Ladies and gentlemen, please welcome Intel's Executive Vice President and General Manager, Data Center Group, Navin Shenoy.
Good morning. Good morning, everybody. Thank you all for coming. Thank you to Abir for that beautiful song to start us this morning. Welcome to all of you in the room here at New Lab in N.Y., and all of you out watching on the webcast. Today is a particularly exciting day for me. This is my first Xeon platform launch. Today isn't just another platform launch. Today, Intel is bringing the industry the biggest data center platform advancement in a decade. A platform that brings breakthrough performance, advanced security, and unmatched agility for the broadest set of workloads, from traditional business applications to cloud and network services, to emerging workloads like artificial intelligence and automated driving.
I took over running the data center business about 45 days ago. One thing that is abundantly clear to me in my first month and a half is that we have only just begun to see the impact technology can have in terms of transforming industry after industry. At Intel for decades now, we've been passionate about the power of technology to really drive massive industry transformation. Obviously, it's well-known that companies that have been born more recently in the cloud era take advantage of technology as a core part of their competitive advantage. To me, what's more interesting perhaps is that the breadth of traditional industries that are now turning to technology to deliver new and differentiated product services and experiences is accelerating faster than ever. Consider a very old industry, for example, farming.
This is an industry that was founded 10,000 years ago. Today, farmers are now using multispectral satellite imagery, in-field sensors, even drones to meet the incredible challenge of increasing food production without increasing their farmland. Another disruption we see happening clearly is in the retail industry. The offline brick and mortar traditional retailers now using sensors and analytics to dramatically improve their inventory management, customer experience, and quite frankly, to stay relevant versus their online competitors. We see in this context three large mega trends that will drive new business opportunities and growth for all of us in the industry. First, the shift to the cloud. In particular, the evolution of the cloud to a hybrid cloud to enable the most efficient and scalable infrastructure, is happening in front of our eyes.
Second, more and more companies will start to use the insights of analytics and artificial intelligence to power their next generation of services and products. Third, of course, the opportunities that we all see as the network moves from 4G to 5G. All of these trends will require an increasingly capable infrastructure. In the cloud, as I mentioned, we are clearly headed to what we see as a hybrid cloud end state. Companies will no longer have to choose between the efficiency and scale that they get from the public cloud and the trust and security that they see from on-premise deployments. An increasing amount of workloads will benefit from the cloud economics that we see of better efficiency and scale, whether they run in the public cloud or whether they run on premise.
Over the last three years at Intel, we've seen the number of Xeon processors that we ship into public and private cloud infrastructure nearly double. We anticipate that nearly 50% of our Xeon shipments by the end of 2017 will be deployed into either public or private cloud infrastructure. Second, the rise of AI and analytics. This is well documented. We've all read and heard about the rise of artificial intelligence, but I think it's important to remember that we are just at the beginning of this trend. While there's a massive amount of data that is sort of at the center of this trend, it's also important to remember that less than 1% of the data that is generated is actually being used, analyzed, or acted upon.
Therein lies the opportunity, huge opportunity to unlock the value of all that data through the use of analytics and AI. Businesses around the world will need to think strategically about the creation, the storage, the movement, the analysis of data, and quite frankly, develop intelligent data practices in an increasing fashion. The final trend, the move to 5G, is really much more than just a faster wireless network. 5G is about connecting things that previously haven't been connected, whether that be collecting data from parking meters or traffic lights to make everyday life easier, whether that be connecting industrial equipment to improve efficiency or safety, or whether it's about cars as they drive around sharing data with each other to enable safer, more efficient autonomous driving.
The sheer number of sensors across factories, cars, roadside infrastructure would completely overwhelm today's cellular technology, and that's where 5G comes in. 5G will provide a 1,000x improvement in the capacity, a 50x increase in peak data rates, and a 10x decrease in latency. This is a fundamental change and will require a fundamentally new network. A network that distributes the intelligence throughout, deploys cloud computing closer to the user, and ultimately, I believe this will lead to a massive build-out in infrastructure over the next several years. In the context of these mega trends, the demands placed on server, network, and storage infrastructure have never been greater. Enterprises will need to think about how they handle these new data workloads, seamlessly move, as I mentioned earlier, between the public and private cloud.
Cloud service providers will need to think about how do they improve performance, security, agility, utilization across their infrastructure. Communication service providers will need to deliver a completely new level of network agility and efficiency in the face of the new service demands as 5G becomes deployed. In the world of artificial intelligence, we will see a broad range of algorithms develop, a broad range of workloads develop, and this will require a broad range of solutions. Ultimately, the data center technology itself must transform on the inside to keep up and deliver the performance, security, and agility that businesses need. That is why I am super excited today to be introducing to all of you in the world at large, the latest addition to Intel's data center portfolio, the Intel Xeon Scalable platform.
This platform represents the best combination of leadership capabilities built on our nearly two-decade history and heritage in developing and innovating in the data center. The Intel Xeon Scalable platform that I'm holding here in my hands includes a completely re-architected microprocessor designed from the ground up for the data center specifically, offering greater levels of integration and workload-specific accelerators. The Xeon Scalable platform is the industry's highest performance per watt platform. It's designed for the hybrid cloud from the get-go, designed for data-fueled enterprises, designed for communication service providers. You will see revolutionary leaps forward in highly efficient packet processing, and security capabilities for virtual network functions. We'll talk a little bit more about that later. It includes huge performance increases in deep learning training and in inference for customers that are beginning their AI journey.
We have, as I mentioned, designed this platform from the ground up for the data center, and it's optimized for those three industry mega trends that I talked about earlier. In light of those mega trends, there are sort of three core functions that we thought about as we designed this product: scalable performance, pervasive security, and agility to quickly offer new services and capabilities as you see rapid growth in the business environment. From a performance standpoint, the Xeon Scalable platform is delivering a 1.6x improvement boost over our prior generation platform on a wide range of workloads. I want to put this in context for you. This change, generation to generation, is our largest gen on gen performance improvement in the past decade. This is a revolutionary change from generation to generation.
It's delivered through a well-balanced platform, 28 cores per socket, 50% more PCIe, and memory bandwidth. For emerging workloads like artificial intelligence, this platform offers a two times increase in hardware performance and a more than 100 times improvement with optimized software. We're hugely excited about the capabilities that this new platform delivers on performance. On security, as I mentioned earlier, we've optimized on a number of levels with a 2x improvement in data encryption performance, optimized particularly for data protection when the data's at rest, when the data's in use, and when the data is in flight. On agility, we are delivering a 4x improvement in greater virtualized workload throughput, which helps obviously reduce the total cost of ownership by a 65% versus the install base or a four-year-old server. This product has been in development for quite a while, five years in engineering development.
A huge leap forward in gen on gen performance, as I mentioned, for compute, for storage, and for the network and workloads that our customers care about. I want to take a minute here and specifically give a huge shout-out to all of the Intel engineers that have been hard at work on this platform, working literally day and night for many years to bring this product to culmination here today. Many of them are watching us here on the webcast. The most exciting thing to me about this new platform is in the way it helps our customers. No one understands the need for a scalable, a secure, and agile platform better than the communication service providers. They, as I mentioned earlier, are in the midst of thinking about how do they transform their fixed network functions to drive faster innovation in today's day and age.
How do they take advantage of the cloud? How do they prepare themselves for the 5G era? About a year ago, AT&T joined us on stage at the Intel Developer Forum, they discussed how they plan to transform their network. How they were setting themselves up on the path to 5G. To give us an update on their progress and how they're using the new Intel Xeon platform, I'm pleased to welcome the Chief Strategy Officer and Group President of AT&T Technology and Operations, Mr. John Donovan. Please welcome John. How you doing, John? Good to see you. Thanks for being here.
Thanks, Navin.
It's great to have you back. We've been pretty busy, as you've heard. You've been busy. What's been going on since we talked last in August?
Well, let me start by saying. Do we have a microphone?
Can you guys hear John?
A little mic malfunction. Can I borrow yours?
I'll walk close. Can you hear him now? Yeah.
Yeah. Okay, great. It helps when they do that. I was sitting there, and it dawned on me that when you guys plus-oned us last year, and I was sitting backstage, I thought to myself, "I wonder if we can actually make a difference. I wonder if I'm here as a partner of yours, walking down the catwalk on the partner parade, whether I would actually make something work." It wasn't that I was worried about your technology. I kind of wondered whether AT&T would execute, whether we could make a difference in speeding ourselves up and transforming ourselves. It's made a big difference for us in our evolution to what we're calling Network 3.0.
Great. Last August, one of the things you asked about from us was to provide you early access to new technology to help you accelerate new services to customers. Grade us. How have we done on this?
Well, I would give you an A. There's always room for improvement with an A+, as I always tell my kids. When you think about what we've done with the Xeon platform, it's unthinkable for us a year ago to have considered that we'd be able to take your latest technology and within a matter of weeks, get that into production. We got the new processor starting in March. We have production traffic already carrying over the platform. We've seen about a 30% performance improvement over your prior platform. For us, the reason that matters is measurably our total cost of ownership. When you think about how the number of servers that we need in a cluster, we're looking at about a 25% real improvement.
Wow.
When you think about that, and you look at your sort of theoretical numbers on what's available performance, what matters for us is when the rubber meets the road. We're making great progress on being able to do more things faster as a result of the program. I would also highlight is that it isn't just AT&T. What we're thrilled about now is software-defined networking is becoming a global phenomenon in our industry, and we've needed that so that our industry could evolve at a more rapid pace. I would say you helping us, help you, help the industry builds this ecosystem of innovation. ONAP, which is our operating system open source, we're up now to 55% of the global wireless subscribers that the carriers that represent that have signed on to this program.
We're seeing a real windfall, if you will, as a result of how we've been able to speed up the technology.
That's great, John. Can you give us a little bit more on how the transformation you're going under is benefiting your customers?
Well, the customer use cases are always where the rubber meets the road. We always, within AT&T, talk about half the stuff we do should be doing things more efficiently, and the other half of the things we do ought to be touchable by the customer. What we've done is we have a platform we call FlexWare, which is taking a lot of the network functions that used to provision in months, and provide that as a software capability to our customers. When you think about a common network operating system for service providers, you think about big iron routers and switches and all of that sort of stuff, and that now is becoming software.
What we've been able to do is take our network innovation and not only speed it up for our customers, but give control over to the customers so they can provision a firewall as software. They can add routing and switching capability as software. It's really transformed not only what we take to market but how the customers consume the products that we offer.
That's great. What's next? Where are you headed from here?
You look at the trends that you had laid out there. We're on trying to enable all that. 5G as a network is a game changer. Always, we would look at our G evolution, 3G, 4G, and we do it based on the economics of efficiency, more economical use of spectrum. What's different about 5G is whole generations of technologies will lift and architect differently because the network will be high reliability, super high speed, but also super low latency. That real-time dimension really takes compute to the edge of the network. The network has always been about compute, storage, and networking. As you think about merging those and pushing those closer to the edge, it's really a big game changer for us.
It's exciting times for us because we're taking not only the new stuff like 5G and AI and ML and putting that into networking But we're also finding we can go back to our legacy environment, and we're now getting the economics to allow us to transform there as well. It's very comprehensive in what we're getting done here.
Thank you, John, for being here, and credit to you for being a leader in the industry and being early on a technology transition. We're thrilled you've seen benefits from the Xeon processor, and we look forward to the continued collaboration.
Great. Thank you.
Thank you very much, John. Take care.
Appreciate it.
Yep. As you have heard, we are super excited about the new Intel Scalable platform. Some of the largest companies in the world, such as AT&T, are already taking advantage of it. There's so much more to share with you guys here today. What I'd like to do now is introduce Lisa Spelman, our vice president and general manager of Intel Xeon processors, to get into much more detail. Lisa.
I got the music. How sweet. All right. Thank you, Navin. It's great to be here today at New Lab and to have the opportunity to share with all of you what's going on and a little bit more about what's under the hood of the Xeon Scalable processor here. Before I dive into the details, I think it helps to put the product into a little bit of context. To deliver that business transformation that Navin and John were just talking about, you have to understand the infrastructure requirements, because after all, the infrastructure is the fundamental or foundational layer and building block for the delivery of next-generation services.
Working with customers who are trying to accelerate their own growth while remaining competitive in quickly changing business landscape, we find that infrastructure requirements are actually have a lot of consistency regardless of the type of customer or industry that you're in. Scalable performance is required. Pervasive security and control has just exploded up the requirements chart, and the agile infrastructure is required for scaling quickly and delivering to business growth that you may not be able to perfectly predict. To dive into each of these pillars and how we're delivering, let's start with performance here. Performance is critical to service providers who want to offer the absolute best experience to their customers. Also to enterprises, researchers who are trying to take the next steps in innovation and get the most out of the data that they already have.
We have a long history of delivering the strong generation on generation performance gains. This generation is no exception to that. Of course, we're delivering on traditional performance drivers like Navin talked about, the more cores, the higher performing cores, more memory, more IO. Intel's taking it far beyond that. We're moving beyond the basics, and we're adding unique innovation that truly address the unique requirements of data center workloads. Let's dig into these. The first one I want to talk about is Intel's AVX-512. It's a powerful instruction set in the Xeon Scalable processor that accelerates data processing. This has traditionally been used or valued in the high-performance computing space, but is actually expanding and growing into new workloads as well. It takes vector performance to a new level by doubling the width of the registers and then doubling again the number of registers available.
This results in a 2x flops per clock efficiency compared to the previous generation, and it makes vector computation a lot more effective and more applicable across those wider workloads. Like I said, if you're doing scientific modeling, engineering simulations, you know that this will benefit you. What's come out as we've done that early testing that some of our customers are already engaged in, you find newer and growing workloads like genomic research, virtual reality content rendering, and especially AI benefit as well. I have a lot of real-world results that I'll show shortly. Next, I want to introduce you to Intel QuickAssist. Some of you have met Intel QuickAssist, especially if you are in the comm service provider space and deliver network functions. This is a capability born of network workload requirements, and that's been its history when we've offered it as a discrete product.
The original capability was driven to reduce the size of the data packets getting more and more throughput through your network infrastructure. As that has been deployed, we've also worked with our customers across the cloud service providers and across enterprises and realized that QuickAssist Technology delivers great speedups for security algorithms as well. With QuickAssist integrated into the chipset in this generation, you get 100 gigabits of cryptography and 100 gigabits of compression that allows you to increase the overall workload performance while freeing up that precious compute capacity on the cores for higher order functions. This is especially important if you are in a business that monetizes or derives value from each of those cores.
I've saved the best for last here in the performance realm, and I want to show you what we mean when we say that Xeon Scalable platform is truly architected from the ground up for the data center. Let me introduce you to our new innovative Intel Mesh architecture. It's a fundamental change to the processor design that allows us to increase performance by optimizing data sharing and memory access between all CPU cores. First, a little bit of history. For the past decade or so, we've used the same core topology in our CPUs called the ring architecture.
Each generation, we added more cores, more memory, and I/O, and based on our workload modeling and looking at future trends in the data center and network requirements, we realized that Xeon Scalable wouldn't actually scale the performance that we wanted to see from the product if we didn't increase the bandwidth of the interconnect between the cores, the cache, the memory, and the I/O. Without change, the interconnect, which adds a lot of value, suddenly becomes a limiter. It's a bottleneck that can increase latencies and overall not deliver to the performance we want for our customers. That's why I'm excited to unveil this innovation. It provides a more direct data path for traveling between cores than the prior ring architecture did.
For example, in the previous design, data would have to go around the ring if you're starting at the last core and need to get to the one on the other side, through a buffer, and all the way back around the second ring to get to the final destination. With mesh, the data simply cuts across the improved and increased number of data pathways and avoids the buffer. Also, CPUs don't actually process one cache line at a time. A real data center CPU needs to have all cores accessing all memory, all I/O at the same time, and the mesh fundamentally provides this. While the ring worked well for the amount of cores that existed in previous generations, going up to 28 high-performance cores made it imperative that we rethink here and re-architect. That's what data center workloads require.
It is one thing to just add more cores, more memory, or more I/O, but entirely another to ensure that it's delivered in a well-balanced system that allows workloads to actually take advantage of those resources and get that performance. How do these features actually translate into benefit? I'll start with the world records. They're delivering a stunning 58 new performance world records and counting. The numbers keep going up. The Intel Xeon Scalable platform has reset the bar for performance in the industry. It's an enormous advancement in the number of world records at a processor launch and indicative of the broad range of workloads and platform configurations that you're going to see in the market. I want to send a huge congratulations to our partners for delivering. The world's data centers don't actually run benchmarks.
Benchmarks are a great proxy, what they run is real workloads that deliver value for businesses. We invest deeply with a broad ecosystem of software innovators in order to optimize their software to take advantage of all of the new features we're putting in. Today, I'd like to share just some of the results that you're seeing, and you can find much more of the content online. In the cloud space, for example, Tencent offers a video stitching application for its users. This allows users to create their own virtual reality content or immersive 360-degree video. Leveraging AVX-512 and the Xeon Scalable cores, they're now able to provide that service to their end customers 72% faster. It just fundamentally leads to a better user experience.
All of us have spent time waiting for video, when you can reduce that, you're going to be able to do even more with the capability. In the analytics space, the SAS application provides an integrated environment for modeling, data mining, simulations, forecasting. I love this one because it's a combination of a few of Intel's technologies in a platform-connected better together story. The Xeon Scalable product, coupled with our innovative new Intel Optane SSDs, delivers a 2X improvement over the prior generation. In one generation, to double the performance of your application is astounding, this allows the SAS customers to run more complex analysis and a larger dataset. When you are able to do the next click of the complexity across bigger sets of data, you actually get deeper insight into the business challenge you're trying to solve.
The improvements are driven by the ability to move large blocks of data more efficiently and to pull more compute power from the Xeon Scalable compute cores. The last one I wanted to highlight in the networking space is from Telefónica. Their NFV lab conducted an evaluation of Xeon Scalable and found that they could deliver a 67% performance increase across a virtualized broadband network gateway. Again, I know so many of you run your own virtualized broadband network gateway. What does that really mean? This gives them the ability to handle not just the current data load of traffic that's sitting on the network as everyone's out there on their phones and using it to transmit across so many use cases, it prepares them fundamentally for a 5G-ready network and the coming deluge of data that they're forecasting they need to address.
Given the new feature advancements that we put in, the benchmark results you've seen, the real workload results that you've seen, I don't think it's a surprise that we had customers that wanted to get their hands on the technology early. Companies that embrace technology as a core business differentiator have an almost insatiable demand for compute, and they want that delivered at a consistent beat rate. To address this demand, we undertook our most ambitious early ship program that we've ever delivered in the Data Center Group, delivering production processors ahead of launch to cloud service providers, enterprises, high-performance computing users, and comm service providers like AT&T. Starting in November of last year, we have sold over 500,000 processors optimized for data center workloads to over 30 customers. What have they been doing with it?
We were excited to see that even ahead of our general availability launch, we had three new entrants into the most recent top 500 supercomputer lists based on Xeon Scalable. Enterprises are ramping. Comm service providers, you heard John demonstrate how he's already moved it into production, and cloud service providers have already deployed Xeon Scalable solutions and are offering cloud services based on this technology. To give a little more insight into the impact of this program, I want to share a bit about Google's production deployment. Google first announced their intent to deploy Xeon Scalable with us in November of 2016. That was actually not the beginning of the collaboration. That was the culmination of a multi-year, deep technical engagement to define the features and the performance that Google's unique, innovative cloud architecture requires.
As a result of this collaboration, Google was our first customer to launch the Xeon Scalable instances this February, and now has customers from retail to financial services, oil and gas, education, running in production. I could tell you more about it, but I thought it would be best to show you what Bart Sano, Google's Vice President of Platforms, has to say.
We were first to launch Intel's Xeon Scalable processor in our Google Compute platform. The customers have been able to run millions of hours of computing workloads on the Xeon VMs. They're seeing increased performance for a variety of applications, such as scientific modeling, genomic research, 3D rendering, data analytics, and even engineering simulations. Our customers have reported up to 40% improvement in many cases versus previous platforms. In some cases, where the customers tuned for the AVX-512, they saw more than 100% improvement.
Bart spoke eloquently about the results his customers are seeing. One of the examples that we were excited to see was Seven Bridges, a genomics modeling company, seeing a 1.8x performance improvement. Again, the ability to deliver that much performance improvement in a generation fundamentally allows Seven Bridges to analyze that much more of the human genomic data, further improving and refining the results that they can deliver to the industry and to their customers. Not a surprise that Google was not alone in their desire to bring this platform to market and one of the best parts of the journey that we've been on has been working with the industry.
For the first time, we are bringing out our two, four, and eight socket and above capability processors at the same time, which we think delivers tremendous benefit for data center operators and infrastructure managers as they work through their deployments. The industry of leading system providers has worked in close collaboration with Intel for the last five years to develop an astounding array of systems across compute, storage, and networking capabilities. With over 450 systems planned, I think that our customers will find they have a lot of choice. I'd like to take just a moment to thank the ecosystem for the incredible work and the innovation that they have built on top of Xeon Scalable. Thank you.
The continued ramp of the product will be supported by our more than 480 builders members across cloud, network, storage, and fabric, bringing out optimized technical documentation and reference architectures to speed deployments. One of the integrations in this generation that we're excited about and the builders ecosystem will certainly help drive is the integration of Xeon Scalable with our Omni-Path Fabric for high-performance computing. We think this delivers excellent latency, port performance, and a great TCO. Our fundamental data center security strategy that Navin talked about is to deliver security without compromise. We want to protect the data, again, at rest, in flight, in use, and secure the fundamental hardware platform. With Xeon Scalable, we believe we have nearly eliminated the performance overhead for encryption of data on common usages. This is achieved through AES-NI instructions, a higher performance core, and the addition of the AVX-512 acceleration.
As you can see, as we talked about, we're delivering that 2x performance improvement gen over gen. We're also layering in new features, such as Intel Key Protection Technology, which fundamentally protects encryption keys when paired with Quick Assist. This makes key storage and access, which is so critical to encryption, more secure. With Key Protection, you're able to prevent exposure of encryption keys from software attacks, and your keys are never present in memory. A simple way to think of Key Protection Technology is a hardware lockbox for encryption keys. The end result of this is that less than 1% impact to performance when you have encryption turned on. We've also invested in generations in securing the fundamental platform hardware root of trust, and will continue to make advancements here through new features that we're introducing today.
From boot to BIOS, to virtualized environments, to cloud, you have hardware security built in. This all comes together and allows for IT administrators and service providers to get ever closer to achieving the goal of implementing encryption and security across all of the data without impacting the user experience or the SLA that they're committed to. To put this into real-world context, I want to spend a minute talking about an emerging data center workload that we're excited about because we think Intel Xeon Scalable is uniquely positioned to deliver a lot of value here, and that's blockchain. Blockchain has the opportunity to be a highly disruptive technology and is emerging in a lot of new industry consortia as we figure out how this will come to market and come to bear.
Briefly, blockchain is a transparent digital ledger of economic transactions, and it eliminates the need for a central authority to verify those transactions. Each time there is a new transaction, the ledger gets updated and it's cross-checked across the array of users all over. Theoretically, it would be impossible to corrupt when there are so many checks and balances fundamentally built into the application. This is accomplished by relying on comprehensive encryption throughout the entire transaction chain. We've delivered blockchain reference software code named Sawtooth Lake as part of our participation in open source Hyperledger project. We're also participating in an industry organization such as the Enterprise Ethereum Alliance to accelerate the deployment of blockchain technologies. Recently, we announced an investment in a leading blockchain provider, R3, and today I'm happy to announce a new technical collaboration with R3 built on Xeon Scalable.
We're going to deliver full platform optimizations of the R3 Corda platform, which is a leading distributed ledger, and we're going to do that with Xeon Scalable features and capabilities built in. This will be critical for the financial services industry. The industry sees the AVX-512 Quick Assist Key Protection Technology and fundamentally the performance that Xeon Scalable delivers as the foundation for a blockchain platform. I'd like to introduce a company that is not only participating in the blockchain advancements, but also has had the opportunity to use our Xeon Scalable platform. Thomson Reuters is a global media and information company.
Many of you may know them as a network of 2,600 journalists delivering global news services, they have many other products and services, including a lot of work they do with the financial services industry to deliver information and analysis about financial markets. To share how they're applying data center technology to meet their growing demands, I'd like to welcome the Head of Core Technology Assets and Real-Time Product Management, Peter Marsden. Thank you.
Thank you, Lisa.
Yeah, thank you for coming.
Nice that I'm actually here today.
Yeah. You're here, you're going to talk to us about this. I am hoping that what you'll share with us is a little bit about the value of data that Thomson Reuters has already discovered. A lot of companies and industries are just beginning their journey here, but I think you guys have a pretty rich history.
Yeah. I guess we've been driving data through the industry, the financial industry, since, well, since pigeons were flying. We started a really deep collaboration with Intel back in the '80s, when we were creating our first PC-based platforms, for desktop use, but also for a growing industry of people using applications to consume data. That's really led to the growth of data in the financial industry, where we now see millions of transactions occurring every second, vast growth of the amount of data being transported around the world, but also being collected and being used for analytics every day. Thomson Reuters has been key to driving that industry to be increasing the amount of attention paid to analytics, as well as the complete flow of data.
I understand, I shared with the audience a little bit about our early ship program that we undertook, and Thomson Reuters actually had the opportunity to be one of our key partners there and to test out the platform. What did you guys see with the work you did with Xeon Scalable?
I think our experience with Xeon Scalable covers virtually every aspect that you were talking about earlier, Lisa. We've seen, just from a pure performance perspective in the analytics space, analytics computation in our velocity analytics scalable platform, around about twice the performance that we'd seen with prior generations-
Okay
of platform. In the real-time space, where we're delivering real-time updating, stock exchange, every type of asset class that you can see, we've seen that at least 25% improvement in that space. We're reasonably sure that we can see further improvement in that space as well.
Okay.
If you think about the way that we're using it, we're using it just purely for performance improvement, but also things like the Optane and FlashArray, the whole gamut of pieces.
That's impressive.
We're also seeing, if we think back to the industry, back in 2002, there were probably about 15,000 messages a second traveling across our network. Today, that's up to 75 million messages a second, and we continue to see that growth all the time. With the tests that we performed, we are already getting a benchmark of about 88 million messages per second coming through, and we think we'll be able to scale that up to 100 million without significant problems.
Okay.
That demonstrates not only are we able to use the technology, the AVX-512 vectors, et cetera, but we're actually seeing fewer cores being necessary to deliver the same benefit. That's very significant to us as an organization. We're also looking at the mesh architecture.
Yeah
which is important to us because not only do people need huge amounts of data, they need that data predictably. They need to know what the latency's going to be and minimize the amount of jitter within that delivery. Reduction in latency, improvement in jitter is a huge benefit to us as well.
Your bucket of use cases there is really, like you said, it's a lot of what we talked about, is that more fundamentally agile and efficient infrastructure so that you guys can do more and deliver more services. I am thrilled to hear that the early ship program was successful for you, and you've had so much success so far. What's next for Thomson Reuters on your transformation journey?
I think continuing to drive these things into production is a key part of that. Also some of the points that were covered earlier, the embedded security features, the efficiency around encryption-
Okay
key to our customers that we're able to deliver securely as well as deliver fast and with huge scale. That improved compression and encryption is clearly important. You mentioned R3. We're also partners with R3, and I think that's proving to be a really powerful blockchain partnership for us, where we're providing API access to financial information through the R3 partnership. We see that really is, blockchain industry as a whole is going to be able to take significant cost out of back office operations within the financial industry. Longer term, we're also expecting to see more work on cloud enabling our capabilities. At the moment, we're predominantly a data center-centric infrastructure. We're expecting to see more and more use of cloud-native technologies as we move forward.
That's excellent to hear. We are excited to continue the partnership, and it turns out we know a thing or two about cloud. We're here to help you on that journey as well. I just think there's so much opportunity ahead in the financial services from where you're at today and getting all the way through blockchain. I look forward to working together.
Lisa, thank you very much.
Thank you. Yes, thank you for joining me. Let's transition now to our third pillar on agility. At Intel, we define agility as just fundamentally the ability to do more and deliver more with your infrastructure. Data center infrastructure needs to be responsive. It needs to be much more well-utilized. It needs to be energy and space efficient and easy to deploy for that agile service delivery we keep talking about. Over a decade ago, Intel invented virtualization technology. In every generation, we add more and more capability and performance to reduce the overhead of virtual machine migration. This generation, we've added mode-based execution that delivers a more secure virtualized environment that hypervisors can more reliably verify and enforce integrity of kernel-level code. We've also added features to run sure technologies as part of our advanced RAS. This increases system uptime and is crucial for mission-critical workloads.
Adaptive multi-device error correction is a resilient memory technology that enhances memory error correction while also balancing performance impacts. In addition, we've extended our advanced RAS feature support to be available now in 2, 4, and 8-socket and above configurations. With Xeon Scalable platform, we made sure to extend agility into the storage space with intelligent capabilities like the volume management device. It's designed to deliver seamless management of PCIe-based NVMe solid-state drives, including enabling hot plug capability that minimizes service interruption during hot swaps. This will work with both 3D NAND and our Intel Optane drives. We also see Xeon Scalable as fundamentally the most agile AI platform, and I'll talk a little bit more about that. We're delivering a portfolio of products in the artificial intelligence space that vary in their level of specialization for AI workloads.
Xeon Scalable is the most agile because it can do both training and inference and do them well with good performance. It offers businesses that have a tremendous amount of Xeon-based infrastructure to take that first step forward into their deep learning inference and training workloads at scale with their existing data centers. For example, Xeon Scalable plus optimized software, you can see 113x performance improvement versus a three-year-old system for deep learning training. If you look at the inference portion of the artificial intelligence workloads, I think it's fairly well known that inference runs on Xeon, but I thought it'd be useful to take a moment to explain why that is. Inference handles a much higher volume of data than training and is generally built into a production real-time workload. Think of something like search. It benefits from the scalability, the performance, and the latency that Xeon delivers.
On inference, we're delivering a 2.4x improvement generation over generation through the improved core performance, AVX-512, the mesh architecture, and improved cache handling. I'd like to share with you how one of the leading AI companies in the world is using Xeon Scalable to deliver AI performance.
Machine learning is going through something of a renaissance because the challenges associated to operating at scale have been removed in the cloud. Today, there are thousands of engineers focused on applying these techniques across all areas of the Amazon business with use cases like Alexa, customer service, robotic fulfillment, drone delivery with Prime Air, Amazon Go. AWS customers are running large, sophisticated machine learning models on AWS, and the computational power of the Intel Xeon Scalable processor lets them use more data to create innovative new products and experiences powered by machine learning. Together with Intel, we've optimized deep learning engines with the latest version of the Intel Math Kernel Library and the Intel Xeon Scalable processors available on the new C5 instance family to increase inference performance by over 100x.
We couldn't be more excited to see how customers put them to use as we continue to pioneer with our friends at Intel.
We look forward to continuing our collaboration with Amazon in this exciting space. As you all know, AI will not be a workload that is just for the cloud service providers. It has the potential to impact all industries and all applications. I'd like to introduce now a leading healthcare company so they can share their analytics and AI journey. Montefiore Health is part of the healthcare industry that's poised for significant transformation. To share their new model in patient care, I'd like to welcome to the stage Dr. Andrew Racine, who's Assistant Senior Vice President and the Chief Medical Officer at Montefiore Health System. Thank you. Thank you for joining us.
Sure.
I would love it if you would just start here and share a little more about what Montefiore is and does.
Thank you. First of all, thank you for inviting us. Montefiore is the academic health system for the Bronx.
Okay.
One of the most diverse, vibrant communities in the United States. It also happens to be the poorest urban county in the country.
Wow.
Many years ago, it came to us looking at our ability to provide care for patients in the Bronx that a fee-for-service model was no longer going to be sufficient. We migrated fairly aggressively away from fee-for-service medicine into risk-based analytics. What that meant is we would go to the payers, and we would say, "We are willing to accept responsibility for the outcome of the patients. That will give us more flexibility if we move up the premium stream with them." Currently, we have about 450,000 patients under some kind of risk-based contracting, but our goal is to move that to 1 million patients.
Okay.
In order to be able to successfully manage the healthcare of that large number of people, you need to have extremely sophisticated ways of processing the information about those patients to keep them healthy.
It's an amazing transformation, just thinking of fundamentally changing how healthcare services are delivered. Yeah, you need to know a lot about the patients or the data set as us in technology might call it. Can you share with us how you've used big data and data analytics and advanced computational platforms to help deliver on this new business model?
Sure. From our standpoint, what's going to distinguish advanced health systems in the future from others is the realization on the part of those systems that what they are is fundamentally information management systems.
Yep.
They're going to need to have the ability to gather, to store, to aggregate, and to analyze large amounts of information from data matrices on vast numbers of patients.
Yep.
We actually have a fairly good track record of that at Montefiore.
Yep.
We've been doing predictive analytics for some time, but it's also become obvious to us that there is a significant difference between human modeling for predictive analytics and machine learning, artificial intelligence.
It's the difference between trying to identify a needle in a haystack from trying to identify a series of needles in an entire field of hay. It's clear to us that if you can get those kinds of parallel processing taking place, you can manipulate vast quantities of data and come to the sorts of clinical inferences that you need. I'll give you an example.
Okay.
In our intensive care units currently right now, we have a system that has been developed by two of our physicians, Dr. Michelle Dunn, who is an intensivist, and Dr. Parsa Mirhaji, who is an informaticist. This process that they have instituted allows us to use real-time streaming data from patients who are in the ICU to predict who among those patients will, 24, 48, 72 hours from the current moment, be at risk for respiratory failure. That is a breakthrough technology that allows us to use information in real time to make clinical decisions that are going to allow us to intervene with patients and prevent them from having adverse outcomes. In order to do that, you need hardware that can catalog streaming data in real time-
Yep
from patients. That's really how we put to use our Intel technology.
I think that's incredibly impressive. I think your patients, ultimately, they won't think of it this way, but will certainly appreciate that investment in technology that you're making to deliver that type of advancement and that fundamentally better care. Your journey resonates with me that so many enterprises are going through. It's just the application that you have is so personal. I know you guys have had the opportunity again to use the Xeon Scalable, your data lake, use your semantic data lake into this new processor. What are you seeing?
Well, what we have found is that the combination of the scalable technology and our semantic data lake has allowed us to essentially do these analytics much more quickly-
Okay
at a much lower cost.
Good.
What that does is it allows us to approach certain kinds of clinical challenges that we otherwise would not be capable of doing. I think it's equally important for us to acknowledge what Navin had been said earlier, is that what we now have is a system with a semantic data lake and the Intel technology, something that has high performance, that's agile, and that's also secure.
Yes.
From the standpoint of a health system, the security of data is actually a paramount.
Quite important.
paramount importance.
I can-
It also allows us to spend less time on the technology-
Yeah
more time on the patients, which is really what we want.
We'll spend the time on the technology.
There you go.
You can spend the time on the patients. That can be our deal.
That's our part.
We're thrilled to be contributing to all the great work that you and the team at Montefiore are doing. What's next?
Right. What's next is we've been applying these technologies to some of the most vulnerable patients in our system, patients in the ICU.
Okay.
Fortunately, most of our 450,000 patients will not end up in the ICU. We need to be able to look at vast numbers of patients who are living their lives, and the real challenge for us is not to be able to predict who's going to get into trouble 48 hours from now, two days from now. For us, the challenge is to be able to predict who's going to develop chronic conditions two years from now.
Okay.
Knowing what we know about them now, even when they feel well and they look well, we need to be able to use the information available in terms of their genetics, in terms of their socioeconomic status.
Yep
in terms of their utilization patterns, in terms of their preferences, to be able to predict who it is that we need to pay attention to now to prevent them from going on to develop the kinds of clinical conditions that are going to be adverse for them.
Yeah.
That's the frontier. That's the golden fleece.
Yeah, truly preventative medicine. That's impressive. Well, thank you so much for coming and sharing your journey with us.
Absolutely. Thank you.
We look forward to continuing to partner on it. It's been amazing to roll out and share with you guys these results and what the product is built on. It turns out that even the best data center innovations cannot be used if deployment is too complex. This is a growing problem as more and more of business transformation is built upon a technology foundation. We've heard from data center managers across the globe that while the industry has made great headway in delivering solution stacks required to implement agile cloud infrastructure, the time and money spent on evaluating different configurations and finding that optimal performance tuning is still a barrier to that fast and speedy deployment. To address these challenges, Intel and our ecosystem partners are taking the work that we have done in the Builders programs, we're moving it to the next level.
That next level starts today. This major step is that we're announcing the Intel Select Solutions. We've worked with the leading ISVs and system vendors around the globe to truly the brightest minds in hardware, in software, and in systems innovation to deliver these new solutions. We're delivering reference architectures that are optimized across hardware, firmware, and software. We've delivered a rigorous validation process, verification process to ensure that solutions meet or exceed predefined system benchmarks. Intel Select Solutions will reduce that complexity that is hampering so many deployments. It will allow infrastructure operators to deploy more quickly and more easily for faster data center communications and communications networks. These verified solutions will come on a trusted Intel architecture foundation.
Today, we're launching three initial designs, VMware vSAN 6.6 for enterprise on-prem and cloud service providers, Microsoft SQL Server 2016, also for enterprise on-prem and cloud service providers, and NFVI on Ubuntu for comm service providers. AT&T, John, who was just here moments ago, is one of our leading partners in the NFVI Select Solutions configuration. You know it's truly industry-tested. Our partners shown here intend to launch Intel Select Solutions with their unique and added innovation. We look forward to bringing this list now and many more to come to the market. I'd like to take a chance to recap the news of the day and show you this beauty in person. Today, we're delivering the world's biggest data center advancement in a decade. It's industry-leading performance across the widest range of workloads.
Innovation and architecture is designed uniquely for data center and network workloads, and an unmatched ecosystem that is highly optimized and ready to deliver today. Xeon Scalable is the foundational platform for business transformation. Whether you've joined us here today on the webcast or you're here in person, the conversation continues. I certainly encourage you to participate in the virtual experience we've developed online. You'll hear thought-provoking videos from leaders across our organization and the industry at large, technical sessions on topics such as hybrid cloud, high-performance computing, network transformation, artificial intelligence, plus technology demos, and a showcase that represents more than 150 of Intel's ecosystem partners. Thank you again for being part of this momentous event. I now invite you to fast-forward to your next.
Ladies and gentlemen, we now invite you to join us for lunch.
I mean, you can feel it. One, two. One, two. One, two. Down. What's that? It's just the bass. Right. What about this? The Bronco? I think the ones back there really helped out though, too. You don't have to push everything as much. Perfect to here. Yeah. Get out of that. Aha. Mute those, Unmute B. These are the Jess. Which is what? One, two. Check one, two. Yeah. This is Jess. Up. Evils. Okay. Evils. One, two, evils three. Not bad. I mean, one, two. Check. One, two. A. A. A. A. One, two. Check. One, two. Mic. One, two. It's always that downbeat. Hey. All right. One, two. Yeah. One, two. One, two. One, two. Good job. One, two. One, two. Three. Four. Five. Six. Seven. Eight. Nine. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25.
26. 27. 28. 29. 30. 31.
Right in the face.
Get some sunglasses over here.
Front row has a few seats available. I see you're leaning in. You're sitting there. I think I have a reserved seat up here. Someone can have it. I think I'm going to be here.
See stats on the formula?
No.
Good.
It's glowing? I haven't even had a minute to look at my phone.
It's good.
Thanks everybody for joining the launch, either live here in the room or on the webcast. We're going to do an investor-focused Q&A, but before we get started, just going to quickly hit risk factors. One quick point, everybody here knows we're in the quiet period. To that point, we're not going to discuss anything related to Q2 trending. Also, because we're in the quiet period, we're not going to discuss 2017 or long-term data center financial guides. All statements made that are non-historical facts are subject to a number of risks and uncertainties, and actual results may differ materially. Please refer to our most recent earnings release, 10-K, for more information on the specific risk factors. All right. With that, we're going to go ahead and get started. Mark and I will be running mics.
We're just going to be informal on the Q&A. Vivek.
Thank you. Vivek Arya from Bank of America Merrill Lynch. Thanks for the presentation today. I actually had two questions, one probably fairly easy. I think, Navin, you mentioned in your presentation that this class of product is giving 65% better performance versus something that was four years old. I'm just trying to put that in context. Is that the usual level of performance gains that people-
Yeah
do every four years? I have a follow-up question.
Let me clarify, maybe I misspoke. What I meant to say was 65% improvement versus the last gen to gen. Broadwell versus Skylake, across a broad range of workloads, we're delivering 65% improvement. It's not versus a four-year-old system. That's from a performance point of view. The other data point that I gave was a 65% TCO benefit.
TCO, I think is probably.
Okay, the 65% TCO benefit was related to virtualized workloads. This platform, this new Purley platform, the Xeon Scalable platform, I got to make sure I say it that way. The Xeon Scalable platform has over 4x improvement in terms of virtualized workload throughput. As a result of that, for companies that are running virtualized workloads, you get 65% TCO benefit. They happen to both be 65, but they're different and different points.
Got it.
The real, I think, important thing to keep in mind is the leap we've made gen on gen from a performance point of view. That is not usual, right? That is why we say this is the biggest advancement we've had in data center platforms in a decade. We don't consistently deliver 65%, right? It takes a lot of innovation to accumulate and come together to deliver that big a leap, gen on gen.
All right. That was the easier question. The other question is more on the competitive landscape, I understand your competitor has not actually released the product yet, there's nothing really to compare it to in real-world workloads. Assuming their claims are right.
Can you be more specific about which competitor? I have a few.
Okay. Let's talk AMD first. We can talk about other competitors as well.
Okay.
What AMD is claiming is that at the low end of the stack in single socket and the lower end of dual socket products, that they can offer a lot more memory, a lot more PCIe lanes, et cetera. They can offer a lot more features. Even if they don't outmatch Intel in terms of raw compute performance, they can provide a lot more of these features at the lower end of the stack. I'm trying to put that in the context of does it matter? Is that an important differentiator? Does it make you do things different? Does it matter? Is that disruption to you in any way?
Sure. I'll start and then maybe Lisa and Sailesh, you guys can add. Look, we've had 1 socket, 2 socket, 4 socket, 8 socket solutions for a long time. We've been in the data center market innovating every year consistently for 20 years, right? Some of our competitors haven't really been around for the last few years, right? I think it's important to keep in mind that our customers, many of whom you heard from today, are very sophisticated in the way they decide what to deploy. Are they going to deploy a 1 socket or a 2 socket or a 4 socket or an 8 socket platform? They know based on their workload, which one of those to deploy.
I'm very confident that based on the information we have so far, that we will be at absolute leadership versus the competitor you mentioned on performance, on performance per watt, and on performance per TCO. Okay? We, as we analyze it, are very comfortable with the position we have, whether end user decides to deploy a 1 socket or 2 socket or 4 socket is really their decision, and obviously, competitors that may have some artificial or cherry-picked benchmark that makes them look good on 1 socket may make an argument that, hey, the whole world's going to move to that. We know better, or I think our customers know better. They're going to pick based on their workload and based on total cost of ownership of their workload.
One other comment I wanted to add on was, it's part of the reason I spent the time this morning going through the mesh architecture that Sailesh and his team actually developed. It's one thing on paper to put down cores, IO, memory, but it doesn't actually deliver performance unless you can feed it and connect it well. I think that if, again, to your point about not being able to buy them, we have done our own internal testing, I think that if they were delivering more performance through those features, we would've seen it through benchmarks or customers that would've tested it and seen that. We'll obviously take our competitors seriously and continue to watch it, I think there might be a fundamental problem with putting four desktop parts together and then trying to address these data center workloads.
Hi, thanks. Kevin Cassidy from Stifel. Part of your presentation you showed that you have new instructions, the vector instructions in particular. Can you say, does that make it incompatible then with AMD as they come out? Since they've been out of the market for five years, what other instructions might there be that they don't have in their core chip set?
Do you want to talk about that?
Yeah. I don't know that incompatible is the right word. AVX512 is an advancement in our unique Intel design by our engineers instruction set that delivers tremendous performance. It would be a mistake to have ever thought that x86 and x86 means the same thing. We've always had unique innovation in our instruction set that differentiates from competitors' technology, this is just another continued advancement of that.
Yeah. Just to add, I think workloads that are compiled with the new instruction set would not work on other architectures which don't support it. You can always compile them without those instructions. It would just have lower performance.
They'd work, they just wouldn't accelerate.
That's right.
Okay, maybe just one follow-up to it. You highlighted cloud and artificial intelligence and 5G. Is there anything in particular here that helps just the basic enterprise market or any other markets?
Sure. I think the examples I gave, and Lisa expounded on the 4x improvement in virtualized throughput, 65% TCO benefit, I think those are very compelling numbers that we're already seeing enterprises. You saw Montefiore on stage, AT&T, talking about their desire to upgrade their infrastructure based on those. 65% TCO is not something that we do all the time. I believe that enterprises are going to look at this new platform and say, "Okay. The economics compel me to look pretty seriously at upgrading my infrastructure." We'll see how that plays out. The other thing I'd say is, I think we've moved from an environment where the only way to get the benefit of cloud economics was the public cloud, to now having more opportunities to do things on premises.
You saw Microsoft, for example, yesterday, announcing their Azure Stack, bringing their cloud solution to an on-prem-like solution. I think as companies like Microsoft and VMware, and even in the open source community, companies like that deploy this platform, they will see benefit in the enterprise as well.
I can add-
Just to add. Go ahead.
I was going to add one more thing just on the AI side of things. Inference performance is greatly increased on this CPU compared to previous generations, and I think that's actually something that's extremely relevant to the enterprise. Some people are going to look at that today, but then in a year or two, you're actually going to see a huge benefit from that.
Versus an FPGA or using a GPU. Can you just talk about the relative performance and why you'll be able to protect your ground in inference?
Sure. I think there's, my colleagues can add, but the first thing to remember about most inference workloads is they're usually part of another workload, right? They're real-time in nature. As cloud service providers or enterprises or anybody is deploying an inference solution, they're usually doing that as a subset of another workload. The latency between your general purpose microprocessor and the inference has to be as low as possible. As companies start to explore that, they say, "Listen, I already have this Xeon-based, scaled-out infrastructure. I'm already running my general purpose workload there." The inference part of that workload, the fact that they're connected and I need to do that in real time, right? It's not a batched process. It's a real-time process.
Makes it such that the economic argument to go build a specially built, purpose-built inference infrastructure is not really there. That's one of the reasons why most inference is done on a general purpose Xeon today. We're not standing still, right? There's going to be competitive dynamics where people will try to develop purpose-built inference solutions, and that's why Sailesh and the engineering team have been working on this product. We spent a lot of energy in thinking about how do we dramatically improve inference performance. Greater than 2x hardware based, and the 100x statement you made was when you add the software optimization. I feel pretty good about the improvements being made and the position we'll have on inference given the dynamics and the new product that we just launched. You want to add anything?
Yeah. I think in general right now, if you look at what's happening in deep learning, it's really a race to low precision dense matrix computation. That's where everybody is looking at how do you get to the lower and lower precision and do matrix multiplication much more efficiently. Those are tricks that pretty much everybody is going to play. Our competitors, including people who are going to be building accelerators, are going to go at it. I think that is a limited set where people will get to all of those, then it'll come down to how fast can you scale your latency response, how do you move data to feed these engines, and so on. It'll still come down to how good of an architecture you have to actually serve all these usages.
I think we have been working on that, we feel confident about what we can do in the future with that.
I just want one follow-up. I just want to ask, it wasn't part of the launch with the scalable launch, I'm just curious, the FPGA strategy with Altera, just how does that fit into this portfolio that you launched today?
Yeah. We didn't say it specifically today, we are building an integrated Xeon Scalable processor FPGA product. That product will come out into the market in early 2018. Okay, we are still building that product. We're still optimistic about the prospects of that. We've seeded it. We've put it in the hands of customers. They're busy sort of thinking about and looking at what workloads benefit the most by having the FPGA and Xeon close together. As we get closer to the availability of that product, we'll talk more about that.
I think you might have asked about FPGA's applicability to AI workloads. I don't know if, Naveen, you wanted to respond to that one, too.
I think we're still in the early stages of that. Inference is specifically, accelerating inference, keeping it very close to host memory is an important consideration. I think that's where this multi-chip package idea really started, and I think that's where it's going to be most applicable in the AI space.
Thanks. It's Chris Caso from Raymond James. Wonder if you could address the question of whether or not customers may in fact be waiting for Purley to come out, if there may be some pent-up demand that you may address, and maybe you could bring that back to past launches and what you may have seen. The second part of that is specifically within hybrid cloud and private cloud, and I know that's an area that you guys have been waiting to accelerate for a while. What is it about this launch that perhaps could act as the catalyst for that acceleration?
Sure. On the first one, the thing to remember about this launch that's a little different than what we've done historically is what Lisa talked about in her section, right? We have done the biggest early ship program we've ever done. Where today was the general availability launch, but as you heard, there's many customers that are already deploying this new platform, right? We've already shipped, sold 500,000 units well ahead of our launch. The good news is that many of those customers are buying the sort of higher-end product, right? If you look at the Google instances, for example, they are the high core count Xeon products. It's a little bit different than normal, where it's tough to gauge how much pent-up demand there is per se, given that we've already been sort of seeding it into the market relatively early.
Your second question was about what is it about the new Xeon platform that makes it more or less applicable to private and hybrid cloud. I think there's sort of two things going on there. One is the availability of private cloud solutions, right? There have been some customers waiting for a viable on-prem cloud solution, and I think the launch of the Azure Stack yesterday is a big step forward in that, right? We'll see how that plays out as we go into the second half. There's nothing per se about the new Xeon that is private cloud specific, but the general gen on gen performance improvement, all the statistics that Lisa talked about in terms of the workload optimizations that we're doing, all come into play as an enterprise decides what to deploy as they consider the public versus private cloud.
Do you want to add anything there?
No, I think that was great. You got it.
Okay.
Thank you. Srini Pajjuri from Macquarie. Just first a follow-up question. I think you're launching most of the SKUs at the same time, unlike before. You said two socket, four socket, and eight socket at the same time. I'm just curious as to, historically, what do you normally see in terms of the adoption of a new product like this? How long does it take for the volume to cross over 50%? How many quarters is it typically? Because of your launch schedule this time, do you expect that ramp to be any different?
Yeah. The two, four, and eight socket capabilities all at once, we do think adds value, again, for data center managers, infrastructure operators, to be able to do that platform evaluation together. As they look at their workloads and assess what the best configuration will be to run it, and if any of their workload characteristics have changed over time, or their business requirements. We actually see that as potentially having the ability to accelerate adoption. I also think the early ship program will fundamentally provide benefit there as well. You have cloud service provider instances that are out there, and we know we have enterprises out doing tests in dev that plan to deploy and now have the information and the performance testing available, so they don't actually need to wait and start their evaluation now.
As well as just in general, enterprises like Thomson Reuters, Montefiore, that had the opportunity to test ahead. All of these we view as ramp acceleration capabilities. When a product crosses over 50% can obviously have a lot to do with different market dynamics and whether the product's more cloud service provider targeted or more comm service provider targeted. It ranges widely across our Xeon product lines, what a ramp looks like. In this case, given all of that coming together, I think we'll see a fairly quick, within a couple of quarters, transition.
Okay, thank you. I have a question on the cost side. Given all the new features and the performance improvement that you guys are delivering, I don't know what the die size is. I don't know if it's public. I'm just curious as to what implications it has from a cost standpoint, and should we expect any change in the next few years in terms of your margin profile for this business based on what you see on the cost side?
Well, we're not being specific about the die size. I'm sure somebody out there will figure it out pretty quickly given that we're shipping now. Look, at the end of the day, margin is determined by ASP and cost, right? What we've focused on is delivering value here. We think that the performance characteristics of this product and the early ship indications are already showing us that people are valuing the high-end part of the stack. If that continues, then I would expect we get some ASP uplift as we transition from Broadwell to Skylake, right? Because we are delivering performance per TCO, and raw performance improvement like we haven't done in quite a while.
If history continues to repeat itself, we find that when we do that, people move, A, quickly to the new platform because it delivers them high value, and B, they deploy the higher-end products inside of the stack because, again, they can get performance per TCO benefit. If that plays out nicely, I would expect that we have ASP uplift. We're not giving a forecast specifically on gross margins or operating margins on this particular product line, but you guys will see that play out as we ramp this over the next year or so.
Thank you for the presentation. It's Matt Ramsay from Canaccord. Some features that were mentioned just briefly in the presentation that are going to be included in this platform, like the integration of Omni-Path and support for Optane and eventual support for our XPoint DIMMs, et cetera. I guess, how important are those to the performance gen over gen that you delivered in the numbers? I guess, maybe if you had to rank order those in terms of dependence of those performance improvements, anything you can give there would be helpful. Thanks.
Want to take that one?
I'll start, then Sailesh can chime in. I think that SAS example that I gave was an Intel Optane SSD plus Xeon Scalable delivering 2x. You can see that that puts it in the above-average performance gain delivered through more of a platform view versus just looking at the processor. For the Omni-Path architecture being integrated, we are excited to bring this to market for our high-performance computing customers because it gives them the choice. They still have the discrete option, they've got the integrated option, and depending on the architecture and configuration of their cluster and fundamentally the problem they're trying to solve, they'll be able to pick between which delivers a better TCO advantage for them.
The last one on the 3D XPoint or Intel persistent memory, we had shared in May that we'll be bringing that to market with our next generation in this Purley platform. I think over time, you'll see us start to share a lot more information about transformational type performance for certain workloads that really require that large memory at really fast speeds. We'll share more on that one over time, but I think that'll be a very big kind of platform-connected story as well. I don't know if you had further comments on that.
I think you covered it. Omni-Path is low latency, high bandwidth network for high-performance computing. Both ColdStream and 3D XPoint are really about big data and how to improve the performance of those workloads.
Hi, Quinn Bolton with Needham. I just want to follow up on the deep learning performance improvements. I think you talked about hardware giving you about a 2x improvement gen to gen, but another 50% from software optimizations. Is this optimization for the new AVX-512 instructions, or are there other optimizations you're doing? If it's other optimizations, can those be applied back to software for the previous generation processors? Then I've got a platform follow-up question.
There are optimizations that are both ways. Some are specific to this processor generation, and some are generally applicable. I think the way it's been stated is typical solutions prior to this were sort of out of the box, general purpose linear algebra libraries that were used. We've really optimized them on IA in general, now we're also taking advantage of new features as part of Intel Xeon Scalable, like the INT8 AVX instructions that Sailesh was mentioning. We have a combination of the two, really.
Yeah. Hardware to hardware is 2x or slightly ahead of 2x. The rest of it is basically, as Navin mentioned, it was really people using out of the box without being optimized either for threading or vectorization or any of that. That just represents the prior software versions were not really optimized for Xeon.
Great. The platform question is just your competitor, AMD, talking about eight memory channels per processor with their EPYC processor, I think. Scalable has six. Can you just talk about the CPU memory bandwidth in the platform and how you think you compare with your competitor?
I don't think we'd say anything different than what we said before. At the end of the day, people buy delivered performance for the workloads they want to deploy, right? Whether it's six or eight or 10 or whatever, all that stuff translates ultimately into delivered application performance. As it pertains to real-world application performance, I don't know what their platform is going to deliver, right? I don't have one. When we get one, we'll be able to compare and contrast. We have a lot of application performance data. Lisa shared a lot of it with you. There's a lot more available already. Based on everything we've seen, we're highly confident that we will have clear leadership on performance per watt, against our x86 competitor.
I think the main thing from architecture standpoint, what we try to focus on is for the number of cores, memory channels, PCIe, and all of that attributes, how does it scale across the entire for the different sets of data center workloads. It's not just the number, but any particular data center workload, when you actually start loading up more and more work on that system, how does it scale in performance? That's where we talked about the Mesh and other kinds of enhancements that we did in our architecture are necessary to be able to scale performance across the number of cores that you actually provision on your processor.
Vijay Rakesh from Morgan Stanley. Follow-up on AI inference. You build up impressive portfolio of products, right? Xeon Scalable.
Could you speak up a little bit?
Xeon Scalable, FPGAs, ASICs. Could you just talk about your go-to-market strategy for not just the data center, as well as devices on the edge?
Sure. I'll start and then Navin, you can add, or Lisa or anyone else. The approach we have, as I mentioned, we think that we're really at the onset of this artificial intelligence wave. The workloads and algorithms are going to change a lot between now and over time. We have the benefit of having a broad set of potential solutions. First and foremost, we're going to make sure that our general purpose microprocessor is increasingly optimized for AI. You see that in the product that we launched today. The 2x improvement, the 100x with the software optimization. Particularly around inference, we know we have a good position on inference today, and we want to make sure that we continue to innovate to keep that position. In new domains where maybe it's less clear what the algorithm will be, people will experiment with FPGAs.
They will use FPGAs to sort of learn. Then as algorithms change, they can in field reprogram. That's the whole purpose of having an FPGA. Then in domains where the algorithms are fairly well understood, in places like training, we're going to go build accelerators, purpose-built accelerators. That's what Navin and his organization are off doing. So, we have the benefit of having assets in all three of those domains, and we're going to play all those assets. It may be that we play different assets over time as we see where the workloads evolve and where we are able to size the market opportunity. It's difficult to predict what things will look like three or four or five years from now because the pace at which the algorithms are changing is just incredibly fast. That sort of benefits our approach to have multiple solutions.
You can add to that, Navin.
Yeah. In addition to that, since we do have multiple hardware platforms, we can really address our specific customer needs. Also, we have a strategy across software to decrease the friction of moving between them. We've launched some various open source projects, nGraph and Neon and a few other things that are out there, that basically try to abstract the hardware to a common representation. A data scientist who's using these tools can actually write their code in one way and reuse that across multiple platforms. Perhaps they start on something very general purpose when they don't really know what the value to their organization is yet. Then as they need more and more performance, they can move to a more optimized platform, for that specific workload. Our software will kind of move with them, decreasing the friction there.
It should actually be write once and run on any of these platforms, going forward.
I think you asked about edge inference as well, maybe. You want to talk about that?
Yeah. We also acquired a company called Movidius last year, that has the leading platform in edge inference, power performance per watt, and also capabilities for computer vision. This is the formation of our edge inference story. Will actually be sewn into an entire family of products we're building for AI. Edge inference all the way to data center training.
Thanks. I apologize ahead of time, too, in the weeds questions, with this platform scalable, you have six memory channels times two DIMMs per channel, the speed doesn't degrade. If you look at the current generation, you can put three per channel, it degrades. I guess the question is, currently, if people are putting all loading up fully, they use LRDIMM. Why would you use LRDIMM with the scalable architecture?
To get to even larger capacity than what you can get with two DIMMs per channel.
Okay.
There are workloads which actually want more memory capacity. There are different solutions. Some people actually go towards 4-socket deployment where they get capacity and more threads. In some cases, they look for LRDIMMs and other kinds of things just for memory capacity.
I guess, would you see the percentage of people using LRDIMMs, like maybe, 8%-10% of the market? Would you see that being the same going forward?
I can't speculate on that. I haven't seen a big demand on LRDIMM, but there is definitely a specific workload sets then that it can address.
Got you. I just want to ask you about Ethernet connectivity, and you've integrated 10G. Just kind of curious what kind of uptake. How are you bundling the chipset with the connectivity versus the Xeon, and what kind of attach rate do you think you would see with the integrated 10G?
Yeah, it is integrated in the chipset and of course our system vendor providers have the opportunity to offer discrete Ethernet, of which Intel has a healthy portfolio there that continues to deliver a lot of great performance. They have a one gig and a 10 gig integrated option. We'll continue to work with the market on adoption. We'll market this capability. We think it brings a lot of value just for, again, a total cost of ownership and then just inherent performance capabilities that you get from having integration. It is our first generation of delivering this capability, so we'll be tracking market adoption along with you.
Hi, Tristan Gerra. How do you see the competitive landscape changing with Arm-based ecosystems? What are you hearing from customers in terms of desire to try Arm-based systems? Also if you could talk about the risk of vertical integration and ASICs.
Okay. You asked a lot of questions there. There's been talk about Arm-based competition in data center for quite a while. As of right now, there isn't anything in production, right? We obviously are paranoid about all possible competitors. The key thing for us is to continue to push the ball forward and to be aggressive about our innovation and not rest on our laurels. One of the main things that, as I've come in over the last 45 days, that I've been assessing is the roadmap and ensuring that we have a very aggressive roadmap as we look out over the next three or four years, that is optimized for a broad range of workloads, from the enterprise to the cloud, to the network, to artificial intelligence.
As I look at that and I look at what our competitor roadmaps look like, I feel very comfortable about our position, particularly with respect to Arm. That said, we're going to watch that space very closely and we're going to see what they go do. Where perhaps there's gaps in our roadmap, we're going to invest to close those gaps. Your second question was around, what was it again? Remind me.
About ASICs and vertical integration model.
I think there are going to be places where various companies in the ecosystem invest in building proprietary ASICs that run their particular workload very well. At the same time, you would expect that, right? You don't make an investment in that unless you expect you can do that better than the industry can. My view is that as we go through time, Navin in particular is investing heavily in building custom ASICs for workloads that customers are asking about. As we do that, he's doing that in close partnership with some of the largest hyperscale customers out there. They're sharing their workload information, they're sharing the evolution of their workloads, and we're in turn building a solution that's highly optimized to that workload. If we do a good job of that, they will buy those products from us.
They have no natural incentive not to if we deliver a better performance per watt, performance per TCO solution. The key here is that we make the right investments and we have the right partnerships. You want to talk a little bit more about your approach there, the way you're thinking about it?
Yeah, I think that's right. You're always going to see holes in the market where maybe some company saw something early. We can talk about numerous examples, but some companies have invested in AI very early, and there frankly weren't products in the market to fill those needs, so they had to go after it themselves. I think this may be what you're referring to. As the market matures and we see broader demand for these kinds of things, a company like us, specifically who wants to hit many different customer use cases out there, is going to invest and fill those gaps as we need. Whether we do that through an ASIC model or through more optimized on our own fab technology is basically a trade-off we have to make depending on the investment required to hit that market opportunity.
Hi. Thanks for taking my question. It's Will Stein from SunTrust, thanks for hosting this informative event today. You've talked about some of the emerging technologies we've heard about for some time, like 3D XPoint, also Omni-Path. One that we haven't heard anything about today is silicon photonics. At least I don't recall hearing anything about that. Is there any relevancy to this product with that technology? Anything to announce there?
Yeah. Silicon photonics continues to be incredibly important to us. We're making big investments in silicon photonics, I think the demand for silicon photonics will only increase over time. The reason we didn't talk about it specifically today is because it's still a discrete solution relative to this platform. Yet, we haven't done any unique integration. You can imagine that as we think about the future roadmap, we're analyzing the workload evolution and we're analyzing are there opportunities to integrate more tightly and more closely. So I would just say watch this space as we go through time. Today, we did talk about the integrations we've already made between Ethernet, FPGA, and Omni-Path. We're integrating all three of those capabilities into this platform.
Don't take at all the lack of discussion about silicon photonics today as an indication that we're not continuing to invest and we don't see a big future for silicon photonics. We absolutely do.
I think just to add on, I think since we've launched the first generation of our silicon photonics, you learn a lot from your customers as you work towards developing a solution, you learn even more once you've delivered one to them. I think that's been the most exciting thing that we think actually with what we've delivered and how it's ramping, we have hit on some unique differentiation that we're continuing to build on in the roadmap. I think we have a lot more good news to come. It's just, I was trying to keep it into that time limit, and the thing started blinking, and we went over, so we were just trying to be efficient. You'll hear a lot more from us on silicon photonics and the roadmap there.
Thank you.
Thanks. Oh, hi. Chris Caso, Raymond James again. Just a question, regarding the networking business. You addressed some of this in the prepared marks, some of it with the guest presenter.
Hey, Chris, do you mind just speaking up just a little bit?
Are you with me?
move the mic closer?
Oh, sorry.
Yeah.
You're perfect.
Just with regard to the networking business and what the platform brings for that business. You referred to some of it in the prepared marks, some of it with the guest speaker, but if you expand on that, please.
Go for it, Dan.
Yes. We're truly pumped up about the networking business and this platform for networking. As you know, we're in the midst of a major network transformation, going from more purpose-built networks to more virtualized, software-defined, and cloud-like. When you think about some of the features in this product, first we're integrating specific network accelerators, cryptography as well as compression. You heard Sailesh and also Lisa talk about the Intel Mesh Architecture a little bit, and what's significant about that is one of the major workloads within networking is packet processing. That's all about getting data and packets in and out of I/O, memory, and the compute subsystem very efficiently. Because of that, we're seeing significant speed up in packet processing, seeing over 1.7x gen to gen, which is impressive.
When you think about the journey of network transformation and you think about Network Functions Virtualization in particular, you think about the use of virtual machines. You heard about the virtual machine density, over 4x increase over a four-year-old server. When you think about all those things for the network, it definitely increases the overall flexibility, agility, and scale to meet the long-term vision of network transformation and preparation for 5G, also offers service providers the assurance they're going to hit the right performance, security, as well as reliability.
I think fundamentally with this product, we talk a lot about some of the compute workloads here together today, the compute storage and network addressability, it really is a culmination of this multi-year journey, and the reason Dan's so excited is because we actually can address more of the silicon TAM opportunity that exists in the network space now that we have Intel Xeon Scalable and that performance level that it delivers. It also doesn't hurt to have the ecosystem leaders like AT&T and John, truly a leader in the industry, setting the bar for the industry about how we will adopt and accelerate the pace of technology innovation in the comms SP space. I think we've got a lot of good momentum in that direction.
All right. We'll make this the last question.
Thank you. Srini Pajjuri again from Macquarie. If I recall correctly, I think you guys talked about a packaging technology on your roadmap called EMIB at your analyst day, I believe that even Xeon will leverage that technology going forward. I'm just trying to understand, how is that different from what AMD is doing today? I guess you used the word stitching together for CPUs using a fabric. It does look to me that it's a package. They're leveraging their packaging technology to kind of integrate more cores. Essentially what you're looking at with our competitor is what they've done is they've built small chips and they're using small pipes to connect them. From a technology standpoint, I'm trying to understand the differences.
Do you want to take that?
Yeah. I think you have to separate the architecture of how everything is integrated from the exact technology that is used to integrate it. The discussion that we had was really from an architecture perspective, all the attributes of how you scale and all that. You can decide to actually take the same construct, cut it, put it on two chips, but essentially from an architecture perspective, make sure that the way they connect has all the attributes of architecture scalability. What you're looking at with our competitor is what they've done is they've built small chips and they're using small pipes to connect them. There's a big difference between that.
All right. Dan, Lisa, Navin, Sailesh, thanks so much for the time. Thanks so much for the folks who joined here and on the webcast. Appreciate your time with us today.
Thanks for coming, guys.
All right. Thank you.
Thank you.