Earnings Call: Q2 2017

Aug 11, 2016

Good afternoon. My name is Desiree, and I'll be your conference operator today. I would like to welcome you to the NVIDIA Financial Results Conference Call. All lines have been placed on mute. After the speakers' remarks, there will be a question and answer period. I will now turn the call over to Arnab Chaddah, Vice President of Investor Relations at NVIDIA. You may begin your conference. Thank you. Good afternoon, everyone, and welcome to NVIDIA's conference call for the Q2 of fiscal 2017. With me on the call today from NVIDIA are Jensen Huang, President and Chief Executive Officer and Colette Kress, Executive Vice President and Chief Financial Officer. I'd like to remind you that today's call is being webcast live on NVIDIA's Investor Relations website. It is also being recorded. You can hear a replay by telephone until 18th August 2016. The webcast will be available for replay up until next quarter's conference call to discuss Q3 financial results. The content of today's call is NVIDIA's property. It cannot be reproduced or transcribed without our prior written consent. During the course of this call, we may make forward looking statements based on current expectations. These forward looking statements are subject to a number of significant risks and uncertainties, and our actual results may differ materially. For a discussion of factors that could affect our future financial results and business, please refer to the disclosure in today's earnings release, our most recent Forms 10 ks and 10 Q and the reports that we may file on Form 8 ks with the Securities and Exchange Commission. All of our statements are made as of today, 11th August 2016, based on information currently available to us. Except as required by law, we assume no obligation to update any such statements. During this call, we will discuss non GAAP financial measures. You can find a reconciliation of these non GAAP financial measures to GAAP financial measures in our CFO commentary, which is posted on our website. With that, let me turn the call over to Colette. Thanks, Arnab. This quarter, we introduced our new family of PASCAL based GPUs, one of our most successful launches ever. We also benefited from both the winding adoption of deep learning and our expanding engagement with hyperscale data centers around the world as they apply deep learning to all the services they provide. Revenue continued to accelerate, rising 24% to a record $1,430,000,000 We saw a strong sequential and year on year growth across our 4 platforms, gaming, professional visualization, data center and automotive. Our business model based on driving GPU compute platforms into highly targeted markets is clearly succeeding. The GPU business was up 25% to $1,200,000,000 from a year ago. The TEGRA processor business increased 30% to $166,000,000 In Q2, our 4 platforms contributed nearly 89% of revenue, up from a year earlier and 87% in the preceding quarter. They collectively increased 29% year over year. Let's begin with our gaming platform. Gaming revenue increased 18% year on year to $781,000,000 reflecting the success of our latest generation of Pascal based GPUs. Demand was strong in every geographic region. The PASCAL architecture offers a number of amazing technological advances. It enables unprecedented performance and efficiencies for playing sophisticated AAA gaming titles and driving rich immersive VR experiences. In our most successful launch ever, we introduced 4 major products. They are GeForce GTX 1080, 1070, and 1060 for the enthusiast market. And the Titan X, the world's fastest consumer GPU for deep learning development, digital content creation and extreme gaming. Wired Magazine called the GTX 1080 an unprecedented piece of electronic precision, one that performs herculean feats of computational strength. Forbes called GTX 1060, which brings a premium VR experience within reach of many, a fantastic product. And hard book Knux describes Titan X as a technological tour de force with frame rates that are simply mind boggling. The GTX 10 80, 10 70, 10 60 and Titan X are now in full production and available to consumers worldwide. VR's potential is on vivid display in a new open source game that we released during the quarter. Available on Steam, NVIDIA VR Fun Health is an open source title created with our GamesWorks SDK. It integrates physical simulation into VR along with amazing graphics and precise haptics that you feel like you're actually at a carnival. Moving to professional visualization, Quadro revenue grew to a record 214,000,000 up 22% year on year and up 13% sequentially. Growth came from the high end of the market for real time rendering tools and mobile workstations. The M6000 GPU 24 gig launched earlier this year is drawing strong interest from a broad range of customers. Digital Domain, a leading special effects studio is using Quadro to accelerate productivity for its work on films and commercials, which requires especially tight turnaround times. Engineering giant AECOM and the Yale School of Architecture are using Quadro to accelerate their design and engineering workflows. Last month at SIGGRAPH conference, we introduced a series of new products that embed photorealistic and immersive experience into workflows incorporating Iray and VR. We launched the Pascal based Quadro P6000, the most advanced workstation GPU, enabling designers to manipulate complex designs up to twice as fast as before. We demonstrated how deep learning is being brought to the realm of the industrial design to create better products faster. And we launched 8 new and updated software libraries such as VRWorks 360 video SDK, which brings panoramic video to VR. Moving to data center. Revenue reached a record $151,000,000 more than doubling year on year and up 6% sequentially. This impressive performance reflects strong growth in supercomputing, hyperscale data centers and grid virtualization. Interest in deep learning is surging as industries increasingly seek to harness this revolutionary technology. Hyperscalecompanies remain fast adopters of deep learning, both for training and real time inference, particularly for natural language processing, video and image analysis. Among them are Facebook, Microsoft, Amazon, Alibaba and Baidu. Major cloud providers are also offering GPU computing for their customers. Microsoft Azure is now using NVIDIA's GPUs to provide computing and graphics virtualization. During the quarter, we began shipping Tesla P100, the world's most advanced GPU accelerator based on the Pascal architecture. Designed to accelerate deep learning training, it allows application performance to scale up to 8 GPUs using our NVLink interconnect. We also announced a variant of P100 based on PCI Express, which makes it suitable for a wide range of accelerated servers. At our GPU Technology Conference in April, we introduced DGX-one, the world's first deep learning supercomputer. Equipped with 8 P100s in a single box, it provides deep learning performance that is equivalent to 250 traditional servers. It comes loaded with NVIDIA software and AI application developers. We are seeing strong interest in DGX-one from AR researchers and developers across academia, government labs and large enterprises. 2 days ago, Jensen hand delivered the very first DGX-one production model to the OpenAI Institute. They plan to use this system in part to build autonomous agents like chat box, cars and robots. Broader deliveries will commence later this quarter. We will be talking more about deep learning later this year at regional versions of our GPU Technology Conference set for 8 cities around the world, among them Beijing, Amsterdam, Tokyo and Seoul, as well as Washington DC. Our grid graphics virtualization business more than doubled in the quarter. Adoption is accelerating across a variety of industries, particularly automotive and AEC among customers added this quarter was Stratoil, a Norwegian oil and gas company. Finally, in automotive, revenue increased to a record $119,000,000 up 68% year over year and up 5% sequentially driven by premium infotainment and digital cockpit features in mainstream cars. Our effort to help partners develop self driving cars continues to accelerate. We have started to ship our DRIVE PX2 automotive supercomputer to the 80 plus companies using both our hardware and DRIVERX software to develop autonomous driving technology. We remain on track to ship our autopilot solution based on the DRIVE platform. Beyond our 4 platforms, our OEM and IPA business was $163,000,000 down 6% year on year in line with mainstream PC demand. Now turning to the rest of the income statement. We had a record GAAP gross margin of 57.9 percent, while non GAAP gross margin was 58.1%. These reflect the strength of our GeForce gaming GPUs, the success of our platform approach and strong demand for deep learning. GAAP operating expenses were $509,000,000 down 9% from a year earlier. Non GAAP operating expenses were $448,000,000 up 6% from a year earlier. This reflects increased hiring in R and D and marketing expenses, partially offset by lower legal fees. GAAP operating income for the Q2 was $317,000,000 compared to $76,000,000 a year earlier. Non GAAP operating income was $382,000,000 up 65%. Non GAAP operating margins improved 6.80 basis points from a year ago to 26.8%. Now turning to the outlook for the Q3 of fiscal 2017. We expect revenue to be 1,680,000,000 dollars plus or minus 2%. Our GAAP and non GAAP gross margins are expected to be 57.8% and 58% respectively, plus or minus 50 basis points. GAAP operating expenses are expected to be approximately $530,000,000 non GAAP operating expenses are expected to be approximately $465,000,000 dollars and GAAP and non GAAP tax rates for the Q3 of fiscal 2017 are both expected to be 21%, plus or minus 1%. Further financial details are included in the CFO commentary and other information available on our IR website. We will now open the call for questions. Operator, could you please poll for questions? Thank you. Operator? And your first question comes from the line of Mark Lipacat. Hi, thanks for taking my questions. First question on the data center business. Can you help us understand to what extent is the demand being driven by the deep learning applications versus the classic computationally intense design applications? Sure, Mark. Our data center business is comprised of 3 basic markets as you're alluding to. One of them is high performance computing. And it's one could say that characterize it as a traditional supercomputing market, very computationally intensive. A second market is grid, which is our data center virtualization, basically graphics application virtualization. You could stream and serve any PC or any PC application from data center to any client device. And the 3rd application is deep learning and this is largely hyperscale data centers applying deep learning to enhance their applications to make them much smarter, much more delightful. The vast majority of the growth comes from deep learning by far. And the reason for that is because high performance computing is relatively stable business, it's still growing business. And I expect the high performance computing to do quite well over the coming years. Grid is a fast growing business. I think Colette said that it was 100% year over year, but it's from a much smaller base. And deep learning is not only significant in size, it's also growing quite substantially. That's very helpful. Thank you. And then last question, on the new so you're just starting to ship PASCAL now. I guess my understanding is that historically, as you're shipping a new product, it's the yields have opportunity for improvement. The more volume you ship, the more you climb down the yield curve. What classically happens to here from the on the yield? And does that positively impact gross margins over the next 3 or 4 quarters? Thank you. Yes. So we've talked extensively about our the way we prepare for new process nodes over the last several years. And for long term NVIDIA followers, you might have recalled that 40 nanometer was a very challenging node for us. And with all of these challenges, it's an opportunity for us to improve our company. And we've implemented a very rigorous process node preparation methodology. And it starts with, of course, with some of the world's best process design engineers, circuit design engineers and process readiness teams. And we have a fantastic group dedicated to just getting process ready for us. And the second part of it is just is how that process readiness is integrated throughout the entire company. And so I'm really proud of the way that the company executed on PASCAL. 16 nanometer FinFET is no trivial task, not to mention the speed of the memories that we use. It's the world's first G5X. We also ramped the world's 1st HBM2 memory and 3 d memory stacking. So the number of technological challenges that we overcame in the ramp of PASCAL is quite extraordinary. I'm super proud of the team. Going forward, we're going to continue to refine yields and that is absolutely the case. However, we came into 16 nanometer with a great deal of preparedness. And so it's too early to guess what's going to happen to yields and margins long term. But we'll guide 1 quarter at a time. And your next question comes from the line of ritoshia hare. Hi, thank you for taking my questions and congrats on a very strong quarter. Your Q3 revenue guide implies further acceleration on a year over year basis. Are there 1 or 2 end markets where you expect outsized growth or should we expect growth in the quarter to be broad based? Yes, Toshi, I appreciate it. We are we're experiencing growth in all of our businesses. Our strategy of focusing on deep learning, self driving cars, gaming and virtual reality, where markets these are markets where GPU makes a very significant difference is really paying off. And I think this quarter was this quarter is really the Q1 where we saw growth across every single one of our businesses. And my expectation is that we're going to see growth across all of our businesses next quarter as well. But it's driven by the focus on these key markets and away from traditional commodity components businesses. I think one particular dynamic sticks out and it's a very significant growth driver where we have an extraordinary position in and it's deep learning. Deep learning, you may have heard is a new computing approach. It's a new computing model and requires a new computing architecture. And this is where the parallel approach of GPUs is perfectly suited. And 5 years ago, we started to invest in deep learning quite substantially. And we made fundamental changes and enhancements for deep learning across our entire stack of technology from the GPU architecture to the GPU design to the systems that GPUs connect into, for example, NVLink, to all the system software that has been designed for it like cuDNN and digits to all of the deep learning experts that we have now in our company. The last 5 years, we've quietly invested in deep learning because we believe that the future of deep learning is so impactful to the entire software industry, the entire computer industry that we, if you will, push it all in. And now we find ourselves at the epicenter of this very important dynamic. And it is probably this if there is one particular growth factor that is of great significance, it would be deep learning. And your next question comes from the line of Vivek Arya. Thank you for taking my question and congratulations on good growth and the execution. Jensen, the first question is tied to PC gaming, very strong trends. I was curious if you could quantify how much of your base has upgraded to PASCAL? And have you noticed any change in the behavior of gamers in this upgrade cycle, whether it's the price or what part of the stack they are buying now and how quickly they are refreshing versus what you might have seen in the Kepler and the Maxwell cycles? Sure. Thanks a lot, Vivek. Let's see on PC gaming, there's a few dynamics. Our installed base represents somewhere around 80,000,000 active GeForce users around the world. And in fact, only about a third has even upgraded to Maxwell. And we only started shipping PASCAL half of this last quarter. And so that gives you a sense of how much and PASCAL is unquestionably the biggest leap we've ever made generationally in GPUs ever. It is not only high performance, it's also energy efficient and it includes some really exciting new graphics technologies for VR and others. And so I think Pascal is going to be enormously successful for us. And it comes at a time when the PC gaming marketplace is also quite different than the PC gaming market 5 years ago. One dynamic that's really quite powerful is that the production quality, the production content is much, much higher in video games today than ever. And the reason for that, I've mentioned several times in previous calls, is that the installed base of capable game platforms that are architecturally compatible, meaning that PlayStation 4 and Xbox 1 and PCs are essentially architecturally compatible. And so the footprint for developers has grown tremendously over previous generations. I mean, this is a dynamic that's relatively new. And so as a result, the quality of games go up, which means that the consumption of GPU capability goes up with it. And I think we're absolutely seeing that dynamic. I'm super excited about the fact that the next generation game console, the big boost, the 2x boost is coming just around the corner. And that's going to allow game content providers, game developers to aim even higher. And I think that that's going to support long term expansion of our gross margins and ASPs of PC gaming. I would say that there's some other dynamics that are quite powerful as well as you know very well, which is esports is no longer just an interest. Esports is a full force global phenomenon and very, very powerful in Asia in just by every developing country and of course the Western world as well. And I think that on top of that, not only is VR off to a great start, we're seeing some great content now, But some of the things that we introduced recently with PASCAL tapping into this grassroots, but rather global interest in using video game as an art medium. We introduced project NVIDIA Ansel, which is the world's first in game photography system and allows you to create virtual reality photographs and it's just really, really amazing. And so you could use your video game, capture your amazing moments, share it in VR or in high res with all your friends. So there's a lot of different ways to enjoy games now and the production value just continues to go up, which is great for our our platforms. And so I think just to summarize your initial question, how much of the installed base has upgraded to PASCAL? Very, very small, of course, because we just started production ramp. But even then, only a third has upgraded to Maxwell. And so there's a pretty large, pretty significant upgrade opportunity ahead of us. And your next question comes from the line of Stephen Chen. Hi, thanks for taking my questions. Justin, the first one if I could on the data center competitive landscape. Earlier this week, we saw one of your data center competitors make an acquisition of a smaller private company. And I was wondering if you could talk a bit more about how you view your position in the data center market as with respect to machine learning, AI? And also kind of how your products are positioned from the high end or low end type of machine learning application performance? Sure. Thanks. Well, as you can imagine, we have a good pulse on the state of the industry. We've been in this industry since the very beginning. And deep learning was really ignited when pioneering researchers around the world discovered the use of GPUs to accelerate deep learning and made it practical made it even practical to use deep learning as an approach for developing software. The GPU was a perfect match because the nature of the GPU is a sea of small processors, not one big processor, but a whole bunch of small processors. And vitally, they're connected by this connecting tissue. This connecting tissue inside our processor, connecting memory, connecting fabric that makes it possible for the processors to communicate with each other all simultaneously. That architectural innovation has been the source of our GPU computing initiative some 10 years ago. That invention has really been groundbreaking. And so the GPU was really quite a perfect match for deep learning where neural nets are communicating neurons essentially inspired by neurons communicating with each other all simultaneously. And so the GPU was really quite a perfect match. If you look at deep learning today, 5 years later, I think it's a foregone conclusion that deep learning is being infused into just about every single Internet service to make them smarter, more intelligent, more delightful to consumers. And so you could see that the hyperscale adoption of deep learning is not only broad, it's large scale and it's completely global. And this new computing approach, we realized, was going to be quite significant long term. And so 5 years ago, we started making quite significant investments across the entire stack of our company. GPU computing is not just the GPU chip. Its GPU architecture is the GPU's design, the GPU system, all of the algorithms that run on top of it, all of the tools that run on top of it, the frameworks, our collaborations with researchers all over the world. And so that collaboration and our investment has improved deep learning on GPUs dramatically in the last two generations. When we started this, we were in Kepler. Maxwell was some 10 times better than Kepler and Pascal is some 10 times better than Maxwell. And so in just 2 generations, just 5 years' time, we've improved deep learning by an enormous amount. And a GPU today is very unlike a GPU back in the good old days because of all the work that we've done to it. And our strategy and this is where we're different, not only the focus in the GPU and the expertise in parallel computing, but where we're really different, I would say, is our singular architecture approach to deep learning. We've essentially put all of our investment behind one architecture. We've made this architecture available from hyperscale to data centers to workstations to notebooks to PCs to cars, to embedded computers, to even a brand new fully integrated high performance computer in a box, we call DGX, the NVIDIA DGX-one. And so there's so many different ways to gain access to the NVIDIA architecture, the NVIDIA platform for deep learning. Just literally all over the place all around you. It's available to you in retail stores, in these retail stores, from OEMs in the cloud or even universities all over the world just in embedded computer kits. And so our approach is quite singular and quite focused. And so my sense is that our lead is quite substantial and our position is very good, but we're not sitting on our laurels as you can tell. And for the last 5 years, we've been investing quite significantly. And so over the next several years, I think you're going to continue to see quite significant jumps from us as we continue to advance in this area. And your next question comes from the line of rahmet shah. Yes, thank you. I had a question on automotive. You mentioned that Drive PX is now shipping to 80 car companies. Jensen, I'm kind of curious, are the wins here sort of similar in size and focus more on prototyping? Or are there opportunities here that could ultimately translate into full production wins and drive the automotive business disproportionately? Well, I appreciate the question. Yes, we just started this quarter shipping Drive PX2. And just before I answer your question, let me just tell you what Drive PX2 is. Drive PX2 of course is a processor. It's the DrivePX2 version with just one single processor with just a partner and our Tegra processor and optionally with discrete GPUs, you could literally build you can build a car with autopilot capability or an AI copilot capability all the way to self driving car capability. And it's able to do sensor fusion. It's able to do SLAM, which is localization and mapping, detection using deep neural nets of the environment in a surround manner, all of the cameras around the car, all feeding into the processor and the DRIVE PX processor doing real time inferencing of surround cameras, all the way to the actual planning and driving of the car, all done in this one car computer, this one car AI supercomputer. And so this quarter we started shipping them to all of our partners and developers and so that they can start developing their software and their systems around our computer and on top of our software stack. We have the intentions of shipping in punting production many of these. And it's hard to know exactly what everybody's schedule is, but it ranges everything from very soon to the next couple of years. Developing a self driving car is no it's a fairly significant undertaking. And so nobody does it for fun, Shirley. And the question is maybe if I could frame the question just slightly differently, Do I expect people to be building OEM cars or do we expect them to be building shuttles that are maybe geo fenced? Do we expect them to be building trucks? And you know how many trucks are on the road and how much of the world's economy is built around trucking products all over the world to services of basically taxi as a service. The answer is that we're working with customers and partners across that entire range from cars that are sold to trucks to vans to shuttles to services. And your next question comes from the line of Craig Ellis. Yes, thanks for taking the question. The first is just a follow-up on some of the gaming strength in the quarter. With the company launching the founders edition availability of gaming products in the quarter, can you talk about how that went? And for those products, how gross margins compare to just chip vape sales that would go into a gaming card OEM? Well, first of all, Founders Edition, I appreciate you asking that. Founders Edition is engineered by NVIDIA, completely built by NVIDIA and sold directly by NVIDIA and supported by NVIDIA. There are some people that some pingers and customers who prefer to have a direct relationship with our company. Its availability is limited and it's engineered just at the highest possible level of quality. And we limit the production of it. And the reason for that is because we have a network of partners who are much, much more able to take the NVIDIA architecture to every corner of the world literally overnight. We have a fair number of partners who blankets every single country on the planet as we know. And so and they can provide them in different sizes and shapes and styles and different thermal solutions and different configurations and different price points. And so I think we believe that that diversity is one of the reasons why the NVIDIA GeForce platform is so popular and it creates a lot of excitement in the marketplace and a lot of interesting different diversified designs. And so, I think those two strategies are harmonious with each other. And but the key point is we built the founders edition really as a way for some customers to be able to purchase directly and have a relationship directly with us. But largely our strategy is to go to the market with a network of partners. As for gross margins, they're marginally the same. And your next question comes from the line of Matt Ramsay. Yes, good afternoon. Thank you. Jensen, I wanted to ask a couple of questions again on the data center business. The first being, we've done a little bit of work trying to estimate in our team what the long term server attach rate for accelerators in general could be and for GPUs within that. So it'd be really interesting to hear your perspectives on that. And then secondly, is there a market there for an APU type product in the data center? I know you guys have Project Denver and then some other things going on from the CPU perspective, but is there a deep learning integrated CPU GPU play that might open up more TAM long term for your company that you guys are considering pursuing? Thank you. Sure. Yes, first of all, the type of workloads in the data center has really changed. Back in the good old days, it largely ran database searches and but that has changed so much. It's no longer just about text, it's no longer just about data. The vast majority of what's going through the Internet and what's going through data centers today, as you guys know very well, are images, their voice, their increasingly and probably one of the most important new data formats is live video. Live video, if you think about it for just a moment, it's live video. So it doesn't stay in the server and it doesn't get recorded, which means that if you want to enjoy that live video, there needs to be a fair amount of artificial intelligence capability in the data center that's running real time on their live video, so that the person that might be interested in the video stream that you're streaming knows who to alert and who to invite to come and watch the live video. And so if you think about data center traffic going forward, my sense is that the workload is going to continue to be increasingly high throughput, increasingly high multimedia content and increasingly, of course, powered by AI and powered by deep learning. And so I think that's number 1. The second is that the architecture of the data center is recognizably, understandably changing because the workload is changing. Deep learning is a very different workload than the workload in the past. And so the architecture, it's a new computing model, It's recognized that it needs some new computing architecture. And accelerators, like GPUs, are really a good fit. And so now the question is how much? It's hard to say. It's hard to say how much, but my sense is that it's going to be a lot. And without any predictions, it's going to be a lot more than we currently ship. And so I think the growth opportunity for deep learning is quite significant. I think every hyperscale data center will be GPU accelerated. It'll be GPU accelerated for training. It'll be GPU accelerated for inferencing. There may be other approaches, but I think using GPUs is going to be a very large part of that. And then lastly, APUs. I guess, for data centers, I guess, my sense is for data centers, energy efficiency is such a vital part. And although the workload is increasingly AI and increasingly live video and multimedia where GPUs can add a lot of value. There's still a lot of workload that is CPU centric and you still want to have an extraordinary CPU. And I don't think anybody would argue that Intel makes the world's best CPU. It's not even close there's not even a close second. And so I think the artfulness and the craftsmanship of Intel CPUs is pretty hard to deny. And for most data centers, I think if you need if you have CPU workloads anyways, I think Intel's eons are hard to beat. And so that's my opinion anyways. And your next question comes from the line of Ian Ng. Yes. Thank you. So earlier you talked about taping out all the PASCAL products at this point. I mean, are you with 3 products in the market, are you seeding the sub $2.50 price point for cards to competition? Or is this something you can serve with older Maxwell product or some upcoming product? Thanks. Yes. Thanks, Lanyon. We have taped out, we have verified, we have ramped every Pascal GPU. That's right. However, we have not introduced everyone. And your next question comes from the line of Steve Smidgey. Great. Thanks a lot for the question. I just wanted to follow-up a little bit on virtual reality. You guys had talked a little bit about investments there. And I was just curious what sort of reception you're getting at this point and what's going to be in your mind the biggest driver getting that going? Is it more headsets or more developers working on that? Thank you. Yes, Steve. I think it's all of that. We have to continue to keep pushing VR and get the head mounts out to the world. I think HTC Vive is they're doing a great job. Oculus, of course, they're doing a great job. And so I think we track very carefully all of the head mounts that are going out there and it's growing all the time. 2nd, the content. The content is really cool and people are really enjoying it. And so we just got to get more content and developers all over the world are jumping on to VR. It really is a great new experience. But it's not just games as you know. One of the areas where we have a lot of success and we see a lot of excitement is in enterprise and industrial design, in medicine, medical imaging, in architectural engineering. We use it ourselves. We're doing a fair amount of design of our workspace and we render everything using our photorealistic renderer called Iray, fully accelerated by our GPUs and then we render it into VR and we enjoy it completely in VR. And it's something else to be inside an environment that's photorealistically rendered and completely enjoying in VR. So architectural engineering and construction is going to benefit from that. So we see a lot of broad based adoption of VR. Now one of the things that we did, which was really spectacular, is the multi resolution, multi projection rendering of Pascal. It's the world's first GPU architecture that has the ability to render into multiple projections simultaneously instead of just one. And the reason for that is because the GPU back in the good old days was designed for displaying into one display. You have one keyboard, you have one display. But that motive of computer graphics has really changed as we moved into the world of virtual reality all kinds of interesting different display configurations and display types. And so multi projection was a revolutionary approach to graphics and Pascal introduced it and you really benefited in VR. And the second thing that we did was we integrated physics real world physics simulation into VR. The benefit is that the loss of physics, as you know, you can't feel anything. Things don't collide. Things don't bounce. When you pick up something, you don't feel the haptics of it. We have made the entire environment physically simulated. And so as a result, you feel the entire environment. When you took a bottle of water over, it behaves like a bottle of water tipped over and balls behave the way balls behave and things don't merge into each other. So that integration with haptics is going to completely revolutionize VR, we believe, and that's physics simulation is another thing. And so, I think our position in VR is really quite great. And I'm super enthusiastic about the development of VR. And your next question comes from the line of Vijay Rakesh. Hi, guys. Thanks. Just on the data center side, Jensil, you mentioned 3 key segments, HPC, grid and deep learning. Could you say what percent of mix are those for the data center? I would say it's about half deep learning. So at the moment, half deep learning and probably call it 35% of a third is high performance computing, maybe more than that and the rest of it is virtualization. And going forward, which is probably your question, my sense is that deep learning would become a very significant part of that. The other thing to realize is that deep learning is not just for Internet service providers for voice recognition and image recognition and face recognition and such. Deep learning is a way of using mathematics, using software to discover insight in a huge amount of data. And the one place where we generate a huge amount of data is high performance computing. Every single supercomputing center in the world is going to move towards deep learning. And the reason for that is because they generate a huge amount of data that they really have very little ability to comb through, to sort through. And now with deep learning, they could discover really, really subtle insights in data that's hyper dimensional. And so that's the way to think about deep learning is it's really mathematics. It's a new form of mathematics that is very, very powerful. It's a new approach to software. But don't think of it as a market. I think every market is going to be a deep learning market. I think every application is going to be a deep learning application. And I think software every piece of software will be infused by AI long term. And your next question comes from the line of Harlan Sur. Good afternoon and solid job on the quarterly execution. You guys had really good growth in professional visualization and record revenues. I would have thought that most of the growth was being driven by the upcoming release of the PASCAL based P5000 and P6000 family. So I was sort of pleasantly surprised that most of the demand was driven by your current generation M6000 family, which means obviously that the PASCAL demand cycle is kind of still ahead of you. Number 1, is that a fair view? And then what's driving the strong adoption of M6000? M6000? And if you haven't already released it, when do you expect to launch the PASCAL based 5,000 and 6,000 family? Thank you. Yes. Thanks, Harlan. I appreciate the question. The team's been working really hard over the years to really change the way that computer aided design has done. Your observation is absolutely right and it's coming from several different places. First of all, more and more design is really about product design, industrial design, where the feeling of the product, the aesthetics of the product is just as important as the mechanical design of the product. And whether you're talking about a building or just a consumer product or a car, we need to be able to simulate the aesthetics of it in a photorealistic way using real material simulations. The computational load necessary to do that is just really quite extraordinary. And we're now seeing one design package after another, whether it's Dassault's leading packages, SOLIDWORKS leading packages, Autodesk, Adobe, the amount of GPU use has really, really increased and it's increasing quite dramatically. I think partly because finally for all of the ISVs, for all the developers, not only is the market demand for earlier views of photorealistic designs an important decision criteria. They can also rely on the fact that great GPUs are available just by every computer. And so the pervasiveness of GPUs allows them to take advantage of the GPUs and to really trust that the software capabilities that they put into their packages, if they rely on GPUs, will have the benefits of GPUs there. And so I think that, virtuous cycle you're starting to see, in design. And so the investment that we made in the photorealistic rendering several years ago, the GPU acceleration of optics, this layer for path tracing that is used by just by every software package in the world, our continued evangelism of GPUs and its general purpose use from computer graphics all the way to imaging is something that I think is starting to see benefits. That's number 1. Number 2, Maxwell was the most energy efficient GPU ever made until Pascal. Maxwell was twice the energy efficiency of Kepler. And the amazing thing is that Pascal is twice the energy efficiency of Maxwell. But Maxwell made it possible for thin and light designs and laptops and more elegant workstations and the ability to put more horsepower, more capability into any workstation because of power concerns. Maxwell made it possible for the entire industry to uplift the level of GPU that it uses. And I think that going forward, your last question is going forward, do we see how do we see PASCAL? PASCAL is in the process of ramping into workstations all over the world. And so I think in the coming quarters, we're going to expect to see PASCAL out there. And my expectation is that the dynamics I just described, which is software developers using more photorealistic capabilities, our invention of GPU accelerated photo rendering, Iray and optics and MDL material description language and then lastly, the energy efficiency of GPUs. Those three factors combined is going to be really healthy for workstations. And then last, VR. VR is coming and in order to really enjoy the type of applications for design, you're going to need a pretty powerful GPU to support it. And your next question comes from the line of Ross Seymore. Hi, guys. Thanks for letting me ask a question. A couple for you, Jensen, on the automotive side. I guess the first part would be, we've seen in the recent months some partnerships being formed with some of your competitors and some of your customers and we've seen some of those partnerships actually dissolve. So I wondered how does NVIDIA play in this general autonomous driving part of your automotive business would actually exceed the infotainment size of your automotive business? Thank you. Yes. Thanks a lot, Ross. Well, we play in a graceful, friendly and open way. And I mean that I guess rather seriously. We believe this. We believe that building this autonomous driving car, self driving car is a pile of software and it's really complicated software. It's really, really complicated software. And it's not like one company is going to do it. And it's also not logical that large grade companies who are refining their algorithms and the capabilities of their self driving cars over the course of the next 2 decades can outsource to someone the self driving car stack. We've always felt that self driving cars is a software problem and that large companies need to be able to own their own destiny. And that's the reason why Drive PX2 is an open stack and it's an open platform so that every car company can build their self driving car on top of it, number 1. Number 2, the Drive PH2 architecture is scalable. And the reason for that is because automatic braking and autopilot on a highway and a virtual copilot and a completely autonomous self driving car, a self driving truck, a geo fenced autonomous shuttle, all of these a completely autonomous taxi, all of these platforms cannot be solved by one chip. It's just not even logical. The computation necessary to do it is so diverse. And the more digits of accuracy or the more digits of precision towards safety that you would like to have in dealing with all of the unexpected circumstances, the more nines you would like to have, if you will, the more computation you have to do. Just as voice recognition, the amount of computation necessary for voice recognition over the last just 4 or 5 years has increased by 100 times. But notice how precise and how accurate voice recognition has become. And image recognition, video circumstance recognition, context recognition, all of that is going to require just an enormous amount of computation. And so we believe that scalable platforms is necessary, number 2, and then number 3. Detection, computer vision and detection, object detection is just one tiny sliver of the entire autonomous driving problem. It's just one tiny sliver. And we've always said that autonomous vehicles, self driving cars, is really an AI computing problem. It's a computing problem because the processors need not just detection, but also computation. The CPU matters. The GPU matters. The entire system architecture matters. And so the entire computation engine matters. Number 2, computation is computing is not just a chip problem, it's largely a software problem. And the body of software necessary for the entire system software stack, if you were, the operating system of a self driving real time computer, real time supercomputer doesn't exist. Most supercomputers are best effort supercomputers. They run a job as fast as they can until they're done. But that's not good enough for self driving cars. This supercomputer has to run-in real time and it has to react the moment it sees that there's danger in the way. And best effort doesn't cut it. You need it to be a real time supercomputer and the world's never built a real time supercomputer before. And that's what Drive PX2 is all about, a real time supercomputer for surround autonomous AI. And so that's the focus that we have. That's the direction that we've taken. And I think what you're seeing is that the market is starting to react to that, that maybe as they go further and further into autonomous driving that they're discovering that the problems are related to the type of problems that we're seeing. And that's the reason why Drive PX is a computer, not a smart camera. And your next question comes from the line of Joseph Moore. Great. Thank you so much. You talked about deep learning in the hyperscale environment, but it seems like you're getting some traction as well in the enterprise environment. I know at least one IT department we've talked to has been doing some implementation. Can you talk a little bit about your progress there and what does it take for you to sort of build that presence within more traditional enterprises? Well, as you know, deep learning is not just an Internet service approach. Deep learning is really machine learning supercharged. And deep learning is really about discovering insight in big data, in big unstructured data, in multidimensional data. And that's what deep learning that's the I've called it, it's Thor's hammer that fell from the sky. And it's amazing technology that these researchers discovered. And we were incredibly, incredibly well prepared because GPUs is naturally parallel and we put us in a position to really be able to contribute to this new computing revolution. But when you think about it in the context that it's just it's software development, it's a new method of doing software and it's a new way of discovering insight from data, What company wouldn't need it? So every medical every life sciences company needs it, every healthcare company needs it, every energy discovery company needs it, every e tail, retail company needs it. Everybody has lots of data. Everybody has lots and lots of data that they own themselves. Every manufacturing company needs it. Every company that has that cares about security, every company that deals with a massive amount of customer data, has the benefit of can benefit from deep learning. And so that when you frame it in that context, I think I would say that deep learning's market opportunity is even greater in enterprises than it is in consumer Internet services. And that's exactly the reason why we built the NVIDIA DGX-one, because most of these enterprises don't have the expertise or don't simply don't have the willpower to want to build a supercomputing data center or a high performance computer. They would just rather buy an appliance, if you will, with all of the software integrated and the performance incredibly well tuned and it comes out of a box. And that's essentially what NVIDIA DGX-one is. It's a supercomputer in a box and it's designed and tuned for high performance computing for deep learning. And your next question comes from the line of Ambrish Srivastava. Hi, thank you very much for squeezing me in. I had a question just one question on gross margin, Jensen. Very big top line guidance, but yet gross margin is guided to flat. What is the reason? And I understand it's not always perfectly correlated, margin should be going up that much. But is it pricing? Is it yield? Because the mix also seems to be moving in the right direction, more ProVis, more HPC and less of the OEM business? Well, our guidance is our best estimate. And we'll know how everything turns out next quarter when we talk again. But at some high level, I would agree with you that as we move further and further and more and more into our platform approach of business where our platform is specialized and rich with software that increasingly the value of the product that we bring has extraordinary enterprise value that the benefits of using it is not just measured in frames per second, but real TCO for companies and real cost savings as they reduce the number of server clusters and real increases and real boosts in their productivity. And so I think there's every reason to believe that long term this platform approach can derive greater value. And so but as for the next quarter, I think let's just wait and see how it goes. And your next question comes from the line of Rajendra Gill. Yes, thank you. Vindra, how are you? Right, exactly. Good. A question, Jensen, on the DRiPX2. So my understanding as you described it, it's one scalable architecture from the cockpit to ADAS to mapping to autonomous driving. But I'm curious to see how that kind of compares to the approach that some of your competitors are taking with respect to providing, I guess, different solutions for different levels of the ADAS systems, whether it's level 1, level 2, level 3, specifically with this V2X communication, where for level 4, Tom is driving, where you're going to need 6 to 20 different radar units, 3 to 6 different cameras, LiDAR. I'm just trying to square how your approach is different from some of the some of your competitors in the semiconductor space. Yes. Good question. There's no way to square and there's no reason to square. And you're not going to find one answer. And the reason why you're not going to find one answer is because nobody knows exactly how to get it done. We all have intuitions and we all have beliefs about how we're going to be able to ultimately solve the long term fully autonomous vehicle that wherever I am, the car I step into, the automobile we step into is completely autonomous and it has AI inside and out. And it's just an incredible experience. But we're not there yet. And all of these companies have slightly not all, but many companies have slightly different visions of the future. Some people believe that the path to the future is fully autonomous right away in a geofenced area that has been fully mapped in advance. Some people believe that you can use it just for highway autopilot as a first starting point and work quickly towards fully autonomy. Some people believe that the best way to do that is through shuttles and trucks. So, you see a lot of different visions out there. And I think all of those visions are coming from smart people doing smart things and they're targeting different aspects of transportation. I think there's a fallacy that transportation in every single country and every single form is exactly the same. It just doesn't work that way. And so there's technology insight and then there's market insight. And there's a technology vision versus your entry point. And I think that's where all of the squaring doesn't happen. And so you're solving for problems that you're solving for a simple equation that won't happen. However, there's one thing that we believe absolutely will happen. We are absolutely certain that AI is going to be involved in this endeavor. That finally with deep learning and finally with AI that we believe we have the secret sauce necessary to break these puzzles and to solve these puzzles over a period of time, number 1. Number 2, we believe unquestionably that depending on the problem you want to solve, you need a different amount of computational capability. We believe unambiguously this is a software problem. And that for the largest of transportation companies, they need to own their own software in collaboration with you, but they're not going to let you do it and keep it as a black box. We believe unambiguously problem, that this is a real time supercomputing problem. That is not just about a special widget, but computation is necessary, processors, a computer system, system software, enormous amount of operating system capability is necessary to build something like this. It is a massive software problem. Otherwise, we would have done it already. And so I think that you're going to see this year the beginnings of a lot of those of some very visionary and really quite exciting introductions. But in the next year and the year after that, I think you're going to see more and more and more. I think this is the beginning and we're working with some really, really amazing people to get this done. And your next question comes from the line of Mitch Hey, thanks for taking my question guys. So just kind of circling back to the data center piece and the deep learning aspect. Is there a change in ASPs you guys are seeing when you enter that market? No. So as such, there's going to be no margin change from the data center sales and I guess the same question on the automotive as well? Automotive ASPs for self driving cars will be much higher than infotainment. It's a much tougher problem. Every car in the world has infotainment. With the exception of some pioneering work or early, the best the most leading edge cars today, almost no cars are self driving. And so I think that the technology necessary for self driving cars is much, much more complicated than lane keeping or adaptive cruise control or 1st generation 1st and second generation ADAS. The problem is much, much more complicated. Your next question comes from the line of Brian Alger. Hi, guys. Thanks for squeezing me in. I think this will be the first. Congrats actually on a pretty darn good quarter and amazing guidance. I want to come back to the difference of PASCAL versus what would be otherwise competition from either Intel or AMD. There's been a fair amount of documentation talking about the power requirements or the power draw differences between Pascal versus Polaris. And one would think that while that's important in gaming and it's gotten a lot of notice, it would actually be more important for these deep learning applications that we've been talking so much about over the past half hour, 45 minutes. Can you maybe talk to that side of the design, not so much the horsepower, but maybe the power efficiency of it and what that means for when you scale it up into really big problems? Yes, Brian, thank you very much. First of all, I appreciate the comment. The team worked really, really hard and over the last several years, the last 5 years, all of the employees of NVIDIA had been pursuing a strategy that took until today really to show people that it really pays off. And it's a very unique business model. It's a very unique approach. But I just want to congratulate all the employees that have worked so hard to get us here. I appreciate the comment also about energy efficiency. In fact, energy efficiency is the single most important feature of processors today and going forward. And the reason for that is because every single environment that we're in is power constrained, every single environment. Even your PC with 7 50 watts or 1,000 watts is power constrained because we can surely put more GPUs in there than 1,000 watts. And so that's power constrained. We're in environments where we only have 1 or 2 watts. It might be a drone and we need to be we're completely covered constrained, so energy efficiency is really important. We might be in a data center where we're doing deep learning and we're training neural nets or we're inferencing neural nets. And in this particular case, although the data center has a lot of power to provision, number of GPUs that they want to use in it is measured in tens, tens and tens of thousands. And so energy efficiency becomes the predominant issue. Energy efficiency literally is the most important feature of the processor. Now from there, there are functionality and architectural features. The architectural changes that we made in PASCAL, so that we could stay ahead of the deep learning research work and the deep learning progress was groundbreaking and people are starting to discover the architectural changes that we put into PASCAL and it's going to make a huge difference in the next several years of deep learning. And so that's a feature in architectural innovation. And then lastly, of course, there's all of the software that goes on top of processor team base. We call it GPU computing instead of just GPUs because GPU computing is about computing, it's about software, it's about systems, it's about the interrelationship of our GPU with the memories and all the memories around the system and the networking and the interconnect and storage. It's a large scale computing problem. It is also the highest throughput computing problem on the planet, which is the reason why we've been called upon by our nation to build the world's next 2 fastest supercomputers. High throughput computing is our company's expertise. High throughput computing from fundamental architecture to chip design to system design to system software to algorithms to computational mathematics and all the experts in all the various fields of science. That is the great investment that we made in the last 5 years and I think the results are really starting to show. And your next question comes from the line of Blayne Curtis. Hey guys, thanks for squeezing me in here and great execution on the quarter. Two related questions. One, I just Claire, just curious your view on the return use capital and buybacks, obviously, an accelerated one, only $9,000,000 last quarter. What's your view going forward? And then Jensen, maybe a bigger question in terms of ease of capital, whether you could talk about you said CPU is not an area that you would want to go into, but obviously GPUs have legs. I was just curious if you had looked around other areas maybe in the data center where you could also add value? Yes. Thanks, Clayton. The return of capital continues to be an important part of our shareholder value message. But remember, it is still 2 parts of it. Part of it is still dividends and part of it has been our purchasing of stock. So as we continue to go forward, the dividend is definitely a long term perspective and we'll make sure that we can watch the dividend yield there to stay competitive and also looking at our profitability. Our share repurchase, we'll look at the opportunistic time for those repurchases and making sure that we're also doing that carefully as well. And long term use of capital, I would say this that what NVIDIA is really rich with is we're rich with vision and creativity and the courage to innovate. And that's one of the reasons why we start almost every conversation with anything by gathering our great people around the company and seeing what kind of future we can invent for ourselves and for the world. And so I think our use of capital is nurturing the employees that we have and a platform to innovate and creating the conditions by which they can be successful and do their life's work. And so that's, philosophically where we start. We're not allergic to acquisitions and purchases. And we look all the time and we have the benefit of working with and partnering with companies large and small all over the world as we move the industry forward. And so we're surely open to that. But our natural posture is always to invest in our people and invest in our own company's ability to invent the future. And your next question comes from the line of C. J. Muse. Yes, good afternoon. Thank you for squeezing me in. I guess two quick questions. The first one, thank you for breaking out deep learning as a percentage of the data center. Can you provide what that percentage was for the April quarter? And then the follow-up question is, I look back over the last four quarters and I look at your implied guide, you're looking at roughly 50% incremental operating margin. And curious if that's the right kind of number you would underwrite here? Or should we be thinking about improving mix as well as maturing process in manufacturing at your foundry partners such that that could actually be higher as we look ahead? Thank you. Yes. Deep learning is a software approach, a new computing architecture, a new computing approach that the industry that researchers have been developing for 20 years. And it was only until 5 years ago when pioneer work was done on deep learning on GPUs that really turbo charged it and gave the industry, if you will, a time machine that brought the future to the present. And the power of deep learning is so great that this capability is expanding and people are discovering more ways to use it and more applications and new deep learning architectures and the networks are getting bigger and deeper and more complicated. And so I think that this area is going to grow quite significantly. It represents a vast majority of our data center revenues recently and my sense is that it's going to continue to be a significant part of it. And so what was the second question? Did I miss it? I think his question was really about data centers and deep learning, right? I think your question was regarding deep learning and the percentage of data center and how is that has moved? Yes. And this vast majority, roughly this vast majority. And your next question comes from the line of Kevin Cassidy. Thanks for taking my question. Maybe I go to the other end of the spectrum and speaking of energy efficiency, are there are you finding new opportunities for Tegra aside from the infotainment in automotive? Kevin, I appreciate your question. TEGRA is at the core of all of our self driving car initiative. And so without TEGRA, there would be no self driving cars. And so Tegra is the core of our self driving car initiative. It is the computing platform for self driving cars. And DrivePX2 includes Tegra as well as discrete GPUs of Pascal. But the core of it, the vast majority of the heavy lifting is done by Tegra. And we expect that going forward. And so Tegra is incredibly important to us. Tegra is also the core of the processor of Jetson. Jetson is a platform that is designed for other embedded autonomous and intelligent machines. And so you could imagine what kind of intelligent machines in the future would benefit from deep learning and AI, but robots and drones and embedded applications inside buildings and cities, There are all kinds of applications. I'm very, very optimistic about the future of Jetson, but at the core of that is also Tegra. And so think of Tegra as our computer on a chip and it's our AI computer on a chip. Okay. I appreciate all the questions. Thank you all for joining us today. Our growth is really driven by several factors. Our focus on deep learning self driving cars, gaming and VR markets where GPU has been vital is really starting to pay off. The second factor is that Pascal is the advanced GPU ever created and we're incredibly excited about it. And we this last quarter, we ramped it with enormous success. And I'm so proud of the team for all of the preparation and the execution this last quarter. And the third is hyperscale adoption of deep learning is now widespread, is large scale and we're seeing it globally. Those are several growth drivers ahead of us. As we go forward, we're also looking to sharing our many developments in deep learning AI with you. We're really just in the beginning of seeing the actual growth of deep learning as we scale out into the market. Deep learning adoption is now widespread and is ramping at every hyperscale data center. It's a new computing model. It requires a new computing architecture, one that GPU is perfectly suited for. And the thing that we've done that I'm really delighted with is the strategy that started 5 years ago to optimize our GPU computing platform from end to end and optimize it for deep learning. At the processor level, at the architecture level, at the chip design level, systems and software and algorithms and a richness of deep learning experts at the company and the collaboration that we have all over the world with researchers and developers has made it possible for us to continue to advance this field and this platform. And as a result of that, our deep learning platform has improved far more than anybody would have expected. If you just projected it based on semiconductor physics, it would be nowhere near the level of speed up and step up that we got from generation to generation. From Kepler to Maxwell, we got 10x. From Maxwell to Pascal, we got another 10x. And you can surely expect pretty substantial improvements and increases from us over the next several years. Where we really shine is not only as a fantastic platform for deep learning and the training of the networks, but it's also a fantastic platform to scale out. You can enjoy our platform, whether it's in cloud or in data center or in supercomputers and workstations and desk side PCs and notebook computers to cars to embedded computers, as I mentioned just now with Jetson. This is a one singular architecture approach. So the thoughtfulness and the care of the investment of the developers and the software programmers and researchers is really our preeminent concern. And as we know, computing is about architecture and computing is about platform and mostly computing is about developers. And we've been quite thoughtful about the importance it is to developers. And as a result, developers all over the world, all over the industry can use the singular architecture and get the benefits of their science and their applications as they scale and deploy their work. So that's it. We had a great quarter and I look forward to reporting our progress next quarter. Thank you all for joining us. This concludes today's conference call. You may now disconnect.