Okay. Well, good morning, everyone. Welcome to NVIDIA's Investor Day. I'm Simona Jankowski with Investor Relations, and it's my pleasure to welcome all of you here today, as well as all of those who are joining us on the webcast. Before we kick it off, I would like to read our safe harbor. We will make forward-looking statements in today's program regarding our expectations and other future events, which may differ materially from NVIDIA's actual results. I'd like to refer you to our SEC filings for a description of our businesses and associated risks and other factors which could cause the results to differ materially from these statements. All our statements are made as of today, March 19, 2019, based on information currently available to us. Except as required by law, we assume no obligation to update any of these statements.
If we use any non-GAAP financial measures, you'll find the reconciliations to GAAP on our IR website. Okay. With that, let me just go over very quickly the agenda for today. We're gonna be starting off with a few minutes with Jensen Huang, our founder and CEO, who I think you all know, talking about our strategy. We will then move over to our gaming business, which will be covered by Jeff Fisher. Following that, we're gonna talk about data center with Jay Puri, automotive with Rob Csongor, and finishing up with financials with Colette Kress, our CFO. We are gonna have about an hour of Q&A after all of that with Jensen and Colette.
After that, we're gonna have lunch, which is gonna be in the Gold Ballroom if you walk out the doors and down the hall to your left. In terms of just a couple of closing items, if you need anything throughout the course of today, just reach out to myself or Shawn Simmons on the investor relations team. You can find us in the back of the room or just email us. Again, I'll like to request that all of you silence your phones. With that, it is my pleasure to now welcome to the stage Jensen Huang.
Thank you. This is gonna be the fifth hour of my keynote. If you missed it yesterday, if you happen to have missed it, you can watch it on YouTube. Put it on 3x speed because it'll take about two hours if you did that. Simona. Oh, yeah. There we go. First of all, welcome. It's great to see all of you. Here's what I'm gonna do. I'm just gonna do a couple of things. I'm going to explain this, for most of you know this very well. There's some new faces in the room. I thought I would do this. I wanna explain what accelerated computing is. Accelerated computing is accelerating. It's not an accelerator. I wanna define the difference for you. Okay. As soon as, Hey, guys.
Hey, guys. Fish. Hey, guys. I'm trying to give a talk. Close that door. It's an analyst meeting. You know, when you guys come to NVIDIA's formal events, it always seems like home cooking, doesn't it? Accelerated computing. Accelerated computing is particularly important today because CPU scaling is no longer happening at the exponential rates it used to. At a time when application workload demand on computing is growing incredibly fast. The question is: how do we extend Moore's Law? How do we extend Moore's Law? Well, we came about this idea called accelerated computing a decade and a half ago. 26 years ago, when we first started the company, we realized that accelerators could help us achieve performances otherwise impossible with a normal computer. Accelerators.
We identified one particular accelerator that has what we called at the time, if you saw one of my presentations 26 years ago, it said, bless you, sustainable opportunity. Sustainable opportunity. Meaning that this particular application called virtual reality, trying to achieve virtual reality, 3D graphics, was going to take nearly forever. The reason for that is in order to create this environment, you have to simulate physics, light physics, particle physics, material physics. You have to simulate physics, and you have to do it so fast that for all practical purposes, it's gonna take forever. The reason for that is because at the time, simulators, supercomputers were doing it. Simulating some fluid dynamic simulation or particle simulation, it was taking a week on a supercomputer.
What are the odds that we're gonna be able to do it at 120 frames per second? To be able to simulate all the interactions with all of the agents performing artificial intelligence capabilities all interacting together. The odds of that happening within a lifetime is approximately zero. We were not wrong. We identified one problem statement that we said had sustainable opportunity. Ten years into it, we discovered that in fact, in order to continue to expand it, we have to expand the aperture, if you will, of the things we accelerated. No longer was it sufficient to just accelerate graphics. We had to first simulate the physics and then accelerated the graphics. Because you have to simulate the water, you have to simulate the leaves blowing in the wind.
You have to simulate things, you know, particle physics as buildings crumbled. It was impossible to have animated all of that. We decided that you had to simulate that. We expanded the aperture of our accelerator, and we invented this idea called CUDA so that we could expand not just accelerating graphics, but the domain of virtual reality. The domain of virtual reality. That time when we transitioned from a graphics accelerator to a domain accelerator, we became an accelerated computing company. An accelerator accelerates a function. An accelerated computing platform accelerates a domain of applications. Does that make sense? An accelerator is a video accelerator. H.264 accelerator. An audio codec is an accelerator. All of the stuff that runs on an audio codec, with the exception of the analog, can run in software.
All of the things, all of the functions in a video decoder or encoder can run in software. In fact, the first prototypes of a decoder is in software, and the first prototypes of an encoder is in software. All of these functions, computer functions, can run in software, and it's possible to design an accelerator for that one function. You would use a video decoder to decode video, but you would not use a video decoder to compute molecular dynamics. You would use a video encoder to encode video, H.264, H.265, or back in the good old days, MPEG-1 and MPEG-2. You would not use a video encoder to do, for example, a recurrent neural net for deep learning. If you designed a recurrent neural net deep learning accelerator, you wouldn't be able to use that, for example, for random forest machine learning algorithm.
If you designed a functionality just for an accelerator for one functionality, it would certainly be very good, but it doesn't have the necessary aperture to accelerate a large domain. The challenge, of course, is if you created a product that has an aperture of infinite domains, what you've done is you've created a CPU. The reason why accelerated computing is so wise and the reason why, although many other parallel computing approaches have come before us, the reason why it has lasted the test of time is because it allowed the CPU to do what the CPU is good at, and it accelerated the domain of applications that we are good at. That discipline of trying to figure out how to expand the aperture while reducing the aperture at the same time, that strategic choice is ultimately the strategies you see at GTC.
There are several things you could do to test whether something is an accelerator or an accelerated computing platform. Of course, the first thing is it has to be a programmable architecture. On the one hand, one day you have to do molecular dynamics, another day you do quantum chemistry, another day you simulate a large climate science program called WRF. Another day you're reconstructing images out of electron microscopy called Cryo-EM, which won the Nobel Prize in physics two years ago. It's hard to be able to do that if it is only designed for one thing. It has to be programmable. The second thing about all computing architecture is that it has to be an architecture, which means this: an application that you wrote for that computer runs on that computer and on that computer, and you buy a new computer tomorrow and the application runs on it.
A computing architecture has some capability of compatibility over time, and it has to have a large installed base. Otherwise, applications can't find computers to run it on. Accelerators don't have that problem. The other characteristic of a accelerated computing platform is it has to have a rich software stack. It turns out the most important thing about our company is our stack. That's why we talk about it all the time. If you look at this is our stack. Our stack starts with the system architecture. I'm not showing the chip. I'm taking the chip for granted. The system architectures, the RTX is for graphics, DGX is for scale up high-performance computing, otherwise known as deep learning or supercomputing. Hyperscale, HGX. AGX for autonomous computers. Little systems that are intended to live at the edge, largely disconnected from the cloud.
Largely disconnected from the cloud. We are currently, because we're artificial, we're somewhat intelligent, we can perform our jobs disconnected from the cloud. I am currently disconnected from the cloud. Okay. I'm autonomous. That AGX is designed to be an autonomous machine. On top of that is our most important layer called CUDA. I call it CUDA here, that layer is really complicated with a whole bunch of stuff. It's not worthwhile to go into, it's basically, if you will, our AWS. It's basically our Windows. CUDA makes it possible for an application that runs on CUDA to run on all of these devices. Yesterday, I announced a $99 computer. A $99 full computer. It runs the same software stack as a $1 million supercomputer, as a quarter of million dollar DGX deep learning system, or a PC.
There's only one computer architecture in the world aside from this that does that, and it's the x86. An accelerated computing architecture has a rich software stack. The other thing about accelerated computing that's interesting is this. Because of what I said earlier, there's only one computer architecture that can boil the ocean. That's called the CPU. It's general purpose. That's its nature. That's its weakness too. Its strength is that it can run everything. Its weakness is that it doesn't run anything super well. Now, during a time when the performance is increasing by a factor of two every year and a half, it was plenty fast enough. The reason why it was plenty fast enough is because software developers take two or three or four years to complete each round of major innovation.
Meanwhile, the computer's already quadrupled in performance by the time that the next build comes along. It's fantastic to just ride that wave, to do nothing and just let the wave take you. That was the whole dynamic of Moore's Law. It was fantastic while it lasted. If that slows down, then all of a sudden you can't solve new problems. If you can't solve new problems, the software industry will suffocate because they can't obviously introduce new ideas. That's why the world needs a path forward. We need a way to go forward. You're not gonna find a way to go forward by coming up with another general purpose computer. You have to find a way to go into it through domain acceleration.
Not a function accelerator, not an accelerator, but an accelerated computing architecture so that we can take the industry forward. Well, this accelerated computing architecture must have vertical domains that it focus on, otherwise known as the counter of horizontal, vertical. We select verticals strategically and methodically so that we can, one, make a contribution by the time that it's necessary. It's sufficiently large to be able to sustain the enormous investment that we put into it. It's not so large it's essentially a horizontal problem. For example, a web browser is so large there's no such thing as a web browser accelerator. The only way to accelerate a web browser is to make every web browser faster. However, video games is a little bit of a unicorn.
We identified the killer app 26 years ago using exactly the same methodology I just described. 26 years ago, we used the same methodology and we said, "If we wanted to be one of the world's most important computer technology companies someday, what is the killer app that we can make a contribution to that will take us all the way?" The killer app we found that we thought of, that we identified and focused on at the time was a $0 billion market. Electronic Arts was 14 people large. $0 billion market. That $0 billion market is called video games. It's a unicorn because it has two characteristics simultaneously. It never happens. It never happens. You have a spreadsheet used by millions and millions of people, but the computation requirement is low. You have a weather simulator.
The computation requirement is enormous, the volume requirement is very low. In both cases, it's unable to justify an accelerated computing platform. There was this unicorn that stood out there. We imagined that if someday there was a such a thing as a video game industry, it would both be large because everybody would be gamers. Who wouldn't want to play? Two, the computation requirement of it would be gigantic. The unicorn. 26 years ago, we found the unicorn. Well, that same method is being applied here, and you can see one segment after another, we're essentially finding verticals that are sufficiently large in domain that could sustain more and more and more and more and more investment as we grow into them. High-performance computing, scientific computing, a very important segment. Artificial intelligence, we're gonna talk plenty about it today.
DRIVE autonomous vehicles, a very difficult computation problem. Not just a computation problem for in the car, but computation problem before you get to the car. That gigantic computation problem I described a little bit yesterday. NVIDIA DRIVE, the whole platform of NVIDIA DRIVE, the initiative of NVIDIA DRIVE is about creating the autonomous vehicle future, not about making a self-driving car. It's a little bit different. One of them is very large in scope. It requires you to be a software-defined company. It requires you to have an ecosystem. It requires you to have developers and tools. Isaac. I've described Isaac in the same way. The ultimate AI problem is both a wonderful opportunity when you solve it in the device at the edge, but getting there is a supercomputing problem.
I've shown a couple of examples yesterday where you're essentially creating a virtual reality world where the robot has to learn how to be a robot. That is a supercomputing problem. Clara, named after Clara Barton, who started the American Red Cross, is our platform for medical imaging, computational medical imaging, turning the instruments of medicine into a software-defined problem. Today, it's a bunch of instruments and widgets and things like that. Today, tomorrow, it's going to be largely software-defined. Algorithms are going to fly, and they're going to be able to do things that are otherwise impossible today. Lastly, Metropolis. The Metropolis name kind of gives it away. It's really about thinking about cities and places as one gigantic robot in the future. Our city in the future will have three characteristics.
Cities of the future, factories of the future, buildings of the future will have three characteristics. The first characteristic is tons of sensors. The second characteristic, a bunch of computation at the edge. Basically, the reflexes of that robotic city doesn't have to go to a cognitive brain in the cloud. The third connected to a cognitive brain in the cloud. Those three characteristics so that it can make decisions and plan. Perception, reasoning, and planning. The three computations of an intelligent being, otherwise known as the computation loop of intelligence or robotics, is going to be used for Metropolis. At this, I think there are talks at GTC between us and Microsoft where Jetson Nano, our edge computing stack, is connected to the Azure IoT stack, and some really, really exciting applications could be made possible. This is the accelerated computing stack. Very different than an accelerator.
We focus on domains, not functions. There's a couple of things that characterizes a company who's a platform company. If you're a platform company, you talk about design wins less. If you're a chip company, a components company, you talk about design wins a lot. When you're a platform company, you talk about your ecosystem a lot. The reason for that is because you created the market or you're creating the market, and you need a lot of partners to work with you to realize the full potential of that market. You have ecosystem partners that work with you on your platform. That platform is rich with software. That software is domain-focused, not function-focused. It doesn't do CNN. It does accelerated data science. Okay?
If you look at the comparatives when you hear us talk, that's the reason why we talk this way, because we're a computing platform company. I announced a couple of things yesterday. First, Fish is going to talk more about this, but the big takeaway here is RTX is off to a great start. It is clear now that ray tracing is here. This week is Game Developers Conference, and all they're talking about is ray tracing. It is clear that ray tracing is here. Remember this, ray tracing is software. The reason why you can tell ray tracing is software is Turner Whitted, NVIDIA researcher who invented ray tracing, iterative ray tracing, recursive ray tracing, did the first implementation on a VAX in software. We know it's software because all the movies that are made is in software.
It's called rendering software, and it runs on CPU farms, otherwise known as rendering farms. What RTX does is not do ray tracing. What RTX does is make ray tracing fast. We love the fact that people do ray tracing. We just want to make it super fast. There's no question ray tracing is here, and RTX is going to make it super fast. Number two, the second thing that I showed you yesterday was the fact that graphics is going to be a new data center workload. This brings us so much joy, as you could imagine. Graphics is going to be a new data center workload. You heard Matt Garman say it on stage yesterday as well, as he was talking about the need, the demand on AWS, and one of the major applications is graphics. He'd mentioned graphics several times.
Graphics in the cloud, we're super excited about that. It's going to be a new data center workload. Yesterday we gave an update, and Fish will talk more about this, about our partnership with regional telcos, global telcos, and it's part of our GFN strategy. We call it the GFN Alliance. He'll explain this, but very simply, they buy these servers from us. They buy these super optimized graphic servers from us called RTX. After they buy the servers, we host a service on top of it. Because the GFN service belongs to us, we can host that service and we share the revenues with them. We share the revenues with them. They buy the servers from us, they share the revenues on the subscription fees on top. Does that make sense? That's called a GeForce NOW Alliance. You could imagine the economics.
It could be quite good. I talked about data center. Graphics in the data center. There's several new workloads in the data center. We already talked about high-performance computing in the past. We already talked about deep learning in the past. We already talked about inference in the past. We're gonna talk more about that today. Some of the new workloads that I talked about this week, graphics is one, and the second one is a gigantic one. This is the unicorn that we've been looking for in the data center. Let me explain to you why. Remember, I explained earlier that there are two types of applications in computing, and that's why there's largely two architectures. If you look it up, it says there are capacity machines and capability machines. That's the way the supercomputer industry talks.
The way the hyperscale data center people talk is they say there are scale-up machines and then there is scale-out machines. Scale-out is hyperscale. You take a cost-efficient computer, and you scale it out linearly so that you could support a whole lot of jobs that are small at the same time. Scale-out. Scale-up says you build the largest computer you possibly can, whether it is the largest amount of computational capability, the amount of storage, the amount of active memory, the amount of networking. You build yourself the largest machine you can so that you can solve the largest problem as fast as possible for one person. Weather simulation, climate science, these things take forever to simulate. That is called a capability machine. A capability machine, a capacity machine. A scale-up machine, a scale-out machine. A supercomputer, a hyperscale data center. Are you guys following me?
All three phrases are identical. Okay?
All three are identical. Here's the unicorn. It turns out a supercomputer, the market size for it is not very large. The computational challenge is great. We're doing fantastic in supercomputers. Here, I showed you a bunch of numbers. You know, people in NVIDIA call this CEO math, and I just gotta concede it is not accurate. It is absolutely right. Okay. It's not accurate, it's right. This is intuitive math, and if you go double-check it, you'll find that it's probably wrong in some area. At, on the large scale, it is perfectly right. All right. If you look at the numbers, each one of the numbers, each one of the ticks, if you will, is three orders of magnitude. Are you guys following me?
This is three orders of magnitude here. This is extreme log-log, okay. Now, extreme log-log says in supercomputing, I need 1 billion petaFLOPS, not in seconds, in units. I need 1 billion petaFLOPS in order to perform some of those simulations. In the case of concurrent users, CCUs, a hyperscale data center has to support hundreds of millions of people at the same time. In the case of supercomputers, only tens, if not, at the very most, 100. In the case of hyperscale, the amount of computation that it takes to perform these neural networks is small in total. It's just there's a great deal of them. The scale goes everywhere from hundreds of gigaFLOPS to maybe hundreds of teraFLOPS, in units, not in time. Okay. What's interesting is this: data science is that bubble in the middle.
When we have more time, I am happy to break it all down so that you guys get a feel for the numbers. The important thing is this, on the upper right to the lower left, or upper left to the lower right. The upper left is three orders of magnitude more computation from the top to the bottom, and to the lower right is three orders of magnitude in volume. The reason for that is this: data science is the only high-performance computing problem we know where there's millions of people. Millions of people in different fields of science, healthcare, financial services, they call them quants, insurance companies, retail, logistics, travel, you name it. Every single industry will benefit from data science.
That's why there's so many people. The amount of computation you need because the amount of data that you're working on is so gigantic, it's simultaneously a large computing problem. That's why the quants have the largest computers. Now imagine there are going to be millions of quants. The reason for that is because there are so many industries where there's domain expertise, and finally, the technology is capable of being used at a large scale. The frameworks, the algorithms are sufficiently robust now, and the schools are teaching it. You guys know that data science is being taught to every single field of science in a university now, from sociology to oceanography to forestry to agriculture. It is the fourth pillar of the scientific method. This new pillar of the scientific method came about, literally made possible in the last 10 years, came about in the last five years.
About the same time that deep learning was happening, the same dynamics was happening to data science. It is going to be a very large market. Data science as a fourth pillar, theoretical, experimental, computational, and now data-driven science. Okay? This is quite a large market. For the upper left. Am I doing this right? Upper left. On the upper left, our strategy is to create something simple for people to use. It's basically an appliance of a supercomputer, because most companies don't have the ability to build a supercomputer. It's too hard. Too much IT, too much system integration, too much software optimization. We containerized it, if you will, turned it into an appliance. On the right-hand side, it's a scale-out problem. We have to turn basically a data center for an enterprise or a hyperscale data center into a high-performance computer.
There, we have to break it all down in a different form factor, build different GPUs, write different software, work with different partners, and our go-to-market is different. The go-to-market on the right side are the world's top enterprise makers, enterprise computer makers. They're all signed up. They're all so excited. On this left side, they're really deep super quants, super data scientists. There, the numbers are not in the millions. They're probably in the order of, call it 50,000, okay, 100,000. They need the best machine, and we reach them through experts in other specialized IT experts like storage companies, because it turns out if you wanna use one of those machines, you need a lot of storage anyhow. A lot of harmony there.
So it has something to do with our go-to-market, and Jay will talk more about that. The second point is data science is a major new market. I mentioned ecosystem. This is our ecosystem in one slide. Now, when I say ecosystem, I don't mean design wins. When I say ecosystems, these are all partners of ours who are taking the NVIDIA architecture to market. They're not changing it. They're not hiding it underneath theirs. They're taking it to market. They might integrate it with theirs. Okay, so this is our platform and their platform coming together. Sometimes this is our platform going to market by itself. These are ecosystem partners of ours, and we're super happy that literally everybody in the larger IT industry is part of our ecosystem today. We announced two new types of computers.
We announced a data science workstation. We announced a data science server. Both of them software fully integrated. You buy them. You should be able to deploy them. Really complicated set of software. However, it's already configured and optimized for you. One of the things that you could see. Then in the cloud, we announced a partnership with AWS, and Matt Garman was very, very, very I really appreciate him coming down and celebrating the moment with us. So we have workstations, servers, and cloud to take this data science platform to the world. Then lastly, one of the things that you might notice is if these workloads, work sets are so large that it doesn't fit on one computer, the connectivity between the computer becomes the greatest challenge.
I showed yesterday the performance of a fast interconnect and the performance of a fast interconnect with the right type of CPU offload, the performance difference is 2x. What that says is that the architecture of the networking, not just the speed of the networking, matters a great deal. The architecture of the networking, not just the speed of the networking. Our vision is that someday, the computing fabric will not stop at the boundaries of the server. The computing fabric will extend out into the network, and the network and the compute will become one large computing fabric, especially for data centers, which as we know, is the most important computer in the future. Lastly, we announced autonomous machines.
Autonomous machines is both an edge opportunity for us, but we wanted to show you that, in fact, the reason why we're part of it is not just because of the edge, it's because getting to the edge is a big opportunity. That getting to the edge is a big opportunity. In order to create the ultimate AI, which is otherwise known as a robot or self-driving car or IoT device, AI IoT, when people say those things, they're saying basically a robot. An autonomous machine. In order to achieve that capability and putting intelligence at the edge, the process of getting there involves deep learning systems, machine learning systems, data analytic systems to develop the software, to simulate the robotics before you deploy it, and then of course, to deploy a very complicated set of algorithms and software.
Our strategy with the autonomous vehicle is to enable the entire world of AVs to become autonomous. Whether it's robot taxis or passenger-owned vehicles or trucks or cars or vans or, you know, forklifts, construction vehicles, farming equipment, they're all gonna have autonomous vehicle capability. We wanna enable all of that. We created an open software-defined accelerated platform. Accelerated computing platform. We wanna do the same for something that's even larger than that, robotics. There are 1 billion cars sold each year, but as we know, based on the things that I described earlier, there'll be trillions of things out in the world someday. They're all gonna connect into essentially large networks that turn buildings, factories, cities, farms into essentially autonomous robots. The future factory would be a factory, would be a robot that's building other robots.
When I say robot, I don't actually mean, you know, necessarily somebody who has limbs and walks around. Robot, the concept of a robot, chatbot. An AI assistant is essentially a digital robot, okay? When I say robot, I just want you to hear something different than what you might be imagining. An autonomous system, an AI system. We announced a family of products, and we announced yesterday that we're expanding our partnership with Toyota tremendously. We were already working with them on some early developments of cars, and now from end to end, from software development, AI development, to simulation, to computing, to algorithms, we're gonna partner deeply with the world's largest car company. They've selected us to be their primary technology partner, and we're very honored by that. That's basically it.
There are three takeaways then, that 1, RTX has taken off. It's off to a great start. Ray tracing is here. There's a whole bunch of new workloads for the data center. Graphics is one, data science is another, autonomous vehicles is another, IoT is another, robotics at a very large scale. The third is data science is the new driver for HPC, every data center in the future will be a high-performance computing data center. I want to thank all of you for coming. With that, I'm going to hand it off to Jeff Fisher. Ladies and gentlemen, this is Fish.
Thanks, Jensen.
You're welcome, Fish.
Is there a clicker?
Fisher, I left it there for you.
Okay.
Fish and I have only been working together for 25 years.
We were children. Welcome, everybody, to Investors Day 2019. I wanna give you an update on gaming. There's a lot going on. We had an exciting year last year, I'm sure you guys all know, we look forward to a even more exciting year this year. Last year was a record year for gaming. We launched RTX, biggest leap in graphics in 15 years. 15 years ago, we launched programmable shaders in our Fermi architecture. Today, virtually, well, every game is based on programmable shaders. With RTX, we launched a brand-new architecture heralding in real-time ray tracing, I'll talk a bit more about some of the momentum behind this next-generation architecture. Max-Q laptops driving thin and light laptops. This year, we've got the thinnest, the lightest, the most powerful laptops driving the laptop market. I'll talk a little bit more about that.
Just past last year, most recently, we brought our Turing architecture down into the mainstream to the $219 price point. We now have a top to bottom stack of Turing GPUs. Millions more gamers coming onto the architecture. Finally, not to be missed, last year we mentioned crypto came to town, and this past year it left town. We see the crypto hangover on track to sell through our channel inventory by the end of Q1. That is moving nicely. This year was a record year, 13% growth year-over-year, and let's dig in a bit more. First of all, the fundamentals of gaming remain very strong. Our basic core business, it continues to be strong. As we've mentioned before, and I continue to say, everybody born today is a gamer. Every child is born a gamer.
The demographics are also working in the favor of gaming. Gamers continue to game longer in life. You start out a gamer, you game longer in life, the total population of gamers continues to grow. What's driving that? Well, esports momentum is still huge. I wanted to, if you don't mind, for a moment, I saw in my inbox this morning, we get a, we get a weekly update from my team on what's going on in the world of esports. I know you guys track that news very closely, but just in case you missed a few things, I'm gonna read you a couple things that came in my inbox this morning just for note. Call of Duty franchise esports to sell at $225 million per team. Call of Duty franchise spots to sell for $25 million per team.
Apex Legends is primed to be the next big esport. No surprise there. Battle Royale blurs the line between entertainment and esports. It's not just for competitive gaming, but it's also for watching. Snoop Dogg's, you probably missed this. Snoop Dogg esports series kicks off tonight. Walmart becomes the first major grocer chain to put esports arenas in the stores. $5 for open play and leagues at night. ESPN announces creation of College esports Championship, of course. Disneyland Paris to host a Dota 2 Major this year in May. As you know, the Dota 2 Major, the final tournament is being hosted in Shanghai this year. China is one of the biggest markets for esports. In the Mercedes-Benz stadium that will hold 185,000 spectators. The Dota 2 tournament is the biggest tournament in the world from a prize pool standpoint.
Last year, it had a $25 million prize pool. esports is obviously getting a ton of attention. The momentum continues to grow, and it's bringing in new gamers in the U.S. but most importantly, in the APAC regions, in emerging markets and in China. The viewership for esports is about doubling over the last four years. Continues to bring in an audience. More people are watching esports online than watching basketball. The number of gamers continues to grow, attracted by the competitiveness, the competitive nature, the social nature of esports and competitive gaming. About 30% over the last four years, more gamers coming into PC gaming. Of course, on the AAA gaming side, the cinematic side, game production value continues to increase. We talk about this year, it continues to grow. Game developers are adding more realism into their games.
It takes about a 5 times more powerful GPU to play today's games at 1080p, 60 frames per second, than games that were released in 2014. Games keep getting more realistic. You need a higher-end GPU to play them. These are the fundamentals that continue to drive gaming, and we see a strong future for gaming. Let's take a deeper look into our business. Specifically within the gaming business, our GeForce GPU business grew 18% year-over-year last year. That is in both desktop and notebook. The contribution is both from units and ASPs. Our five-year CAGR for units and ASPs continues at about 14%. Total revenue growth over the past 5 years, about 29%. If you look at laptop is outpacing desktop in terms of growth for reasons we'll talk about later.
Both ASP and units contributed to our laptop growth that drove about a 59% year-over-year increase. Looking more specifically at RTX. Jensen mentioned that RTX is off to a great start. I would say RTX is on to a great start. You see what I did there? RTX is on. Yeah. Okay. RTX is on to a great start. We've now released RTX GPUs down to $349. At CES, we launched the RTX 2060 at $349. I mean, time flies when you're having fun, but that was just about eight weeks ago. I look back at our estimated sell-through of RTX from $299 up, that's a 2060 up, compared to our estimated sell-through of Pascal from $299 up, that's about a 1070 up, starting from time zero of each of these devices.
Turing sell-through, RTX sell-through is outpacing Pascal by about 46% in revenue, normalizing the time 0 first eight weeks of sales. RTX Turing is definitely off to a great start. Estimated sell-through. If you look at our installed base, the installed base is ready to upgrade. About half of our installed base is Pascal, the other half is older architectures. Turing is just getting its toehold in at 2%. If I look at the performance of the installed base, 90% of our installed base is below one of our most recent GPUs we announced, the 1660 Ti, is below the performance of the 1660 Ti. I'll tell you in a little bit of why I picked 1660 Ti and why that's relevant. Another fact, digging into our sales.
The Turing buyers that we're able to track that are upgrading from our install base are buying up. 90% of the GeForce RTX buyers are buying up from a lower price point. They had a lower price point GPU in their system. They bought a Turing and upgraded. 90% are buying up. Let's take a look at what's driving that. There are two types of games, not necessarily two types of gamers, but two types of games. esports, simply put, I'll say esports, which is competitive gaming. Cinematic or triple-A gaming. Within our installed base, there's about a 50% overlap. About 50% of our gamers will play both. Some will play one, some will play the other. They all value the performance of the GPU. esports gamers value frame rate.
Faster FPS means faster response time, and faster response time means more wins. We see in our install base gamers that are playing esports titles want to play at 120 FPS or higher. Interestingly enough, looking at our ecosystem, and specifically Fortnite, we can see that gamers that play at higher FPS have a higher, what we call a KD ratio. I'll call it a win-loss ratio. Gamers that play at 60 frames per second relative to gamers that are playing Fortnite at 240 frames per second, win roughly 2.5 times more often. Their KD ratio increases about 2.5 times more. They win about 2.5 times more often at higher FPS. It's natural. Faster to point and shoot is the one who's going to win.
There's definitely a relationship between more FPS and more wins, and the pros know this as well. There's a popular site called prosettings.net, if you've not been to it. prosettings.net has about 900 gaming pros and streaming pros online, where they enter what all of their gaming hardware is, including their system config and GPU. 98% of those on prosettings.net, the pros on prosettings.net, are powered by GeForce. Interestingly enough, over 2/3 of those pros are playing on systems that have the performance of an RTX 2070 or higher. Of over about 1/3 are the performance of an RTX 2080 or 2080 Ti. The pros know that the better the rig, the faster the system, the more often you're gonna win and the better the gameplay.
Looking at our GPU stack, I mentioned the GTX 1660 Ti, 90% of our install base is below GTX 1660 Ti. GTX 1660 Ti is what's required to play Apex Legends, as I mentioned, fastest-growing competitive gaming title on the planet right now, at 120 FPS, 1080p, high settings. Gamers, serious gamers as well as pros, value FPS. GTX 1660 Ti is what we see as a starting point, but they don't stop there. As with the pros, they will upgrade their rig to get the best possible performance. Also mentioned Triple-A gamers. Triple-A gamers have different priorities, or gamers who play Triple-A games have different priorities. Their priority shifts to image quality. These games are designed for cinematics, the best possible image quality. They'll play it, they want it at a smooth frame rate, say 45-90 frames per second.
In order to get 45 to 90 frames per second on modern game, say Metro, Battlefield V, you need to start with an RTX 2060. This is at 1440p. If I were to benchmark this at 4K, you would need to start at about an RTX 2080. There is definitely upward motivation for gamers to upgrade to play the latest esports titles at very high FPS and to play triple-A games at the highest possible image quality. Now is where the fun really begins. That's with RTX and ray tracing. It was this past November that Microsoft launched DXR. Like I said, time flies when you're having fun, but it was just about four months ago when the gates opened for ray tracing in games. Microsoft launched DXR. This week, the next big shoes drop.
Jensen had mentioned it as well. Epic is announcing that Unreal Engine 4, the number one engine for AAA games, is integrating DXR and RTX support in their game engine, and it will be shipping to the thousands of game developers in the next couple weeks. I think tomorrow they're actually gonna give a specific date and show some demos. If you didn't notice from the keynote yesterday, real-time ray tracing is definitely the next big thing. The demos that you saw were unbelievable. If you aren't convinced then, take a look at the Control demo from Remedy that was posted last night on YouTube. Control's an upcoming game that looks amazing ray traced. Unity also announced that Unity is the number one, powers about 50% of the world's games.
Unity announced at our keynote yesterday that they're integrating RTX support and DXR into their game engine, they're gonna be handing out builds to developers starting on April fourth. In addition, most of the first-party engines, including Frostbite, Remedy, CryEngine, engine from Crystal Dynamics, 4A Games, are all supporting DXR in real-time ray tracing. I'm heading up to the Game Developer Conference after this show. My team is booked solid with devs talking about real-time ray tracing coming into games. Real-time ray tracing is the most exciting technology we've rolled out. We've seen the best response that I have experienced at NVIDIA from developers to implement real-time ray tracing in their next-generation games.
We're tracking, let's say about a dozen games that are coming later this year, early next year, to implement real-time ray tracing, and those are the ones that we just have visibility into. With the release of Unreal and Unity, I expect that to accelerate. As Jensen mentioned, ray tracing is a software algorithm. It will run on CPUs, it's accelerated by GPUs, but we designed RTX to further accelerate, to make real-time ray tracing possible in fully interactive games. If you look at Metro, we can run Metro on Pascal. I don't know if you've seen some of the coverage, but we announced that we're gonna be adding DXR support to our drivers for all of our GPUs, so gamers can play with it. The fact that it'll run doesn't mean it's gonna run interactive.
In fact, with Turing RTX, with RT Core, and RT Core plus DLSS, RTX will accelerate over ray tracing over Pascal about 3x. In order to get fully interactive ray tracing, you need an accelerator, you need a next-generation architecture, and that's what GeForce RTX was designed for. Arrow should stop at the green bar, 3x. We're super excited about the future of RTX. We're super excited about the momentum behind real-time ray tracing. Game Developers Conference this week is, I think, when it all really kicks off in earnest, and we're gonna see a ton of momentum coming out of the show. Let me talk about notebooks now. Students want mobility. Students want to game.
Starting at CES, we launched RTX coming to notebooks, and we launched the next generation of Max-Q, thinner, lighter, and more powerful notebooks than the world has ever seen. Max-Q has been driving the growth of the notebook business for the last several years. This year, for the last several years, we estimate the OEM end market revenue of notebooks to be about $12 billion. It's grown about 10x in five years. This is what the OEMs are seeing in terms of their total revenue and revenue growth from gaming notebooks. It's easy to think of a gaming laptop as the fastest-growing game console. OEMs are so excited about gaming and notebooks in particular, they're rolling out more and more models. This year, we expect the number of Max-Q notebook models of thin and light gaming laptops to double to about 45 models.
Within each model, there's going to be multiple GPU configurations. You could easily double or triple that in terms of different notebook configurations that will be in the market this year. Max-Q thin and light gaming laptops are taking over and driving the growth of the laptop market. Jensen also mentioned GeForce NOW at his keynote and the next billion gamers that we can address. Today, we have 200 million GeForce gamers. If you look at the entire population of gamers playing on underpowered notebooks, playing or want to play on underpowered notebooks or Macs can reach another billion gamers. GeForce NOW has been around for about two years now in earnest. We've been perfecting the experience, quality of service, number of games onboarding. Got about 500 games now available on GeForce NOW, 15 data centers, 300,000 monthly active users.
About 1 million people on the waiting list because we can't service them. The demand among gamers who are on underpowered PCs appears to be pretty huge. Within our current monthly active users, about 90% are playing on PCs that are underpowered, do not have GeForce GPUs in them. What is GeForce NOW? GeForce NOW is a GeForce gaming PC in the cloud. Give users access on low-end clients to a high-performance gaming PC in the cloud. Fully interactive gaming, we're rolling out VR. It's a simple game launch. We are not a store. It's a PC in the cloud. It's a simple game launch. You launch a game off your desktop just like you would any other game, and voilà, it's playing in the cloud. We offer an open ecosystem. Publishers, developers, direct to gamers. We don't intermediate. We are not a store.
The stores, the publishers, the developers keep 100% of their revenue. We are a service. Scaling out. We've had a ton of interest. We've seen a ton of interest from telcos who are interested in interactive gaming and VR. It's a perfect use case for 5G. It's a perfect value-added subscription to their broadband customers. We created a program called GeForce NOW Alliance. What GeForce NOW Alliance is, as Jensen had mentioned, we've developed a server that is optimized for cloud gaming. We're using that in our data centers. We are packaging it up as an end product for GeForce NOW Alliance. We'll sell a complete server. On top of that, we will run our GeForce NOW service. License the telco, share revenue as it scales out.
This gives us the opportunity to hit markets that we don't currently address, and it gives telcos the opportunity to bring in more value-added customers into their ecosystem. We announced two partners yesterday at Keynote. SoftBank focused on Japan, bringing their 6 million broadband customers, and ultimately 30 million mobile customers. LG Uplus in Korea. As you know, Korea is a big gaming market, as is Japan. Bringing their 4 million broadband customers, 4 million cable customers and 13 million mobile customers ultimately into the ecosystem. We expect to see the alliance services starting to roll out in the second half of this year. That's gaming for me. I hope I touched on some of the things you wanted to hear about. Our growth levers for this year. RTX is off to a great start.
46% initial ramp revenue, sell-through revenue, Pascal to Turing. GeForce laptops, fastest growing game console. It's the way I think about it. Students, gamers, kids want mobility, they want high performance, they want thin and light. Max-Q is driving this growth. GeForce NOW, we can reach another 1 billion customers. We're super excited about the alliance partnerships. I think our service is awesome. If you haven't tried it, you can log on. I'm sure that Sean or Simona can get you a code to jump the 1 million gamer wait list. You can check it out. It's really, it really is amazing. The interactivity will blow you away on your Mac or enterprise notebook. GeForce Alliance will let us scale out. We announced LG and SoftBank and expect to have more announcements coming over the course of the year.
That's my story for gaming. Look forward to speaking with you all later if you have any additional questions. Thanks so much. I think Jay is up next for data center.
Hey, Jay, before you start, I gotta make a quick announcement.
Okay.
make a quick announcement.
Where are you?
We are in historic grounds. It turns out this cozy room is the location of the world's first GTC Developers Conference. This is how many developers we had. This is how it all started. This was the first one. So excited to tell you that.
Wow. All right, great. Good morning, everyone. Welcome. It's nice to see you all. My name is Jay Puri. I am responsible for NVIDIA's Worldwide Field Operations, and it's a real pleasure to be here. Today, I'm gonna talk to you about our data center business. We had another record year. We grew over 50%. The business is now $3 billion. You know, the computing approach that we pioneered is just really taking off. Our business is driven by applications, and you're at GTC, and you can just see the excitement that all the developers have about NVIDIA's platform. In fact, the number of developers grew more than 50% just last year. The momentum is really terrific. Of course, we are number one in deep learning. We are the de facto platform for deep learning training.
We are getting real traction in inference now also. In fact, our inference business last year was a few hundred million dollars. You know, things are actually going very well. There was a bit of a pause with some of the large hyperscalers towards the end of last year as they digested some of their big purchases earlier in the year. That is temporary. You know, the amount of traction we have with them and all the announcements you heard yesterday with Matt Garman here, with T4 and, you know, NVIDIA's RAPIDS platform now being incorporated into all of their machine learning platforms and so forth. I mean, the amount of stuff we are doing with these customers is actually quite mind-boggling. I'm sure the business is gonna follow as it has to. Okay.
Let me talk a little bit about the size of the market. The overall server market today is about $100 billion. We feel that, you know, $37 billion of that is ripe for high-performance computing, as Jensen described it. About a decade ago, a little more than a decade ago, we introduced CUDA to scientific computing, which was our first segment. Of course, at this point, we have a commanding position in that market, right? All of the super computing centers, every major university, all the research centers, they are now deploying NVIDIA's accelerated computing model.
About five years ago, when deep learning came to the front, and the hyperscalers like Google and Microsoft and Facebook and all, you know, quickly realized that artificial intelligence, deep learning was gonna transform their business and they needed a fast computing platform, and CPUs were just not gonna cut it, you know, they all migrated towards GPUs. We quickly saw that opportunity and leaned into it big time. We took our CUDA architecture, widened its aperture a little more, as Jensen put it, and we had libraries such as cuDNN and so on. Very soon working with all of the framework developers, you know, we had the best platform for deep learning, and we are doing really well with the hyperscalers there.
A couple of years ago after that, I think even the traditional industrial companies in automotive, healthcare, retail, financial services, you know, the leaders began to realize, "Hey, AI is going to transform my business." They all wanted to start using deep learning, and we introduced DGX, which is a supercomputing appliance that allows you to do AI really quickly, you know, get off to a good start. We are starting to make real progress in the enterprise now. Okay? This is just the start. As Jensen mentioned, you know, data science is a new workload that is going to have a major impact on all of these segments. It's going to mean that the high-performance computing part of the server market is going to more than double over the next five years.
We believe that NVIDIA's addressable opportunity there, our TAM, is going to be, you know, $50 billion, give or take. We're really excited about that. Okay. Let me talk a little bit more about the platform, you know. Jensen did a great job of explaining to you that there's a real difference between an AI computing platform and just an accelerator, you know. I think all of the computer science world has now understood that Moore's Law is at an end, and domain-specific acceleration is the way forward. Obviously, this is a big opportunity, as I pointed out.
Many companies want a part of it, there are all types of accelerators that are being announced, and perhaps some accelerators like FPGAs and so on that would like to be platforms, but frankly, they're pretty far from that if you use the proper definition, as Jensen pointed out. Now look at our platform, right? We've been at it for over a decade, 12 to 15 years. We saw this opportunity a lot earlier than most companies. We've been investing in it for a long time. You know, at this point, our platform is software compatible from the Jetson Nano to the largest supercomputers in the world.
We have been really disciplined about making sure that we maintain backward compatibility through all this time as we continue to innovate at a furious pace every year. You know, the number of applications that are here is growing at a really rapid pace, and they span multiple domains, as Jensen explained earlier. Nobody has the maturity of this platform. Just think about the investment. We have made tens of billions of dollars worth of investment in this platform ourselves, and that's just the tip of the spear. It's really about our ecosystem. The 1.2 million developers that we have on the platform now, all of the scale-out partners. If you count the total investment in NVIDIA's platform at this time, it's gotta be, I don't know, hundreds of billions of dollars.
It is not easy for someone to come in at this point and try to duplicate this, right? We have all the important applications in the domains that we are addressing now, whether it is scientific computing, AI, going forward in data science, and the performance is incredible. You know, when you have a new domain like AI, it is important to have some industry-specific benchmarks that everybody can look at to compare different options that they have. Google led an effort to come up with a set of industry benchmarks called MLPerf recently. It's a very comprehensive set of benchmarks. You know, they did a very good job. They are tough.
In fact, we have lots of companies that are part of the MLPerf Consortium, only about three or four companies could even submit results that met the requirements of the benchmark. I'm so pleased that, you know, NVIDIA was the leader in all 6 of the important benchmarks. Not only that, you know, we beat the competition by a fairly healthy margin. It shows that not only do we have a very widely adopted platform, but it is the most performant platform that is out there for artificial intelligence. Okay, our value proposition. Actually, if you understand accelerated computing, I think you understand why the value proposition is so compelling. You know, we are able to accelerate applications manyfold. If you can accelerate applications manyfold, obviously you don't need as many servers.
If you don't need as many servers, you know, the acquisition cost is going to be less. If you don't need as many servers, the energy cost is going to be less. I hope you know that in most data centers, the energy cost actually is more over a five-year period than the acquisition cost. As a result, you know, our value proposition is extremely compelling. Now, of course, we have machine learning or data science. That's our latest workload, and it's a huge opportunity. You can see, you know, our advantage in TCO is 80%. All right. Let me talk a little bit about, you know, our business model. What do we sell? Okay. We sell, we have two types of products.
We make our own systems, the DGX line of products that goes from about $40,000 to over $400,000. You know, it's stacked up in racks, in pods, and so forth. We work with our storage partners and our networking partners to develop a complete solution for our customers. We also take our technology to market through our OEM partners, through our Tesla product line, for example, where our Tesla cards go from $1,000 to about $10,000. Also our architecture is available through every single cloud service providers. Let me talk a minute about why do we do this. You know, why do we have this product line, and what is our business model?
We do our own systems for a couple of reasons. One reason is, of course, it's all about the full stack, as Jensen mentioned to you. If you're going to innovate on the complete stack, we have to have a reference architecture, and that is our reference architecture, right? It's important for us to continue to innovate and move the technology forward. It's also very important for our development partners, all the developers. They have to have a gold standard, if you will, for NVIDIA's architecture, NVIDIA's accelerated computing platform. A second reason, just as important from my perspective, is it's a great tool for business development. You know?
We have to go and create these markets, which means we have to go and engage with all of these lighthouse customers when we're first getting started, and we need a way for us to be able to engage with them. Having our own product line that we can go in with and work with them on creating the first solutions and so forth is very important. That's another reason why we have our own set of products. I have a small sales force, and really we want our platform to be ubiquitous. The real go-to-market strategy actually is through our OEM partners and through our cloud service providers. Every single OEM in the world, every single system builder in the world, is now using our platform to build their solutions, and we are available in every cloud provider. Okay.
Finally, we have NGC, the NVIDIA GPU Cloud, our software hub, that sort of unifies everything because it is available, our accelerated applications and know-how and so on is available there, and that can be deployed, whether it's on our systems or it's on systems of our OEM partners or even in the cloud. That is sort of what we sell and how we sell it. Okay. A minute on our go-to-market strategy, right? Of course, the foundation is our platform. That's where we add all the value, and, you know, that's what we're really proud of. Because it is about domain-specific acceleration, it's not about general purpose computing, what we do is we go and pick those domains.
We go into vertical industries, whether it's transportation or healthcare or financial services or retail or what have you. We look at those industries, you know. We go meet with the leaders in those industries. We try to understand what are their pain points, what are the applications if you could accelerate, would have a major impact on their business, and then we work with them hand in hand and see what we can do about accelerating those. Our track record, of course, is very good. That's kind of how we go to market. We go and look at specific domains in specific verticals, and then we go and accelerate those.
Once that is done, then we have a tool such as we have our Deep Learning Institute, whereby we can use that capability to go and explain that to all of the other customers in that industry that, "Hey, we have a fantastic solution for you now." Of course, we spend a lot of time enabling our partners and our ecosystem, other ecosystem players, to allow us to scale out then and really go and make these solutions available widely. That's it. Pretty simple, actually it's a lot of hard work, but it's fun. It's a lot of fun because, you know, it's fun when you can offer that kind of a transformative solution to the industry.
The types of discussions that you have with people, you know, you can just see the joy that we are bringing to people and how impressed they are with what our platform can do. Okay. Let me just go back very quickly into the three segments that I talked about, scientific computing, hyperscale, and enterprise, and just tell you why, you know, our opportunity in each of these markets is actually growing very quickly. Scientific computing, that's of course, you know, our beachhead. That's where we got started. We are very proud of the science that is possible on our platform.
It was a moment of pride for us, when the Summit supercomputer is the fastest supercomputer in the world, fastest supercomputer in the U.S. There's already such fantastic science that is being done on it. You know, I was just reading some articles about what they're doing about some cancer research, some medical, other medical research around addiction and so on, nuclear energy, fusion, types of things for renewable energy, weather prediction. Just fantastic work is already being done on these supercomputers. We also have the number one supercomputer in Europe, with the Piz Daint.
Last summer, when Japan wanted to have a really great AI supercomputer, they wanted to have an AI supercomputer available for all of their industry to be able to use, they chose NVIDIA, and so we power that supercomputer also. The number of applications that we are accelerating is going up. We accelerate the top 15 applications that are important in high-performance computing and scientific computing. At this stage, we actually accelerate over 600 applications. From 450 last year to over 600 applications. At this point, you know, almost all of the applications that account for the vast majority of the cycles in supercomputing centers are accelerated by NVIDIA.
The other reason that this market is actually gonna become even larger is because it's not just about simulation anymore. People wanna do AI at the same time. Because as you can imagine, you know, it's all about getting to your answers fast, getting to, you know, scientific results fast. If you can use AI to predict where your simulations are going, you know, you can get results faster. So forth. In every field, important domain of AI, whether it's precision medicine, renewable energy, or all this climate weather science that's important, they are now doing both simulation and artificial intelligence.
Again, because we don't have an accelerator, we have, you know, a domain-specific accelerating architecture, when they using our product, they can just do both types of workloads simultaneously, no problem, and it grows the overall TAM for us. All right. Next is hyperscale. We are the leader in deep learning training. Everybody knows that. But, you know, sometimes I get the question, is that saturating? Actually, nothing could be further from the truth. Just look at the numbers. You know, the amount of petaFLOPS per day of training that is being done is just going, you know, straight and up to the right. Not only that, but the complexity of the networks that are now being developed as people wanna do more and more sophisticated AI is increasing.
Today, when people are benchmarking training, you know, it's usually ResNet-50, which they're looking at. Well, that's about 25 million parameters, okay? The interesting AI that is gonna happen is around, you know, the AI assistants and so on. For that, you need networks like automatic speech recognition, Jasper. That's 200 million parameters. BERT for natural language processing. That's 350 million parameters. I'm sure we're just getting started. There is no question that the need for training is going to just continue to increase. In fact, you can just look at the cuDNN downloads, and they just, you know, continue to go up. As I mentioned before, MLPerf is proof, if you any needed any, that there is no better training platform than NVIDIA's.
We are available in every single hyperscaler. I still believe that the big opportunity for us, in addition to training, is inference. We are starting to get traction, but I think it's just gonna accelerate, let me tell you why. You know, in the past, when people were doing inference, a lot of that was images, and it could be done in batch mode on idle CPU cycles at night. For example, if you have, you know, these Google cars roaming the streets, mapping an area, and later on, they want to label their maps with the names of businesses on the route, well, they can do that at night. You know, there's no urgency to that. If there are plenty of idle CPU cycles, they can use that.
If you want to do the types of interactions with AI assistant, like I think Jensen demonstrated yesterday with an example of that with Microsoft, well, then it's a totally different story. You know, you ask a question. The first thing you have to do is you have to go from speech to text in neural network for that. Then, the text, you have to have some natural language understanding. What is the meaning of this text? You know, what is the context? You need some kind of a natural language processing network of you're running inference on that. After you've done that, then you will do whatever is needed, you know, get a result back or search or whatever.
Once you've done that, you may display it as an image, or you may need to go ahead and take that and put that back into speech. You get it back into speech, that speech sounds like a robot. You don't want that to sound like a robot. You need another network to make it sound more natural sounding. You know. The complexity. Plus, not only do you have to do all this stuff, you have to do this stuff in a few milliseconds so that it's useful. Not gonna wait for idle CPU cycles to do that, right. You need GPU acceleration for inference going forward in a big way as AI becomes more sophisticated. Inference is gonna be a big opportunity for us.
You know, here is examples of many companies that are already using NVIDIA for inference. You know, ByteDance, I don't know if you know TikTok. It's started in China, but now it's everywhere. It's short videos. They're just exploding, maybe one of the fastest growing company. They use us for video moderation, you know, make sure that the content there is safe and nothing that we wouldn't want to have on there. PayPal is using us for fraud detection. Billions and billions of transactions, right? By using our technology, they can reduce fraud by 10%. By the way, as I was talking to them, it's pretty interesting, the types of fraud that people think of, it's pretty amazing.
All kinds of collusion between buyers and sellers and, you know, fake stores being set up and whatnot. People can be pretty creative, but they can find it out now, in, in pretty much real time, using our technology, and they can save 10%. Not only can they save 10%, they said they can use, you know, 8x fewer servers. Again, the TCO is just pretty incredible. I don't know, WeChat, Tencent, I mean, this platform is just absolutely incredible. Does everything. A lot of the inference on that including, by the way, if you end up using WeChat with somebody in China, you know, it'll do all the natural language understanding and provide subtitles in your native language, the results are really great.
There's a lot of inference going on already, but as people do more sophisticated inference, you know, I think it's gonna be a very big opportunity for us. Finally, in the hyperscale and in the enterprise, you know, as Jensen said, data science is the big opportunity. It is the unicorn that only, I think, NVIDIA's platform is gonna be able to address in the proper way. Already, all of the cloud providers, whether it is AWS SageMaker or Azure ML or Google ML, they have all adopted our RAPIDS acceleration into their platform, and it's being, you know, it's gonna be available to their customers.
Hyperscale, I feel, you know, very confident that our business in this space is gonna just keep growing. Finally, the third segment is enterprise. This was new, first we started working with people, you know, as I said, the leading companies is starting to work in deep learning. We realized, these guys are actually already doing a lot of data analytics. I mean, everybody, we've been talking about the digitization of the enterprise, and, you know, they all know that to be competitive in this space, they have to collect data about their customers, about their suppliers, about their processes, and they have to get business insight from that in this extremely competitive world that we operate in.
So far, there is a lot of open source software, right, for doing the data preparation, you know, the ETL part of it, and then, you know, pandas and scikit-learn, and scikit-l earn and so on to accelerate the models and then display them in graphs and so forth. There was a ton of open source software already that these people were using, but frankly, you know, they were just not able to be effective enough. Again, the reason is tons of data, but by the time you actually prepare it, Jensen gave the examples, you know, yesterday, about the, I think it was the Verizon network, whereby, you know, it takes eight days to actually massage the data.
You know, by the time they do that, it's already not not current enough before they can even run the models on it and so forth. It really needs acceleration. It needs NVIDIA's accelerated computing platform, and that's what we've been doing. We've been working with all of these open source, you know, the whole open source community, all these algorithms and so on, and making sure that they can all be sped up using RAPIDS so that you can actually work in more of an interactive way. You know, if not interactive, at least, you know, get results in a couple of hours rather than days and months, so you can really improve your decision-making and start making a real difference in your enterprise.
Data science is gonna be huge, so that is the big, big opportunity for us in the enterprise space. I have a few examples here, you know, of some of the work we're doing in deep learning. I mean, there is great deep learning work being done today. Continental in the automotive industry, for example, they're a big Tier 1 supplier to almost all the major car manufacturers, and they are, of course, embarked in trying to build self-driving cars, a great partner of ours. They are using lots of DGXs to do everything from, you know, the data factory, deep learning training, simulation, and so forth.
We have a great relationship with them, whereby not only do they buy our products, but we help them setting up the end-to-end flow for, you know, for doing, you know, building, these networks for self-driving cars. That's just one example of, I don't know how many companies in the automotive industry that we are now engaged with. Similarly, Siemens Healthineers, they're a leader for medical diagnostics, and they have lots of AI experts. They have about 40 AI applications that they are ready to deploy, and they run hundreds of AI experiments today on their DGX supercomputers. You know, I'm pretty sure that every instrument company is going to need to do that and follow their example.
You have, we have wonderful stuff going on in the deep learning space, machine learning and data analytics, data science. That is the big opportunity. Already we do have some what we call lighthouse accounts of, you know, early accounts that we're working with to understand their needs and improve our platform and so forth. Uber is using our GPUs to, you know, just to match the supply of their drivers compared to the demand of their riders. I'm using these phrases even though, you know, that's kinda how they talk about it. I think about customers and, you know, drivers. Anyway, they're trying to match the two, make sure you're gonna get picked up at the right time quickly.
You know, they also use AI, data analytics and machine learning for things like pricing your ride and so forth, and fraud detection, and all of those things. Uber is a great account we're working with now. Walmart is another account that is very excited about our, you know, our platform. They're using it for things like forecasting. You know, you can just imagine Walmart is the largest retailer out there.
The hundreds of billions of dollars business that they do, if they can improve forecasting just by a little bit so that they have less spoilage or something that you'd want when you go to their, one of their stores is actually, you know, is not out of stock, you know, that has an impact of hundreds of millions of dollars to them. They need to do that in as much real time as possible, you know. Today, they definitely use machine learning for that, but, you know, it's days behind. They don't have real-time information, and this would make so much difference for them, and they're very excited. You know, you can see why it's very evident that our opportunity in all of these segments is going to be larger and larger as we go forward.
I'm very, very excited about that. The next thing that we wanna work on, you know, we wanna make sure it's easy for people to adopt our technology, right? The easier I can make it for them to buy, deploy, purchase, the faster our business is gonna grow. One of the key elements of that, I think, is NGC. NGC is really great. It's the NVIDIA GPU Cloud. We started by having it as a depository for, you know, our containers. Now it's more than that. Now it, you know, we call it our software hub. Of course, we have, now 50-plus containers that has, you know, our HPC applications that we've accelerated. It has all of the DL frameworks.
It has many of, you know, all of the RAPIDS algorithms, et cetera. There are so many different algorithms, and we want it to be, you know, end-to-end. This is not a simple thing to be able to pull these applications together. We make it easy for people to use because we just take it all, use all the best libraries to optimize the full stack, and then we just containerize it and put it on NGC cloud, right? That number is just gonna continue to grow, make it very easy for people to go get at AI computing, data science computing, okay? We're not stopping there. We have not just the frameworks and trainers, but we have the training scripts for these frameworks.
We even have pre-trained networks so that you don't have to start from scratch. You can use, you know, start with these and then do transfer learning on your own data and come up with networks that are optimized for your own work. Finally, we are even putting some of the key industry workflows up in the cloud for our customers. Two of them around medical imaging, you know, Clara, some of those libraries, train models for that. For Metropolis, some of our IVA applications and so forth, they are intelligent video analytics. Those models are available in the NGC cloud now.
I think this is going to make it a lot easier for our customers to actually start doing real AI work, and that will be good for our business. By the way, you can deploy these, you know, NGC cloud, I just wanna reinforce, is available everywhere. You can do it on-prem or you can do it in the cloud. In fact, you can do it in any of the clouds there or, you know, on-prem on any of our OEM systems that are certified for NGC, and of course, you can do it on DGX. Okay, some of the other things that we are doing to make our technology easy for people to deploy, one is, I mentioned it earlier, these reference architecture partners.
You know, when we first got started, we introduced the DGX appliance, and we said, "That's great. You know, we've got the whole stack all optimized, people can get started right away." We would find, you know, they would put our DGXs in one room, and they would put the storage, you know, in another room and connect it by 1 Gb Ethernet or something. They would say, "Hey, the performance is not very good." We, you know, we quickly realized that we can't just solve the compute part of the problem. You know, we've got to solve the overall data center problem so people can deploy our technology. We started having discussions with the leaders in storage, such as Network Appliance, Pure, EMC, IBM and so on.
You know, then the networking companies, such as Mellanox, Arista and Cisco. Together, we have developed these pods, these reference architectures. By the way, the reference, it's about, you know, domain acceleration. It's not that you can have one reference architecture that does everything. This is a pretty, you know, important work and takes some effort. We have these ref pods for different workloads. You know, it may be a different pod for training versus simulation versus what have you, like data science or whatever. We are working with them on actually putting these pods together to accelerate, not just, you know, to accelerate at the data center level. It's pretty exciting to have these applications.
They're all pre-configured, and once it's done, we can show people what the what the results are going to be. You know, these proof of concepts that frankly drive you crazy, you know, they'll buy one, and then it takes six months for them to prove it out. All of that stuff hopefully can be, you know, condensed into just a few days, hopefully, or at least a week or so. Then you can prove it out that, yes, you're gonna get this kind of performance improvement in your workload and, you know, people the customers are very happy about that. We have some alliances now with the colo data center providers.
There is, if you're doing scale-up computing, there are certain requirements that traditional IT data centers are not used to handling, in terms of, you know, the amount of power that is required and just the density of the computing and so on, the cooling systems that may be necessary. You know, if our customers are having some difficulty working with their IT department, well, just go deploy it in one of these colo centers, right? They're now available, and they know exactly how to build this out for you. Finally, in terms of, again, you know, Jensen talked about we have two types of computing that we're focused on. Early on in scientific computing sector, we've been focused on those capability machines, scale-up computing, the supercomputers.
As we go into deployment, whether it is inference in the data center, or, you know, for data science, I think it's both, as we said. Today, a lot of the people are just doing inference on all of these volume servers that they have, right? By making T4 available in all of the high-volume servers from these OEMs, we can allow them to do inference and data science right in their current data center. You know, millions, I don't know, 20 million or something of these servers are sold every year. Of course, you have, you know, Spark and so on to try as an attempt to make all of these distributed computing environment work as one, and we're accelerating Spark. That's all great.
Then over time, I think as the workloads get bigger and bigger and they wanna do it faster, people are gonna realize, yeah, it's, you know, we can do it in the distributed environment, with traditional servers and with T4 in them, or there are many times when people are gonna want, data scientists are gonna want, you know, the fastest supercomputer with lots of memory and so on to go do data science. We're gonna be able to address both of those capabilities. T4 is now available from all of our all the major OEM suppliers, and we are no longer limited to just this, you know, the capability machines. We also have the capacity machines, the scale-out machines, which really widens the market, and that's available for us.
It's again, all of this stuff is NGC certified, and so we know that it's gonna support our platform and all the applications that have been developed on it. Okay. That's it. You know, lot is happening in our space. The data center market opportunity is a big one for NVIDIA. I'm very excited about it. At this point, there's no question accelerated computing is the path forward. If somebody ever talks to you about a new ASIC that came on, you know, please remember it's about the accelerated computing platform. It's not about accelerators. I, you know, I feel quite confident in our position when it comes to that. It is all about the acceleration stack and data science that is the next big opportunity.
Not only is the opportunity big, but we are taking a lot of steps to make it easy for our customers to purchase and deploy our solutions so that the business can grow faster. Okay. Thank you very much.
Ladies and gentlemen, you will now have a brief 10-minute coffee break.
[Break]
Good morning, ladies and gentlemen. We will resume our program in five minutes. Five minutes. Thank you.
[Break]
Good morning, ladies and gentlemen. We will resume momentarily. Please take this opportunity to find your seats and silence your cell phones. Thank you so much.
Good morning, ladies and gentlemen. Next, we have Rob Csongor with Automotive.
Hi, everyone. I'm Rob Csongor. I'm gonna talk to you about Automotive. I'll give you an update on our strategy. You guys know our strategy in Automotive is an end-to-end platform. It's an open platform. It's for building autonomous cars. What I'll do in the, in my presentation is I'll give you an update on how that business is growing. I'll give you an update on the market drivers, what's driving the business, what are the things that are important. I'll give you an update on our strategies and what the size of the opportunity is. Then I'll talk to you about our progress, what are the things you can look at to see whether or not we're making progress towards our objectives. Okay?
First of all, in terms of growth, there's a lot of different growth factors that you can look at in our business, but there's a couple I'll touch on. On the revenue side, I guess it was another record year for Automotive, but we're looking at a much bigger opportunity. The thing that I think I would touch on and highlight as something that was very significant this past year. If you remember last year at Investor Day, we were just launching Xavier. We had announced that Xavier is gonna be going out. We said that Xavier was the processor to power the autonomous vehicle, to power the self-driving algorithms, to power the cockpit, and we were just launching it.
You know that as part of our open platform, we work with literally hundreds of companies. Sensors, Tier 1s, car makers, truck makers, all sorts of different vehicles. During this past year, we basically went from 0 to order over 80 companies that are now building on top of the Xavier platform. This is, I think, 1 of the most significant things. The Xavier platform, of course, is software compatible to the previous platform, yet people who are looking to drive an autonomous vehicle and get it out soon made the move from our previous generation to the Drive Xavier platform, and that's really important. Another thing that's important if you look at this past year, is that it's not just that people started developing on Xavier, but the different kinds of vehicles that are now being developed using Xavier as the base platform.
Of course, Toyota this past year, and you saw the announcement yesterday, but separately, just individually, Toyota has selected Xavier as the platform. We announced that. I highlight a few other examples just because they're interesting. Volvo selected Xavier as their platform, but they selected it Level 2+ translates to mass-market vehicles. NVIDIA is not only being used, or up until then you had seen us in mostly robotaxis or high-end Level 4 type vehicles. Now they are Level 2+ as an important first platform to engage on to get it out soon, and Level 2+ is a very high function, fully featured AutoPilot solution that we call DRIVE AP2X. There's a number of reasons Level 2+ became more important this year, and I'll talk about that.
At the other end of the spectrum, Level 2+, we announced that robotaxis are being built using Xavier as the base platform. Xavier or Pegasus, as we call the development platform for Xavier. We announced this past year that Daimler, we're working with Daimler to develop their robotaxi solution. Not just cars ranging Level 2+ to robotaxis, but also different kinds of autonomous vehicles. This past year, you'll now see that there are forklifts, autonomous forklifts being developed on Xavier. There's construction equipment, earth movers now being developed on Xavier. There's last mile delivery vehicles, delivery bots, UAVs, UGVs, a whole world of autonomous vehicles being developed. That really brings up and illustrates the fact that the world of autonomous vehicles is much bigger to us today than it was last year. It's not just about cars and trucks.
We believe that every vehicle will be autonomous, and the reasons are compelling. They're different for every type of vehicle, but in every case, they're not being developed just because this is a new feature that you'd like to add on. There's usually a critical problem or something happening where autonomous vehicles can uniquely solve a problem. For example, in cars, of course, we're all aware that 3,000 people die every day in the world. We literally have a 9/11 every day in the world due to human-caused accidents. In trucking, it's a little bit of a different problem. We live in the Amazon era. Today, there's a shortage of 60,000 truckers in the U.S. that's expected to triple by 2026.
Furthermore, with electronic logging devices that are now required on truckers, that limits the amount of time that they can drive per day and further reduces the amount of productivity that can be brought out into the road. Self-driving Level 2+ solutions for truckers allow the truckers to extend the amount of miles that they can drive on the road, because the amount of miles where they are not driving but resting doesn't have to be logged as driving time. This is a significant game-changer for people in that industry. In the trucking industry, of course, you're feeding the demand for delivery. There's 120 million households in the U.S. Half of them, 60 or 130 million, sorry. 65 million of those households are Amazon Prime subscribers. This is placing a demand on delivery.
Mobility services, robotaxis, buses, the cost of ownership. We have an entire generation of people, young adults today, who don't want to own a car. The cost of ownership for using services is lower, and also the footprint on the planet, the amount of parking lots can be reduced. The whole world of autonomous vehicles. It turns out that there's 1,000 accidents, 1,000 fatalities, sorry, that occur every day related to workplace accidents. 20% of those accidents are specifically related to construction. This year you saw Komatsu announce that they're using Xavier to develop an earthmover that can look around with cameras placed on the earthmover and make sure that you can detect workers that are around it and make sure that no harm comes to them. Forklifts, delivery bots, tractors, agriculture.
In countries like Japan, farming has become a crisis. The average age of a farmer in Japan is 67. The amount of farmers in Japan has dropped in half in the last decade. Autonomous vehicles to help with agriculture and food production are not just a good idea, it's a strategic imperative to develop. The result of all of these things and more is that you are seeing, and this is projected autonomous vehicle shipments by 2025, 30,000 heavy trucks, 750,000 agricultural vehicles, 2.5 million commercial robots, 1.1 million UAVs. Okay? The world of autonomous vehicles is much bigger than it was. To address this market, there's really three growth opportunities.
There's three areas where NVIDIA has developed a platform solution and what we call the end-to-end solution. The end-to-end solution really consists of, number one, you have to build computers that go into the vehicles so that you can autonomously drive. Number two, you have to train and develop deep neural networks to create the algorithms for those cars, both in the cockpit as well as in the car to drive. Third, you have to test and validate those algorithms to make sure that the vehicle that you put on the road is safe. These are not three individual, separate, random pieces of equipment.
The reason why NVIDIA decided that we will build a car end to end is so that we could deeply understand the problem. In the process of doing that, we of course learned that all of these things are essential to building an autonomous vehicle. You cannot build or deploy or test an autonomous vehicle without these things. If we need them, other people need them. That turns out to be true. The end result of this is that on the DRIVE computer side, given all of the market dynamics I just described, we have a $25 billion TAM opportunity, driven in the short term Level 2+, Level 5.
We have a $3 billion opportunity on the DGX side, just in terms of how many car makers are there, how many cars do you need to develop algorithms for, millions of images you have to collect per DNN, 10 plus DNNs that have to be developed per car, from that, you can do the math. This is only just getting started. Imagine all of these vehicles, more models, more cars, more vehicles coming out. Of course, this will just grow for us. Finally, the testing and validation. The testing and validation really has to do with you need a way to accelerate your testing and validation, because otherwise you are going to be spending hundreds of billions of driving miles for hundreds of years to test adequately, make sure that the car is working.
Just based on the kind of engagements we have now, the type of miles that have to be driven, we believe that this is a $2 billion opportunity for us. At the high level, these are the opportunities. Specifically, there's a couple things that are really driving the market for us. Over this past year, I think you're aware that Tesla Model 3 became the best-selling premium car in the U.S. In those cars, in Model 3s, Model S's, these different models, the autopilot function has an attach rate of close to 80%. They are selling that autopilot and generating an estimated roughly $1.5 billion of incremental revenue based on the fact that it is an excellent autopilot. It is operating with multiple DNNs, surround cameras, and it has high performance computing that's powering the whole thing.
In contrast to very simple ADAS solutions which can provide assistance, there's nothing wrong with them, but they just simply are not a full function driving autopilot. This is creating a market for a very full-featured Level 2+. when you look, for example, at our announcement with Volvo of Level 2+ solution, you notice that this solution is being targeted at mass market. It's not just for a premium car, it's for top to bottom vehicles. We think this is an important market driver. On the training and development side, of course, you have to collect data, you have to label data, you have to train. Not only that, you heard Jensen talk yesterday, you heard Jay talk about the new opportunity of data science.
Car makers also collect enormous amounts of data, not just the ones that are about training and developing autonomous vehicles. They collect data on customer behavior. They do pricing analysis. All of these things we believe are going to be opportunities for us in the data center of the automakers. Finally, validation. Simulation now is not just viewed as an option for deploying a car. You see increasingly articles coming out now that say that simulation is the key to accelerating the safety and arrival of autonomous driving, and we believe that. We also know if you're aware that RAND Corporation issued a report where they said that they did a mathematical analysis of what it would actually take to test and validate a self-driving car, and they came to the conclusion that it would be just about impossible.
You would have to drive billions of miles with thousands of drivers for hundreds of years. Therefore, you need an alternative solution where you can test for corner cases and a lot of the things we announced here at GTC about DRIVE Constellation are going to be the solution for that problem. All of these things, the opportunity, the market drivers, of course, form the basis of our products and our strategies. When we say end-to-end, it means from driving, training, and validation. When we say open, it means that we have a massive ecosystem of hundreds of partners that can plug in. They can develop solutions on top of our platform. Our customers and partners are welcome to use as little or as much of our solution as they like. For example, let me illustrate. We announced DRIVE AP2X yesterday.
This is our full autopilot Level 2+ solution. We have three Tier 1s that have announced three auto suppliers that have announced that they're building on the Level 2+. One is Continental, one is ZF, and one is Veoneer. Those three actually are the perfect example of how the ecosystem has the choice and flexibility to develop on our platform. ZF uses our software top to bottom, not just the DRIVE OS layer, not just the API layer, DriveWorks, but also all the way through into applications. Continental uses part of our software stack. They use our perception, and then they supply a lot of their own path planning and a lot of their own parking solution. Volvo and Zenuity, using Veoneer, are developing.
They develop on top of the DRIVE OS software layer, CUDA, cuDNN, TensorRT, and then they develop the software stack on their own. Perfect illustration of three different partners building on top of the NVIDIA platform. On the driving side, our solution starts with our platform DRIVE AGX, with our software, and of course, all of the complexity, everything having to do with perception, localization, and path planning, and all of those break down into a whole bunch of different algorithms and solutions. Very complex, very compute intensive, and an enormous amount of software. Yesterday, you heard us announce, I think we've shown previously that NVIDIA has a world-class perception stack based on our artificial intelligence. We've also shown world-class localization to HD map, working with every mapping company in every continent.
Zenrin in Japan, NavInfo in China, here, TomTom across North America and Europe. Yesterday, we announced Safety Force Field. We announced the mechanism for doing world-class path planning and creating a computationally safe methodology for a car to navigate in a dynamic world with lots of moving objects, and then to take that methodology and transform it into driving software that will allow an autonomous vehicle to drive safely. The end result of all of this, together with our tools and then an ecosystem on top of it, makes up our driving strategy and our driving platform. An enormous amount of work. On the training and development side, in the last 2 years since we first started engaging automotive companies, we went from basically a handful of customers to over 60 automotive companies today that are training and developing using DGX for automotive.
This is obviously a significant increase. Collectively in that number, there's 25 car makers, 15 tier 1 truck makers, mobility service providers, mapping companies, and startups. By the way, it's not just customers of NVIDIA DRIVE. For example, here at GTC, you can go listen to BMW present on training on a DGX at GTC. BMW, of course, in their current generation, are using Intel. DGX represents an opportunity and a product that the entire world can use to train and develop their self-driving cars. Finally, on the validation and test side, we announced DRIVE Constellation, it's really a three-pronged approach to how we test and validate a car. First of all, we do what we call component-level SIL or software in the loop.
Imagine that you can take the data that you have and play it back to your computer, then you can do regression testing, you can change things. You can say, "Hey, let me remove a radar. Let me have a camera fail. Let's see how the algorithm responds." You can do this in super time, okay? For example, you can have several months of driving that occurs within a fraction of the time. We also allow you to do DRIVE Constellation HIL or hardware in the loop. The DRIVE Constellation box or solution is 2 different boxes. 1 box that is simulating or synthesizing the world. It creates the world. The other box is where you put your driving computer. It thinks that it's in a self-driving car.
There are leads that come in that represent sensors. The sensor images and feedback that comes in are simulated. The drive computer there drives and sends out actuation signals back to the synthesis box, the simulation box, and as a result, you're now able to test it. When you're driving normally, you know, there's companies that are talking about driving, you drive millions of miles. You know that most of the time nothing's happening. You're driving on 101, and everything's fine. It's a sunny day, and you're in the lane, and you go. Obviously, in simulation, you can create challenging scenarios much more quickly than waiting for them to occur in real life. What we show here at GTC, and if you have a chance, go check it out, it's amazing.
At a touch of a button, we can make it rain, we can make it snow, we can make it nighttime, we can make it foggy, just confuse the bejesus out of the car. All of this is, I think, essential to accelerating the testing and validation. This strategy which we came up with, was born out of our needs to develop our platform. As I said earlier, if we need it, then why wouldn't somebody else need it? Up until yesterday, you might say, "How can you prove that?" Or, "How can you show a validation point that this actually is true or this hypothesis works?" Today, the best way I would illustrate it is to just highlight the announcement with Toyota. The Toyota announcement is exactly that engagement model.
It is a recognition of the fact that all of these things are essential for the world's largest automaker to recognize that first we need the computational power in the car to drive the algorithms. Second, we have to create the simulations to test and validate it. We have to have the computer, we have to have the development vehicle, and of course, we have to have AI for the AV vehicles. This is NVIDIA's automotive business strategy applied to the world's largest automaker. It is the model for our engagement, and it is the end result of what we intended with our strategy. We were excited to announce Toyota. We obviously believe that this is what's needed in order to scale, to create lots of vehicles across all of this different world of autonomous vehicles.
Then, of course, we look forward to making more announcements in the future. Aside from this, if you ask, what are the key things that occurred this year that you could look at that are key individual milestones or accomplishments towards our goal, I would really break it into our innovation, our product milestones, as well as partners. If you look at them, a lot of them I mentioned. You know, Constellation, our simulation solution Safety Force Field, which is NVIDIA now moving to the third part of what's required for a self-driving car. We've shown world-class perception. We've shown world-class mapping, localization HD map, and now we're showing world-class solution for path planning. DRIVE AP2X, we believe Level 2+ is now important.
I think you'll see a lot of car makers make decisions on Level 2+ this year. The reality is that HD maps don't exist everywhere. Where they don't exist, NVIDIA will create a personal map for you. It'll be generated by the car based on where you drive so that you can drive safely. All the things we show, 50-mi loop Pegasus, we now have taken our graphics expertise and now leveraged it into creating the confidence view so you can trust the self-driving car. It's not enough to just have the car drive. The car has to communicate back to you what it sees so you can trust the car and believe that you're safe. DRIVE Hyperion, which is the extension of our development strategy from our SDK, we can now put our SDK into a car.
The second we make a change on our software stack and do an OTA, you as a partner get it instantly, and that's part of our strategy. Our simulator, TÜV SÜD. If you know TÜV SÜD, they are one of the safety experts in the world. They certify various processes of developing a car. This past year, they certified NVIDIA as being passing a certification for being able to develop a silicon semiconductor solution for a car, and this is we're the only semiconductor supplier to be able to reach this. In addition, we are now the only non-car maker company certified to drive self-driving cars in China. China, of course, very important, not just to us, but to a lot of our customers and partners, global brand companies, as well as the local companies in China.
Of course, global mapping. On the partner ecosystem side, you notice I won't go through every one, but you notice that they're grouped into not just cars, but trucks, not just Level 2+, but robotaxis, autonomous vehicles, Yamaha, Komatsu, and of course, Chinese companies. All of these are announcements that were made this past year, I believe validate our approach. Okay? Just to wrap up, our strategy is simple. We believe NVIDIA is the only company that is delivering an end-to-end open platform for building autonomous solutions, as evidenced by the things I've talked about. On the driving side, we believe the world of AV is bigger than ever. It's not just about cars and trucks, I've shown you some of the design wins on these new types of autonomous vehicles.
It's a big opportunity. I believe the strategies that I talked about are game changers for a lot of the car makers. Certainly, you see some of the evidence of that, especially with the announcement with Toyota. For training and development, we're just getting started. Collecting training and analyzing data are essential for autonomous vehicles. We've now grown to over 60 automotive companies on our DGX business, and like I said, it's just getting started. Finally, on the validation side, DRIVE Constellation simulation systems are now available, and the DRIVE simulation system, like every other part of our platform, is open. We have multiple partners, from IPG developing physics models and sensor models to Cognata, who's developing traffic scenarios, existing simulation solutions that already exist in the market that can now tie in to our platform because of our open platform strategy.
Okay. Thank you very much. At this point, I'm going to introduce Colette Kress, our CFO.
Okay. Still morning. We're a little bit behind, but we can catch up. I'm gonna try and just summarize in total what you've heard throughout the teams. Then we'll take that time afterwards to open up for Q&A. Let's just talk through a couple numbers. How about that? All right. Another record year. This is actually our 5th consecutive record year in terms of revenue, as we finished fiscal year 2019 at $11.7 billion and growing more than $2 billion year-over-year. A growth rate of about 20%, fueled by all of our different platforms, which we'll talk about. Our gross margin, also a record in terms of its overall growth in reaching 61.7%.
Keep in mind, there is still in there, we would've been higher except having to write down some of the overall inventory later in the year. Since the absence of our overall IP licensing, our value-added platforms continue to drive our overall gross margin up. Our operating income, also a record year and reaching $4.4 billion and growing faster than our overall revenue at 22%. Overall profit, whether you look at overall net income or EPS growing significantly faster at approximately 35% as well.
Now, when we think about the market platforms that we just addressed, throughout the room, you heard from three of them, four of them in terms of here, all reaching overall record level, and this is in a view to look at our overall growth rate over the last three years and the compounded growth rate that we have seen. First, starting with gaming. Gaming, in terms of its long-term growth rate, has been growing 30% over this period of time, even this last year growing 13%. As you think about this going forward, you should think about the overall gaming as being an overall entertainment industry. Fish was up here talking about what you should see in terms of the growth drivers as we move forward.
RTX is here, a new overall architecture to take us forward for the next couple of years, and we now have a full portfolio of RTX available. Talked quite a bit in terms of the overall ASPs and how they have overall helped our portfolio in the past, but as you can see, there's even more opportunity as we move forward. The overall unit growth in terms of gaming is definitely there as well as we think about the refresh opportunity of our existing gamers, and as we know, there are more gamers coming on board every single day, those in terms of starting at a younger age and also staying in terms of longer in terms of, well, in terms of their 40s. This, in terms of we'll continue as we hope, moving forward.
We also talked about new opportunities and things that we have seen most recently, the growth of overall notebooks and the use of notebooks and the mobility to continue their overall gaming experience. Additionally, we talked about streaming gaming, and now we have an opportunity to again address this very wide and growing market in a new form factor and for gamers that have not actually been in touch with it. Now, Professional Visualization. Professional Visualization, also extension in terms of the graphics that we see on the gaming side, but taking that to the overall enterprise. We've seen an expansion of this market as well, largely focused in terms of the mobility of their overall workstations.
The thin and light, the overall performance improvement has expanded. You see in terms of the growth rate that we see in Pro Viz 15% over the last three years, growing quite nicely. You also have RTX coming to overall Professional Visualization. You also have heard in terms of yesterday, our focus in terms of the creatives out there and how they can improve the overall rendering process with Pro Viz. Data Center. A business over the last three years has pretty much almost 10x increase. Just three years ago, this was a $300 million business, and we're now approaching $3 billion. I think the whole day today, as well as yesterday, was really focused about the breadth and depth in terms of the overall solutions that we have for overall Data Center.
That means in terms of focusing not only on High-Performance Computing, something that we've been working on for 10 years, the addition of hyperscales over the last couple of years, but now the growth that we can see in terms of the enterprise. That focuses on many different types of workloads, focusing in terms of deep learning, which you know us very well by in terms of overall training, also what we have been able to do in terms of expanding to overall inferencing, our growth in terms of High-Performance Computing and adding overall AI and acceleration in there as well. Lastly, we're focusing on many of the different workloads that the overall enterprise uses in the expansion of the market from data scientists to the overall focus in terms of rendering as well. Automotive.
On, on the surface, in terms of we're just getting started, we're still looking at a three-year CAGR of 26%. That 26% is largely due to our base of overall infotainment systems. Over the last couple years, you've seen us also grow in terms of incorporating AI within terms of the cockpit and our initial overall work in terms of what we can do for autonomous driving. This is gonna be broad and far in terms of where we can actually address the market using our solutions in terms of automotive, not just thinking about what will be inside of the car, but what will be in their data centers and what we will do to help them as they continue to have these cars on the road in terms of the testing, the validation, and other pieces.
Again, our overall portfolio, all in terms of growth opportunities as we move forward. Our gross margins. Our gross margins continuing to grow over this three-year period of time, and our value-added platforms continuing to be the most important part of our overall gross margin and what has driven that. We'll talk about this further in terms of the need of overall software in terms of our platforms to bring them to market to allow people to overall use that. As you know, the software is not necessarily included in terms of our gross margin. That will be incorporated in terms of our OpEx. Overall growth in terms of our gross margins and definitely an opportunity to continue overall growing. We broke out here our gross margins in a slightly different view in terms of our overall gross profit.
Where do we get the majority of our overall gross profit? More than 70% of our overall gross profits stems from gaming and overall data center, which obviously takes up a good portion of our overall business. Keep in mind, one of the highlights that we talked about on our last earnings calls was the impact of inter- and intra-overall segments in terms of there. Mix is the largest driver in the near term of our overall gross margins, mix both in terms of between our overall segments as well as in our overall segments. The black lines here indicate in terms of the ranges that we can see based on the portfolio that we could sell in those two major overall segments.
These overall drive our gross margins as we continue to build a larger and larger proliferation of products in terms of the data center, as well as the different overall gross margins and ASPs that we have in terms of our gaming business. Operating expenses. Our operating expenses business, excuse me, our operating expenses here grew about 27% this last year, trying to keep up with the growth that we have in terms of our product portfolio. Very well structured overall OpEx because we can have an overall architecture consistent across, and that unified architecture allows us to be quite efficient in terms of the amount of spending that we need to do.
Our outlook for fiscal year 2020 as we move forward is a slightly lower rate in terms of what we had seen in this last couple years. We're expecting about a high single-digit growth rate or a little bit over $3 billion, $3.1 billion overall growth. Our operating leverage. We talked about this a bit in terms of what we have seen in terms of the leverage that we get from having a single overall architecture. Just five years ago, our engineers that we had were mostly focused in terms of on hardware, meaning we had a larger organization in hardware than we did in terms of software.
As you've seen us talk about the overall software over the last couple days, you'll see now in fiscal year 2019, we have a larger percentage of software engineers, a significantly larger amount of overall software engineers than we do overall hardware. When we think about our R&D, therefore, by those platforms, starting at the bottom in terms of the underlying architecture, the GPU architecture, that makes up 40% of our overall R&D costs. Our software layer is therefore about 30% of the overall cost as we string that across all of the different GPUs and all of the different systems that we have.
On top of that, we just have a small percentage, about 25%, that allows us to go industry-specific, market-specific in terms of building out our individual solutions, whether that be for automotive, whether that be focused on AI, or whether that be focused on, in terms of what we need for graphics as well. Our operating margin expansion has been focused on this unified model. It allows us to overall expand our margins quite nicely over the last three years and continue to effectively invest in our businesses without having to worry about the overall margin increase. We'll probably see this continue as we go forward, as well as we look at this as a very key area for us to focus in terms of growth. Our cash flow and overall cash balances.
Our cash flow has grown quite about 3x over the last three years, we're reaching about $3 billion or $3.1 billion of this last year. That's allowed us to produce an overall cash balance of $7.4 by continuing, though, with our overall capital return program. Our capital return program is an integral part of our overall shareholder value and delivery. Since 2013, we've delivered more than $7 billion to shareholders or approximately 70% of our free cash flow. What this has allowed us to do in this last year is we started out the year with a little bit smaller in terms of capital return. We initiated our intent for capital return for the new year and started that at the end of fiscal year 2019.
What we have remaining in terms of our intent for fiscal year 2020 is about $2.3 billion to return to shareholders over this period. Where and use of our overall cash. As we look backwards in terms of 2019, very in line with where we had talked about the last time we had met, we'd focused primarily in terms of investing back into the business. You can see this with $2.8 billion invested back. We focused in terms of also CapEx. A lot of that CapEx is focused on our engineers and allowing them to give the tools, the supercomputers that they need to build in order for them to eventually sell them. Also our focus in terms of the capital return is the key areas that we that we focused on. As we move into fiscal year 2020.
Fiscal year 2020, you'll see about the same side of overall OpEx, a little bit higher, maybe about $100 million-$200 million more. You'll see about the same amount of CapEx of approximately $600 million, focused not only on our internal engineers, but also in terms of the facilities that we need. You'll see a large amount that we'll be able to take the cash that we have on the balance sheet to execute our overall transaction for $6.9 billion. We'll continue with our capital return and finish that out as well of the use of our overall cash. Highlighting here, the title says, Our Outlook Remains Unchanged. We're in the middle of Q1.
Just to remind you that our Q1 was not necessarily about a normalized and, in terms of overall returning, back to where we believe we have in terms of the growth opportunities in front of us. $2.2 billion in overall revenue. We are still working through the excess channel inventory that we have in gaming. We indicated back in November that we thought that would take about 1-2 quarters to work through. We're on track, and we feel confident by the end of Q2 that we will be completed with our overall excess inventory that we have in the channel. You've seen the initial signs of that as we've continued to start selling in our newer platforms into the market from the 2060, the 1660, and the 1660 Ti.
Our overall gross margin for the current quarter is at 59%, which is up 300 basis points from where we just finished this last quarter as well. Our operating expenses will remain flat with last quarter. We'll see that slightly uptick in the next couple quarters as we go, but that's what you'll see to get and reach that overall growth rate for the full year. We get questions quite a bit that says, "You're often giving us overall full year guidance on overall OpEx to help steer us on something that you can definitely control." We provided our full year in terms of operating expenses, in terms of looking at high single digit overall growth over the prior year. We also took this opportunity to provide full year revenue range of overall guidance.
We look at that to be slight to flat to slightly down. The flat to slightly down was to help the teams understand what we saw in fiscal year 2019. We took this opportunity after overall cryptocurrency to find a quarter that was not tainted with cryptocurrency to come up with what we believe is a normalized run rate for overall gaming. That means we took Q2, Q3, Q4, as well as our Q1 guidance and looked at that in terms of the overall desktop business and concluded, on average, we'd look at about a $900 million quarter. On top of that, we have our overall console and notebook business, which equates to approximately $500 million. That's a $1.4 billion normalized gaming baseline for us to start.
Again, remember Q1 doesn't necessarily reflect our overall normalized as we're still working through that excess inventory. That allows us, as we move forward, to grow from this point forward. It allows us to look at the back half of the year as reaching some of the growth potential of the great opportunities that we have produced today. That's what we have in terms of our full year overall guidance for revenue. Mellanox. We're excited to announce that we have signed an agreement to acquire overall Mellanox. This is part of the overall transaction summary and the key points of that. We will purchase it for $6.9 billion in overall enterprise value. We expect this deal to close at the end of our overall calendar 2019.
Right now we will work through the overall regulatory approvals that we need in terms of in the U.S. and overall China. We're excited to bring the company on board, and we'll be working now to better understand how we'll overall integrate them forward. Again, we'll have to wait in terms of the overall regulatory approval. At the time that we close, we'll have a discussion in terms of what we expect in terms of guidance afterwards, how we will incorporate overall Mellanox in terms of our reporting structure. Okay, that was our short summary, and we are here for Q&A.
I'm gonna invite Jensen up here. We will open up, hopefully turn on the lights out here, 'cause right now it's a little dark for us to take questions from the group.
Hey, good job, Colette. I enjoy listening to my team talk. Whoa.
[crosstalk] Is it possible for us to turn on the lights so we can see their eyes? I just see. There, much better.
Thank you for the presentation. Toshiya Hari from Goldman Sachs.
Thanks. Over here. Hi, Toshiya.
Hi, hi, Jensen. In one of Jeff's slides, I think he showed the trailing five-year CAGR for the gaming business, both in terms of units as well as ASPs. It was encouraging to see the ASP number, I think it was 14% accelerate from what you had showed last year. You know, more importantly, considering all the things you guys talked about in terms of the e-sports momentum, the Max-Q initiative, the traction you've seen so far in Turing, how do you think about the next five years for that business, both in terms of units and ASPs? Related to that, does Intel's intention to reenter the market over the next couple of years impact how you think about or how does that impact your thought process, if at all? Thank you.
I'll answer the second one first. You know, we have to pay respect to all of our competition. I mean, we stay alert and we've competed with 120 graphics companies in our company's history. At one point in time, we competed against 35 at the same time. There were large companies, there were small companies. We're quite adept at competition. You're looking at a company that's incredibly focused and incredibly intense. From the leadership all the way down, there's just so much technical depth and so much passion for this business that I think we're gonna remain quite competitive.
Nonetheless, we always should stay alert. In terms of growth rate, here's the way I think about it. There's a couple of-- There's some numbers that should inform us. On the one hand, it is recognized that the PC is a gaming platform, a host for a gaming platform, and that GeForce is essentially a game console. A game console has a reasonable price point in people's head of somewhere at the end of a life, at the end, about $300, and at the beginning, around $400-$500. That's kind of a ASP in the head of a gamer. Does that make sense?
If you're a gamer and you're going out to buy a game platform to play games, in the case of a PC, because it's a good host for the game console, they can upgrade that host several times with a new game console. Every couple of years, they could buy into a new GeForce, and they can imagine paying some $300-$500, somewhere in that range, for something that delivers performances much better than a game console would. It's a very logical, sensible thing for them. That informs it. There's a couple of other ways to inform it. Unlike a game console that's largely for playing games, PCs could be used for e-sports. There's two types of people. Not two types. There are many ways that you can play sports.
You can play sports because you enjoy it, you can play sports because you wanna win. I think that another way to think about ASPs is for the people who are athletes or aspirational athletes, or they just really love to win. They need to have better gear. That's one of the reasons why you see in e-sports the high-end GPUs are like 2080 Tis. They want 2080 Tis because they wanna run, never missing a heartbeat, at 120, 150 frames per second. Many gamers can click 300 clicks a minute.
So when they click, they wanna make sure that they get a shot off before the next person. So at that kind of frame rate, you're not gonna miss a click. So this Buying the best gear is another reason for that. The rest of it is production value is increasing all the time. Max-Q increases ASPs. Max-Q increases ASP because you're using much higher-end GPUs running on a much lower voltage to deliver a great performance. Max-Q's great innovation is really about running, you know, using silicon in a way that is about running it slower at the most energy-efficient point. Max-Q increases ASPs as well.
production value increases ASP, Max-Q increases ASP, competitive gear increases ASP, those kind of factors don't play into game consoles. The game consoles, you know, that sensibility provides for me, I think, the long-term floor, the long-term floor. I don't know if these numbers all help you, but it's kinda in that space for us. Those dynamics is what's causing ASPs to grow over time. Okay?
Yeah, thanks. Aaron Rakers with Wells Fargo. Great presentation. I think one of the most interesting things that we heard is this idea of revenue sharing, this GeForce NOW Alliance. I'm curious, you know, kind of first question, how do we think about the proliferation of your partnership ecosystem, or how are you thinking about it in terms of the service providers? Can you help us understand the attributes of the revenue-sharing model, how we should think about that from a financial perspective? Then one real quick follow-up question. Any updated color on kind of your visibility on the data center side would be helpful. Thank you.
Sure. Every country has a different telco. From many countries in Eastern Europe to Western Europe to Asia, Southeast Asia, Latin America, India. This is the first time we've been able to create a game platform that could scale out to those regions, the other billion gamers. Most of the gamers we've been able to reach are in the Western and China. There are so many emerging countries that would love to have access to PC gaming. PC gaming is particularly great because it's free to play. It's social. It's easy to access. It's open. They want a PC anyways. They need a PC anyways. There are a lot of characteristics about PC gaming that makes it vibrant and unique. Using the GeForce NOW Alliance, we can reach them.
You buy a server from us, then we operate the network on top of it for you. You buy the server from us, we operate the network on top of it. We take, in terms of, relative to the When we go into a subscription model, Right now it's in beta. When we go into a subscription model, say, out of a few dollars, you know, call it $10 a month of subscription fee, maybe, they'll keep, you know, more than half, and we'll keep less than half. The reason for that is because they bought the server. They're operating it. They're running it on their network. Does that make sense? All of the capital investment is theirs.
On top of it, we're bringing the network, we're bringing the service, we're developing all the software, we're operating it for them, we're enhancing the QoS, we're onboarding all the games, we're doing all the marketing because NVIDIA is the gaming platform. They get to benefit from it as well. One of the things that's really great for them is in order to capture, Gosh, that's a terrible way of describing it. In order to win a new customer, the economic benefit, as many of you know, is quite significant lifetime. For them, this is a pretty fantastic way to differentiate their service over somebody else's service. The way that we see it is we'll probably enter into at least one of these relationships per country.
For the larger ones, maybe two or three. This is quite a scalable approach. That's one of the reasons why we built the whole stack. We could do this. Nobody else on the planet can. We built the whole system, architected the whole server, developed all the software. Everything is in one, everything is a one-shop. Of course, we've been operating the service now for a couple of years, and we're getting quite good at it. You had a second question.
Data center.
Data center visibility. Our data center business is in a grid in my mind. There's high-performance computing. There's high-performance computing. There's CSP for training, CSP for inference, CSP for cloud, and now enterprise high-performance computing. Enterprise high-performance computing, for example, data sciences. Cloud computing, all the things that we do on CUDA today. Deep learning, you know very well. Inference, Jay talked about. Last year, we kicked it off. We're doing fantastic. Supercomputing, you know very well. That's one way to think through it. You have all of the go-to markets by industries, and you overlay that across. We monitor the intersections of this for every one of those grids because how they use it and how they go to market is different.
Jay told you our way of going to market, basically several ways. One is, of course, direct sales to the cloud service providers. Second, basically a high-performance convergence, hyperconverged high-performance computing solution. We call it reference architectures. DGX POD reference architecture. We also go through the market through enterprise partners. We have all these different ways of going to market, and we just track the pipeline for each one of those. Some of those we get better visibility, and some of them we get lesser visibility. For example, last year we had a little bit less visibility in the hyperscale data center because they, in retrospect, we all realize now, they brought too much capital earlier in the year, and they had to slow down.
We didn't know about that at the time, and by the time we found out, it was well into the quarter. Some areas we have less visibility, but we try to have as much of a pipeline as we can and monitor the pipeline on a weekly basis. We feel pretty good about the year.
Yeah, John Pitzer with Credit Suisse. Jensen, thank you for the presentation.
Yeah.
Couple questions. One kind of near term, one longer term. On the near-term front, you spent a lot of time yesterday and today really focusing on your investments in software platform and ecosystem. There is one of your competitors that places a bunch of emphasis on process technology and line width nodes. Would love to hear you kind of talk about where that sits in kind of your quiver of IP, and maybe talk about the path to 7 nanometer for you. That is the near-term question. I guess longer term, last week you clearly demonstrated that you think interconnect is going to be very important going forward in data center architecture. Wondering if you can make the same sort of comments around memory, because clearly there is another one of your competitors who is looking at memory and persistent memory as perhaps a way to really lower the TCO.
How do you view that as a competitive threat, and what could you do on the memory side of things, to help out?
Yeah. You don't hear us talk about process technology, packaging technology, memory technology. The reason for that is even though we are world-class at using it, and oftentimes the earliest, for example, 3D packaging, the world's first is SXM, the largest chip that the world makes. HBM, we used it before anybody else. The reason why we don't talk about it that much is because we are just as good at buying all that stuff as anybody else. I don't find it particularly differentiating to be able to buy 7 nanometer. It's available for anybody who wants to buy it. They wanna sell it to you. That is not a point of differentiation to me. What is a point of differentiation is architecture efficiency.
For example, the fact that RTX 2080 Ti or RTX 2080 or Turing is so much more energy efficient compared to somebody's 7 nanometer GPU is shocking to everybody, but not to me. Not to me. That's the whole point. To be able to use something cost-effective so that and cost-effective, cost-efficient, and get the most architectural innovation out of it, that's what we hire our engineers for. TSMC hire their engineers for building 7 nanometer. Our job is to get the most efficiency out of any silicon that we purchase. Our goal is to be able to deliver the best energy efficiency, the best performance, the best functionality at any given point in time. Turing is just crushingly good. Just gotta measure it. It is that good. That's one of the reasons why it's off to a great start.
In terms of the data center, where you see us really differentiate is, of course, we buy all the best. We're one of the world's largest consumers of HBM2. In fact, we are the world's largest consumer of HBM2. We're the world's largest consumer of 3D packaging at TSMC, CoWoS. We ship more 3D packages than anybody. We just don't talk about it because our customers don't care. What they care about is the functionality they get, the efficiency they get, the performance they get, the TCO they ultimately get. That's what they care about, and that's what we focus on. In order to overcome the slowing Moore's Law, in order to overcome it in a dramatic way, I don't mean improve it by 10%.
If you wanna overcome it by X factors, which is what we're about, if you overcome CPUs by 10%, you might as well just wait for the next CPU. Because accelerated computing requires software optimization, you would only do so if there's an X factor in there, and I mean 10x factor, because it's a fair amount of work. That 600+ applications, all of those frameworks, all of those deep learning neural network models we now accelerate, engineers worked on it really, really hard. Ours, theirs, the ecosystems, everybody working super hard. It wasn't because of the pervasiveness of CUDA, nobody would lift their finger to do it. Now that they've done it, they want to achieve the promise that ultimately accelerated computing delivers, which is 10x, 15x, 20x, 50x. That's how you move the needle.
You can't do that by stacking up chips differently. You can't do that by just buying a special node. You can't do that by just getting memory, you know, pay a little premium for it and do that with memory. You've gotta do that only in one way, the good old-fashioned way, which is software. Rewriting, refactoring, coming up with new algorithms, good old-fashioned software. That's where the computational magic of our company is. It's well known in the industry, we have a very large team of computational mathematicians. When you see all of those breakthroughs in computer graphics, or you see all these new algorithms in the libraries that sound strange like cuGraph and cuBLAS and cuFFT, well, it turns out inside it is an enormous amount of expertise in refactoring mathematics in such a way that it's both accurate and fast.
You know, it's no different than people talking about a breakthrough in sorting algorithm. It's no different than MapReduce. MapReduce is for Hadoop, what essentially we just announced called RAPIDS for accelerated Hadoop. Think of it that way. Okay. Hadoop comes into memory. It's called Apache in-memory. On top of it is called RAPIDS. RAPIDS, the way to think about RAPIDS is essentially MapReduce except accelerated by GPUs. Well, you don't build that unless you have a great deal of, you know, computer science expertise, and that's what NVIDIA is. That's our differentiation. That's why we're not addressing a percentage share of a market someone else created. That's why our company is always talking about new markets that we're creating. Those new markets tends to be tens if not hundreds, billions dollars large industries.
You can't do that unless you go and reshape it, refactor it, come up with new algorithms. You can't build faster chips to do that alone.
Hi, Mark Lipacis, from Jefferies. Thanks a lot for the.
Yeah. Hey Mark.
Presentation. I found the accelerated computing platform framework and vision particularly compelling. But it seems like some of your customers, your biggest customers also use that same lexicon, platform and they also have lots of resources. I was wondering if you could help us maybe share with us a framework for thinking about the platform that some of your customers are developing. Is that the NVIDIA platform, is it, is your customer platform sitting on top of the NVIDIA platform, or is it sitting next to the NVIDIA platform, let's just say five or 10 years down the line? Thank you.
Yeah, excellent. Excellent. The reason for that, Mark, is, if you look across CUDA-X, two of the squares are horizontal platforms. In that case, a partner of ours, ecosystem partner, would tend to jigsaw puzzle and interweave with it. Parts of our platform will stick out. Parts of our platform will not stick out, but accelerate parts of their platform. Okay? Let me give you an example. In the case of cloud machine learning platforms like Google Machine Learning cloud or Amazon SageMaker or Azure Machine Learning. Okay? In those cases, our XGBoost library sticks all the way up to the top.
Our RAPIDS, which is essentially the modern version, accelerated version of MapReduce, goes all the way to the top. Our cuDF is basically like Pandas for one user or Spark for data centers. cuDF is basically like Spark, accelerated in the Python ecosystem. cuML is basically scikit-learn. Okay? Our platforms go all the way to the top in some cases. In many cases, like TensorFlow, our Tensor Core architecture, Tensor Core AMP, basically, and cuDNN, CUDA, cuDNN, Tensor Core AMP, it sticks into and is deeply integrated with TensorFlow. What you see is TensorFlow. It just depends. The way we come at it is this: We try to create a platform where if the ecosystem prefers another platform supplier's approach, we would integrate into theirs.
If one doesn't exist and one never will exist, for example, if we didn't write RAPIDS, the MapReduce of GPU-accelerated data centers would never exist. Nobody knows how to do it, nobody has enough body of engineers to do it, nobody has the will to go do it. It's too much work. MapReduce sitting on top of YARN, sitting on top of Hadoop is very complicated stuff. GPU accelerate that is beyond comprehension. Nobody's gonna go do it. That's why we had to go do it took about four years to go do that. The first part is, when there's a platform like data science, we integrate into it depending on how they like, okay? Google has some of our stuff sticking out.
Notice RAPIDS is now in virtual machines on the Google Cloud for their machine learning. It sits next to TensorFlow. In the case of SageMaker, some of it more of RAPIDS integrate into SageMaker, and some of it sticks out. In the case of Azure, the vast majority of it sticks out. Okay? That's one answer. The second answer is, in some vertical markets, like for example, large scale medical imaging, computational software-defined medical instruments, the future of medical imaging, multi-modality, image reconstruction, AI visualization, segmentation in 2D and 3D, multiple disease, multiple sensor modalities. We've created a platform for that because one doesn't exist on the planet. We call that Clara. We will now integrate that platform into our partners. For example, GE has their medical imaging platform. It's very, very good.
Siemens has an excellent one. Sony has parts of it. Canon has some of it. Toshiba has some of it. Philips has a lot of it. We will integrate Clara in pieces into those. We'll integrate all of it into Nuance, which is the text annotation standard practically of radiologists. Okay? That's a Clara example. We also gave you a DRIVE example. We developed a whole stack from top to bottom, end to end. Everything is open so that if somebody would like to use our simulation platform but not our physics platform, they'd rather have their own car physics simulator, for example, IPG, we're delighted to plug that in.
If somebody would like to have our visualization and our physics simulation, but they would like to have somebody else's traffic AI simulator, Cognata, we're delighted to plug that in. We create APIs all over our platform so that the ecosystem can adapt to it. The positive way of thinking about it, which is the way we think about it, is of course, we would like to enable the ecosystem to shape our platform in the way they like to use it. Okay. The benefit to us, of course, is we're a more central part of the ecosystem. If you look at the transportation ecosystem, every day that goes by more and more and more and more people have some of our stuff all over their company.
Whether they're buying our chips for the car or not buying our chips for the car, they have our development system. Sometimes they built their own development systems, but they have our chips in the car. Sometimes they have our software in the car as well. All kinds of ways of working with people. Mark, the answer is this: There's nothing more powerful than a platform of platforms. That's why the NVIDIA ecosystem is sticky. That's why the platform is sticky, because we have other people's platforms integrated with our platforms. Our platforms are also, you know, out on its own, and together we're helping the ecosystem, helping that industry move forward, okay? Simplistically, that's how.
Hi, it's Tim Arcuri at UBS. Thanks. I had two questions. First, in gaming, Jensen, if you read a lot of the websites, they sort of talk about the fact that most of the gamers that are playing triple A games, they have pretty old monitors, you know, three to five-year-old monitors. How do you think about maybe whether the display technology becomes a ramp or a gate on how fast, you know, Turing might ramp? Number one. Number two, in terms of manufacturability, you're already reticle limited on a lot of your designs. How do you think about how to combat that? Do you move to a chiplet design? You know, Intel's already sort of moving in that direction. Can you talk about that too? Thank you.
Sure. The vast majority of the world's gamers are currently at 1080p. The first thing that they wanna do is in the world, once Once the market is at any given resolution, in the case of 1080p, the first thing that they want, they get to 1080p, but then they wanna increase their frame rate within that 1080p, and they wanna increase their frame rate, then they wanna increase the beauty of the images at 1080p. Okay. Increasing their frame rate is not just about seeing it smoothly, it's about reducing latency. A hundred frames per second is much, much lower latency than 30, right?
30 frames per second is 33 milliseconds, which is quite a large number of milliseconds in the world of competitive sports. So, that's within 1080p . Once they achieve over 100 frames per second and the visual fidelity, all the options are turned on, the next thing is they would like to go to the next resolution, which is 1440p. When you go to 1440p, everything gets cut in half. Now you've got to increase your graphics processor so that you can start getting your frame rate back. Meanwhile, we just added ray tracing. We're gonna keep on making their game experience better. Every two or three years, the resolution of monitors kind of clicks up another 2x.
The next, you know, one after that is 4K, but right now people are at 1440p. I don't find that monitors are an obstacle at all because there are, as you know, I just mentioned there are four factors. There's monitor resolution, there's latency and frame rate, there's visual fidelity, then there's new features. We've got some really great new features coming. I'm sorry, your second question? I just turned 56, and it's like, boy, you know? Yeah, right. Chiclets. Chiclets. I think Chiclets are good. They're yummy. I like the orange version. We are at reticle limits. We are at reticle limits. Pascal, the P100 was near reticle limits. Volta reticle limits.
Volta is reticle limits. It is the reason why we invented NVLink, so that we could take 16 GPUs that are reticle limits and connect them all together. Okay. That's number one. There are limits to 3D fast fabrics because the data size is so gigantic in, you know, today, you can't fit it in 1 node, no matter how big that node is. We have to find a way to connect it through smart interconnect, and that's the reason why we decide-
unit growth in total. We'll probably announce that as we work through the rest of the year.
The, sorry, just the inventory flush, Q1 or Q2?
Yeah. We indicated in one to two quarters, starting back in November, that we would work through our overall inventory. That means at the end of this quarter, that's the second quarter, 1 to 2 quarters that we get through. Not Q1, Q2-
Starting back with where we started in November.
Got it. By the end of Q1, then it should be done.
That is correct.
Thank you.
Stacy, you might have an assumption that isn't quite right. Whereas or we've not been very clear. We have a GPU at every price point. One of the confusions that we created for ourselves is the price point of 1070 to 2070 . There's an impression that those two numbers should always be the same price. That somehow a BMW 5 Series would be exactly the same price over the history of time. Unfortunately, that's not possible. 1070 was higher priced than 970 . 970 was higher priced than, you know, 570 .
It's just that because people are paying so much attention now these days, nobody paid any attention to it in the past. These numbers were only numbers to us in the past. It has become numbers to society. It's a little bit like the 3 series, the 5 series, the 7s. It has gotten that kind of notoriety. I think that that was, you know, partly surprised me too. We have a GPU at every price point. There's a GPU at $299, at $399, $499, $599, just like before. At every single price point, it's way better. If people were buying the 45%, the simple answer is it couldn't have been just for all from ASPs.
It has to be from, a lot from units because our ASPs didn't go up that far.
Do you think they just wore out? [inaudible]
The simple answer is yes. I mean, the simple answer is yes. Yeah, the simple answer is yes.
Hi, Ambrish from BMO. Colette, I had a question on the Op model. As we think through gross margin for this year, I wanted to go back to, especially in context of qualitatively, you've talked about the positive impact from the crypto business. You gave us a delta between the inventory and then the 300 basis points improvement. As we think through the year, nearer term, what's the right way to think about gross margin trajectory? A little bit longer term, when you talk about the Op margins focus to get those margins up, OpEx is growing in line with what you've said consistently that you would be investing in the business. Is it going to be more on the margin front? Mix?
What's the right way to think about it? Thank you.
When you think about our gross margins, as we move forward, moving from the conversation we just had about gaming, as we look at the ranges of overall gross margin in that business, as we see people upgrade and upgrade higher into a higher overall GPU, that helps us in terms of overall margins. Additionally, when we think about, when we think about our overall data center business, we know that there's a significant amount of software that is incorporated in the platforms that we sell. Now you have an opportunity again to improve the overall gross margins as our data center business becomes more and larger percentage of our business as a whole. Let's not forget our overall automotive business.
We're continuing that transition, from just our overall infotainment business to move to AI within the cockpit, move to the overall, development services that we are working with them. Then long term, when we think about the overall production, piece of it as well. All of these things continue to change both the mix of what we're selling and overall improve our overall gross margins. Okay? Your focus in terms of on the OpEx and where we are focusing on the OpEx, is that the nature of the question?
Sure. Just more getting to your OpEx.
Meaning how do they, in terms of the till pieces. In this most current year, we take a look at this on a yearly basis to say, "What is the appropriate amount of spend?" We have great opportunities in front of us, that we need to make sure that we have properly invested in, but we have the uniqueness of that unified architecture to probably get the most out of the spend that we do in terms of OpEx. Going forward, we're not here at a model that says we look at OpEx as a percentage of revenue. It's a little bit too massive company and numbers-focused. What we actually do is look at the workloads. What is it gonna take us to get that work done?
We focus in terms of redeploying even our internal headcount towards these projects, so that we can better utilize our workforce for more greater things. Right now, I would look at we will always keep OpEx front and center as a key area of investment, but keep that in mind in terms of focusing on operating profit, in terms of how we can produce the best overall profit and leverage that we can as well. Those are the two things that we keep in mind rather than just an absolute overall OpEx or an OpEx as a percentage of revenue.
Yeah. Go ahead.
Hey, it's Matt Ramsay from Cowen.
Hey.
A couple of questions. I guess first, Jensen, you guys made a bid from Mellanox last week, and no big secret that there were a couple of other folks that were also involved in that bid, and I think you guys came at it a bit late. I wonder if you could give us a little bit of an update as to the industry and partner reaction to you attending to acquire that business and what steps you're making to keep the InfiniBand standard open. Then Colette, maybe you could talk to us a little bit about the infrastructure your group may be putting in place to monitor inventory levels across the business and across the channel.
You may be getting a little bit less granular information now from the GeForce software stack as to when GPUs are actually activated for gaming. Whatever infrastructure you put in place there to monitor inventory, an update would be helpful. Thank you.
We are super excited that Mellanox decided to accept our offer. Wow, was it competitive. The reason for that is because they're such a unique company. 20 years in the making, 100% focused on high-performance computing networks, a software stack that's integrated into high-performance computing software stacks all over the world. You know that this is. When you're building a high-performance computing system, this is a great company to work with. They have a lot of expertise. It is the only. You should also highlight that when you look at all these press releases of systems being built, they seem to be the only other company aside from us mentioned. It's actually kind of interesting.
The reason for that is because their engineers work hand-in-hand at the data centers on all the software engineering that's necessary to get the performance, lowest latency processing, the best performance, the offloading necessary. The data structures being moved around in these, in distributed computing these days is really, really complicated stuff. They're really a super special company. The customers and the industry is just delighted. They're delighted because they really feel that this important company is going to be well cared of in our hands because we understand computer architecture. This is a computer architecture question. This is a system architecture question. This is not a chip question. It's not a components question. It's an architecture question. They understand that we care about this area very much.
From the highest points of leadership all the way through this company, Mellanox knows, the industry knows this is something that we're very good at, something we care very much about, and that we're gonna continue to invest in this, and we're gonna invest in it, leveraging many of the things that our company has. They can take advantage of all that to accelerate their development. This is an area that the industry is just delighted by. We're gonna, of course, keep it open and our whole platform, as you know, is an open platform. What NVIDIA is about is creating open platforms that everybody else can build, their companies, their market, their applications, their data centers around. This is an open platform company.
The comments on the overall inventory, and our process that we have done both in terms of our inventory that we have on hand. With the sudden drop-off in terms of overall cryptocurrency, many of the work that we had started for the demand that we felt followed that had started as early as six months prior in terms of the work with our overall fabs, our works in overall purchasing, the components and the pieces that we need to put that together. At the time that Q3 and Q4 came around, and we had seen the drop-off of crypto, it became that opportunity to look through primarily just the components and the over amount of components that we had associated with there.
We feel that's a thorough process that we do from time to time, and this was even more of a thorough process to make sure we fully understood. Again, looking in hindsight, probably nothing we could do, given a lot of those purchases were done more than 6 months ago. The other focus is focus in terms of on our channel and our focus in terms of where they are in this process in terms of channel. What we've done is looked at not only just the weeks or what can we get in terms of reporting in terms of the weeks, but where they are in the life of the overall product. Where are they before it launches? Where are they at the time that it launches? Where are they six, eight weeks into it?
To assure they have the appropriate amount to both feed the market and that we have enough inventory and that we haven't gapped out, but also on the side that says, is there the right amount levels if we are a year or 2 years down. The overall cadence is continuing, but the rigor in terms of at the life cycle at any stage is probably where we've put more of the focus. Jensen, you probably have more to add here.
No, I know every chip by name now, and we have a relationship with every one of them. So I monitor all of them from birth to their next life.
Good morning. Thanks for, thanks for hosting this presentation. Harlan Sur, J.P. Morgan. We had the head of your healthcare team, Kimberly Powell, present at our healthcare conference recently. The team is doing a lot here, right? Medical imaging, patient diagnosis, drug discovery, genomics, and they're leveraging all of the systems platforms within your ProViz and data center portfolio and driving healthcare-specific platforms like Clara that you mentioned, Jensen. I know that you're targeting the platform approach across other verticals, industrial, retail, agriculture. Wondering if you can just size these vertical targeted businesses with your TAM outlook of $50 billion. Could the vertical focus represent 20%-30% of the overall $50 billion TAM? That's my first question. Second question, if you could just give us an update on China.
Have you seen demand fundamentals starting to improve with the more relaxed government stance on gaming bans?
I know the second question better. Yes. The first question, the way we do it is this: We never talk to you about TAMs until we have clear sight of it. Notice we've not one time talked to you about Clara TAM. We just assume it's zero. Until we really understand deeply, like DRIVE, and we're engaged deeply with the ecosystem, that's when we start sizing it. Otherwise, we go to zero. Industrial, we assume zero. There's no question it's not zero. There's no question it's not zero. We largely assume it's zero. Let's see. What others? Robotics, we assume zero. There's no question it's not going to be zero. There's no question it can't be zero. It will very likely be the largest AI market.
Everything is sensors literally everywhere. Temperature sensors, vibration sensors, camera sensors, microphone sensors, unfortunately sometimes, and, you know. It's gonna be everywhere. I think we assume it's largely zero until we have a really clear sight of it. Then we can talk to you about it with some amount of expertise. Until then, we just assume it's very large based on intuition. Well, most of the markets we go into in the beginning, it's all based on intuition. Let me give you an example of the intuition that led us to Clara that is very clearly the right intuition. The intuition was that in the future, healthcare, its most important instrument, imaging, medical imaging of any modality will be software defined. That was the intuition, that it will be software defined in the future.
That was spot on. Now we had that intuition about five years ago, and we started working on it. We try to not overinvest in it in the beginning so that we could do a lot of discovery work, and do some prototyping work. If you take a look at some of the early versions of Clara that I showed you, I mean, it was rickety, you know? It at least gave us the opportunity to engage with doctors and research universities all over the world and get a lot of feedback. Now, you know, we're in deployment. So that's how we. That's kinda how we do it.
10 years ago when I started working on DRIVE, the early version of it was kinda rickety. I knew that there was no question in my mind that a self-driving car was going to be a software-defined problem. You're not gonna connect 17 chips, you know, separate chips from 14 different vendors together into a, what is apparently a self-driving car. That's just not how it works. There was no question in my mind it was gonna be software defined. We just kinda take it methodically, and the timing has to be right.
There's some other things that we're working on, that I don't think the timing is quite right, and so we're we underinvested slightly and, but I keep an eye on it and dabble on it so that this company has a future, beyond what we currently describe to you, and we have a future 10-15 years out, you know, that we're working on at all times. Okay? That's the thing I really love about our company is this ability to on the one hand execute incredibly well on today's work, realize the dream for tomorrow, and start to explore the day after that, and to find the right balance of all of that. That's, you know, I just, I just love working with a management team on this.
I think we have time for one last question.
Sure. Oh boy, the pressure is high, sir.
Yeah, thanks. Mitch Steves with RBC Capital Markets. I just want to turn back to the gaming. I really had two questions. First, it's good to hear that the Turing launch has gone well for the beginning, but how do we get comfortable, I guess, around the content increases going forward without any visibility into kind of the games being made? Secondly, if I recall, about a couple of years ago, you guys used to really emphasize VR, the unit opportunity there with the ASPs would be, but I notice now it's kind of like not as topical. I'm wondering why that is and kind of what the unit opportunity's going forward since that was supposed to be somewhere around a 20/20 opportunity.
Yeah. Great. Let's see. Why am I so absolutely certain? I'm as certain about ray tracing as I am that this is the last question. I'm in total control of it. No, you said so. Simona said so. Number one, the reason for that is this. There's no question that ray tracing is the right answer because it was always the right answer. Mimicking the physical behavior of light, the physical modeling of light, is what computer graphics is all about. The issue with ray tracing was never was it the right answer. Was it the more elegant answer? Is it the simpler answer? It's all true. It's just that it just was too computationally intensive.
We found a way to use this hybrid rendering approach of some rasterization, some ray tracing, and that's what RTX means, mixed mode rasterization and ray tracing. We invented this technology, invented this approach, and we evangelized it to the ecosystem, to the world. You saw some of the things that happened. Microsoft with DXR, Vulkan RT, engines built on top of it. Epic's engine is now 4.22 is now DXR-ready, RTX-ready. Unity's next build coming out on April fourth is also RTX and DXR-ready. These are the engines of the game industry. This is the operating system of the game industry. If the engines has it and it works fast, you just use it. That's how it works. You just use it.
You don't have to invent it. You just use it. It's in the toolkit. There's no question in my mind that it's going to happen. I'm absolutely certain of it. Okay. Games keep coming out. Games keep coming out. They come out on almost monthly cadence. Several hundred games a year, as you know. Not to mention China. I mean, there's a whole bunch of games being in Korea, a bunch of games being made. There's lots of games being made. There's no question in a year's time, ray tracing will be literally everywhere. This conversation is worthwhile to capture. In a year's time, we'll come back and say, "Gosh, you were right." It was the last question. That is world-class humor, sir.
What was the second question?
VR.
Oh, VR. We don't talk about VR very much, but VR is really still very important. It's particularly we work with Microsoft on HoloLens. We because industrial design in a professional market between the two of us, we do really great work there. VR is used in industrial design all over the place, styling, architectural engineering. It is a very important part of our Quadro business. It's probably one of the reasons one of the drivers that's causing ASPs to go up. In the consumer world, I think what we really would love to have is a VR headset that is less cumbersome with less cables.
Turing has a special connector that comes out of it that it's called VirtualLink, that connects into a head mount display that reduces the amount of cables and the weight of the cable tremendously. A whole bunch of new head mount displays are coming out. It is starting to show up now, and I think you'll be surprised. I think there's no question that the experience is fantastic. The next step beyond that will likely be some form of head mount display that is VR/AR-ish and streamed from the cloud. If somebody could figure out how to stream VR from the cloud, and you might have seen some of our work in this area, some collaboration we've done with AT&T and Verizon to test NVIDIA's wireless VR from GFN, from GeForce Now.
We could stream VR directly out of the cloud. This technology is still in development. We're still very early, you know, I would say beta quality. The experience is really quite phenomenal. When you take that and you connect it up to a head mount display now that's wireless, then you have no cables at all. If it's semi-translucent, then, you know, where AR starts and VR starts and ends is gonna be quite interesting. Okay. Don't take your eyes off. Keep asking me this question. We're continuing to work on it. The ability to mix reality and virtual reality is gonna come. It's absolutely gonna come. I'm excited about it. I wanna thank all of you guys for joining us today.
GTC, this is all where it started, and it's kind of fun to sit up here and chat with you guys where it started. Now you guys know what GTC turned into. Last year, we had over 30,000 GTC attendees, and I'm looking at 200. That's fairly fast growth in a matter of 10 years. I wanna thank all of you for your support. Have a great GTC.
Lunch, if you head out the doors, turn to the left in the Gold Room, and we'll be here as well as with the executives from the company, and we'll join you for lunch. Thank you.