Nvidia Volta Presentation at GTC 2017

we've dedicated our company's lives to advancing this form of computing to enable breakthroughs in science all over the world and this is what GTC is dedicated to do well we have a lot to talk to you about today so let's get started there are two fundamental forces that's shaping the computer industry today the first is what is known as the end of Moore's law as you know very well Moore's law was propelled by three fundamental forces over the last 50 years the first is that semiconductor physics has made it possible for us to scale transistors making them smaller and smaller as a result making them faster as well as denser by making a denser we have more transistors and when we have more transistors we can apply them to improving the architecture in the design of CPUs over the years we've made CPUs data pass wider we made it hype lines deeper we invented all kinds of great technology like superscalar technology and out of order execution each and every one of those generations we applied those transistors the increased number of transistors to those design techniques meanwhile the frequency of those transistors continued to increase as a result over the course of 50 years the microprocessor to CPU increased in performance by 50 percent per year or twice every couple of years while the power of something that just naturally gets faster and faster over time without you doing anything about it is really powerful software developers develop their software once and every single year that software got faster meanwhile they load a more software onto the microprocessor as a result features got more rich so you either got you either received higher performance with the same software or more functionality at the same speed over time the computer industry advanced in that way for 50 years and all son over the last several years that freeride has come to an end transistors got so small that essentially these switches these light switches are always on to the point where the amount of leakage sub-threshold leakage this current that's going through that transistor whether it's in a state of 1 or 0 made it impossible for those transistors to continue to deliver on the performance that has been promised and relied upon for 50 years those transistors no longer make it possible for us to increase performance without increasing power dramatically meanwhile as the number of transistors continue to grow microprocessors aren't able to apply those new transistors to new design techniques we ran out of ideas as a result microprocessor architecture advancement has slowed the transistor performance no longer increased the combination of those two factors caused what is effectively at the end of Moore's law while simultaneously something else was happening simultaneously this new approach to computing called deep learning emerged into the world it's been around for quite a long time but deep learning had a really interesting handicap now first of all what enabled you to do what deep learning enables you to do is whenever you converse with a computer whenever you load your photos into into into the cloud and it magically discovers how it ought to be sorted by the vacations you've been to by the people that you you took the picture of somehow it's smart about collating grouping finding those photos for you it's the technology that's making it possible for us to develop self-driving cars and potentially detect cancer earlier that fundamental breakthrough has been around as it turns out quite a long time and in fact one of the most researchers AI researchers is was born right here in Munich jurgen schmidhuber was one of the earliest researchers to have started to work on artificial intelligence and artificial neural nets and he was one of the first to apply GPUs to deep neural nets and he was his lab was able to win the roadsign detection contest his lab also was one of the first to started to use recursive neural nets recurrent neural Nets and created this new idea called long-term short-term memory this neural net architecture that's making it possible for us to detect sequences and recognize speech that was done right here in Munich well this deep neural net has had a had a had a handicap and this handicap was in order for this deep neural net to be effective it needed an enormous amount of data it needed to Train this layer these layers and layers and layers of neurons and synapses inspired by the human brain it needed tons and tons of data to train this network otherwise it was over fit or otherwise known as ineffective well in order to train this neural net with huge amounts of data you need to do huge amounts of computation trillions and trillions and trillions of operations necessary to effectively for this new network brain to learn how to perform an intelligent task like image recognition or voice recognition while that handicap that handicap was overcome with the discovery of the GPU and so these two factors the end of Moore's law and the emergence of a new software development method a new software technique happened almost exactly the same time and these two dynamics turbocharged the adoption of GPU computing this form of computing we've been working on for literally a decade and a half in fact you could say that ever since the founding of our we focused on only this form of computing no company of our scale and size has ever dedicated itself does this one singular field of computer science GPU accelerated computing was all we do now as a company of 12,000 people we've now dedicated and invested some close to 30 billion dollars in the pursuit of advancing GPU accelerated computing well if you look at the chart if you look at the chart if we're able to turbocharge these applications and we can continue to advance Moore's law at the rate that it was promised at the rate that we had relied on basically a proxy if I could just summarize it to approximately 10 times every five years every five years it would improve by another factor of 10 every 5 years have improved by another factor of 10 in 15 years the difference the deficit of Moore's Law no longer continue to advance is 1,000 times you just got to imagine what 1,000 times give you 1,000 times of almost anything I'm going to put some of that into perspective today but 1,000 times and this is the promise of GPU computing because we do computing in a fundamentally different way and we rely on a large number of transistors not faster transistors but large number of transistors and because we optimize each and every application with a specialized function with a specialized processor and the invention of CUDA made it so that developers all over the world and all different fields of science could access this form of computing we've been able to accelerate applications in fact far faster than Moore's law while deep learning researchers discovered that and they realized that using our GPUs they could overcome the handicap that deep learning had incredible effectiveness if trained on a large amount of data which requires trillions and trillions of operations well this GPU this GPU here this is the Nvidia Volta GPU this is the most advanced processor the world's ever done imagine nearly 10,000 engineer years to design this one chip several billion dollars of R&D the largest single processor the world has ever made the Nvidia Volta 21 billion transistors all effectively put to use there's just no way to build a microprocessor that's 21 billion transistors large because there's just you run out of ideas you don't need that many more transistors the only way to make that CPU go faster is with higher clock speeds that's not true with GPU computing we could apply parallelism in a very specialized way to solve algorithms that are very important to researchers and scientists so this is Volta 21 billion transistors and a whopping staggering 120 teraflops well to put a hard xx teraflops into perspective that's effectively a hundred CPUs this right here replaces essentially a whole rack of computers and I'll illustrate that more in a moment so the end of these these two factors these two forces the end of Moore's Law and the emergence of artificial intelligence where software writes itself where software writes itself by torturing itself with huge amounts of data learning effectively from digital experience these two forces when it emerged turbocharged the adoption of NVIDIA GPU computing and the numbers show it over the last five years the number of people who've attended gtcys around the world has grown by a factor of 10 the number of GPU developers has grown to 600,000 there are 600,000 developers in the world now who are learning CUDA from the several hundred universities all over the world that teaches CUDA from the hundreds of text books that are written about it 600,000 developers can surely develop amazing software for a few billion people to consume so we've now reached critical mass in GPU computing development engineers the number of people who downloaded CUDA the number of times that it was downloaded in the last few years almost 2 million but the shocking statistic is this half of it was last year it took us 15 years to get here and literally in one year's time the number of CUDA downloads increased by a factor of two well the number of applications that we've been accelerating started out with only one in the beginning called namby molecular dynamics but over at that time if you take a look at the applications we now accelerate of course computer graphics computer graphics is a rear really rare type of application this is the highest computing demanding application in the world that is also simultaneously high-volume because it's a it's a medium for artistic expression it is a medium for one of the most popular applications on the planet videogames it is the medium by which hundreds and hundreds of millions of people enjoy their pastimes it also happens to be incredibly technologically demanding and so as a result these two combinations high computation intensity technological technological intensity and extremely high volumes came together to gave us this propellant this enormous ardi budget that allows us to fund the advancement of GPUs there's actually not another processor that I know where the computational demand is simultaneously high and the volumes are absolutely incredible that continues to propel the development of that technology well video games today not only look good but the amount of Technology inside the simulation of the virtual reality which is basically physics simulation has really benefitted many other industries the first of course is scientific computing the advancement of science all of our lives are made better for it society is made better for it and just about every single field of science that we know whether it's quantum chemistry material science fluid dynamics molecular dynamics imaging why their simulations energy discovery seismic processing just about every single field of science has now benefited from the advancement of GPUs and then of course deep learning our architecture is so pervasive it's so accessible that every single developer every single deep learning researcher has jumped onto this platform and then recently even databases have been accelerated sequel databases are going through a complete reinvention if you could access video games as fast as you can on a PC why can't you take the entire enterprises data put it into one computer and play your company's database as you would a video game and surely the developers who are working in these companies whether it's map D Connecticut blaze blazing DV screen technologies they've completely re-engineered the database now the reason why the database is so important is this if you can figure out a way to access the data and refactor to data and recreate its graph and it's interactive it's all of its relationship your ability to analyze and seek out dependencies and relationships and correlations among all of the things that's happening in the world is dramatically improved and as a result on top of these companies these amazingly fast databases a whole bunch of analytics companies are starting to emerge and whereas deep learning requires an enormous amount of data many of us don't have as much data in any particular area in our company and yet we would like to discover insights from them their traditional machine learning algorithms that have now been accelerated instead of requiring a high performance computer a supercomputer to run your machine learning algorithms whether it's clustering k-means clustering or generalize linear regressions or gradient boosting or support vector machines all of these machine learning algorithms are now being accelerated on GPUs and so from media and entertainment scientific computing deep learning which is revolutionizing just about everything that we do on the Internet to enterprise computing is an analytics the GPU is now find itself with a rich rich suite of applications that it accelerates I think there's something like 450 applications now serving just about every important industry the industries that we find ourselves in are some of the most important industries in the world high-performance computing of course internet services of course just about every time you talk to your phone just about every single query that you make every photo search that you do every recommendation that comes your way has somehow been accelerated by GPUs we're in the process of revolutionising the transportation industry we've always been part of the design and simulation part of their workflow now we're inside the car we also find ourselves solving one of the greatest problems in computing planning imagine hundreds of millions of cars hundreds of millions of people and we got to figure out which car and which person to assign them to ride hailing medical imaging being revolutionized by artificial intelligence as we speak and of course logistics trillions and trillions of dollars of Commerce are going to be recognized by artificial intelligence and robotics we find ourselves solving some of the greatest challenges in computing today solving the unsolvable for these industries where our company is all about creating tools to enable the breakthrough of scientists researchers and developers enabling breakthroughs is the ultimate purpose of our company we make tools and so you could just imagine our delight when two recent just last week announced Nobel Prize winners the work that they did the revolutionary work that they did that we were able to make a contribution to won the Nobel price of Chemistry cryogenic electron microscopy they could literally freeze the molecule in mid motion and take a picture of it now taking a picture of a molecule at atomic scale whether it's an antibody or it's a virus so that you can understand the mechanics of biology taking a picture at the atomic level you've got to imagine how incredibly tough that is and so these researchers these international researchers was able to do that using this technique called cry OEM and Crowl em now makes it possible for us to understand the biology and molecular biology at a level that nobody could ever imagine if you could see something in action imagine our ability to understand it a hundred years ago Einstein predicted that gravity was a wave well as a result of the advances of some American scientists Rayner Wiese Barry bearish and Kip Thorne they detected the world's first the for the very first time gravitational waves a disturbance in the universe measured the slightest signal of a gravitational wave as a result of two black holes colliding because of our GPUs they can reconstruct that image just about every modern instrument just about every modern instrument whether it's the astronomic or the atomic scales just about every modern instrument scientific instrument has a GPU in the back and the reason for that is this just about every form of instrument every form of measurement every form of detection we know today because we're trying to do it at astronomic and atomic scales requires computation just about every single instrument is a computational instrument today and without instruments science can't advance what you see what you see can give you a great deal of insights so from the astronaut astronomic to the atomic now we bring you back to the human scale our GPU it makes it possible for us to imagine being somewhere else well one of the things that we do of course is beautiful computer graphics and I'm about to show it to you but today we're announcing a brand-new product that's called the NVIDIA holodeck this has been a dream of ours since the beginning of computer graphics we've been pursuing this dream for a very very long time and today we've taken a very giant step I'm delighted to show it to you we combined several things in this holodeck we think of this is the design lab of his future we combined several things photorealistic computer graphics of course but once you go into a holodeck into a virtual world you would like it you would like this world to behave according to the laws of physics if you touch something you would like to know you've touched it if you drop something it should fall to the ground if you lift something it should seem heavy unless you don't want it to be heavy it should obey the laws of physics it should allow us to all be in there you should be able to come into it I should be able to come into it from anywhere we are in the world enable virtual collaboration and then lastly as we know if there's going to be virtual there if there's going to be artificial intelligence in our world there should definitely be artificial intelligence in the virtual world and so we wanted to create this world we call it the holodeck this virtual world where we can all be part of it and so why don't I show it to you now guys hey Jenson hey guys Sean you're the one in the white right yeah I'm right here in the middle and I'm here with Lars and hey now the truth of the matter is you guys aren't together this is the benefit of holodeck exactly you could be almost anywhere and we could Network into it and you can now have a shared experience in virtual reality now ladies and gentlemen what you're seeing here what you're seeing here of course is going to be a little bit limited by the projector when you get a chance during this show I would love for you guys to try a holodeck yourself the graphics that you're seeing that they're seeing the graphics that they're seen runs at 90 frames per second okay the graphics that they see runs at 90 frames per second as a result they can move at will inside this environment and feel really really natural they don't have to be in the same place they could be in different rooms they could be in different states and most of the time when we're back home and I'm enjoying holodeck with them the team is in st. Louis in another building and yet we all seem like we're together okay guys show us what you want to show us sir just because it's a holodeck of course that means we can load in any environment that our imagined nation could could bring us so we've we've created this environment to do a study of another object if we want to see something represented we can bring it anywhere and really check out the design wow that's a beautiful lab that's certainly not bad but a dominant could you could you bring the car in for us please [Music] so Jenson I've seen the inside of your garage and you're much more of a car guy than I am so so what do you think well I don't know it's pretty proud to drive the brand new McLaren that's a that's beautiful okay so because the whole model is here we just imported it straight from CAD all of the details and materials are present in the design the leather the leather a carbon fiber look at that the rubber steering wheels looks like rubber steering wheels the car paint everything is the original design the benefit of this of course we're not trying to build a video game we're trying to build a lab of the future and so the lab of the future has to seem real here's what we imagine some day you're going to go into this lab and you're going to create a product and you're gonna have a is helping you and the AI could be handing you things the AI could finish the job for you for example you might be able to design the basic shape of the car and you say you know what I want to use off-the-shelf inventory that I've got in the company and I want you to finish my car based on the inventory of parts that we currently have and then the AI goes and finishes the job comes up with the blueprints you might also work with your AI to create the factory that's going to build this car and so all the robotic arms and how their program that robotic arms are learning how to be robotic arms inside this environment and you're helping it along and so when you're done when you're done not only did you design an amazingly beautiful car you also created the entire factory necessary to manufacture the car okay so that's how that's how we imagine this as a collaboration tool as a design tool going forward wow that's just incredibly beautiful exactly as you mentioned we're able to go in and change designs and materials and really explore what this you know objects about in this case you know this McLaren is beautiful but we can also take a look inside of the car in a way that would be difficult to do with a physical product yeah take this cool tools wrong thank you so we can use this geometry clipping tool to go inside and take a peek through the engine cover into the engine itself and really explore the objects that we can see Wow and it's just a marvel of engineering and of course there's over 30,000 individual components represented in here inside of the holodeck with us and we can explore all of them individually or take them apart that's every part that's every part now the thing that's really amazing is when you're in virtual reality right now you're looking at it from a third-person view but when you're in virtual reality this car is gonna look like a city right in front of you it's just utterly shocking the idea of a clay model is gonna be utterly unnecessary and when you walk up to the car you touch it you touch it in virtual reality you're gonna feel like you've touched it because there's collision detection and with with haptics it would touch back to you okay you've made a physical contact with that car what else can you guys do so that's what we wanted to show you it's kind of a sneak peek into the Nvidia holodeck now when you guys are inside the virtual inside this virtual reality environment you don't have to be in the same place but you feel like you're in the same place right exactly I'm right here with my buds [Laughter] yeah you guys look great all right guys thank you very much good job the nvidia holodeck the original the CAD designs the original CAD designs completely photorealistic is physically simulated so it based the laws of physics you really feel like you're interacting with the environment it has virtual team collaboration it doesn't matter where anybody is you could put on your headset come into that virtual environment and you could be basically have the benefit of virtual presence and then lastly with AI integrated into the system you could create future AIS and I've got some examples I don't wanna show you later the system works basically like this you could take your original CAD designs without changing it at all as many P starts as you like the original CAD design unmodified because this isn't a video game this is a design tool you take that entire CAD database you could add all kinds of materials to it if you like you could add all kinds of materials and shaders and all kinds of other other textures to it using my in 3ds max and then you import that directly into our into the holodeck now today we only support my own 3ds max however this is a plug-in architecture and when we're intending to support literally every single CAD package in the world and on the other side of it just to illustrate that you could be literally anywhere the NVIDIA holodeck early access is now go to our website the Nvidia comm slash holodeck website and register for early access ok the NVIDIA holodeck super excited about that good job you guys [Applause] well AI is the killer app the new killer app of GPUs whereas computer graphics was the killer app that really propelled our R&D and really propelled us over the course the last 25 years we've discovered a new killer app and this new killer app as a result of deep learning is now being felt all over the world the momentum behind the adoption of deep learning is really quite staggering if you look at it from startups with just a few tens of millions of dollars of funding just a few years ago five years ago NGO has now grown to six point six billion dollars the number of papers that are published around the world on deep learning has grown to three thousand this year now three thousand the way to think about that is if one out of a hundred researchers work is worthy of being published imagine the number of people who are doing deep learning research and you could just see that in the neural network conference the International neural network conference nips is neural information processing systems basically neural network neural network artificial intelligence conference over the years look at the attendees when they register this year long before literally a month before the conference starts a general-purpose panic to register I think I think I think only David Hasselhoff in Germany can attract this level of enthusiasm huh guys David Hasselhoff he might actually be in the back I just have haven't come on as a guest so so you could just see the the adoption of deep learning and the momentum around around AI and and that the type of work that's being done just spans so many different fields of computer science and different fields of science there's a healthcare enormous amount of research being done there robotics enormous amount of research being done there natural language understanding speech recognition is taking the words that come out of my mouth the sound and translating it to text sound detect speech to text let's speech recognition but what did I say and what did I mean that's natural language understanding natural language understanding obviously is much much more complicated a very deep area of AI lots and lots of work being done there if natural language understanding could be cracked could you imagine how it's going to change the way customer service is done how we interact with computers incredible incredible opportunities the number of research work is just really shocking and people are solving these problems that we've for the longest time felt was impossible to solve and so today I want to show you a few that are just kind of fun it's easy to understand and when you think about it is there's the the oh my gosh moment the first example is done at a video research what's happening here is a form of computer graphics that's photorealistic but takes a long time it's called ray tracing we're simply following literally following a ray of photons as it bounces all over the room eventually coming into our eyes well it takes a long time to literally trace each one of those photons and that's why when you see those white dots and black dots that's when the computer graphics is not complete well we taught an artificial intelligence network how to finish the job and the way to think about that is if I showed you this incomplete picture I bet all of you could figure out what is the best color to put into that dye somehow we figured it out in our brain well we taught a computer how to do the same thing and it literally filled into noise imagine what you could do with this thing using the same autoencoder architecture you can now take an image that is very low in resolution and enhances resolution from absence of information we can create information that fundamentally breaks information theory from the absence of information we can create information which is what's happening there and so you can increase resolution if you have a photograph that's ruined parts of it was ruined we can go back and fix it just by you looking at it if you could figure it out we can probably teach at artificial intelligence how to do the same thing here's another one let's say you and your men the avocado is a pear-shaped fruit with leathery skin smooth edible flesh in a large stone and so this is what the apricot Oh is a pear-shaped fruit with leathery skin to smooth edible flesh and a large stone so this is this is what's happening if I could just get you to stop so this is what's happening imagine imagine lip-reading you learn how to look at the animation of somebody's face and you could imagine the words that is being said this is the inverse of lip-reading we had the artificial intelligence network look at a whole bunch of videos of somebody talking and from it it can now infer what is the animation of the face from the words spoken what is the animation of the face from the word spoken okay like lip-reading except backwards this one is called pose estimation pose is your 3d posture how are you posed how are you positioned in 3d space well we have to teach if you were to look at a video you can probably infer what is the pose of that character in 3d space however how do you teach an artificial intelligence network to do that how do you write software to look at a bunch of ones and zeros in a video and recognize where is the person and what is that person's three-dimensional pose from 2d to 3d pose if we were able to do that we can recreate that character in 3d world looking at the video so shining a camera on top of me I can now capture me and beam me into virtual reality so imagine the holodeck next time there'll be a video video camera just looking at me and somehow I warp into this virtual reality world wrench is doing amazing work in pose estimation this is really cool this is a 3d animation character this is a virtual reality character and based on whatever you throw in front of it whatever you throw in front of it it figures out what's the best way to animate to navigate the course you see what's happening so we've learned how to climb learn how to hop off all completely based on artificial intelligence and smart enough not to walk into that and as a tiptoes through this little tiny bridge just like we would do a virtual character taught AI and then this last one before I show it to you it's a robot that learned from an example this is Peter reveals work at UC Berkeley he goes into virtual reality he shows the robot how to stack bricks just a few times with just a few examples this is what I want you to do stack bricks and now it doesn't matter where the bricks are located robot figured it out one-shot learning imitation learning basically how we learn so this is the actual robot stacking actual bricks and this is how it learned it in virtual reality okay just some examples of what AI is able to do none of these examples have we been able to write software to overcome for the longest time just imagine writing software to solve these problems and now finally we use this artificial intelligence approach called deep learning we feed it tons and tons of information and somehow these deep neural Nets learned well the work that is being done in deep learning is literally all over the world there are just so many researchers doing deep learning work and we're just delighted that we are the deep learning platform for modern AI there are all kinds of platforms that are being used all kinds of frameworks and tools that are being used PI torch and paddle paddle of China chainer of Japan MX net that Amazon uses cafe to that that Facebook uses Microsoft's cognitive toolkits ent K theano and tensor flow from Google the number of industries that are advancing AI in research and now in startups all over the world we're working with some close to 2,000 startups all over the world in all these different industries all developing on the Nvidia platform now our strategy and our commitment is to continue to advance with all of our might the ability to develop better smarter more complex a is faster now the reason why even though we've advanced computing now advanced AI research by a factor of a hundred in the course of the last four years we've improved performance by a hundred times in four years time deep learning has been completely revolutionized because of the GPUs however we believe that we need to improve the performance by so much more and the reason for that is simply this we want to make it possible for these AIS to learn how to write software itself but to have a eyes create a eyes itself and therefore we need these massive supercomputers just churning away trying all these different types of algorithms and having software write software by itself just two weeks ago just about the entire world's computer industry adopt the Volta our latest generation GPU Volta is now available in every single cloud on the planet from Azure to Google Cloud to Amazon Web Services Alibaba Baidu $0.10 it's available everywhere anywhere on the planet as a result all of these startups that are working on in videos platform no longer need to build their own supercomputer they could go rent one in the cloud and as they grow and their workload continues to increase at some point they might decide to buy their own supercomputer and every single computer company in the world now offers the NVIDIA voltage GPU for their GPU servers from IBM HP and Dell Cisco Lenovo long way basically the NVIDIA GPU is everywhere it's everywhere in every computer every single PC and every single region available in every single cloud and so we're going to commit to continue to advance this computing platform the fact that we have one architecture makes it possible for your software and for the company that you create to be able to enjoy growth every place on earth using one architecture using one singular architecture you could develop on NVIDIA and scale on on NVIDIA all over the world okay so the world's AI platform well the next major challenge the Mex next major challenge is after developing the AI you have to run the AI developing the AI is called training running the AI running the deep neural net is called inference to infer from input some insight about the future whether it's classifying the image recognizing things recognizing voice it could be predicting what futures the Securities are going to be it could be detecting fraud somebody who is trying to enter break into your company so these deep neural nets are trained on NVIDIA platforms now the big challenge for us is to make it run incredibly efficient across all of the world's computing platforms now this is this is an area that some people call I oh t but I see the world developing in three different ways the first is that the AI network will run in the cloud and these computers are massive and scale 10 megawatts 20 megawatts data centers and they're running and inferring and basically running queries from billions and billions of people who are accessing the cloud at the same time there's another class of device way on the other side that we're actually quite familiar with today things like fitbit's nest thermostats your phone all of these little tiny devices that have a little bit of intelligence inside and that little intelligence network could be voice recognition it could be image recognition it could be recognizing your heart rate measuring your blood pressure measuring to tremor on a piece of equipment and recognizing that that tremor is gonna lead to a failure in about a week's time and so instead of waiting when that machinery goes down you might send somebody out to repair it in advance and so these this is called trail it's called IOT and I believe over time there'll be trillions of these devices and then there's a class of machines in the middle they're not exactly super computing clouds they're not intelligent sensors but they're autonomous machines these machines needs to operate and be intelligent and do smart things even when they're not connected to the cloud at all for example a self-driving car you would really really appreciate it if your car drove safely by itself even without internet connection it could be a drone that flies out to seek to find somebody who's lost and lost in a forest it could be the drone inspecting pipelines or going through and looking at looking at a force fire or it could be somehow somehow inspecting your your plot of land and all the food that you've all the farming that you're doing and looking for whether there's a disease that are broken out it could be manufacturing robots it could be these little drones that people are starting to develop for the last mile of delivery as a result of the Amazon effect where more and more of us are basically buying products online and hoping that it would deliver be deliver to your house the number of truck drivers in the world can't possibly serve all of us buying products online and hoping for it to be delivered that same time and so companies all over the world are billing these last mile delivery delivering pizzas delivering soy sauce in China delivering a hamburger to your house okay so autonomous machines all of these are going to need AI inside them and so the question is how are we going to be able to support a world where you only have a few Milind watts of power on the one hand to something that has to be completely autonomous on the other hand and then of course data centers the number of different types of devices that we have to support is exploding the networks that we're supporting is exploding starting with CN n CN NS convolutional neural networks is a neural network that magically figures out what are the important features that represent a particular object or a particular pattern and it has this ability to to learn the features by itself and to generalize generalize that information so that even if you change the color the shape stuck slightly the orientation slightly it still recognizes it to be the thing that it learned so you could change you can have a Cappy partly occluded and it still recognizes it you could have a cat changed colors wear hat and it still recognizes it the recurrent neural networks learn sequences speech etc and this is particular one is the revolution that came out recently from ian Goodfellow he created this thing called the generative adversarial Network you have one network that's trained to be a master detector you have another network that's trying to fool this master detector these two neural networks one of them generating fake fake ones the other one detecting the real ones so it could be an art piece of art for example one network generating fake art fake Picasso while the other one has been trained to recognize real Picasso and because the two of them are competing against each other one trying to foal it the other one try not to be full and as they as they learn and learn and learn and learn and learn and develop over time one becomes just amazing at recognizing Picasso's the other one becomes amazing at generating Picasso's two networks now incredibly powerful the adversarial network and then lastly the recurrent neural net the reinforcement learning neural net that was used by Peter Biel and this is basically trial and error the network is trying over and over again randomly initially every time it tries there's a value function that determines whether it's doing it well or not so well whenever it does it well it's encouraged to move in that direction if it does not so well it's encouraged to move away from that direction a value function determines that reward and Punishment recurrent neural networks is how we're going to use is the basic method for us to Train robotics explosion of the type of networks the number of and this is just a just a brief view of it there are so many different designs and so many different architectures coming out well the complexity of the network is growing as well this little tiny dot this complexity chart the size of the the area of the circle represents the number of operations times number of operations and I just simply multiply the two of them together number of operations times the amount of memory that it has to access ok GOP's times bandwidth that little tiny dot is the original alex net just a few years ago that beat every single computer scientist in the world in image recognition that's that little tiny dot that little tiny dot overnight without learning how to do computer vision learn computer vision find enough to be every single computer vision algorithm expert on the planet and one image net completely revolutionize computer vision if it wasn't because of image net the work that we're doing here in self-driving cars and robotics wouldn't have nearly progressed as fast as it has and yet look at that over the course of just four or five years it has grown than these networks have grown three hundred and fifty times this is a speech network has grown 330 times in the last several years this one's really exciting language translation I really believe that in a few more years I'll be speaking into this microphone in English and it will come out in German and it will do it completely in real time and so as a result that Star Trek universal translator is about to happen the Nero machine translation Network goes in English comes out German the network complexity has grown so much and so what we've done is this two weeks ago we announced two weeks ago we announced a new type of optimizing compiler it takes neural networks in it takes newer networks in and it uses advanced compiling technology and optimizes the network for all of the target devices that would run some of the target devices are very high performance GPUs like the one that I showed you Volta a hundred and twenty tops the 250 watts powers of data center all the way to something that's a self-driving car to a little neural network that goes into a little little robot we call Jetson just a few watts it takes input from any framework any architecture of network from any framework runs it through this optimizing compiler and it targets machines that ranges all the way from several hundred watts to several hundred milliwatts the NVIDIA tensor RT 3.0 it basically does this it takes this new neural network that's incredibly complicated this is just one layer and all the math that goes through it and we do wait an activation precision calibration first and then we fuse the layers some of the layers could be combined together some of the tensors could be combined together some of them can be unlimited all together and so we fuse the tensors we do kernel optimization with dynamic programming tracing almost every single path to figure out what is the single best path for optimizing the the speed and the size and the accuracy all at the same time and we also of course have to support multiple streams because sometimes the machine has multiple cameras well the result is pretty amazing so that's what it does and this is how well does it and so if you take the CPU we train the neural network which is done largely on NVIDIA GPUs today however once the network is trained it's largely run on CPUs today and the reason for that is because the entire world's cloud is powered by CPUs and so you train that Network the resident 50 now this is 50 layers deep resonant the CPU plus tensorflow which is the Google which is Google's framework runs at a hundred and forty images per second without tensor RT without this optimising compiler the output of tensorflow as we move it into this volta v100 this GPU here this does 300 frames per second however if you optimize it with tensor RT all of a sudden that performance just skyrockets 5700 frames per second because tensor RT optimizing optimizes for CUDA and optimizes for the architecture of our GPU takes advantage of the instructions that we put in here called tensor cook tensor core which is optimized for deep learning for open nmt neuro machine translation basically language translation for four sentences per second 225 without tensor RT - 550 but this is just one dimension of the speed-up 40 times speed-up in imaging a hundred and forty times speed-up in language but when you look at the latency how long it takes in addition first is how many you could do at the same time the other is how long does it take for you to do any one of them in the case of images we have the latency in the case of language a hundred milliseconds versus several hundred milliseconds will make all the difference in whether you can talk to a computer or not having a reasonable conversation so tensor our t3 our brand-new optimizing compiler now what does it look like in a data center this is what it looks like in a data center so suppose you had had no GPUs in a data center no tensor RT this is a hundred and sixty CPUs a hundred and sixty CPUs will do 45,000 images per second and this is basically what's happening in the cloud today whenever you ask for a photograph whenever you upload a photograph this is essentially what's happening every single one of those transfers go through some kind of a deep neural network and it's run on the CPU cluster like this in this case 65,000 watts so 1500 watts per rack for those racks about 600 thousand dollars without cables in the top of rack switches and all the power supplies about six hundred thousand dollars and this is what it looks like if you bought a GPU I know I practically fell out of my seat you see this before after before after the more GPUs you buy the more you save just think about that the more you buy the more you save 1/6 the cost 120th the power 120th the power this will improve your total cost of ownership a factor an order of magnitude before after this is the best demo in the world this is the best demo in the world this is the only demo that IT department goes here do that again our IT department goes on chess and I love that I love that I'm gonna make that a screen saver just go back and forth back and forth just like that my kids always every time like you know my kids know how cheap I am but this is this is a they call this save the money just won money save the money this is save lots of the money ok so that's what it looks like at a data center but let's take a look at that let's look at that in real time what does it really feel like ok so this is this is what inferencing is now the thing is the thing is what's really remarkable about deep neural nets is this this is what you see what the computer saw what the computer saw was a whole bunch of once a zeros in three different layers are G and B they're just a whole bunch of ones and zeros a whole mess of them and somehow from that whole mess that's not a cloud from a whole mess that's not a Beatle from a whole mess that's not a cactus and that's not a butterfly from that whole mess of ones and zeros somehow it's got to figure out what these things are and so Ryan why don't we show sir here we go okay so this is this is what's what's happened is this is inference on a CPU and you could see up there where up in the corner up in the work so so it's running on a cpu it's doing 4.8 images per second but what are these flowers though oh let's have a look that's a sword Lily okay there's a bird of paradise all right I believe it moon orchid that's a moon orchid okay that's what a daisy that's an octave a Peruvian Lily okay lots of lilies a Lenten rose a sad flower oh I bet that one smells good looks like it smells good this is a bit okay that's what it looks like a Japanese what antimony okay Ruby lipt cattle oh yeah I don't know half of these a blanket flower okay Boggan so guys the computer saw a bunch of ones and zeros and it classified these images to be those things now here's my theory 100% of you failed it I wouldn't have called any of them I would he just said flower flower flower flower that's a flower that's still a flower you see no I'm saying that's a cloud I don't know what that is that's an animal myth that's Christmas that's right and so 100% of us would have failed it and here's this computer looking at ones and zeros figuring all out utterly superhuman imagine applying this technology to medical imaging imagine applying this technology to self-driving cars imagine applying this technology to robotics we can finally write software that allows software to write itself so that it can solve the otherwise unsolvable well the problem is four images per second looks fast here but if a billion of us are using it that's gonna be a data center so large it will melt icebergs and so obviously we can't do then and so so I write show show us what what joy looks like sure that's one GPU and so this is the difference between the fastest CPU in the world versus one of these babies now that's this is a that's a fast fast car okay here in Munich here in Germany this would be a supercar right here look how fast that is I just hope it's still right and so these these GPUs that's a California Poppy I'll finally yellow iris okay so this is what it looks like on a GPU with tensor RT incredible speed up [Applause] 562 in this particular Network we sped it up 100 times this GPU has replaced 100 servers that's the shocking amount imagine all the money you're gonna save run out and get one right away ok so that's four images let's take a look at another one well and so the next one the next one is a technology that's done by based on the technology by deep brown there's startup company their startup company in Silicon Valley and and what they've done is this they've taken basically sonic information and they mapped it into a computer graphics into an image and they use CNN's use image recognition on sound information and as a result their deep learning voice recognition is not only incredibly accurate it's also incredibly fast their application is for companies who are who are with customer service or they need to make sure that all of the transcription or all the phone calls are done within regulation oh it could be in healthcare could be in finance it could be in law it could be in customer service hotels it can be retail so many different types of companies in the world that can't afford to do voice recognition in the cloud they have to do it inside their company and not only that they want to have to ability to retain the data the observation from all of the speech so that they could apply that information to other AIS to enhance their business and so the thing that deep Graham does is a voice recognition engine that is just lightning lightning fast now the thing that's really great is this and so that we created this demo it goes so fast so super real-time that you could literally watch any movie and search for words just search for the sound of those words it's gonna listen at super real-time it's gonna watch and listen the super real-time a movie and find whatever scene we need okay so this Brian's gonna show you that go ahead run so you might have got there but we're gonna show it today so this is a we've transcribed a whole bunch of Game of Thrones and now we can just go and search through it so we can find our favorite scenes no one will take my dragons this one's really cool is Monty up here hey Monty so so the folks at Columbia Cambridge Kinsella Cambridge consulting yeah Cambridge consulting created this this incredible artificial network and and I'm not T's going to tell you about in just a second but we're gonna run it so fast the inference is done so fast that this artist this trained artist is going to be able to collaborate with you so Monty first of all tell us tell us what you guys did sure so this is Vincent here it's a combination of many different deep neural network technologies using generative adversarial network technology but several of them combined those processed 8,000 masterpiece artworks from Western art from Renaissance through to current day 8,000 is nowhere near enough traditionally for deep learnings do anything useful we're going to generate megapixel images here and you would estimate we should need at least a thousand times as much but because of the generative adversarial nature this challenge that Jenson spoke about between different networks competing and improving each other we can squeeze so much more out of the data than might traditionally be viewed possible and so so just very very simply Monte and his guys has trained this network how to basically draw how to basically draw how to generate art like the great artists okay how to draw and they use this gain network to basically learn how to draw and then they used another network there's a whole there's some different networks but the other ad network that they have to create is something that would enhance the resolution of the of the drawing so that it's many megapixel so it looks beautiful okay and so why don't you show it to us oh okay I'll give you a short demonstration so if I were just to touch the pen onto the surface initially we get kind of random noise there's no possible way you can guess what I'm trying to draw there now looks like inherent in the net where that that's 8,000 artists all merge together but if I so that went for perhaps a kind of jagged landscape something like that you can see in real time as I do that it's beginning to perhaps you know coloring it's getting a hint there that the sky is unlikely to be rock coloureds I can begin to draw more detail it comes through more more clearly like that it's a very very simple picture and for example I've got a couple it's going from the yellow lines but I drew a better one earlier that looks a bit more like that and what's interesting there if I bring the lines up for you can see very very simple sketch so this isn't recognizing objects this doesn't know that it's a hasn't been trained just on landscapes if I go and do something damaging to the picture you can see it's completely changed its mind on what that is that's no longer a landscape and even if I add some kind of super real detail like that moon there we've now got more of an illustration out of it it's it's the same picture and go wait a second I could do this yeah go on alright so I I could be an artist please get going okay all right I think you can see how fast the inference is running here so this trains overnight on a DD x1 supercomputer it takes about 14 hours but it's now running you know sort of real-time frame rate inference on it on a small box here we've also trained the network with slightly different parameters so one thing we can try some slightly different styles so it looks like it's a little bit like Picasso right absolutely sign that yeah Wow JH yes right there okay there'll be a thousand dollars press the heart let me just let's let's try a few of those different networks Patties so there's something that'll see that perhaps as a lie no prints or that have seen more modern pieces Wow colorful yeah that's fantastic look at that actually effect incredibly fun okay good job Monty thank you thank the guys for me that's so amazing alright so so you saw some examples of us using deep neural networks to solve problems that otherwise would have been unsolvable in the past we we created a platform that makes it possible for researchers and developers to train their network and develop their network wherever they are it could be in the cloud it could be in their data center it could be on their computer we created this new tensor RT optimizing compiler that targets the entire range of all the processors that we make and as a result sped up deep learning inference by a factor of 50 a hundred 150 as a result saving enormous amounts of money well creating this deep Learning Network and solving unsolvable problems has led us to solving some of the most challenging and most exciting and in our belief the most transformational opportunities in the world transportation as you know is an enormous industry it's the fabric of society it makes it possible it gives us freedom makes it possible for us to get around of course to fundamental technologies to fundamental technologies has now inspired car designers all over the world to do amazing redesigns of what's possible in the future this is one example this is the work that has done an Airbus but if you could imagine the wing is autonomous the wing is completely autonomous there's an autonomous machine inside the wing as a battery-powered wing the car is autonomous and his battery it's an electric vehicle if you want to have a flying car you simply summon your wing the wing flies to you you get disconnected from the electric drivetrain the wing picks you up and holy cow imagine the experience and so so people are starting to think about these ideas people are starting to think about these ideas from the extraordinary from the extraordinary to figure out how can we revolutionize transportation for the last you know for the 50 mile range which is what this is intended to do - every single day's transportation and so what we did was we applied ourselves to create a driving Network a driving platform and create a platform by which the entire autonomous the entire transportation industry could adopt to create autonomous vehicles our driving platform consists of several layers the first layer of course is the computer the way you compute for an autonomous machine is fundamentally different second is an operating system an operating system that allows it to take information at real-time from all the different sensors and process it and algorithms and libraries we call drive works that exposes the capability of the computer to applications on top we call drive a V drive a V for autonomous vehicles and the type of algorithms and the type of applications we have to develop it's basically in three categories one you have to figure out where you are it's the perception networks where are you in the world where's everybody else around you what do you sense the second is localization based on everything that you sense you have to figure out where you are where everybody else is where they're moving to and track everything around you and then third figure out based on all of that reason about what to do you might have to stop you might have to divert to mind have to turn follow a route accelerate pass another car there's a whole lot of different actions that you could potentially take and so we created this platform this completely platform that is basically drivable several thousand engineers at Nvidia are working on this this is an open platform it is fully functional safe when it's done it is auto grade and it employs three fundamental computing approaches one is deep learning another computer vision and then the other is parallel computing we have to do these algorithms in a whole lot of different ways and the reason for that is because we want diversity these cars are going to allow us to not pay attention and still drive us safely but if something were to happen it could be a systemic problem something in the algorithm it could be a computer failure maybe it just got old maybe it broke maybe it got bumped into even if the computer failed and even if the algorithm failed we still have to have a backup that allows us to drive the passengers to safety just imagine that how do you create a computer that when it fails it still operates perfectly and so the answer of course is both having redundancy as well as diversity we have to solve the problem we have to solve the same problem in multiple ways we have to solve the same problem in multiple ways one simple example is this one way for you to avoid one way for you to avoid hitting something around you is recognizing and detecting everything around you and figuring out where they are and then simply driving where they're not ok figuring out where everything around you is and then driving where they're not the other way of solving that problem is to teach a neural network to recognize where is it safe to drive so one way is to detect what is not safe the other way is to detect what is safe if we have these two basic approaches for perception then that provides us some amount of resilience algorithms like this over and over again are stacked all over this computer so that we can provide redundancy and resilience well let me show let me show let me show you a home movie the engineers are working really hard and the car that we integrate all of this into is called bb-8 and they're driving all over California all over the East Coast there's some 20 cars driving around and we made a home home movie for you let me just let's play it now we created the holes we created the stack so that we can understand how deep learning works so that we could help help the industry solve one of the greatest challenges in the transportation industry and then this platform is completely open you could use you could use all our parts of our platform whether it's calling on the libraries that we have developing the libraries completely yourself programming directly to CUDA or using our autonomous vehicle applications that's on top of it one of the areas that is going to revolutionize society in a just an incredible way is right handling you guys know this very well the utility of a car is dramatically higher the convenience of moving around is dramatically higher more and more people are moving into the urban areas and so it's less and less convenient to own their cars and so we see we see all of us see this revolution that's happening with ride-sharing or right hailing the number of companies that are working on the next generation autonomous vehicle transportation services it's just all over the world there's just some really really exciting work that's being done here in Germany howdy the work that's done by Daimler in United States uber lyft of course way mo in China Baidu Ford with our go AI GM with cruise startup called zukes doing really great work on autonomy in Singapore and Yandex in Russia just about every single geography in the world companies are trying to figure out how to make these cars completely drive by themselves without a driver inside which basically means this in the worst case scenario if a car computer were to fail the algorithms were to fail the backup computer will have to complete the mission the backup computer will have to completely commit complete the mission even if a sensor were to break even if somebody were to bump into your rear your rear and somehow dislodged the sensor so it's no longer calibrated you still have to finish the mission because otherwise you would have to say in a very gentle voice get out you would have to tell the passengers to get out because there's nobody driving the car you can't take over there's no steering wheel and so the level of computation that's necessary to solve this problem is just enormous and so just give you a sense this is a level two sensor suite with cameras and radars and here's a level five sensor suite as cameras and radars a lot more of them because they're all completely redundant and then there's light ours that are surrounding this car this car is just packed with sensors if you think about the work that I has to do it has so much more resolution because it has to see farther it has more redundancy that has to do deep learning on it's got all these surround light ours that has to do point out point cloud processing these cameras and lighters have to localize in a redundant way they're tracking all of the objects around the card because it's likely this car is driving in urban areas and there are people crossing the street these cars also our mapping cars because it's got to update the HD map that the car is driving within as the roads change there is no driver and so every single time it has to have complete confidence that it can complete its mission has not have enough horsepower so that the path planning and the controls of the car is as smooth as possible it turns out that in the old world high-performance engines makes cars smoother smoother for acceleration smoother at idle smoother at cruising in the future the more computational performance you have the smoother your ride will be okay there's a very similar analogy there and then of course sensors and computers have to have it has to be fail operate and this computer has to have enough capacity for software to be added to this car hundreds of software engineers are developing these cars at the same time we simply can't guess accurately enough how much computational capacity is necessary and so this is what they do this is what it looks like in the trunk brand-new AI this brand-new AI is an SDK we call drive IX intelligent experience we've basically done this we've fused we fused the AI that drives the car and all the perception layers that senses what's around the car with an AI that's in the car that senses you the driver and all of the passengers in the car the sensing the perception of what's around the car and what's inside the car and some basic neural networks that we've already pre trained to track where your eyes are gazed where your head is posed read your lips understand your words all of those type of capabilities gesture recognition combined with the perception of what the car sees okay the combination of that the fusing of that is going to allow our customers to write applications that are really quite magical you'll be literally be able to walk up to the car and the car knows exactly who you are and it already adjusted the seats opens the car and if you're a passenger and knows who you are and adjust the seats and changes everything according to your desires you're walking up to the trunk because you're carrying a whole bunch of bags it recognizes that you just went shopping it's you that food is gonna end up in your trunk which now has a plate has plenty of space because of Pegasus and you're gonna put your food in a trunk okay and so all kinds of different scenarios are going to be possible and so this SDK this SDK is going to be available in early q4 and basically runs on drive our drive platform and as a result hopefully these cars becomes an AI your car is an AI okay so it's an AI card knows who you are understands what you like it recognized the situation around you recognizes that maybe you're you're dozing off and you should be maybe you're not looking in the direction of some kids playing in the street and it warns you say says it to you nicely that you should be looking to your right okay whether it's an autonomous driving mode or not in an autonomous driving mode your car AI will look out for you creating the car computer is just one of the extraordinary endeavors that these companies have to take in order to create the autonomous vehicle fleet of the future it is an act absolutely extraordinary undertaking software development like it's never been done computer deployment like it's never been done the amount of software were writing the amount of AIS we have to develop and the fleets that are going to be tested all over the world is an extraordinary undertaking and it has a lot of different pieces today I just wanted to show you what it takes to build this self-driving car not just a prototype building a prototype is fun and easy but getting ready to deploy the fleet is incredibly hard first of all you've got to collect a whole bunch of data and you have to label them just perfectly for every single country every single Road structure every single road condition you do it because it's so important and you do it because you can we have hundreds of people doing this at Nvidia you take that Network and you train it on the supercomputer that we designed the supercomputer here is called the DG x and this GPU goes in here this GPU goes in here can can we see it and then can you guys see it okay I can't see you guys seeing it and but this is one of the GPUs there eight of them here one petaflop s-- one petaflop so that's like that's like that's that's basically a whole wall of computers fit into this one NVIDIA dgx this DG X is used for training the network okay so we use it to train the network and run training on dgx but there are so many road conditions so many scenarios we simply can't collect data on because we like not - we don't want to simulate somebody jumping in front of your car at the moment there's no sense doing that and so we decided to create a 3d simulation imagine harnessing all of Nvidia's capability and photorealistic rendering and we created a simulator for the car to try all of these different scenarios that otherwise would be impossible to do we call the 3d sim on dgx I've already shown you drive a V on DP X Drive px and then lastly resem Rison is like gigantic undertaking and here's the reason why it's so important to do it every single time you update your network every single time you update your network before you put it on the road you would like to know that it's going to pass in simulation all of the roads in the world all of the roads in the world and over a period of time we're gonna collect up enough data enough videos for literally every single road in the world and in the world and so we call this super real-time simulation with D GX and using the new tensor art III this is what we can achieve with just eight Nvidia DG X's which is eight and video DG X's okay so this is a picture of our Saturn 5 supercomputer the reason why they're all intermixed like that is because each one of our D GX is consumed about 3000 watts ok that 3000 Watts of course replaces a whole wall of supercomputer of servers and so we're delighted by that but we still need a whole bunch more in order to simulate the world and so each one of the racks can consist of about 20,000 kilowatts 18,000 watts and that allows us to hold about about 6 of them and so what's going to happen is this we collect and we collect miles and miles and miles of road data and in this case 300,000 miles we can simulate on eight of those in five hours 300,000 miles represents 10% 10% of all of the paved roads and United States all of the paved roads in the United States and so basically if you could simulate that in five hours we could simulate every single paved road in the United States in basically two days so when we finally develop our software and we want to release it we can simulate the whole thing in two days now of course that's if you're perfect the whole point about simulation is to find bugs and the whole point about finding bugs is sometimes you do find them and so we have to do this over and over and over again and that's one of the reasons why we have so many of them so that we could reduce the risa Malaysian time down to essentially hours let me show it to you hey Justin listen this is um this is basically us running in real time this is running on one Drive px one Drive px is doing a really great job it ignored the wiper snowy road recognizes all the pedestrians the signs it's doing a great job detecting things detecting the road okay now let's go to suppose we want to do this on our DG axis and this is what's happening and this is how we take a brand new network the entire software stack and because our architecture exactly compatible between drive px and dry and DG x the supercomputer and the car computer is architectural II in software compatible we take the software stack we put it on DG X and we run it at super realtime okay and so so that thank you thanks guys good job and so that shows you give you a sense of what it takes to drive to create a self-driving car from the training of the neural network on a supercomputer to the simulation of extreme cases on a supercomputer to driving on a supercomputer to testing on a supercomputer that entire flow that entire flow is extraordinarily computationally intensive and so we've plumbed this entire platform and to end and it's open to our partners and developers all over the world there are now a hundred and forty five startups working on the nvidia platform they're building trucks they're building shuttles they're building cars some of them level four some of them are level five some of them - for long-haul reasons some of them for the last mile there are autonomous machines coming from just about every corner of the world everything that moves in the future will be autonomously assisted some of them will still have drivers but all of them will have AI and so hopefully it will be a lot more efficient in the long term and transportation is gonna get reinvented I mean there's just some really really cool things mapping companies that are using our platform they use our cars for mapping and they use our platform for map generation so autonomous machines the next generation of autonomous machines is super exciting let me finish by talking to you about one of the things that we're working on that's super exciting and this autonomous machine is unlike the autonomous vehicle and the reason for that is the autonomous vehicle is designed for collision avoidance autonomous machines are designed for interaction which means they largely collide collision is one of the most important things that autonomous machines do well we can't afford to these are the tennis machines learn how to interact with the world by being in the world and so there are two fundamental technologies we have to create the first is a whole new type of processor we call it project Xavier project Xavier is in the process in the process of being taped out we'll have silicon here pretty soon it fundamentally changes how computing is done it's really about real-time sensors it's really about computing in parallel using computer vision using sensor fusion using deep learning and it has to be incredibly energy efficient 30 ter ops of operations in 30 watts that's the first piece the second piece is a simulation environment we need to create a world a world a virtual world where robots can learn how to be robots they go into this world and they use reinforcement learning and they try it over and over and over and over again until they finally figured out how to be a productive and useful robot basically there are three different things we have to train the network just like we do with the self-driving car then we put that network into a virtual environment where the robot is in and that virtual environment is powered by a supercomputer so it could run at super real time and when you're done with that network then you could put it on this processor we call it Xavier Jack the AI right into the head and this robot will literally stand out and start walking that's our goal and so in order to make this happen there are two pieces of technology we have to create one is Xavier and one is Isaac we're going to show you Isaac real quick guys this is the Isaac world this is Isaac's lab and this is Isaac learning how to play hockey now you could see Isaac has two eyes and those two eyes are basically deep learning eyes so Isaac's world is accelerated AI because Isaac has to have AI to start it has to be able to recognize Percy perceived the hockey-stick perceived the puck perceived where the net is and then it's got to figure out somehow without us writing any code what to do and we gave it a value function and the value function basically says if you put the puck in the net that's good if you don't put the puck in the net try again okay and so we allow it to try and try and try and notice it's doing it right now it's learning how to it's there it's learning right there now in order to speed this up a little bit you Gordon is that you back there yeah yeah we're just going to switch to fill screen and then we're gonna switch to robot wait just trained a little bit more so the robot we just saw there it learned to basically hit the puck and no more than that and then we've got another robot which we've trained for a little bit longer and this robots learnt to hit the puck it's just further along the training program and then we can hit the puck but it's not very good at getting it in in the goal so to speed up the learning even more what we really want to do is train it in parallel and learn the aggregate experience server lots and lots of robots and so Gordon here in this case what's happening is we replicated Isaac and we let them all train and then we stopped it at some point we take the smartest Isaac we take the smartest Isaac and we put the smart as Isaac and everybody's head and then we started all over again okay you repeat rinse and repeat this over and over and over again running on top of a supercomputer and all of the Isaac's they see the puck they see the net they see they see the the they see their stick and they have to figure out how to sync it now ideally if this world this world looks like our world if this world looks like our world ideally when we're done we could literally take the AI that's in Isaac in the virtual world put it in the physical Isaac and the physical Isaac will recognize exactly what to do into it okay alright so let's uh how about one more example so this is this is sorry this is Isaac as trained as much as we as we could get he's quite good at getting and the goal he's not absolutely a hundred percent perfect but he's pretty good at playing hockey and he's much missus Wayne as much more convincing and he's pretty good yeah all right not bad good job okay so autonomous autonomous machines the next adventure for our industry and we're seeing so many so much excitement around Isaac every research organization we go to they can't wait to get their arms on Isaac so we're looking forward to releasing it in the near future okay so today we talked about several things we announced the Nvidia holodeck the design lab of the future merging fusing some of the most advanced technologies and computer graphics today photorealistic rendering virtual reality physically real simulated environments and artificial intelligence the second thing we talked about is Nvidia the AI computing platform and how we have established ourselves in AI put our AI platform all over the world in every single cloud in every single data center and recently we announced tensor art III which allows us to take these networks that were trained on these platform to run in real time with super levels of efficiency the third thing I talked about is the Nvidia drive the entire vision of drive is far more than building the chip it's the chip it's the operating system it's all the libraries and algorithms and all the applications I'm making possible for autonomous vehicles however it also includes developing the entire software stack for a fleet of cars that we would like to deploy in the near future and then third at fourth I talked about the Nvidia DRI px Pegasus the world's first computer that is able to allow all of these robot taxi developments happening all over the world to go to production and then lastly project Issac our AI world for robots so that they can learn how to be robots ladies and gentlemen I want to thank all of you for coming to GT C it's going to be a really great GT see they're looking at all the papers that are that are here this is a brand new computing era the work that we're doing is just so exciting the computer industry is now able to solve problems that historically was not solvable thank you for coming and have a great GTC you

Channel: You-In-Vision

Views: 25,626

Rating: 4.6536798 out of 5

Keywords: nvidia volta gpu presentation, nvidia volta gpu launch, nvidia volta presentation, nvidia volta presentation gtc 2017, nvidia volta presentation october, nvidia volta launch, nvidia holodeck launch, full nvidia gtc 2017 press conference 10/10/17, nvidia gtc 2017 presentation, nvidia pegasus unveil, nvidia unveils pegasus, nvidia holodeck, nvidia gtc 2017, full nvidia gtc 2017, nvidia robotic taxi ai, nvidia ai robotaxi, nvidia ai, nvidia

Id: fM4JTm9E5os

Length: 101min 45sec (6105 seconds)

Published: Tue Oct 10 2017