet spot of DRAM, which is essentially where the cost per bit is the cheapest and what typically everybody consumes, is at 32 gigabytes today in comparison. The most that you can get with DRAM is 128 gigabytes, it's really difficult to get, and the price goes up exponentially as you go up the capacity curve. Most people find it incredibly difficult to procure and use. It's just not cost effective to use that. I will tell you today about, or talk about a few examples, why having more memory is really a big deal in the data center. I actually would have liked to have a lot more hours to spend with you.
Unfortunately, that's not going to be possible. I'm going to show you what happens to your performance if you run out of memory, and what kind of problems that represents, and what we are solving. There's other applications that we've talked in the past about how you can actually remove some of the bottlenecks in your system of having not enough memory, such as being able to launch a lot more VMs, such as running more containers in your system. All those are things that essentially, when you solve those problems, you're removing these bottlenecks from your system, and what we see is the CPU utilization goes up, and essentially, what our customers are realizing and what they're really excited about is that they're able to get a lot more out of their hardware investments.
If you happen to be an infrastructure as service provider, for example, being able to run a lot more VMs on the same investment level is essentially meaning that your ROI on your investments are going up. This is exactly what excites some of our biggest customers today. We talk about the product being persistent, which means we can actually build storage devices with it. It turns out because of our speed, you can build world's fastest performance storage devices. I'll show you a few really interesting areas of how we could do that. The product itself is DDR4 pin compatible, which means that you could just insert it into your Skylake platform without any modification needed. It will be just coexisting with your other hardware, so you don't need to change your motherboards, your platform.
Being also a storage device, of course, we had to make sure that it's secure, it's hardware encrypted, and we have to put a lot of reliability features in this. That's like error corrections, sparing, et cetera. This is one of the most reliable memory products you will find in the market. Of course, we've talked about this. Now I can proudly tell you that all these three capacities are in production, and we started the first shipments as of yesterday. All right. Thank you. Oh, how wonderful. I have spent roughly the last 18 years working on new memory technologies, totally new materials, pretty much everything you can think of. I believe 3D XPoint is nothing short of a miracle. I love it.
I couldn't emphasize enough how important the rest of the innovation that we brought to the table here is, and how intimately intertwined these different pieces on the platform come together to make this thing work. You've heard today earlier from Sailesh. He talked about the CPU, how we had to rearchitect a lot of the elements in the CPU, in the core, in our memory controller, optimize how the cores are interacting with each other in the presence of this memory. Immense amount of architectural and design work went into optimizing our CPUs around this. On the module itself, we have actually a lot of innovation. We have media controllers that manage non-volatile memory at unprecedented latency and speed levels. There is no other memory device or storage device out there today that has these kind of requirements that has to work at DDR speeds.
We're delivering that here. At the platform level, we had to rearchitect our firmware, BIOS, UEFI. All of them had to be redone so that we could integrate this new memory type within our platform. On top of that, when you look at the software stack that sits on this, there's a lot of innovation that went into that. To enable all the software innovation, we have first invented a new programming model called persistent memory programming model. Andy Rudoff is the father of that. He's sitting here, so for your questions later, he's available. We made this an open standard and created libraries so that software vendors, our partners, had a very easy time adopting and developing software around this new technology. We have been working intimately with a lot of operating system vendors and ISVs to develop a very robust ecosystem around this product.
When you look at this, all these different components coming together, this is what it took for us to deliver this technology. We strongly believe that Intel is uniquely differentiated in being able to bring all these skill sets and expertise into the picture. This is going to be our long-standing differentiation and competitive advantage that I don't think other people will be able to replicate. We talked about the ecosystem. I'm quite proud to share with you that over time, we have been able to create a vibrant and diverse ecosystem around this product line of our customers, our partners. With the operating systems in the ISVs applications, they have really come along with us for quite some time.
They saw the vision behind what persistent memory could do for their businesses, for their customers, and co-invested with us, worked very closely with us, and developed this really fantastic ecosystem so that by the time we go out there's going to be a strong ecosystem that people can use. I'd like to address a quick point. I think earlier there was a question, do applications work without any modifications using this memory? You got the first level of answer that, yes, we allow people to definitely seamlessly work without any modifications to their application. If you want to really take super advantage of some of the special or optimize it to take advantage all the features of the product, you may have to rewrite your software. Majority of the software or applications that run on the enterprise data centers actually don't access data directly.
They use a middleware, a data access layer, as we call it. That could be databases, that could be frameworks, whatever analytics tools. The applications that thousands and thousands of people write actually interface with these data access layers through standardized APIs. That's why you're seeing here a long list of companies who are delivering today some of the most popular data access layers. They are optimizing their code for Optane persistent memory so that the applications don't have to change. The application will still see the same standard APIs. They can still access the data without any modifications. Just the underlying software below it has moved, and it's not going to have an impact for the rest of the industry. In May 30th, when we had a conference, we had announced a developers cloud.
As much as we are working with all these customers, we also wanted to make sure that other people who have great ideas or want to just check it out, how a persistent memory works and test out their ideas and do development, they can come to our development cloud, play with the new hardware, because we can't just access everybody, unfortunately, as much as we did. Within days of us announcing this, we got 50 applications coming in from startups, from academia, from other ISVs. We were very thrilled. I'm also happy to announce today that we are going to, I think, declare a challenge for some of these developers. We're going to grant some awards for innovative ideas, new proof points, and as well as use cases that we didn't think about.
We want to incentivize our partners in the ecosystem to come up with even more ideas very soon. I'm going to jump and talk about few use cases that are of interest, and you've seen some of these before. The first one is going to be running analytics using a very popular tool called Apache Spark. This is very common. A lot of people use this for their analytics as well as machine learning. What you're going to see here is two machines. One of them is essentially running with DRAM and storage, and the other one with our persistent memory. Now, what you're going to see here, the one that's running with DRAM storage, and we're running essentially several queries. I think there's five queries we are running.
What you're going to see here is that as the data size grows, the data stopped fitting the memory available. You can see here in the middle, the disk activity, the disk traffic. In the lower group bar, you can see that there's a lot of activity going on to the disk. Essentially, the application ran out of memory. The data set sits in the SSD. I have to constantly go back and forth within the SSD. If you remember the long cycles of operating system accesses that we talked about earlier, this is happening constantly here. In the other machine, we put our Optane persistent memory. There's plenty of memory. The application doesn't run out of memory, and it just runs smoothly. We are seeing about eight times improvement in analytics with this application.
You may say, "Hey, this is just benchmark. It is not real." It actually turns out to be an extraordinarily strong pain point for a lot of our customers. Just couple of weeks ago, just when we were looking at this example, we had a visit from one of our very large e-commerce companies. They described us this very point. They happen to be run, actually, Spark. What they do is they run millions of transactions on a daily basis, and at the end of the day, they push them through their analytics tools and in batch more do their work, and in the morning, they're ready to go with the next day of shopping day. Everything was working perfectly well, but recently, as their data size started growing, they started running into this very problem. Their analytics jobs don't finish on time. They're constantly behind.
They don't know how to do that. They're excluding data. They came up to us and said, "Hey, can you please help us?" We showed them this data, and they were super excited. I think this is going to be fantastic for them. I think we're going to solve their problem. All right. The next data point we're going to talk to you about is this demo that I was running. If you could switch back to this, see where we are. All right, thank you. It's been about 17 minutes. Our system with DRAM is still running. What I like to do is start my system with the Optane persistent memory, and let's see how fast it's going to take.
What happens here is the system comes up, the database tool is going to rebuild some of the tables and indices it needs to run during the normal operation, and then comes back and says, "I'm ready to go." Right. Thank you. Thank you. Yes. This looks cool. Actually, the first time we demonstrated this, somebody commented that Intel must be doing something fishy here. This can't be real. There is no magic here. It's just the nature of the product. It's just persistent. Data sits there. We didn't lose it. You don't have to do anything. It's just there. You just bring it up and boom, you go. Actually, I'm not going to continue this. Let's just go to the other side, if we could, please. It's going to run for a while. It turns out this thing is going to run for 35 minutes.
Of course, as the data size grows, the gap changes. You're seeing significant gap between the two. I had not quite appreciated what a big deal this is to some of our mission-critical customers. Raj talked about this earlier in a very eloquent way. We found actually customers who for 20 years did not bring offline servers because they didn't dare touch it. They cannot do software upgrades. They cannot do kernel upgrades, security upgrades. They keep living with the old stuff, and they are in constant fear that, God forbid, something goes down, they won't be able to restart and their business is just going to come to a screeching halt. This is a fundamentally game changer technology for all those guys, and they're very excited about this. Okay.
Let me jump into yet another example of how persistent memory can change paradigms, something that you couldn't possibly do with anything else that persistent memory allows you to do that. I'll talk about data replication. Replication essentially is a mechanism for storage to create copies of data so that you can have reliability in case something happens to your original data. This is a very standard feature for all storage subsystems. You use it every day, whether you know it or not. This is happening in the background all the time.
In a traditional system, the way replication happens is that, essentially, the CPU of the original device will read data from the storage device, do all the things we talked about before, going through the kernels and software, and push it to its DRAM, shove it over to the next device's DRAM, and then the receiving CPU does the reverse, reads all the data, and pushes it back to the storage device. It's highly software involved process that, as you can imagine, takes a long time. With the persistent memory, we can avoid all of those. We can take the RDMA protocol and tell the RDMA protocol, "Please push this data to the other side without any CPU involvement whatsoever." Data just goes from one server to the other one, and voila, you're done. You couldn't do this. This was not possible before.
It just opens up such tremendous opportunities. When we build such a system, the latency difference between two replication paths that we are seeing is 14 times. You translate that into an IOPS. It's how many operations I can run on a similar system. You're seeing 14 times improvement. I'm not talking 14%, I'm talking 14 times more throughput you're putting through. I'm pretty confident that in the very near future, we're going to see more and more storage subsystems start looking like this. Actually, one of our great customers, Oracle, already talked about this, and Corey, who's a good friend and VP of Exadata Development, he's like, "I can't believe this." Just like zero CPU involvement. Essentially, this allows the CPUs now to go do what they're designed to do, compute, not just shove data around. Right?
This gives people such a great utilization of all the infrastructure that they bought and built. Really game changing. All right. We talked about the memory storage hierarchy. You all know that persistent memory is not the only product that we build in this domain. We certainly have a great portfolio of storage products as well in the SSD domain with our partner organization, NSG. We do have a set of products that are called Optane DC SSDs that are bridging the gap of performance. These are the highest performing SSDs in the market, as well as we are now introducing QLC SSDs, which are bridging the gap, the cost gap between SSD and the hard disk drives.
Just today, during his keynote speech, the NSG leader, Rob Crooke, talked about first of all New Optane DC SSDs that are hitting right now 60 drive writes per day specs. This is pretty unique in that market. As well as talk about our first data center PCIe, NVMe, QLC quad-level or four level per bit SSDs deployed at Tencent, achieving essentially 10 times more customers served per server compared to their previous solution. As Navin mentioned, when we have this entire portfolio of products and solutions, we can think about a bigger picture. What's next? How do we integrate these different technologies together to create a bigger data center vision? When we do that, the vision we have for a data center is that we see Optane replacing SSD and SSDs replacing hard disk drives.
We are seeing a paradigm where random IOs are now memory bound. We don't do those things over storage anymore. You're going to have, at rack scale, petabytes of very cheap solid-state-based storage. That's the vision we are driving. We are essentially doing the disruption that SSDs did to hard disk drives with the new technologies and improving and allowing actually a totally different levels of performance to come to fruition at improved TCO levels. That's the vision that we're driving to.
In order to make all of this happen together, when you look at the skills that you have to bring to the table, we talked about these before, all the way from the deep semiconductor materials research to processor architecture, computer architecture, changing the platform, being able to create software, being able to influence the ecosystem and understand how a data center works that we have shown to you over and over again. This is what it really took for us to bring this new product to market. At a personal level, I started working at Intel more than 25 years ago as a CPU designer. I left Intel in 2001 to go work on new memory technologies, I came back about two years ago to work on this project because I was convinced that there was no other company on this planet that could do this.
I strongly believe that there is no other company who will have the scope, who will have these capabilities, and who will have this level of differentiation. Being able to have shipped this first product and achieved a milestone yesterday is hugely gratifying to me. At the same time, having worked with a team that was so incredibly committed to this, and made so many personal sacrifices, has been an incredibly humbling experience for me as well. I like to thank really the entire team for getting us this far and working so incredibly hard. I'd like to acknowledge all of them here. Thank you so much. Finally, this is just the start for us. This is just the start for us. We're going to keep building new products.
We're going to push our vectors of differentiation around capacity, around performance, around cost, and around persistence and new use cases. We have a totally new roadmap that is aligned to our platform roadmaps as well. We'll keep pushing on this. Our entire team is deeply committed to this. In summary, we've delivered the big, persistent, affordable memory, and it's a good step, and we believe that it's going to be a catalyst for Intel's platform growth in the future. Very exciting day, and thank you so much for being part of it and sharing this great news with us. Thank you.
Thanks for the presentation. Really appreciate the distinction between rewriting at the middleware level versus the application level to take advantage of Optane. How involved is that rewrite, and I guess where is the ecosystem? To what extent have you guys already been working with these companies to do that? Where are you in the process?
It's a good question. Thank you. As I mentioned to you, we have certainly different levels of engagement. The first step that we have taken is created a standard around which everybody could agree on. This is essentially the persistent memory programming model. It's an open standard open for everybody else. You had to have that, and people agreed to it so that the underlying operating systems, the virtualization tools, and all the development tools line up to that, which is already done. Today, I can report to you that Windows or Linux versions are all lined up and supporting this model, so ready to go. On top of that, the second layer of engagements we had are with ISVs in this data access layer. With companies like SAP and Oracle and several others, we had direct engagements.
We have enabled them by delivering them libraries, development kits, sample systems, development vehicles, and all of that has happened over time. We are essentially going down a priority list one by one and sort of growing the wave of enabling, if you will. The idea is that a large number of companies will be ready for us when we go.
It sounds like with some of your commentary that SAP is probably the first?
SAP is one of the first. Of course, I'm bound with certain non-disclosures. SAP has talked a lot about this technology. They've been very excited for a very long time. The memory limitations have really become a limitation for their business. They are super aligned to this. I'm allowed to talk about that. Some of the others have done less so my apologies. I can only show the logos and you figure it out. Thank you. Anything else? Any other questions? All right. Thank you so much. I think we're breaking right now for 15 minutes and enjoy the rest of the presentations. Thank you.
Are we on? Good. Okay. Good afternoon, everyone.
Hi, Dan.
It's cavernous. Thank you all for staying here. I'm told that Lisa and I are in between you and drinks, we'll keep it lively. My name is Dan McNamara. I lead Intel PSG, the former Altera, it's been about 30 months since we joined Intel. I routinely get asked, I mean everyone I meet, "How's it going? How's the integration? How has this gone?" I usually go through a long story about why it's going so well, and I do believe it's going well. All I had to do is sit here today and listen to all my colleagues, hopefully you all saw that FPGAs are proliferating across virtually every business group we have and finding their way in integrated products across Intel, which is really important. Then I'll show you some of our results also from an external marketplace standpoint.
It's going extremely well, but that's not the point of the conversation today. Today, I really want to talk about how FPGAs and their multifunction capability help and add value to all of these data-intensive workloads that we've been talking about most of the day. More importantly, though, FPGAs is one in a very large set of assets at Intel, and I believe an unmatched set of assets. When you take that FPGA and you marry it, whether it's in a package or on a board with a Xeon CPU or any other SoC technology, and you put an overarching software stack on it with IP, it's clearly unmatched in value, and I think you saw a lot of that through the course of today.
If you think about the acquisition, if you recall back in June of 2015, one of the key strategic imperatives was to really take the sum of the parts and create a multiplier in value for our customers and our shareholders, and we really believe we're well on our way to that. One thing that would help would be someone ask a question at the next earnings call to Bob about FPGAs, because I heard a lot of good questions here, that will just cap it off for me. If you guys are looking for the question, send me an email. Let's get started. You heard everyone talk about the data era, I'm not going to go through I think they did a fantastic job of talking about why and what's happening here.
The big thing for me is this is driving a massive innovation phase, at lunch, we were talking about this. I mentioned that it's probably the biggest innovation phase in my career, it's across every part of this cycle, which is super exciting. The part I wanted to talk about is whether it's the edge or the network that Dan did such a good job talking about or the cloud that Raejeanne talked about. There's going to be some unintended data bottlenecks, that's where the FPGA comes in. The real-time aspect. As I look out at all of you, we know a couple of things. People are looking for interactive experiences, we have a very distracted society today. People want things very quickly. You want responses if you're searching online, you want a very quick return at intelligent search. There's just so many examples.
Smart city infrastructures, smart factories, it's just all about real time. That's really what this talk is about, is really how when you take an FPGA and couple it with a Xeon or other CPU or SoC technology from Intel, how we not only remove the bottleneck, we actually accelerate it. That's really the point of this, and we'll go through some examples of this. Before I do that, what is an FPGA? Let's start there. At its simplest form, it's a massively parallel processor. It's got thousands of processing units that are very obviously reprogrammable. Very high speed I/O. I like to call it a data shovel. I routinely call it a data shovel and have for many years, even before the data era. It can be used one of two ways with a processor. One is on a data ingest.
You take large amounts of data, you pre-process it down to a smaller set, hand it off to a processor, the processor makes decisions and things like that. You do it in a co-processor mode, where the processor will hand off either portions of a workload or portions of multiple workloads, hand it off, the FPGA will hand it back. Regardless of how you use it's very low latency, again, for real time. It's very high throughput, which we talked about with the network. Most importantly, it's multifunctional. It drives all these business outcomes. If you think about some of the examples we used, if you think about real time at the edge, think about identifying a person in a crowded city street in the city. It's real time. It does not have time to send it back to the cloud.
It needs to do all the processing on-site and have a decision. Dan talked about the network. Imagine when NFV and 5G are deployed. Comm service providers will be able to change services in milliseconds to a customer and thus drive new revenue streams. That is infinite flexibility between hardware and software, we'll talk a little bit about that. The cloud, Alexis talked about this. The cloud is always looking to provide more and more services, deal with all this data, be flexible to react to all of these different workloads, it's always looking to manage their TCO. When I show you some of the examples of how FPGAs are being used in the data center today, you'll see that it really helps them on the overall TCO.
All of this is driving what I've been calling for probably the last year, the largest adoption cycle ever for FPGAs. We have massive innovation and a large adoption cycle for FPGAs. There's no better time to be at Intel. That's why the timing of the acquisition is just perfect for us, the ramp and the momentum we have right now is just tremendous. The innovation phase and the adoption phase is driving our results. As you see here, these are the 3 years that we've been part of Intel. If you look at this year alone, in the first half, we're up 140% in the data center. Our advanced products, which is really a good indicator of what we're doing with the newer products, 28, 20, and 14 nanometer from Intel, they're up 50% year-on-year.
Overall revenue is up 17%. That's not really the story, even though that's a good story. We have broken 2 consecutive records for design wins, that's not as part of Intel, that's for our 30-plus year history. That's really about the solution sets that we're putting together that I talked about earlier. CPU, FPGA, software stack, IP, I'll talk a little bit about that. That's really what's driving this, that's why we're so excited. It's the things that don't really show up in the reportable numbers. It's the design wins, that's what our business relies on as the engine for growth going forward. I wanted to talk about the solution set because it's very different from our competitor. I want to just show you the stack, I'll jump into each of the verticals.
It fundamentally starts with the best hardware, and we believe we can do that as part of Intel. That could be a board, that could be a discrete FPGA, that could be an integrated CPU or SoC with FPGA. It doesn't matter. We need to deliver the best silicon and hardware. The next step is really marrying traditional FPGA software flows that build productivity with a broader software stack from Intel. Dan Rodriguez talked about the DPDK and some of the software flows there. We have a computer vision software development flow. There's a Xeon acceleration stack. There's so many different software flows from Intel, which by the way, is probably the best-kept secret at Intel, is how much true software capability we have.
Marrying those to really attract new users, software developers that can actually abstract away the hardware of an FPGA and really use the FPGA and accelerate some of these workloads. Then the third one is really where it gets interesting. That's where we marry all this IP and our ecosystem of partners together and deliver full solutions to our customer. The best example of this is what Alexis just talked about an hour ago in her Smart NIC. If you think of what that is, that's a board, that's an FPGA on it, that's some of PSG's IP, that's some of her IP, that's some partner IP, that's a software stack from Intel. That's what we're talking about, and that's what we're trying to accomplish, and that's what creates the multiplier in value for our customers and our shareholders.
Now I'm going to talk a little bit about each one of them and run through them and just give you a feel for what exactly people are using the FPGA for in each of these verticals. You have to start with the cloud. There's two main areas of the cloud. By the way, I'm going to try and answer some of the questions that came up through the course of the day. First one is what we call infrastructure acceleration. Think about this. Cloud service provider wants the beloved Xeon and the cores there to be working on value add to the customer, their customers. They want to generate revenue and services with Xeon cores. The mundane tasks of running a network or doing security or accelerating storage, that's not generating revenue.
They want to offload that or accelerate that with other devices. FPGAs are perfect for that because the question came up, why FPGAs? Because of the flexibility to react to all the changing needs within the data center. That's one main area. On the other one we call lookaside, and that's everything from. We talk about this sort of as, okay, these are high cycle apps, where they chew high cycles on the Xeon, whether it's AI that Naveen talked about or video transcode with all the different data formats that Dan talked about, or just database acceleration or search. You've heard Microsoft talk a lot about intelligent search based on FPGAs. What's interesting about all of this is if you think about. Someone asked the GPU question. If you look at this slide, GPUs only fit on the right side.
They can't do infrastructure acceleration. If you look at some of the SoCs and the custom ASICs out there that someone mentioned earlier, they can do one or the other, but they can't do both. Fundamentally, the FPGA is the only one that can do both, and I'll give you an example about it, and this is really the reason why people are using it in a data center. If you take that picture that I just showed there and you slice it down the middle and one half of it is running just network offload, so a Smart NIC type application. The other half, and we can do this because in our software, you can partially reconfigure a device to do completely two different things. The other half is doing search, intelligent search. If you think of Cyber Monday and then Singles Day in China, right?
Those are the two largest online retail days in both countries. Everyone's shopping all day. The FPGA is accelerating search and trying to create a really good customer experience. After everyone goes to sleep, that same FPGA and that same right half of the FPGA, meanwhile the left is still running the network offload. That right half gets changed in milliseconds to do data analytics and maybe create some form of report for the vendors to come back and do some ads the next day. That's the beauty of FPGA. That's the multifunction capability, and that's why we're seeing virtually every cloud service provider pulling us in and talking to us about how they can leverage FPGA. It's a really exciting time in the cloud, and you saw the results that I just showed.
Raj talked a lot about the enterprise, which is another really great area, and it's new for us. This is really a new strategy since we joined Intel. This is, again, leveraging the strength of Intel and the scale of Intel. We're partnering with folks like Dell EMC and Fujitsu, and we're building out what we call our pack and stack strategy, which is a Programmable Acceleration Card with a complete software acceleration stack. Any number of different workloads can be accelerated on premise. That's exactly what Raj was talking about, how there's a mix, hybrid and public cloud. What's interesting about this is we're building out a new ecosystem of partners that are building IP libraries for this stack. There's 30 today, and we hope to double that in a year.
What we're doing, again, is while we're creating new developers of FPGAs, we're also providing really a complete solution with our partners. You name any one of the enterprise verticals, and we can walk in there and do financial risk analysis with someone like Levyx and a PAC card. This is really exciting for us because this is a completely new expanded TAM for us as part of Intel. The network, Dan did a great job on this. FPGAs have played really well in the network for many years. It's probably where we grew up, if you think about it. 5G brings a tremendous amount of challenges and incredible amount of compute capability, right? Thousand X capacity, tremendously low latency, and obviously even the standards aren't even finalized yet. We play here, this is a classic end-to-end, better together story.
The FPGAs typically will fit in the baseband in the physical layer and do DSP filtering, forward error correction, IO expansion, then just like I said earlier, hand off to a Xeon or other SoC technology. Similarly, in the radio head, FPGAs have been used for many years on linear predistortion and crest factor reduction to really optimize the amplifier and the antenna. Again, couple that with some SoC technology from Intel, we have a complete end-to-end story in the infrastructure of 5G. The numbers there are projections that are public. More importantly, Dan mentioned this, we're in trials right now through the back half of this year and all through 2019 with a number of our comms customers, it's going extremely well. Again, it's better together. It's complete Intel technology married to each other.
This is yet another area where while we were there before, this is probably 10x stronger situation for us and our customer than it was before. Dan also talked about the transition and what I always call is NFV is really the foundation for 5G, this is, as he told you, is taking old proprietary routers, switches, load balancers, moving to an x86 type cloud-based infrastructure. Here again is an interesting area where the combination works extremely well. As Dan talked about, you virtualize the network function onto the Xeon, we talked about how comm service providers really want to leverage this flexible architecture to create new businesses for themselves.
With the FPGA, the FPGA can do some of the more onerous packet processing tasks, maybe traffic shaping or some kind of QoS system to deliver to a particular customer and thus hopefully allow the comm service provider generate more revenue for that particular service. Again, the combination is really strong. Dan mentioned a bunch of trials, one of the more notable ones for us is China Telecom. We're doing a virtual vBRAS solution with them today, which is a broadband random access server. They're the largest cable provider in China, this is for their cable subscribers, it's going very well. Again, they're using the FPGA for QoS and traffic shaping. Again, very strong value add to the customer as one Intel, it's going extremely well.
At the edge, the edge is interesting because FPGAs have played there also in industrial and things like that, there's a whole new, Dan talked about this too, is really the compute at the edge and the real-time capability that's needed at the edge is really, really important. Again, I'll give you an example of a surveillance system. Dahua is a leading OEM in China for surveillance, they're building a system called DeepSense. What it basically is leveraging Movidius at the cameras, then as an aggregation point, they have FPGA next to a core processor. The FPGA is taking just what I talked about earlier, large ingest, processing, doing deep learning, acceleration for image recognition. It's a perfect example of a solution that's going on in the edge. Industry 4.0, we're doing a lot of acceleration also.
These are kind of examples of where we go to market together now, not as an independent company that really is better together. Let's just talk about AI. Naveen Rao did a great job, I wanted to talk just a little bit about FPGAs because I think a question came up about this. We do believe that one size does not fit all, I believe that our complete assets will provide tremendous value to the customer, whether you're in the cloud doing training or inference, or you're at the edge doing inference. It depends on the topology. It depends on what your latency needs are. There's so many questions that have to get answered before we provide what we believe is the optimum solution. I think as that plays out, Navin, I completely agree with Navin, it's probably the top of the second inning.
We do have solutions today. Xeon's great. FPGAs are here. FPGAs are really good at batch 1, very low latency inference. Think of anything where you need to make a decision on a frame-by-frame, image-by-image basis. Same things, I'm getting repetitive, very low latency, high throughput. What's really more interesting about FPGAs on this is the flexibility to react to changing topologies. The network topologies are crazy in terms of when you go to 10 customers, it's a completely different neural network topology, FPGAs can continuously evolve with that. Lastly, OpenVINO. Navin talked about these, OpenVINO is a great edge-based computer vision toolkit. Basically, it's a complete open source kit that allow you to do computer vision neural network optimization on any one of our pieces of silicon.
I think in the first month it was released, there was hundreds and hundreds of downloads, just to show you the pent-up demand for the edge inference market. Navin also talked about nGraph, which is going to support all of our silicon. We really feel good about AI overall. We have a lot of assets. Hopefully, you got a more clear picture today, I just wanted to kind of throw in that answer because I think someone mentioned it, like where would you compete with GPUs on inference? Here's the best, I think, quote that probably ties this all together about large data set, how do we operate, what can an FPGA do? AI for Earth is a tool developed by Microsoft, basically capturing data of imagery of the Earth.
It's a tool for anybody studying to solve any one of these environmental issues that we have today. You can predict floods, you can do agricultural yields, you can do any number of different things, which is a great piece of work in my mind. What's really cool about this, Microsoft announced some time ago, Project Brainwave, which is their inference solution with Intel FPGAs. They took this data set, which is 200 images and 20 terabytes of data, they did a full land cover mapping, which is a score. They did a score in inference in 10 minutes, which is just unheard of if you think about that amount of data. If you look at it, if you use that service in Azure, it would cost you $42.
Just think of that, how basically inexpensive it is to do all of that work. This is a very good partnership with us across many levels with Microsoft, and I thought this was a really good way to just tie it together in terms of what this all means with throughput and latency and inference. I talked about massive innovation, I talked about a large or the largest adoption phase. There will be a point in my mind that customers may not want all the flexibility. There'll be a design that they say, "Boy, half of it doesn't need to change anymore," or all of it. One of the things we just announced was the acquisition of eASIC. eASIC is a structured ASIC vendor, been around for many years with lots of patents and a really good technology.
The idea is to really offer a complete end-to-end life cycle solution for our customers. Get to market with us. If and when a customer wants to reduce power and cost, they can. Even target traditional ASICs where a customer is saying, "Boy, this needs to be an ASIC," and they could target it out of the chute. Either way, it's a much lower NRE and a much faster time to market than a traditional ASIC, and it also will give you up to 50% cost and power reduction for FPGAs. Again, it's another tool in the toolbox to deliver this unmatched value. The deal hasn't closed yet. We'll be able to talk a lot more about what we're going to do going forward, hopefully after it closes, and we'll get back with you on that.
Bring it all together, and then I got a few minutes for questions, is the versatility, the multifunction capability is the key. It's really providing a lot of value in this data era. The IA plus FPGA, I hope, if that didn't come through today, I'll have to try again. The value proposition of tying Intel SoC and IA processors together with an FPGA in any form with the software stack is really unmatched at our customers. Lastly, we really think that eASIC will expand our TAM and provide another level of value to our customers. With that, I think I have a couple of minutes for questions. Thank you.
Thank you, Dan. Question about your 14 nanometer product. Can you give us an update in terms of where you are in terms of the ramp, and maybe give us some numbers? Your competitor has already talked about their roadmap for seven, and just curious about your roadmap, if you can reveal anything to us. Thank you.
Sure. Okay, 14, actually, we just talked about this at lunch also. 14 is going extremely well. We're in production and we're filling out the family. At FPGA, you get out with a couple of devices, you get that to production, then you do derivative in terms of densities, right? We're kind of filling out the family this year. In terms of traction, this quarter, going forward, that will be the dominant design win vehicle. I talked about Q1 and Q2, where we broke records. We believe that the traction will continue this year, and Stratix 10 has now taken the lead from our Arria 10 20-nanometer family. It's going quite well.
That family, it's a heterogeneous solution and we're leveraging Intel's EMIB technology, so we have embedded ARM cores in it, and then we have transceiver tiles and we also have an HBM device in there from Samsung, so an integrated HBM. We have leadership in SerDes with 56 gig, and we also were first to market with that HBM solution. We feel very good about the competitive situation with our 14 versus our competitor's current solution. On 10, I've talked about this very recently. We're full steam ahead on 10-nanometer. We're going to deliver samples next year, probably mid-next year, and we feel very good about what we're doing in terms of performance per watt and cost model.
I'm pretty excited about 10-nanometer, and keep in mind, we'll sample our first devices and obviously hit some very key customers that are in a mode for some of the key segments that I just talked about.
Dan, just going back to your competitor Xilinx, they're putting a lot of time, money and effort into the software ecosystem around SDAccel. Does having sort of the Intel AI and that software stack mean you don't have to do that work, or are you doing that work? Kind of help me understand how we should think about the software ecosystem you need to engender as an FPGA player and what being part of Intel might give you that you wouldn't otherwise have.
Yeah, it's a great question. That solution stack didn't show it, but we're still doing SDAccel, think of, we have two different flows, OpenCL and then what we call HLS for lack of a better name. It's just high-level synthesis, which will take C code. In that example where I said the FPGA is a co-processor, you take a block of C code and just accelerate it. We're doing those, I call them pipes, into our tool, and then the overarching flows like computer vision or Xeon acceleration stack overlay that. No, our investment software is very heavy at this point, and we're going to continue to do that. It's really about generating more developers on FPGAs, right? Both of us have the same challenges. If you're not a chip designer, you really don't get FPGAs, so these software flows are really important.
One more, maybe? Okay. Well, listen. Thank you. Really enjoyed the time, now I'm going to have Lisa Spelman come back up on stage. Thanks a lot.
Thank you, Dan. Yeah. Thank you. Good job Boy, he's really gunning for those questions, guys. He wants you taking notes and sending them to Bob. Hey, we did it. Home stretch. Okay, end of the day, last presentation. We're going to kind of bring it all home. Thank you again all for joining us. You've seen a full update today on the portfolio of products and the breadth of what we're bringing to bear on the market and for our customers. I think you've seen us demonstrate that we have had and continue to have sustained leadership across a really wide data-centric infrastructure set of offerings. We're going to keep working to have all of those applications run best on our Intel architecture family.
What's going on now in our environment, though, is that all of our customers are facing an ever more complex IT environment, and it is not enough to hand said customers a quote-unquote "bag of parts." I think Dan's presentation and comments on software flows certainly illustrates that point. As our customers' needs evolve, so do Intel. We're evolving to keep up with our customer needs and ensure that we're continuing to satisfy an ever higher and increasing bar of their deployment infrastructure requirements. We've been working through, obviously, the silicon portfolio that we spent a good deal of time today exposing you to. Then we talked a little bit about Intel Select Solutions, which is solutions frameworks targeted at some of the fastest-growing or most difficult solutions to deploy.
If you look one level above that, think of those as at the node or at the individual component level, you go up to Rack Scale Design architecture, which thinks about solutions in terms of an entire rack in a data center or across an entire data center worth of capacity. We're looking at solutions at multiple levels, and again, just focusing on delivering customer value and ease of deployment. I want to spend a few minutes here talking about those. As those solution stacks have grown in complexity, and the demands of IT have grown, IT in all industries has become not a support of the business, but it has become the actual business.
We've worked with our ecosystem to find ways to not only solve their deployment issues and struggles that they face, but also from an Intel perspective, it's a way for us to accelerate that TAM that we've been talking about all day. Taking the guesswork out of what optimal configurations are and what optimal software setups are. Some of you that I haven't met before or talked to might not know that I spent five years in Intel IT working in infrastructure engineering and operations, I've had that real first-hand experience of really fast escalating customer demands being your employees and your users, and at the same time, a definite lag between when you recognize a technology trend or product launch and you can get it out to market. It's with the frame of that type of experience that we've invested in these solutions.
We're going to talk a little bit about how we decide which ones we work on, and I'll give you a couple of examples. The first thing that we look at is, of course, our Intel Silicon Foundation and what it's best tuned to address for. We deliver across software optimization. Those two points of work are things that you should be very familiar with seeing from us. It's where we've moved next into that system configuration and validation and that documentation that is really driving that acceleration in deployment. We're looking at this as what has the largest market opportunities and again, what represents either uncertainty and difficulty to deploy. I'm going to walk us through an example of one of our most widely adopted Intel Select Solutions and see just kind of how this plays out, and this is VMware's vSAN.
Software-defined storage emerged first at hyperscale data centers as they moved to drive more and more of their storage workloads onto industry standard servers, differentiated by just a different software stack. It gained a lot of traction because of its efficiency, total cost of ownership, and ability to scale. Now it has started to play out across all market segments. It's not just for the cloud service providers anymore. We had been working with VMware on a hyperconverged solution here. Of course, we have the Silicon Foundation that we put there, Intel Xeon Scalable processors, the Optane SSDs, and Intel Ethernet portfolio.
What we're driving for on the silicon side is that in all cases, you get a better together experience, we drive from the engineering level, the work that Sailesh's team does with Rob Crooke's team on the SSD development leads to integrated features that deliver ever-escalating levels of performance. We work with VMware on how their software stack reaches down into that silicon and pulls through the value. In the case of this particular configuration, it was important that vSAN be able to allow for AVX-512 acceleration, and they needed to do specific work and tuning to their stack to allow for that. They also, with Optane SSDs, their application stack was not built to take advantage of that performance. Again, all of their work had been optimized on lower performing, higher latency SSDs, AKA everything else that had been available in the market.
We worked with them to optimize on the SSDs, and with VMD, which is the volume management device, it offers more reliability features and manageability and allowed them to add into the hyperconverged solution the ability for hot swap capability of drives. Now we've done that work, which is where in the past some of our ecosystem enabling has stopped. What we did on the configuration side was work through the most common use cases and figure out what the optimal configuration across cores, cache, memory, storage is in order to deliver that best performance possible. We took what might have been considered a theoretical best performance, and we're able to put it together in something that customers could actually replicate with very little work.
Again, if you think about a deficit of technical talent in the industry to support how fast IT demands are growing, you think about or I think about my experience in IT, of all of the testing we would have to do to get the recipe right, this is where the massive customer value comes in. You're removing the entirety of that testing and tuning, and we're giving them confidence because it's built on incredibly trusted partners like Intel, like VMware, and like our OEM partners that they know we've done that rigorous validation work. This stage of it is what really drives through to that acceleration. The end result in the case of this example was a 13x more VM supported, which again, is in an infrastructure service provider that's just flat out more money as they monetize on those virtualized machines.
Even in an enterprise, it's the ability to do more with less capacity and also be able to increase your utilization of your assets. At the same time, an 8x better price performance allows you to have the $ freed up to invest to continue to support the growing workloads. Let's hear from a few of our customers that have been working with these solutions.
Information technology dominates the world of healthcare now. In every facet of how we care for patients, there's elements of information technology.
At Youth Villages, IT has been an integral, crucial part of what makes this organization effective at what we do.
If you're in this business a long time, what you realize is that every customer's got a horror story for things that were supposed to work well together but didn't work well together.
Doing your own sizing and validation is very expensive in terms of time and, of course, money. You have to get in lots of components. You've got to put them together. You've got to make sure that they're all optimized and running together.
I look at it very differently than a straight IT issue. We reached out to Intel a couple of years ago and said, "Look, we have multiple partners, but Intel is uniquely capable of actually providing end-to-end solutions, which makes them a particularly interesting group to work with.
You're taking industry-leading products and technologies from Intel, and you're pairing those with industry-leading expertise, in the form of a reference architecture, you're putting those together into a really powerful package, in Intel Select Solutions that we can offer to clients.
When we implemented Select Solution, we were very surprised to see that the wait time of our enterprise storage went down 59 times. This is a tremendous amount of speed increase, and it allowed us to use a lot more data in a lot of areas.
It's allowing our people and our staff that work directly with our kids to have the best-in-class technology that makes their work more effective and more efficient.
Select Solutions for me means that we're actually developing very specific software technology that actually allows us to be able to approach our deep learning problems in a way that a healthcare provider would need to think about them.
The best case scenario for us is we don't have to worry about the platform, we don't have to worry about the infrastructure, that's really what the Select Solution allows customers to do and allows us to do as an integrator. It's just a great solution for our customers.
Okay. The example I had given and that some of the customers in there had used was around VMware vSAN, a very horizontal, large market opportunity focused on storage applications. I wanted to talk about one of our more specialized ones that's targeted at a very fast-growing segment. Personalized medicine is actually not a medical problem. It's a data problem. The ability to personalize requires that you have a tremendous amount of data to personalize off of. Currently, I don't know if all of you know this, but drug treatments are recommended and built off the assumption, and this is around the globe, that we are all 5 foot 10, 185 pound white men. We're not. That doesn't actually reflect who we all are. Well, a few people are checking. Okay.
The point being, it doesn't reflect who we all are, and it doesn't reflect how we are all going to respond to different medicines and medical treatment recommendations. In the future, we'll have this opportunity to analyze every person's genome at an affordable cost while protecting patient privacy. You'll be able to create databases that allow practitioners to reach in and seek for like patients to their patients and understand how drug interactions actually worked. First you have to deal with that data challenge. If every single genome sequence is 90 gigabytes of data, that actually adds up quite fast. We've been working for several years now to try to take that time to sequencing down from weeks into one single day, you also have to address the cost challenge that goes with that.
Time is of the essence when you're facing a life-threatening or critical disease. The increases in compute performance and storage capacity over the past decade have driven that time and cost of sequencing down by 10,000x. It's now at this sub-$1,000. If you think about the vast majority of the citizens of the world, that's still not an approachable number. There's work to be done. Intel worked together with the Broad Institute and MIT to create a Select Solution based on their genomics toolkit and drive further increases in the performance of genomics analytics. This solution not only has Intel silicon foundation of Xeon, plus Dan's FPGAs, Intel SSDs inside Omni-Path Architecture, but it's built on top of open source Linux-based software toolkit.
The results that the Broad Institute was able to see, then are making available to other users of the Broad toolkit, is going from a six weeks down to two weeks to analyze 2,300 genome sequences. They're seeing both a time reduction of six weeks down to two weeks, and in that two weeks, they're able to handle 5X more genomes. If you think of trying to build personalized medicine, again, off of 2,300 genomes, it's still not enough. They're now at the point where they're at the 11,000 amount that they can handle every two weeks. We've driven an incredible acceleration that's starting to make this possible, but we still see so much opportunity. We'll continue to invest in this over future generations in order to hit those goals of truly personalized medicine.
It's the end-to-end hardware to software view that has the opportunity to really make a difference in a lot of people's lives. Now that we've talked through a couple of the examples of what we're doing, I want to talk about the breadth of the portfolio. You guys have seen this slide a couple times today. You can see the variety of different areas in which we've invested analytics and high-performance computing, AI, network optimization, hybrid cloud. All very good focus areas. When we started Select Solutions a year ago, we had three solutions. Now, in the last year, working with over 30 industry partners, we've brought the number up to 16 as we've announced today, or 15, excuse me. The customers benefit, I talked about their faster time to deployment and time to market.
What we're starting to see in these first deployments is a 10% higher mix of Gold and Platinum CPUs. When you are having a conversation with customers about the purchase intent of solving their problem versus just a CPU conversation of purchasing, you can get into that total cost of ownership. It's oftentimes more and more realized that a higher performance CPU will actually deliver you a better overall total cost of ownership with very quick returns. In addition, we're seeing a doubling in deals that are driven by Select of adjacency attach rates. This isn't a surprise to us as all of the Select configurations, again, are built on that promise of one plus one equals two as you add in further components based on the Intel Silicon portfolio. Our customers are seeing value. We're starting to see market and financial benefit as well.
We've added some new ones to the portfolio. Naveen talked through the SAP HANA appliance, which gives us a great opportunity to help accelerate Optane memory's time to market. When you think about actually re-architecting the storage and memory hierarchy and the testing that one would want to do in order to ensure that that new hierarchy really works for their application, this will be one of the ways in which we will help accelerate customers' comfort and adoption and show them a ready-to-go offering on HANA, one of the most critical and targeted workloads for our Optane memory. This will be timed to be ready to come to market with our next gen Xeon Cascade Lake launch, and it will be our first 4-socket and above Intel Select Solution. The next one I wanted to talk about was the BigDL one.
I love this example because, again, if you think about a lot of IT and cloud service providers, they've put tremendous investment into their Apache Spark infrastructure and building out their data lake. Some of them have realized big data analytics capabilities on top of it, and some of them are just using it as a more efficient storage mechanism. The nice thing about BigDL, which is an open source deep learning framework that we developed with a lot of really technical expertise in the AI space, folks from Microsoft, Alibaba Group, Amazon, Telefónica, et cetera. The nice thing about BigDL is that it works on top of your Apache Spark existing infrastructure. You don't need to go buy any specialized hardware. It sits on top and allows you to start deriving data from all the work you've done to get your data lake together. It's a logical flow.
The configuration that we've built here is built off of the work that we've done of over 100 deployments of BigDL with customers around the globe. Taking all of their experiences and all of their feedback and putting that into an optimized configuration. I think there's just a tremendous amount of value that it is hard for any one company to gather that level of expertise, and it builds a lot of confidence and trust. This one will be built on the existing Xeon Scalable product, and we will be bringing it to market shortly, and then follow-ons, of course. The last one we will talk about for a minute is the blockchain Hyperledger. This is a more speculative workload as enterprises, cloud service providers, are starting to think about how blockchain plays in their infrastructure. Again, this isn't really targeted at any cryptocurrency type of use.
This is more about just distributed ledger management to be used across manufacturing and automotive and media rights management, a variety of different use cases. Any time you're working on something new and speculative, having the opportunity to start with a reference architecture that's been tested and validated takes the guesswork out and lets you focus on deriving value for your business out of whatever the blockchain promise fits into your solution stack. Those are our 3 new ones. Now that we've covered what we're doing at the node and multi-node level, and you heard Raejeanne and Naveen talk about the growth of the cloud and the massive efficiency and performance and scale that that has delivered for so many customers, we want to talk a little bit about how we're bringing those types of efficiencies from the architecture to even more data centers.
Intel's Rack-Scale Design is an innovative architecture concept and a design guide that enables pooling of your resources across your infrastructure. You have compute, you've got storage, you've got acceleration pools, and you have the ability to compose them, as we'll say, on the fly. If you think about a traditional server, it would have its own compute and storage and network acceleration, all that attached in one single sled. In this case, you have compute pools, and your application looks not at a particular piece of hardware, but at the whole of the hardware that's available. It is able to say, "Because of my needs, I pull extra compute and a limited amount of storage." As that workload gets run, it releases the capacity back to the pool, and the next application can be pooling.
It has a very sophisticated open source orchestration level that allows for multiple applications to be calling upon the same sources of hardware and pooling the resources that they need, all at a much more efficient and higher utilization than any one piece of infrastructure would be able to achieve on its own. This delivers much higher actual workload performance because your workload is literally getting the hardware it requires, and greater utilization of the entirety of your resource pool. We're investing in this to bring out the value of the entirety of our silicon kit and offer it out to customers. We've done a tremendous amount of ecosystem work, and one of the great values for IT is that the Rack-Scale Design architecture compliant hardware that they purchase and systems they purchase will be able to be managed across a common orchestration layer. Okay.
IT will see value, we think, from both of the Intel Select Solutions and from the Rack-Scale Design. With that, I wanted to show you guys the ecosystem. Yes, this is an aligned ecosystem, excuse me, across the Rack-Scale Design and Select Solutions. You guys have heard it a bunch today. As always, our ecosystem is one of our greatest investments and one of our greatest strengths that we'll continue to work so closely with to derive value. With this, we look forward to continuing to support our customers' data-centric innovation, and I want to just take one minute to wrap up the day. We have spent a good chunk of time together.
If you walk away with anything, I hope that you have a sense for how we're viewing the market, and that in this new era of data-centric compute, we really have an opportunity in front of us that is unprecedented, and that we have the right strategies, tools, skills, assets, and team to go out and not just address that, but win that all across the entirety of this data center infrastructure. I want to thank you for your time today, and we're going to wrap it up here and head off to the reception back where we had lunch. For those of you on the webcast, thank you for joining a long day of listening in, but we appreciate your time.