원문 보기
Now, the part that's really incredible is this. This is the first AI supercomputer we made. This is in 2016 when I delivered it to a startup in San Francisco, which turned out to have been OpenAI. This was the computer. And in order to do that create that computer, we designed one chip. We designed one new chip.
In order for us to do co-design now, look at all of the chips we have to do. This is what it takes. You're not going to take one chip and make a computer 10 times faster. That's not going to happen. The way to make computers 10 times faster that we can keep increasing the performance exponentially, we can keep driving cost down exponentially, is extreme co-design and working on all these different chips at the same time.
We now have Reuben back home. This is Reuben. This is the Vera Reuben. And and uh Reuben. Ladies and gentlemen, Reuben. This is This is our third generation NVLink 72 rack-scale computer. Third generation. GB200 was the first one. All of our partners around the world, I know how hard you guys worked.
It was insanely hard. It was insanely hard to do. Second generation, so much smoother. And this generation, look at this. Completely cableless. Completely cableless. And this is This is all back in the lab now. This is the next generation Reuben. While we're shipping GB300s, uh we're preparing Reuben to be in production, you know, this time next year, maybe slightly earlier.
And so, every single year we are going to come up with the most extreme co-design system so that we can keep driving up performance and keep driving down the token generation cost. Look at this. This is just an incredibly beautiful computer. Now, the So, this is amazing. This is 100 petaflops. I know it doesn't mean anything.
100 petaflops. But, compared to the DGX-1 I delivered to OpenAI 10 years ago, 9 years ago, it's 100 times the performance. Right here versus 100 times of that supercomputer. 100 times 100 of those, let's see, 100 of those would be like 25 of these racks. All replaced by this one thing. One Vera Reuben.
Okay, so this is This is the compute tray. And this is So, Vera Reuben superchip. Okay? And this is the compute tray. This up right here. It's incredibly easy to install. Just flip these things open, shove it in. Even I could do it. Okay? And this is the Vera Vera Reuben compute tray. If you decide you wanted to add a special processor, we've added another processor.
It's called a context processor because the amount of context that we give AIs are larger and larger. We want it to read a whole bunch of PDFs before it answer a question. Want it to read a whole bunch of archive papers, watch a whole bunch of videos, go learn all this before you answer a question for me.
All of that context processing could be added. And so, you see on the bottom, eight ConnectX 9, new SuperNICs. You have CX You have uh CPXs. Eight of them. You have uh BlueField 4, this new data processor. Two Vera CPUs and four Reuben packages or eight Reuben GPUs. All of that in this one node. Completely cableless.
100% liquid cooled. And then, this new processor, I won't talk too much about it today. I don't have enough time, but this is completely revolutionary. And the reason for that is because your AIs need to have more and more memory. You're interacting with it more. You want it to remember our last conversation, everything that you've learned on my behalf.
Please don't forget it when I come back next time. And so, all of that memory is going to create this thing called KB caching. And that KB caching, retrieving it, you might have noticed every time you go into your your your AIs these days, it takes longer and longer to refresh and retrieve all of the previous conversations.
And and the reason for that is we need a revolutionary new processor, and that's called BlueField 4. Next is this the ConnectX switch. Uh excuse me, the NVLink switch, which is right here. Okay? This is the NVLink switch. This is what makes it possible for us to connect all of the computers together.
And this switch is now several times the bandwidth of the entire world's peak internet traffic. And so, that spine is going to communicate and carry all of that data simultaneously to all of the GPUs. On top of that, on top of that, this is the This is the Spectrum-X switch. And this Ethernet switch was designed so that all of the processors could talk to each other at the same time and not gum up the network.
Gum up the network. That's very technical. Okay? So, um so these are the These three combined and then this is the Quantum switch. This is for InfiniBand. This is Ethernet. We don't care what language you would like to use, whatever standard you like to use, we have great scale-out fabrics for you. I have a second announcement.
This is the single fastest enterprise enterprise company in the world. Probably the single most important enterprise stack in the world today. Palantir Ontology. Anybody from Palantir here? I was just talking to Alex earlier. [clears throat] This is Palantir Ontology. They take information, they take data, they take human judgment, and they turn it into business insight.
We work with Palantir to accelerate everything Palantir does so that we could do data processing data processing at a much, much larger scale and more speed. Whether it's structured data of the past, and of course, we'll have structured data human-recorded data, unstructured data, and process that data for our government, for national security, and for enterprises around the world, process that data at speed of light and to find insight from it.
This is what it's going to look like in the future. Palantir is going to integrate NVIDIA so that we could process at the speed of light and at extraordinary scale. Okay? NVIDIA and Palantir. NVIDIA invented a new computing model for the first time in 60 years because we wanted to solve problems that general-purpose computers, normal computers could not.
We also observed that someday transistors will continue the number of transistors will grow, but the performance and the power of transistors will slow down. That Moore's Law will not continue beyond limited by the laws of physics. And that moment has now arrived. Dennard scaling has stopped. It's called Dennard scaling.
Dennard scaling has stopped nearly a decade ago. And in fact, the transistor performance and its power associated has slowed tremendously. And yet, the number of transistor continued. We made this observation a long time ago, and for 30 years we've been advancing this form of computing we call accelerated computing.
We invented the GPU. We invented the the programming model called CUDA. And we observed that if we could add a processor that takes advantage of more and more and more transistors, apply parallel computing, add that to a sequential processing CPU, that we could extend the capabilities of computing well beyond well beyond.
And that moment has really come. Telecommunications is the backbone, the lifeblood of our economy, our industries, our national security. And yet, ever since the beginning of wireless, where we defined the technology, we defined the global standards, we exported American technology all around the world so that the world can build on top of American technology and standards.
It has been a long time since that's happened. Wireless technology around the world, largely today, deployed on foreign technologies. Our fundamental communication fabric built on foreign technologies. That has to stop. So, today we're announcing that NVIDIA has a new product line. It's called the NVIDIA Arc.
The aerial radio network computer. Aerial RAN computer, Arc. Arc is built from three fundamental new technologies. The Grace CPU, the Blackwell GPU, and our ConnectX Mellanox ConnectX networking designed for this application. We're going to create for the first time a software-defined programmable computer that's able to communicate wirelessly and do AI processing at the same time.
So, what that means is we're going to take this new technology and we'll be able to upgrade millions of base stations around the world with 6G and AI. Now, 6G and AI is really quite fundamental in the sense that for the first time, we'll be able to use AI technology to make radio communications more spectral efficient.
Doing using artificial intelligence, reinforcement learning, adjusting the beamforming in real time in context depending on the surroundings and the traffic and the mobility, the weather, all of that could be taken into account so that we could improve spectral efficiency. Spectral efficiency consumes about 1 and 1/2 to 2% of the world's power.
So, improving spectral efficiency not only improves the amount of data we can put through wireless networks without increasing the amount of energy necessary. The other thing that we could do, remember, the internet enabled communications. But, amazingly smart companies, AWS, built a cloud computing system on top of the internet.
We are now going to do the same thing on top of the wireless telecommunications network. This new cloud will be an edge industrial robotics cloud. Cloud computing will be able to go right out to the edge where data centers are not are not because we have base stations all over the world. Let's talk about quantum computing.
1981, particle physicist, quantum physicist Richard Feynman imagined a new type of computer that can simulate nature directly. To simulate nature directly because nature is quantum. He called it a quantum computer. 40 years later, the industry has made a fundamental breakthrough. It is now possible to make one logical qubit.
One logical qubit. One logical qubit that's coherent, stable, and error corrected. We now realize that it's essential for us to connect the quantum computer directly to a GPU supercomputer so that we could do the error correction, so that we could do the artificial intelligence calibration and control of the quantum computer, and so that we could do simulations collectively.
Working together, the right algorithms running on the GPUs, the right algorithms running on the QPUs, and the two processors, the two computers working side by side. This is the future of quantum computing. So, today, we're announcing the NV NVQ Link, NVQ Link. And it's made possible by two things. Of course, this interconnect that does quantum computer control and calibration, quantum error correction, as well as connects two computers, the QPU and our GPU supercomputers to do hybrid simulations.
It is also completely scalable. It doesn't just do the error correction for today's number of few qubits. It does error correction for tomorrow where we're going to essentially scale up these quantum computers from the hundreds of qubits we have today to tens of thousands of qubits, hundreds of thousands of qubits in the future.
So, we now have an architecture that can do control, co-simulation, quantum error correction, and scale into that future. We are announcing today 17 different quantum computer industry companies supporting the NVQ Link. Just about every single DOE lab has engaged us working with our ecosystem of quantum computer companies and these quantum controllers so that we could integrate quantum computing into the future of science.
Let's talk about AI. What is AI? Most people would say that AI is a chatbot, and it it's rightfully so. There's no question that ChatGPT is at the forefront of what people would consider AI. And this is a a profound understanding, a profound observation of artificial intelligence. That the software industry of the past was about creating tools.
Excel is a tool. Word is a tool. A web browser is a tool. The reason why I know these are tools is because you use them. The tools industry, just as screwdrivers and hammers, the tools industry is only so large. In the case of IT tools, they could be database tools. These IT tools, it's about a trillion dollars or so.
But, AI is not a tool. AI is work. That is the profound difference. AI is in fact workers that can actually use tools and help us be more productive. The list of opportunities go on and on, which is the reason why AI addresses the segment of the economy that IT has never addressed. Which then brings me to this one.
I just saw this literally yesterday. This is uh the CSP CapEx. People are asking me about CapEx these days. And um this is a good way to look at it. In fact, the CapEx of the top six CSPs, and this one this one is uh Amazon, CoreWeave, Google, Meta, Microsoft, and Oracle. Okay, these CSPs together are going to invest this much in CapEx, and I would I would tell you the timing couldn't be better.
But, what's happening under underneath is this. There're two platform shifts happening at the same time. One platform shift is going from general-purpose computing to accelerated computing. Remember, accelerated computing, as I mentioned to you before, it does data processing, it does image processing, computer graphics, it does com- computation of all kinds.
And so, even without AI, the world's CSPs are going to invest into acceleration. Nvidia's GPU is the only GPU that can do all of that plus AI. An ASIC might be able to do AI, but it can't do any of the others. Nvidia could do all of that. This is what's going on in our company's business. We're seeing extraordinary growth for Grace Blackwell for all the reasons that I just mentioned.
It's driven by two exponentials. We now have visibility. I think we're probably the first technology company in history to have visibility into half a trillion dollars of cumulative Blackwell and early ramps of Rubin through 2026. And as you know, 2025 is not over yet, and 2026 hasn't started. This is how much business is on the books.
Half a trillion dollars worth so far. Now, this is out of that, we've already shipped 6 million of the Blackwells in the first several quarters. I guess the first four quarters of production, three and a half quarters of production. We still have one more quarter to go for 2025, and then we have four quarters.
So, the next five quarters, there's 500 million 500 billion dollars, half a trillion dollars. That's five times the growth rate of Hopper. Let's talk about models. Open-source models particularly. In the last couple years, several things have happened. One, open-source models have become quite capable because of reasoning capabilities.
It has become quite capable because they're multimodality, and they're incredibly efficient because of distillation. So, all these different capabilities have become uh has made open-source models for the very first time incredibly useful for developers. They are now the lifeblood of startups. Lifeblood of startups in different industries because obviously, as I mentioned before, each one of the industries have its own use case, its own use cases, its own data, its own use data, its own flywheels.
All of that capability, that domain expertise, needs to have the ability to embed into a model. Open source makes that possible. Researchers need open source. Developers need open source. Companies around the world, we need open source. Open-source models is really, really important. The United States has to lead in open source as well.
We have amazing proprietary models. We have amazing proprietary models. We need also amazing open-source models. Our country depends on it. Our startups depend on it. And so, Nvidia is dedicating ourselves to go do that. We are now the largest the largest, we lead in open source contribution. We have 23 models in leaderboards.
We have all these different domains from language models to physical AI models I'm going to talk about, robotics models to biological biology models. Each one of these models has enormous teams, and that's one of the reasons why we built supercomputers for ourselves to enable all these models to be created.
We have number one speech model, number one reasoning model, number one physical AI model. The number of downloads is really, really terrific. We are dedicated to this, and the reason for that is because science needs it, researchers need it, startups need it, and companies need it. Let's talk about physical AI.
Physical AI requires three computers. Just as it takes two computers to train a language model, one that's to train it, evaluate it, and then inference it. okay? So, that's the large GB200 that you see. In order to do it for physical AI, you need three computers. You need the computer to train it. This is GB, the Grace Black NVLink 72.
We need a computer that does all of the simulations that I showed you earlier with Omniverse DSX. It basically is a digital twin for the robot to learn how to be a good robot and for the factory to essentially be a digital twin. That computer is the second computer, the Omniverse computer. This computer has to be incredibly good at generative AI and it has to be good at computer graphics, sensor simulation, ray tracing, signal processing.
This computer is called the Omniverse computer. And once we train the model, simulate that AI inside a digital twin, and that digital twin could be a digital twin of a factory as long as well as a whole bunch of digital twins of robots, then you need to operate that robot. And this is the robotic computer.
This is This one goes into a self-driving car. Half of it could go into a robot, okay? Or you could actually have, you know, robots that are quite agile and quite quite fast in operations and it might take two of these computers. And so, this is the Thor Jetson Thor robotics computer. These three computers all run CUDA and it makes it possible for us to advance physical AI.
The factory is essentially a robot that's orchestrating robots to build things that are robotic. You know, this is the amount of software necessary to do this is so intense that unless you could do it inside a digital twin to dis- to plan it, to design it, to operate it inside a digital twin, the hopes of getting this to work is nearly impossible.
Um these factories will have future robotic systems. And one of the most advanced is Figure. Brett Adcock is here today. He's just He founded a company 3 and 1/2 years ago. They're worth almost $40 billion today. We're working together in training the the AI training the robot, simulating the robot, and of course, the robotic computer that goes into Figure.
Really amazing. Uh I had the benefit of seeing it. Uh it's really quite quite extraordinary. It is very likely that humanoid robots and and uh my friend Elon is also working on this, that this is likely going to be one of the largest consumer new consumer electronics markets and surely one of the largest industrial equipment market.
And of course, the cutest robot ever, the cutest robot ever, the Disney robot. And this is this is um uh something really close to our heart. We're working with Disney research on an entirely new framework and simu- simulation platform uh based on revolutionary technology called Newton. And that Newton uh simulator makes it possible for the the robot to learn how to be a good robot inside a physically aware, physically based environment.
So, these digital twins of factories, digital twins of warehouses, digital twins of surgical rooms, digital twins where Blue could learn how to manipulate and navigate and, you know, interact with the world, all completely done in real time. This is going to be the largest consumer electronics product line in the world.
Some of them are just really working incredibly well now. This is the future of humanoid robotics and of course, Blue. Okay? Now, humanoid robotics is still in development, but meanwhile, there's one robot that is clearly at an inflection point and it is basically here. And that is a robot on wheels.
This is a robo-taxi. A robo-taxi is essentially an AI chauffeur. Now, one of the things that we're doing today, we're announcing the NVIDIA Drive Hyperion. This is a big deal. [snorts] We created this architecture so that every car company in the world could create cars, vehicles could be commercial, could be passenger, could be dedicated to robo-taxi, create vehicles that are robo-taxi ready.
The sensor suite with surround cameras and radars and lidar make it possible for us to achieve the highest level of surround cocoon sensor perception and redundancy necessary for the highest level of safety. Hyperion Drive, Drive Hyperion is now designed into Lucid, Mercedes-Benz, my friend, um the folks at Stellantis and there are many other cars coming.
And once you have a basic standard platform, then developers of AV systems, and there's so many talented ones, Wave, Waymo, Aurora, Momenta, Neuro, there's so many of them. WeRide, there's so many of them that can then take their AV system and run it on the standard chassis. Basically, the standard chassis has now become a computing platform on wheels.
And because it's standard and the sensor suite is comprehensive, all of them could deploy their AI to it. And in the future, a trillion miles a year that are driven, 100 million cars made each year, there's some 50 million taxis around the world. It's going to be augmented by a whole bunch of robo-taxis.
So, it's going to be a very large market. To connect it and deploy it around the world, today we're announcing a partnership with Uber. Uber, Dara Dara K, Dara is going to We're working together to connect these NVIDIA Drive Hyperion cars into a global network. And now in the future, you'll, you know, be able to hail up one of these cars and the ecosystem is going to be incredibly rich and we'll have Hyperion or robo-taxi cars all over the world.
This is going to be a new computing platform for us and I'm expecting it to be quite successful. Okay. So, this is what we talked about today. We talked about a large large number of things. We spoke about Remember, at the core of this is two are two platform transitions from general-purpose computing to accelerated computing.
NVIDIA CUDA and those suite of libraries called CUDA X has enabled us to address practically every industry and we're at the inflection point. It is now growing as a virtual cycle would suggest. The second inflection point is now upon us. The second platform transition, AI from classical handwritten software to artificial intelligence.
Two platform transitioning happening at the same time, which is the reason why we're feeling such incredible growth. Quanta quantum computing we spoke about. Open models we spoke about. We spoke about enterprise with CrowdStrike and uh Palantir accelerating their platforms. Uh we spoke about robotics, a new large potentially one of the largest consumer electronics and industrial manufacturing sectors.
And of course, we spoke about 6G. NVIDIA has new platforms for 6G. We call it Arc. We have a new platform for robotics cars. We call that Hyperion. Thank Thank you for for allowing us to bring GTC to Washington, D.C. We're going to do it hopefully every year. And thank you all for your service in making America great again. Thank you.
Comments 0