Well, welcome to New York and the One World Trade Center, a perfect venue for us to do a very exciting announcement from Micron. I think we've got a fun-packed morning with a lot of information in it here for you. My name's Darren Thomas. I'm the general manager of the business unit at Micron, and I'll be your host for this morning's event. I also want to tell you that we've got a panel of very interesting customers with examples of how they've transformed their industry, and we've got an esteemed group of people from all over the industry that will be doing our presentations for the first half of this morning's event. The second thing is I'm going to be walking you through a setup here to give you a little bit of a taste of what's coming so you'll see what the whole story is.
I'm going to keep you from having to guess till the end what the message is. The message is going to be very clear. It's about insights. I'm going to start with the ubiquitous slide on how big the storage business is and how big it's getting. I've done this slide probably about a million times in my career, but I've never seen this number before. I've got to tell you, the numbers are getting astronomical in their size. Matter of fact, 163 zettabytes. A zettabyte is a billion gigabytes. I'm sorry, it's a trillion gigabytes. This is a very, very large number and it creates a daunting challenge for our IT professionals as they try to upload and analyze this amount of information because it's coming by 2025. In 2016, we were only at 16 zettabytes.
This is a 10x growth in a very, very short period of time that is being driven by a couple of new trends. One of the trends is one that's been around for a long time, mobile phones and CPUs. People have been creating content in this cloud space and content people have had to analyze for a long time. What you're seeing now is machines are creating content. This is the new wave. You're seeing devices. These are IoT devices. You're seeing a significant number of the population of the world, 75% of people are connected now, and not just with their cell phones. The kind of connectivity we're talking about is refrigerators and cars and thermostats and security cameras. What they have in common is these are machines creating content without a human interaction much at all, if any.
As a result, the data is going into a storage solution, and what's happening is it needs to be analyzed. The data, in order for it to be valuable, it has to be analyzed. Because it's real time, it needs to be analyzed real time because people are going to make decisions off that information in a very short period of time when the information's not valuable. We're entering this age where now machines are creating content faster than humans are creating content, and that content needs to be analyzed and it needs to be real time. We've got to go from raw data to insightful real-time analytics. Some good examples of this are like a good banking example would be you have a bank.
Well, in the old days, your bank, you filled out a form when you opened the account, and they never assumed you changed. In this new world, you might start off as a high school student saving for college. You go through college, take out loans. You might go get married and have a life and a family and have children. Now you're creating savings for them, saving for your retirement, ultimately you retire. You go through this life cycle even within a year, your requirements change. A bank that had this kind of information about you would be able to keep up with you. Another example is the internet. Everybody searches the internet, all these search engines. Now there's all this content that even you don't know what you're going to search on most recently.
What happens is the search engines now need to be able to do almost predictive analytics. They need to be able to look at what you've been doing, whether you're wearing a watch or whether you own a thermostat or you have a security system, and understand what you might search on next and be prepared for that search because the data is too great for you to wait while the search is going on. Another example is content providers. As you've seen, content providers have pretty much dominated the entertainment industry. Now, with this kind of content, they could tell by your watch or your cell phone that you just moved from your hometown to New York.
If you've been downloading something, you might want to download it from here instead of there, and they might push that new content that you're downloading out to a site here that's not where you were last night. Lots of opportunities for this analytics to be analyzed real time, and those are just some simple examples. Why Micron? I know we're all here. We've invited you. The question you might ask is how does Micron fit in this? In this world of insight, it helps a lot if you understand the DRAM and the NAND. These devices are different. How you address them, how you talk to them are different. We've made subsystems out of them, and inside those subsystems, we have put abstraction layers to make the data more accessible as a subsystem.
As you get into these operating systems that are used to seeing hard drives, not SSDs or not NVDIMMs or some other form As you get into that, what happens is when you take those abstraction layers back out. In order to do real-time analytics, you can't have all these abstractions. You need to take it back out. What company knows that better than a company that owns NAND and DRAM? What you're going to see today is the level of insights that we bring into this, as well as our teams understanding the operating systems, the applications, and the workloads our real customers use. The answer to why Micron is we are the company that has the insights from 40 years of developing memory technology and now flash technology and the subsystems that go with that.
Now we're moving into the next phase, which is to help with the analytics and the insights about the solutions. That includes us going completely from raw data all the way to this real-time analytics insight. With that, it's my great pleasure to introduce Laura DuBois. She's the Vice President of the Enterprise Storage, Servers, and Infrastructure at IDC. She's going to walk us through a brand-new study that we work together with, and then we're going to have the rest of this august group of people come up here, and then I'll come back at the end and close. Laura.
Thank you. Good morning. It's a pleasure to be here with all of you today. Thank you so much for joining us. As Darren said, I'm Laura DuBois, and I run the infrastructure systems and software research for IDC. Today we're going to be talking about the impact that data is having on traditional infrastructure and the challenges with data growth and the pressure that's increasingly placing on operations and line of business to accelerate performance to meet increasingly business imperatives. Business imperatives of transforming business models, business imperatives of improving and transforming customer experience, and lastly, in driving new types of operations internally and improving and driving greater efficiency in operations. It's increasingly a focus on data that makes the difference between a company just merely surviving and a company that's going to thrive and outperform its peers in the industry.
Increasingly we see the focus on data as being a differentiator and the need for increasingly real-time data analytics, real-time analysis, and mining and intelligence that's applied from the understanding of data that is driving this increased importance in performance. I'd like to walk you through four examples of companies that are taking a data-driven strategy. We see data-driven strategies being embarked upon by large banks, some of which are in this audience. We see data-driven strategies being embarked on by large retail organizations, by folks like Starbucks. Starbucks is increasingly looking to provide the customer experience on the digital platform the same way they provide the customer experience in the retail location.
They're increasingly using data combined, their customer insights, their customer retention data, and customer frequency data applied with content from partners, whether that be The New York Times or music sources from Pandora and others, to provide a digital experience for the customer that really makes them want to stay with Starbucks. That's a large company example. We certainly have smaller company examples as well. Someone like Pure Fishing. Pure Fishing is a $700 million company that sells bait and tackle to both professional and consumer or individual fishermen.
What they wanted to do is collect data from the site, which is the fishing site, and be able to provide a variety of different data, whether it's the depth of the cast or the temperature of the water or the reach of the cast or the weather that day, pulling in historical information as well and providing this information to the customer to improve their customer experience. Another example, Briggs & Stratton, a $2 billion engine manufacturer, very conservative company in the Midwest, a physical product. How are they going to transform and become a digital company? Well, they've input intelligence onto their engines, and they're collecting this information, a variety of different telemetric factors, and they're able to analyze and provide this information in a dashboard through this intelligence they're capturing on the engine itself and be able to provide this information to their customers, which are resellers.
Those resellers can then stand up new services to be able to service that, say, home generator before a spike and a big snowstorm. You know that home generator is going to work before the event. Last example is McCormick, the spice company. They're in the process of transforming and being a food experience company. They've taken over hundreds of different vectors of taste preferences, and they've combined these taste preferences, obviously for their particular products and spices, and created an algorithm and an API called FlavorPrint. That they now leverage and sell this information to downstream food manufacturers. This has been so successful, they actually spun off a separate company from it. These are examples of four companies taking a data-driven approach to transform their business.
Now, if data is at the center of this transformation, you might ask, what does that mean in terms of infrastructure strategies? Well, increasingly, CIOs are telling us, in fact, 72% of 1,000 global CEOs told us that digital transformation and business and data growth are having a fundamental and the leading factor in driving infrastructure strategies and decision-making. This is very consistent in what we see CIOs want to do, certainly first and foremost, is security and mitigate risk. Secondarily, reduce cost and increase efficiency, and lastly is align with the business and support new business initiatives, such as the ones that I provided examples for. Now, if this is the objective, is driving and enabling growth, and the growth is coming from the importance in data-driven approaches and strategies, you might ask, well, what is the volume of data growth?
Darren already articulated this, and I'm just going to go through it briefly, which is, there's never been a lack of data, and there never will be. By 2025, there'll be over 163 zettabytes of data created. Just to give you an example, that's growing by a factor of seven off of a 2016 baseline. That's at an industry level, in terms of consumers and individuals creating data, businesses creating data created by applications, data created by use of social media, data created by increasingly the digitization of content, machines creating data. You might ask, well, what's the growth in data for a given organization?
Well, we embarked on a study in partnership with Micron to understand across 800 global companies, across nine countries, what their challenges are around data growth, what pressure this places on traditional infrastructure, and what their objectives and strategies are to do, or what they're doing to address those challenges. Lastly, how that the role of flash impacts those strategies and how they're using flash. Let me share some insights. Pretty much everyone is facing data growth. Only 1% said they didn't have any data growth. On an annualized basis, the growth in data is 24% annually. Now we've seen this moderate a little bit, certainly in the last 10 years, because of the efficiency of storage technology, of technical capabilities such as deduping compression. Nonetheless, it's growing at a very significant double-digit rate.
There's variance in growth of the data across company size, but irregardless, this growth in data is coming from a variety of factors. Certainly, it's coming from business growth. Certainly, it's coming from increasing numbers of users. Certainly, it's coming from the consumerization of IT. Certainly, it's coming from the increasing intelligence out in the endpoints and in the edge locations and capturing more and more data information, whether that be about a user, a system, or a refrigerator, as Darren talked about earlier. It's capturing all this information and it's analyzing it. It's analyzing it to drive better business outcomes. All of this data growth is driving challenges. We wanted to understand, okay, this data growth, what does this mean to your environment? How are you able to keep up with this data growth?
As you can see here, the top three challenges in keeping up with this data growth are around performance. Number one, leading one, can't keep up with performance expectations. This could be latency-specific, this could be bandwidth-specific, this could be throughput-specific. Secondary factors around keeping up with backup, data protection, right? You start your backup, can't get it done before you need to do the next night's backup because the data's growing so much. That's a bandwidth-intensive kind of scenario. The last one is around data governance and the ability to understand and be able to make decisions around what data to keep and what data to dispose of. With these as the challenges because of the data growth, what are strategies that firms are embarking on to address these performance issues? Well, certainly we see investments in new architectures.
These new architectures could be hyper-converged infrastructure. These architectures could be software-defined infrastructure, certainly the rise of cloud computing and the use of either public cloud or private cloud. There's also this massive double-digit adoption of all-flash arrays. All-flash arrays with mixed workload consolidation, the move of workloads onto these all-flash arrays to drive up levels of performance, drive greater levels in IOPS, drive reduction in latency. Lastly, you can certainly add more capacity into the mix, right? If you're adding more HDD-based capacity, that certainly helps you with your growth issue. But if you're using storage efficiency technologies like deduplication, you may want to think again in terms of what impact that has on your performance, certainly as it relates to TCO. These are the actions that companies cite that they're taking.
Certainly, flash is first and foremost, I would assert in the leading challenge, investing in new technologies. Increasingly, we see flash embedded in all of these new technologies, whether it's on-premise or off. Because of this, we see flash being used in a variety of different places in the I/O stack. When we first started looking at this market, flash was being used in the host to accelerate small pieces of data, database indices. Excellent performance you got from that, you move the database to all flash, the full database. Now we're seeing many workloads move to shared flash-based storage. Certainly, we see flash being used in cloud computing strategies and infrastructure-as-a-service, high I/O instances in AWS, and flash being used as a mechanism to drive greater levels of innovation. Now we're beginning to see the rise of rack scale architectures.
The disaggregated and composable infrastructure that increasingly has a low latency, high-performance flash tier in these rack scale architectures. Lastly, a software-defined infrastructure. That could be open-source based, it could be a commercial or proprietary software stack that creates a storage pool out of individual systems with local or direct-attached storage. Increasingly, these environments are being used with new types of applications and to new types of workloads. We refer to as cloud-native workloads that tend to be horizontally scalable in nature, stateless, based on microservices, tend to make heavy use of open-source technologies and the like, and often use flash as a persistence tier. Because of flash is really perpetuate or influence all of these technologies, it's really why we see by 2020 that 77% of storage system spending will be flash-based.
We wanted to understand what the adoption of NVMe was because as we see customers either augmenting or displacing HDD-based technologies with flash, we see we're now in the next wave of adoption with lower latency protocols and higher performance flash arrays based on NVMe and NVMe over Fabric. We wanted to understand what's the level of adoption of NVMe today, and this is among existing flash users. 48% say they're using NVMe flash, NVMe or PCIe flash. This is predominantly in the servers, I would say almost exclusively. When we look at the drivers for why they're using NVMe, it's largely a function of economics, of performance, greater concurrency, higher bandwidth and throughput, and greater business outcomes that they're looking to try to achieve through use cases such as real-time analytics.
Let me give you an example of that. A large sneaker company, probably one we all know, wanted to do a collection of customer sentiments across millions of customers over social media during a global sporting event, the Olympics or World Cup. Collect all this customer sentiment information, tens of petabytes of information that they're collecting, and analyze that in concert with past purchasing behaviors, combined with micro events going on within the broader sporting event. Within the next 30 seconds, what's going to happen if Joe Smith makes a goal, we're going to make this offer. The offers are individualized to specific and individual customers. What the offer is that Cindy Smith gets is different from the offer that Joe Blow gets.
To be able to deliver this kind of real-time analytics, they need to be able to deliver performance in the sub-200 microsecond range. That's an example of where we see NVMe flash systems getting deployed. This is very early stage in NVMe over Fabrics, it will definitely be a game changer and drive the wave and accelerated adoption of next new flash architectures. We see flash strategies being evaluated across a number of key vectors. First one being adding more capacity. More capacity in terms of maybe it's 3D XPoint or 3D NAND, certainly devices getting more capacity rich. We see certainly the use of flash and SSDs across traditional interconnects, SAS and SATA. Now we increasingly are looking at starting to track the adoption of PCIe and NVMe, both in the server and in a shared array.
With the use of flash among these data-driven companies, you'd ask, "Well, what's the business outcome that they're getting from the use of flash? What does it yield for the business?" Certainly there's operational benefits, let's talk about business benefits as well. We ask, what was your rationale for investment and what was the outcome that you got? The leading factor here was efficiency. Right? Efficiency in terms of reduced power and cooling costs, reduced floor space, greater levels of density, doing more with less, right? Taking multiple racks and consolidating it down to half a rack, for example. Right? You get further efficiency because you're able to use write minimization technologies to reduce the amount of data you actually even persist and store.
In terms of the average cited gain in efficiency, 20% increase in efficiency on a mean or blended average basis. The second area is around Customer satisfaction. One call center I talked to that put their call center app on an all-flash array was able to reduce their call close rate by 30%. A very big deal in a call center where closing and keeping those calls short is everything. You can also, from an operations perspective, it's not uncommon to reduce your operations time internally in terms of troubleshooting performance issues by 15%-20% on a weekly basis by investing in flash. That's a very common statistic I hear. 32% said greater customer satisfaction. Whether that's the end-user customer or an internal customer. Certainly there's a correlation between page load time and performance on the back end.
Lastly is reducing cost. 55% cited a noticeable increase in customer satisfaction or customer experience because of the investment of flash. The last one is around reduced cost. Reduced cost in terms of the amount of capacity you purchase, reduced cost of power cooling, floor space, reduced license, may potentially reduce software licensing costs. The average cited was 45% said they saw a decrease in cost in the range of 10%-25%. A very real impact from a business perspective, starting with the imperative around data and using data to transform or drive better business outcomes.
In closing, I just wanted to say the use of flash, yes, it's about performance and IOPS and reduced latency, certainly with the rise of NVMe and NVMe over Fabrics, you can get the same latency that you would get with a local persistence, but the benefits of shared storage using NVMe over Fabrics as it continues to evolve. It's understanding the business outcomes that you want to achieve, and you all should be evaluating how you can begin to determine what workloads are best suited for future architectures that are based on NVMe. With that, I'd like to thank you for your time. Feel free to reach out to me online or after this. With that, I'd like to hand it over to Eric Endebrock, who's going to talk about the strategy that Micron's embarking on. Thank you.
Thanks, Laura, and thanks, Darren. I get the best job here. I always love this part. I get to go talk about the new stuff. We set the stage, a lot about data, what we're doing. Before I did that, I thought maybe it might be a good time to thank everybody sincerely here in the room and online. We have people watching, hopefully all over the world, tuning into this. I thought if you guys would just humor me for one second, a guy from Austin via Boise, Idaho, these days, I'm never going to get to do this anywhere else, and I've always just wanted to say, "Live from New York." You know. Yeah. Thank you. You know what? The funny thing was, they didn't get it on the subway at all last night. I don't know what was going on.
Of course, it was pretty crowded. I hit somebody doing it. Anyway, hey. We saw the insights. We see it's about the data. We see it's about a new economy. If there's any confusion, hopefully not with Micron being up here, we think flash is really that technology foundation that's going to take this to the next level. I want to build up a little bit of crescendo. It's not going to take long. Just wanted to walk through, well, what are we doing around flash? How did Micron get here? Our 10 years perfecting our flash technologies and even more perfecting memories in general, and what does that mean? We spend a lot of time focusing on the workload.
If you've talked to any of our team anywhere, hopefully they haven't started with, "What kind of SSDs do you want?" Hopefully, it started with, "What are you running? What are you trying to accomplish? What are you doing in your solution?" We spend an awful lot of time on that, and it's really focused. If you think about our portfolio, number 1, Laura talked, it's always about capacity and storage, of course. Over the last year, we've seen capacities double, even triple in some cases. The use of capacities is really jumping. Maybe best evidenced by our 5100 SSD line, where today I think we're the only vendor who's got eight terabyte SATA SSDs in the market. We obviously have NVMe and other portfolio products moving well past that, even with line of sight to 25 terabytes plus.
We see this massive capacity adoption and increase in the market. You think about, well, flexibility. Through our FlexPro architecture, we're making it easier to adopt SSDs, to use them, to optimize them, and to connect them into your workloads. What that means is I can control the capacity real-time through its software. That's the beauty of memory. It's not a spinning media. You don't just get 15K, 10K. You get whatever we want it to be. As you can start tuning that flash and the media through your SSDs, you can adapt it better to the environment and the workloads that you want to run. Obviously, there's always performance. NVMe we see as the future. It's a very low latency, high speed technology.
We're always going to be very focused on bringing solutions like that to market to solve some of the customer problems we've discussed. You get into this interesting one with security. There's table stakes, things like encryption and TCG. There's beyond that into FIPS. We're rolling across our portfolio. How do we harden our portfolio for very ruggedized environments? Making sure that it's supporting of the most demanding pieces. You take that even one step further. Micron is a company with a recent announcement we're doing with Microsoft, where we're connecting IoT at the endpoint through secure connections all the way back to the cloud, in this case with Microsoft Azure. Spending a lot of time focusing on: how do I put this together in a way that matches your security needs? You bring all that together.
Obviously, we think that flash brings the most valuable economic advantage. That's sort of our SSD path, and we said, all right, we've got the great technology foundation that we build. It always amazes me, turning sand into something exciting like flash. You put it into SSD as a portfolio. Our logical next step, last April, we launched our reference architecture series. We called these Micron Accelerated Solutions. We had companies coming to us saying, "Look, we get that flash is faster, but that's all we get right now." We started working with those companies to say, "Okay, how do I take it to the next level? How do we start illustrating what flash can do?" These have been great solutions, and we learned a lot through it as we helped other companies learn about it. VMware worked with us.
We created a vSAN solution, outstanding performance for taking advantage of the infrastructure that our customers already had, a very heavily virtualized environment. That was great. We learned a lot. In Ceph, and we're continuing these, by the way. Next week, if you happen to be at OpenStack in Boston, we're revving our Ceph solution at the storage level to an all NVMe flash offering, working with our great partners, Red Hat and Supermicro here, doing some great work around, I guess, extending out what we can do in some more cutting-edge storage workloads. We learned a lot through this, the next step is it says, all right, well, where do I go from these reference architectures?
You took a step out and you say, "Well, that's software." The next frontier there was all of our software and ecosystem is still written for hard drives. There's tremendous inefficiencies throughout the operating systems, throughout the file systems, the block layers. All of this is causing so much, I guess I'll just call it a barrier. It's blocking taking the really true advantage of what flash can do in some of the media. We're spending a lot of time with the software ecosystem, our application writers, and even ourselves, rewriting and tuning all the way from the top down to the bottom through our firmware. How do we take advantage? How do we strip out some of the inefficiencies that are preventing us from taking advantage of flash? Very excited about that and moving it forward.
You say, "Okay, I'll take my flash and my SSDs and some solutioning and some software, and I put all that together" and start thinking about platforms. Even that doesn't get you all the way. We bridged through this last year of really accelerated learning, and you start saying, well, the problems that Laura was describing, the connectivity that Darren talked about, when you think about those kind of challenges facing us, you say, you know what? Just flash alone isn't going to solve that problem. What I have to do is I need to connect it. I need to use a fabric to connect that to reach the insights that I'm looking for. My data velocity, my variety, my volume, all of those are just coming together in this confluence that isn't addressable.
NVMe over Fabric, we think, is one of those key solutions that is going to connect and enable all of these pieces together through the software. Maybe rather than me talk about it, we produced a little video I'd like to show you, just a couple seconds, and then we'll come back in and talk about our solution to some of these challenges. Well, hopefully, that was more exciting than me explaining it. NVMe over Fabric. I'd like to introduce, with that in mind, our architecture based on NVMe over Fabric and a lot of collaborating. We call this SolidScale. This is an architecture designed to address the solutions or the challenges, sorry, that we're seeing and facing in the industry. The extreme latency challenges, the consumption of data, how to bring all that together into a solution.
As you look at this, let's dive into SolidScale and its architecture. SolidScale is really meant to be an architecture overall. It is both a platform, you can see the pictures of it in a scale out mode here. Sorry, scale up mode. It's designed to be both scale up and scale out. When you go out into the lobby, you'll get a view of this. You can stop by, you can look at it in action, and see what's happening with our solution. You'll basically get a view that says, look, whether you're extending your existing IT investment and you need to scale up, or if you're already modernized and you have the right kind of applications and scale out is your focus, we've designed an architecture that's supporting of both of those.
Really targeting low latency, high performance to start with. We figure you can go fast, and then you can add a lot of different capabilities to it. In today's world, going fast with throughput and performance makes a lot of sense. Whether you're doing real-time analytics, you're accelerating databases, you're doing IoT consumption at the edge, needed a solution that solved those kind of data problems. What we're announcing today is our early adopter program. We've built this architecture in conjunction with our partners. We focused and put a lot of effort into solving and pathfinding many of the technical challenges that go with it, and you can talk to us about that out in the lobby. There's still more work to do, and we're putting out a call.
We want to work with our partners, both in the software world, our hardware world, our customers, and we'll be hearing from a few of the customers testing these solutions real-time right now. Here and online, we'd love to invite all of these different groups to come be part of this as we take this forward to the next level. Diving under the hood a little bit, let's just describe what this solution is. Foundationally, the fabric is a key piece of this. Our great friends at Mellanox have been working with us on putting together a very low latency fabric that puts this together. We're using 100 gigabit a second Ethernet, and to connect up within the rack, all of our storage solutions is coming from the device I showed you, the SolidScale platform.
To coalesce it around, to centralize the management, we're working with our friends at Excelero. Great software technology that basically is ground up built for the lowest latency, the highest performance, the most simplified management experience to bring this together. We feel like putting all of these pieces together into a complete solution that we can deliver to you in a rack or into component form based on the building blocks, enables us to really accelerate adoption and use cases, moving this into real-time production. We're doing that. Again, you'll hear more about what we're doing in real-time with our customers here in a minute. What's under the hood here? A little bit more. Let's get into some of the performance. Performance isn't everything, but you think about, I talked about putting flash into any product doesn't necessarily solve the problem.
Today, flash, most systems, an all-flash array, typically maybe 4 million IOPS. Out of our three-node system that's highly redundant, fully available, we'll get 11 million IOPS, and we see this number maybe even increasing. Definitely the power to feed any of the most demanding processor systems and moving it away from direct attached storage. I guess maybe that's probably a place to stop. Our design goal, I probably should have said this up front. Our design goal was how do I build a system that breaks you free of having to put direct attached storage in a server to be close to that CPU, but being able to get the shared aspects and the economics of pooling that together and putting it together in a solution and really being able to manage it as one. We've tested this to 256 nodes. It's a very scalable solution.
We think it's architected to go well over 1,000. I think it's a very scalable solution for moving forward into the future. Obviously, IOPS, just one measure of performance. The other design goal here, and I think we heard Laura mention it, this 200 microseconds. Microseconds being very important to us. We want to have latency. Our goal was very little, if no overhead to being local, and we've kept it within 1%. If you think about what that means, we're around 200 microseconds end-to-end transaction within the solution. The round trip through the fabric, very little overhead, very predictable. This is now unleashing this supply of storage to be a multi-tenant environment. Really think we're very compelling in terms of producing the raw performance of the solution.
What that looks like in an application, just take like Cassandra, Apache Cassandra here, you can see, and these are different views of running it local versus running on a SolidScale solution. I won't go through each one, but the main message here is that on any given transaction, whether it be updating a database or loading it or reading from it, you'll see that we produce very similar results to being local, but you're not throwing away 50% of your capacity, 50% of your IOPS that we estimate are stranded in any given local server. You've unleashed that much freedom within your environment by doing this. Lastly, Microsoft SQL. This is in the Gallagher zone here. Microsoft SQL, and this is all on Linux by the way, we're looking and we fully saturate our 100 gig links.
We can produce over 11 gigabytes a second throughput. If you think about the most demanding modern workloads, the consumption of storage data sets that are so large that I can't even push them close to the processor with a solution set like SolidScale, I'm enabling those workloads now in a way that has never been possible before. We're very excited about what this brings. This is just the beginning. SolidScale is an architectural umbrella. Today, we are focused at a Linux solution and very high performance. We absolutely expect we'll be expanding our operating system support here shortly. We have ideas about how we do more extended memory type solutions for in-memory databases and really tie our other part of our legacy of Micron in terms of the DRAM side into this. There's capacity views as well. This isn't the last step.
This is a performance step on the road to much bigger things that we think address the future of storage problems and the challenges Laura described. Out of this, you can hopefully see our technology, our path from where we've been to where we're going, and this is our attempt as we see these challenges to step up, take advantage, and start bringing solutions like this that have meaningful relevance to the challenges that IT managers are facing. That's today's problem, but you can't just think today, you got to think tomorrow. It's going to be my privilege to bring up Brad Spiers . Why don't you get up here? Brad is like many of you in this room, he's from Wall Street. Hold the claps one second. He's from Wall Street. He's an expert.
This is the kind of thought leader that Micron has been bringing into our company to help solve these challenges. Somebody who's walked in the real shoes, who's solved real problems. I'll let him talk. He can do a better job evangelizing him than I can. Thank you all very much for being here.
Thanks, Eric. Good morning. As Eric mentioned, I spent about the first 20 years of my career on Wall Street, trading a bit, but also building systems. For those of you who are not aware of some of the things that I've accomplished. Back in 2003, we can take a little walk down history lane. As some of you know, the main challenge on Wall Street used to be valuing the derivatives. How do I get enough compute into my data center in order to get the right valuations by the next morning's open? Back in 2003, I actually suggested that we start looking at graphics card processors instead of traditional compute in order to be able to accelerate that. That's one example in which we looked at how to match the silicon to the problem.
Once we delivered that into production, they said, "Brad, why don't you go take a look at external cloud? Go make that secure." I partnered with the security team. We worked together on some of those challenges, and we were awarded three different patents on how to make that secure. The underpinning of those patents, how we would do that, was called a hardware-based root of trust. What that really means is that, once again, we were actually matching the silicon to the problem that we faced. After that, they said, "Tell you what, why don't you take a look at the internal cloud.
Start to take a look at how we would use software-defined infrastructure, software-defined storage, software-defined compute, and also software-defined networking." Well, to do that, what we really wanted to do was run as many containers or virtual machines as possible in the smallest possible footprint. Well, there we learned that the real bottleneck was in fact DRAM and fast storage, not compute. We loaded up our servers, which was a major change. We went from 64, 128 gigs per server to more like 512 or 768. Once again, what you're seeing, this pattern of we were matching the silicon that was in our devices to the problem at hand. Now, as we start to take a look forward, we start to take a look at machine learning, or as Darren and others have been mentioning, how do we derive insights from this deluge of data?
To help me talk about that, I'd like to bring up Steve Pawlowski. Some of you might know, Steve spent 31 years at Intel. Steve was part of the teams that did the very first PC chip, the very first server chip, and he was a senior fellow while there.
Thank you, Brad.
Of course, Steve.
It was really 32 years, who's counting at this point?
Can you tell us a little bit, Steve, about this wall, the memory wall. Why is it that maybe traditional architectures might not enable us to get across some of the challenges that we face as we start to look to provide some of the insights or things from our data?
Sure. A little bit about my background. I spent most of it doing microprocessor design, research and design. After spending 12 years in Intel Labs, I was asked to go back to the product group. Turned it down, got a call from the CEO who said, "You might want to rethink your decision," which, a day later, I ended up in the product group. The first meeting I had was with an HPC company that basically looked at me and said, "We hate your multi-core strategy." I said, "Well, it's nice to meet you, too. This is my brand new job." They said, "No, we hate your multi-core strategy because you're adding cores, but you're not improving memory performance. You're not increasing capacity, and you're not increasing bandwidth per core." I kind of let that go because, HPC, these were big systems.
They were 2% of the market. Most companies are losing money. They're kind of like the Ferrari, and then you have the Yugo, which is the standard high-volume server. 2008, started spending a lot of time on Wall Street, and the reason was is because there are certain individuals within Wall Street that were getting insight into what was happening with platforms and with the competition, and were giving us that insight up to 2 years before the mainstream OEMs were giving us that information. I asked them if this was okay. Jeff Birnbaum cornered me in 2008, which he was likely to do many times, and said, "Let me show you this diagram." He put up 2 slides.
One of them was the current generation server processor with the Fusion-io flash cache, and the next one was the next generation of IO processor, and he was showing the IOPs. The Fusion-io just overwhelmed. We saw greater improvement with the greater number of cores in the new microarchitecture, but the Fusion-io solution was overwhelmingly better, and he said, "Where am I going to spend my money?" You don't have to tell me. My dad used to say you had to tell me something 3 times before I figured something out, but I heard it twice, and we started focusing on what are the real aspects of performance that need to be, and what has been neglected over the years, and it's clearly been the whole memory and storage subsystem.
Thanks, Steve. I guess one of my questions is, how big is this difference in terms of the memory and storage subsystem? If we're to think about the difference in energy between compute and fetching the data, how big a difference would that be?
Well, we spent a lot of time with the national labs, and especially the guys that do the nuclear codes. For years, they've profiled that for floating point instruction, there are seven instructions. One is floating point, the other six are for data movement. When you go ahead and profile what it takes to bring in a memory device, this is what I learned about going to Micron, I really gained a lot of insight into memory. When you're doing a read, even if you want a byte, you're reading two kilobytes, and if you don't use those two kilobytes, they effectively get written back, and that energy is effectively wasted. When you profile doing a floating point operation, the ratio is almost 1,000 to 1 in terms of energy of doing the compute versus energy of doing the data movement.
Hmm. It might be better to move the compute to the data rather than the data to the compute.
Yeah, where I happened to be at that time, that probably wasn't the most popular position to take. As we started looking at improving the architectures, we started getting to architectures that look very much like, in fact, on the drawing board, and I still have those, Micron's Hybrid Memory Cube. Where there's a logic layer on the bottom, and we stack memory on top, and that memory could either be DRAM, we were looking at DRAM, or at the time flash and maybe 3D XPoint.
That coupled with, in 2013, I spent a lot of time in the Supercomputing world. I was at Supercomputing in Denver, and I was given the paper on the Automata Processor, and my specific task was, is this something that we should actually worry about from a competitive standpoint? My read was no, because there was a tremendous amount of software that was a different programming paradigm. Looking at that problem from the memory perspective versus the processor perspective gave me insights I would've never gotten.
It was clear to me that if we were going to change the game, we had to do it starting from the memory and storage subsystem and working out.
Interesting. Now, the Automata Processor, it worries about pattern matching. Right? It's a bit of a different architecture. Can you say a bit more about the style of architecture or maybe some of that different style's benefits?
Well, it does benefit pattern recognition very much, there's people that are looking at them, there are certain agencies and certainly individuals here. When you're looking at text, you want to look at certain, you want to do some distance matching and character recognition in terms of looking at those levels of text. I would say it probably was the beginning of what our concept would've been a rudimentary artificial neural network.
I think there was a lot of work that we would have to do in terms of making them as robust in terms of the classification. It was the beginning of that style of computation, which now everybody's talking about machine learning and artificial intelligence. A lot of those are really based on the artificial neural network technology that's been around since 1980s.
Thanks, Steve. I think there's a number of people here in the room as well as online who are worried about classification.
A number of inference problems. As we start to look forward then even further to then apply that, right? What are the ways in which our partners and customers can then derive insights from that? Can you describe some of the different ways that people use that?
Well, we actually, in 2015, purchased a small company in Seattle, Pico Computing, who builds FPGA modules. The purpose for doing that was as we were coming out with new technologies like Hybrid Memory Cube, like different flash architectures. We could actually integrate them and get them in the hands of users. We actually had a solution that we could sell with a system. When I first started with Intel, I worked in the solutions group that would take those processors. Their job was to get people to start using those.
It was a pretty successful model. That's what we did with this company, and we actually have an FPGA card implemented with HMC that we use for machine learning. Different companies and different elements are using that in order to build. We've got a machine learning algorithm that sits on top. There's a demo of it outside, but we're looking at different things like, we're starting a collaboration with a health sciences university that's doing genomic sequencing, and they want to use machine learning to understand and get some insight into the different genomic patterns of people that happen to share common diseases. We're using that, and we're using it as a test bed to see how the memory is used because it's understanding the application, how the application uses the memory will give us greater insights into how memory can actually be more beneficial.
There's actually, for a wide range of customers, right? Whether they're doing things with genomic sequencing, right? If it's actually even people or ID tracking, right? Things like traffic and navigation. We should also say a word, Steve, about how we actually use it ourselves, right? We at Micron are sort of with you, our clients, we also use machine learning to increase the yield within our own fabs. We have more than a petabyte of data that we analyze, as Darren was mentioning, we need to do it in real time. As you think about the wafers that produce these SSDs that are part of the SolidScale, right? Each of these wafers and the chips has dozens of different layers, and we need to make decisions in real time about whether or not to move wafers on within our fabs. Very interesting.
It is. One thing I did want to point out is, as we've been going through these, the acquisition of Pico Computing is kind of interesting because their founder is probably one of the most brilliant people I've ever met. Someday I'll give you my treatise on the average use of energy in the brain is 20 watts. Some people theorize that people that are socially very capable use the majority of that for social networking, and people that aren't use the majority of that for computation. I've got some examples where that might be true. Anyway, he actually came to me and said, "Sort is a big problem." When I talked to the guys at Oil and Gas in my previous life, they said, "Compute is great. I have a problem because I can't sort.
We came up with an architecture of building a key-value store and a sort table inside the DRAM. As an element's being written, we can sort it at the same time it's actually being written. We're not incurring that overhead of constantly having to update the sort table.
We can sort by 50x, save a significant amount of power, takes very little die area. What's unique is when you look at the classifiers for doing it, they're the exact same classifiers we can use for character recognition and machine learning. Now we can kind of overlay that technology inside a DRAM device and a NAND device as well. In fact, NAND is actually better because we could store the weights in the floating gate, and they can be persistent.
What you're saying in terms of my workloads is, back when I studied computer science, right? I was sort of suggested that N log N was pretty fundamental for how long it would take to sort or even, and if I knew something about the data, but you're saying you can do it at DRAM speeds now.
We can do it at DRAM speeds, and we're not eating our own dog and falling in love with our idea.
We're actually collaborating with a university in Europe that is going to take that and start running the prototype that we have.
FPGA prototype, running it through a range of algorithms to make sure that, yes, we haven't missed anything. The advantage is that with a simple library call, people can use it with their software today. Because of the capabilities there, software can evolve over time. A brand new architecture, this is one of the things that worried me about Automata is it needed new programming paradigm. It takes, in this industry, two Olympic cycles, summer Olympic cycles, eight years.
for new hardware to actually be picked up by the software community. You want to have something that works in the standard implementation, and when that hardware becomes pervasive, the software starts to take over, and then they can start using the new capability. At the same time, you want to get paid for it, too, so that's always a difficult trade-off to be made.
Well, thanks, Steve.
Sure.
Now I'd like to turn it over to Mark Glasgow, who's going to walk us through a discussion with some CIOs. Thanks.
Thank you, Brad. Thank you, Steve. Well, these are certainly very exciting times for the industry and for Micron specifically. The nearly 40-year history of Micron is proof that this is a company that knows how to not only evolve but change and innovate. We continue to do that today, as you just heard from Eric Endebrock, some pretty exciting news with regards to some storage solutions that we think are going to be fairly game-changing in enterprises of all types. We are very excited about that. One of the things that, as Micron continues to evolve, one of the things that they have done for nearly 40 years now is our go-to market has been through the large OEM partners, and that is not going to change.
We are going to continue to work with our OEM partners. One thing that we saw a need for about three or four years ago was to create a sales team that is customer-focused, that is out having conversations with CIOs of major corporations that are trying to solve these really big, tough workloads and do it in an optimal way that not only optimizes power and cooling but real estate space in the data center, all the while keeping up the performance and keeping the end user happy. We are excited about some of the opportunities. One of the things that, like I said, happened in the last three years is the number of customer conversations that we are having that we can bring back to the engineers that continue to innovate and make these exciting solutions.
The sales team that I run is the one having those conversations, and when we create that demand for Micron, we pull it through the OEM channels that have been there for now, like I said, almost 40 years. Very exciting stuff. Today, we have a number of panelists that I am going to invite up here in a minute that can tell us directly from their worlds what are they seeing from a flash perspective and how it is changing their data center, how it is enabling them to do different things with regards to some of the really super demanding workloads that keep coming at us in many different forms. Before I bring them up, I want to take you through a couple of quick slides. I have been afforded the opportunity to fly really around the world and talk to customers on almost every continent.
These are some of the things that we see out there that is very interesting. As Darren mentioned, huge amounts of data being shipped in different storage formats across the globe. In 2015, the total flash penetration was 5%. In 2016, it was 10%. A couple of exciting things inside this first data point. Number one is almost 100% growth year-over-year for flash. That is exciting. The other exciting part is it is only 10% of the total market of storage being shipped on what is the total available market on planet Earth. These are exciting numbers for us. We feel like we at Micron have a lot of room to run and a lot of expansion opportunities.
Again, you'll continue to see us innovate and create solutions that allow us to grow these numbers and to expand and to build upon this legacy of a great company. The second thing that we see are the workloads. Boy, they are changing. If you look at the super rigorous demands that are created by multi-tenant architectures, cloud in every form, public, private, hybrid, these are big technology opportunities for companies to gain efficiencies in the data center, but they are not easy to solve for. I love what one CIO said when he placed a set of our NVMe drives into his data center. He said it was like aspirin for his data center. Everything just sort of felt better. The CPU utilization improved. The IOPS went up. The latency reduced. His power and cooling reduced. The data center real estate being used was reduced.
All of these things are good things. It was like aspirin for his data center. I've stolen that. It's good. The third thing, continue to drive and bring data closer to the CPU. Again, the workloads are demanding this. It's very hard to achieve some of the machine learning and big data items in environments that are running on a storage area network where the request has to go across a network and hit an actuator arm. Just a lot of latency there. Micron continues to exploit these opportunities as these new workloads come at us and the random IO patterns get even more variable in the mixed workloads. These are all very exciting opportunities for us. Micron specifically is very uniquely positioned.
In fact, of all the storage companies that I've known in the nearly 3 decades that I've been in the storage business, I don't know that I've seen a company better positioned with the market coming at that company like the market's coming at Micron. Many opportunities. In fact, I heard just the other day, the CEO of SoftBank, Masayoshi Son, mentioned that our tennis shoes will have NAND and DRAM in it so that it can measure not only the steps that we're taking but our weight, our hydration, all sorts of telemetry data about your body. This is outstanding stuff, and Micron stands to benefit from the Internet of Things. What's amazing is that there's only 4 makers of NAND flash on the planet, we are one of very few.
The barriers to entry into this business are pretty significant, we don't see that growing by a whole lot. There's 3 makers of DRAM on the planet, and we're one of them. In fact, that's what we've been doing for now 38 years. Super excited about that, our ability to continue to innovate in a market that is driving demand, that is actually having explosive demand. There's only 2 companies that do 3D XPoint and that also do volatile and non-volatile memories under one roof, and that's Micron. Micron's one of the 2. I won't mention the other one. This isn't their show. That makes us very uniquely positioned.
The last one, there's only one that does NAND, DRAM, 3D XPoint, and now solutions, actual storage solutions, that our partners, our OEMs, can benefit from and again, take to market, and it's like aspirin for the data center. Everything just sort of starts to feel better. Uniquely positioned. We're excited. Let me go to the next step here and bring up our panelists. First up, Justin Stottlemyer, where are you? There he is right there. Justin has a very interesting background as an IT professional. He's been at eBay, PayPal, Facebook, Shutterfly, and he now works for Intuit. Justin, thank you for joining us. Next up, Don Duet. Don, thank you for being here. Don was partner at Goldman Sachs for 10 years. He was at the company, the firm, for 28 years, and in charge of their global data infrastructure.
Don, we appreciate you and look forward to hearing from you. Last up, Trevor Scholz, CIO for Micron. Before joining Micron, Trevor spent time at Broadcom, AMD, and Cisco. Trevor, thank you.
Thank you.
Let's kick this off. Don, share with us, with your time on Wall Street and at Goldman, surely you've seen a lot of things and had a lot of challenges thrown your way. How has flash changed things? I'm not talking just about the increased IOPS and lower latencies. What are some of the other things that you saw that had a really major impact?
First, I'd just say thank you for being here. I'm really proud to be here. I think what we have seen over the many decades has been that innovation, particularly at the hardware layer, has really created fundamental change in the way that we think about technology, the types of solutions and services we could do. Obviously, the presentation made earlier today showed that at a macro scale. When you step back a little bit and you think about Flash, you think it was only a couple of years ago that it was still thought of as very nascent technology, that it really We were worrying about write ratios, we were worried about the durability of it, we were worried about would it actually continue to work in a high availability setting, and could you really get your mind through that?
I think if you snap forward to today, certainly at Goldman, I think from many other customers, it is now just embedded in so much of what we provide. It's part of the core infrastructure, part of our core architecture. We've been deploying internal clouds that cover almost in the high 90s of our application use case globally around the world. At the center of that, when you drill down to the servers that are facilitating that, there's Flash as the disk and the storage. It's really gone from being what felt like a really innovative and quite creative solution that had a lot of concerns about how do you enter the market, to something now that's become just a core part of the stack and very much in place.
On the business side, I think we really saw the benefits of that early in places where the risk/reward was high, low latency trading businesses, places where we had congestion in architecture. Now I think we're seeing both the opposite. We're seeing not just those businesses having benefited from the performance and those gains, but we're also seeing that by having this now standard as just part of our offering, that it's just creating new applications, new products, new designs. We're getting to higher order access to use of information in real-time basis that in the past would've been constrained, absent this being a generic part of our architecture. It's changed a lot.
I think, again, it's another example, NVDIMM, we're hoping will follow a very similar story in terms of starting off early, ultimately really helping enable substantial amount of business change and growth.
Outstanding. What I heard there, NAND flash, obviously costs have been dropping over the years, making it more available to a wider and broader set of workloads to be able to justify using it on a broader set. Also on the other side, that's the push side of it. On the pull side of it, you've got applications that demand a better, faster, more efficient storage solution.
Whenever you can offer better, faster, cheaper, ultimately, people will adapt to that, and that becomes the new normal. Of course, people then innovate on top of that and create even harder problems that need to be solved. We're constantly in that cycle, but it's really been driving that growth, which is very impressive in the short time that it's been in place.
Yep. Justin, how about you? Same question.
I've been deploying Flash now for six years.
Okay.
Since the very early days it was launched. Looking at, to Don's point, better, faster, cheaper, that's my Twitter tagline, right? It's always about improving application performance, unblocking engineers, providing business value. Flash does all of those things. In some cases, I've replaced old spinning rust with new Flash arrays or in server memory Flash with 10 to 1 ratios in the data center. That type of consolidation just provides massive amounts of business value moving forward. The instant that one set of developers or business unit gets results in microseconds rather than minutes, everyone is demanding the same levels of performance.
Right.
You just can't back away from it once you've delivered it.
It's like aspirin.
It's like aspirin. Everybody wants it.
Trevor.
I think everyone put it really well. There's a tremendous pressure to innovate, and if you can take advantage of an infrastructure change like this going to flash, the better, faster, cheaper just keeps going, and it is the new normal, like Don put it. As a CIO, there's this constant drive to get more analytics to people faster and in real time, figure out how to reduce costs, figure out total cost of ownership
That constant drumbeat of innovation, the flash piece, I think, is core to a lot of what we're deploying right now.
Great. Justin, you've spent a little bit of time with our brand new NVMe over Fabrics product, so I want to throw a question out there for you. You've been kicking the tires on that. Tell us a little bit about what you're seeing so far, why you think that might be important to your enterprise, and the business value derived from that.
Absolutely. There's a few pieces there that I'm looking at, and it's really about what the next level of performance looks like in the data center. Just as flash has really disrupted disk, it still has the limitations of some of the previous architectures sitting behind it. NVMe is really looking to uncork a lot of those problems, especially with the way that you're showing your SolidScale architecture. The same way that software-defined networks are taking over today very quickly and having broad scale, I see that happening with software-defined storage, and NVMe is sort of the underlying platform that allows that to happen. Just as was called out earlier, I do have isolated storage. I do have performance left on the floor. I don't want that to happen. I need to be able to drive innovation, drive business value, and drive performance.
I think NVMe is where it's going to be with massive bandwidth and low latency. I don't even measure IOPS anymore.
Right.
Massive bandwidth, low latency, get results to users. I'm also looking at deploying this in every piece. Everywhere I have an analyst blocked by performance, I want to throw NVMe at it.
Yeah.
I want to unblock them. My data center is expensive. My headcount is way more expensive, and wasting time with them is just useless, right? I have to get them uncorked.
Okay. This is for whoever wants it, I'll just throw it out there. One of the things that I see when I run around the world and have conversations with CIOs just like this is, companies that are able to drive costs out of the data center, in other words, drive better forms of efficiency, seem to win. They seem to win big. One of the biggest inefficient things that I see, so much so it's now become a game for me, because I'll go around and I'll visit with In fact, I was with a head of IT, worldwide infrastructure for a major bank who has several 100,000 servers deployed, and I asked him, "What is your average CPU utilization?" With a dead pan straight face, he said, "2%." This is a massive inefficiency, and we know that flash helps with that.
This is also a big deal because most of the big software database companies charge by cores. If you have to buy more cores and only use 2% of them, that becomes a very expensive problem. Any one of you, tell me how not only flash, but specifically NVMe over Fabrics might help reducing, I guess, the overall cost by way of increasing CPU utilization.
Well, I think that's been pretty well proven that you can get denser, ultimately you get more utilization out of your assets. That story at scale plays out to be meaningful economics, and I think that that's important. I come back to though, I do believe that ultimately that for most CIOs, that may be a critical thing for them, but the more important thing is the enabling aspect, like getting to agility, getting to where you can change because the world is changing so fast and companies are digitizing so fast. These become, again, this is becoming much more table stakes. You just can't solve problems without having this type of infrastructure in place.
I think that is as much of a driver as the economic savings that you can get by getting denser and cheaper, because ultimately that's what they're there to do, is help lead their business forward and help them really grow top-line revenue and really improve the functioning of whatever companies they're at.
Agility and flexibility.
I agree with Don. The cost is important. Really, if you look at most of my peers when I talk to them, the mantra is fast IT. Speed is the new currency of business. Data is the new currency. It's great to have the cost component, but that business enablement is so critical because the business is being forced to make decisions faster, and they're seeing larger data sets. The business is getting more complicated across the board, regardless of digitization. Cost is important, but the real winner is when you can get that business value and the cost combined.
Right.
I would also say, I think the Fabric story is both fascinating and very important because I think that part of the challenge, we've all learned this the hard way, and part of the reason why someone would have a 2% CPU utilization is that we tend often to build things fast for bespoke problems, and you get a very artisanal solution set in your data center. Starting with much more of, no, here's an integrated fabric that you can get the same degree of capabilities that really enables your developer to solve this problem for that place and this problem for that, but having uniformity at the infrastructure layer is so powerful. If you were only shipping a 1U box, you'd probably have lots of snowflakes of those 1U boxes.
Having it in a form factor that really can be deployed, I think, in the way that the Fabric is going to be incredibly powerful and hopefully reduce that problem of I've got lots of artisanal solutions that are hard now to really get efficiency out of.
Yeah.
Yeah, to both of their points, the snowflake is a big problem in the data center. Having something ubiquitous that allows orchestration, automation is key, right? To getting my operation staff out of the way of my developers and out of the way of my business analysts as well. One of the other things that you'll see happening, of course, is the migration today from internal data centers to public cloud, and both of these people are really competing against each other, internal IT against the public cloud, for providing best levels of service and affordability.
Great. 11 million IOPs in 6U are some of the early numbers that we're seeing. I probably wasn't supposed to say that, but yes, we're seeing some very large performance numbers in a very small space. We know that NVMe does wonders for reducing latency. Shared NVMe over Fabric just opens that world up. Do any one of you see new application areas, things that we just couldn't do before? Obviously, big data, we know there's stuff happening there. I know, Trevor, you're using that in the Micron manufacturing operations. Maybe tell us a little bit about that, what Flash is enabling you to do from an application perspective that you couldn't do before with a SAN.
Yeah. Maybe some context. There's just to simplify things, in a manufacturing fab, you have recipe and scheduling management, which is a certain set of business problems, then you have your process control. Then you have the big data sets of analytics that you're working with. My team is finding, we run hundreds of different applications, and just across the board, each workload you look at and go, "Does this make sense?" More times than not, it does. A tangible example in that middle layer on the process control, in sort of an IoT type of scenario, we have tool data, we have machine data, we have application data all streaming into a, this isn't big data, but about a 20-terabyte database.
It used to take us about five hours with the old tiered storage that we used to have to get to the answers that we're looking at. Are things working as they should? Just by making this infrastructure change to the flash, that takes 20 minutes now, and we're trying to figure out how to get that down to five. We're talking about orders of magnitude of change potential. Why does that matter? It's because now my process engineers get access to data faster, and they can make faster decisions. They can actually change things and increase the output. If you look at the big data piece, it's another great example of where, as Brad mentioned it earlier, it's one petabyte of structured data, but it's another 14 to 15 petabyte of unstructured data.
Right.
People are digging into this information, this data, and finding things they never saw before. The faster that people can get to that information, make analysis on that data, the more business value we bring to a CAU type of business. We're seeing tremendous business value just moving to flash.
Well, some of the tolerances that we have to deal with in the fab are so tight and so tricky, the faster that you can get that information and see those insights to make those changes, to your point, increases productivity.
Absolutely
output.
Gets more wafers to all the people in the room.
Yep, we know the world wants more wafers. Justin, any final thoughts on the NVMe over Fabrics? I know you've been staying to this for some time.
I've been looking at NVMe for about a year and a half.
Yeah.
The implementations are just now kind of hitting the ground, I'm really excited about what this means for, as you talked about CPU efficiency, agility is key, but CPU efficiency and driving that end to end is really important for me as well.
Yeah.
Excuse me. From a data analytics standpoint, right now, I have a hard time driving greater than 50%, 60% CPU utilization out of my Hadoop cluster, for example. If I really want to be able to crank that up, I need to be able to balance my IO, balance my CPU, balance heat and cooling across the data center floor, and on top of that, density. The densities that we're starting to see come in can only be taken advantage of by something like NVMe. Without that, they'll be useless, and we'd be siloed just like yesterday, just like today.
Right.
These things are really uncorking what we're going to be able to do across the data center floor and across all these analytics platforms.
Exciting stuff. Don, any final thoughts?
Yeah. Well, I think similar to Steve Pawlowski had mentioned before about the two Olympic cycle, I think in business applications, it probably feels that way, too. I think what's really amazing, and it's going to be a very interesting story that's going to play over the next couple of years, is how are people going to really change their design patterns for the business applications, particularly around things like transactionality and asset architecture, and how does NVMe in particular really just give you a completely different way to think about that from a concurrency transactional perspective. I think, again, you'll see where the early adopters will be people who will see immediate value in getting to much more higher bandwidth type streaming architectures for their business problems that today are constrained in more of an asset database-driven architecture.
It'll be really fascinating to see how the use case patterns and the design patterns emerge over the next several years now that this technology is going to be in people's data centers and in the hands of their developer teams.
It should change things quite dramatically.
Big time. I think that dream of base architecture, but the realities of this is a business and in finance, obviously the criticality of records and getting record retention perfect. I think you have a very different landscape now to be rethinking how you build the next generation of applications.
Right. I'm not sure I really want telemetry data on my body and my tennis shoes, so while that's a cute application, this is real-world stuff from Wall Street and massive pressures on you in your old job from the users to get things done and do it in the most efficient way possible. Trevor, any thoughts, final thoughts?
Final thoughts are, as I talk to my peer set, that this is a foregone conclusion. I think most everyone is moving to flash. It's just the economics around it are a lot of people have moved early. I think there's no confusion in people's mind about the business challenges and where this technology really solves a lot of them.
Yeah. Well, we certainly feel it at Micron. Again, we're afforded the opportunity to have so many of these conversations. There is literally oceans and oceans of spinning media out there that, as we speak, are melting polar ice caps because they really, really do require a lot of energy. We do see a massive opportunity for not only Micron but for the user market at large to create completely brand-new applications that just were not possible before. So exciting times. Gentlemen, thank you very much for your time today. We appreciate it. Next up will be Darren Thomas. Thank you.
First of all, I want to say, I hope I answered the question, why Micron? You could see from our customer set who understands this at a technology level but has real business situations, business concerns that Laura pointed out. Our customers are reaching back down in their designs, they're reaching back to us. You have this movement where large corporations want to understand exactly how the NAND and DRAM and the flash products and all, how they can solve for them. You also saw when Steve and Brad talked, for those of you who could follow it. You can see that the details about the NAND, the details about the flash, the details about the networking and everything that connect is what makes this possible.
Having deep insights from memory to storage through the infrastructure all the way to the workload, that's what it takes to have this insight. The product we are announcing and the product that you'll see here in just a second has that kind of insight built into it. That's the Why Micron story, and I hope you'll see that. I would like to thank Kirsten and her team for bringing us all together for this at this great venue. I also want to thank, sitting over here to my left, Tom Eby, the General Manager of our Compute and Networking team is here, as well as Mark Durcan, our presiding CEO for now. The reason why I want to call him out, it was his vision. It was his vision to take a memory company to where we are today as a solutions company.
Most of my team and the people you saw talking tonight wouldn't be here if Mark hadn't had that vision. I think tonight or today and last night might be a little bit of the fruition of many years of his hard work to get from a company that was just a DRAM company, which is pretty amazing, to a company that is a full-on enterprise solutions and partnering company. I want to thank our partners, everybody in the room. I want to thank those on the webcast for being with us. For those in the room, we will have a gallery that's just out these doors and down the hall. Lunch will be served there, and we will have an enormous amount of my staff here. The design team who built this product is here.
There'll be demos where you can ask the people who actually run those demos or the people who actually develop that technology. Mark will be here, I'll be here, Tom will be here, the team. We have a good number of our sales team here that can take an order. I know Steve is here if anybody wants to do that. With that, I want to close by saying thank you very much for coming. I hope we have demonstrated to you why Micron is the company that can take you to this level of insight and help our customers get to the answers. Thank you very much.