At CES 2025, Nvidia CEO Jensen Huang kicked off CES, the world’s largest consumer electronics show, with a new RTX gaming chip, updates on its AI chip Grace Blackwell and future plans it digs deeper into robotics and autonomous vehicles.

Here it is. Our new GForce RTX 50 series, Blackwell architecture, the GPU is just a beast, 92 billion transistors, 4000 ends, 4 petaflops of AI, 3 times higher than the previous generation Ada, and we need all of them to create the pixels I’m Showing You. 380 ray tracing teraflops so we can do the pixels we need to compute, compute the best image you possibly can and of course 125 shader teraflops. There is actually a concurrent shader teraflops as well as an integer unit of similar performance. So two dual shaders, one for floating 0.1 for integer. G7 memory from micron 1.8 terabytes per second, double the performance of our last generation, and we now have the ability to mix AI workloads with computer graphics workloads. And one of the amazing things about this generation is the programmable shader that can now also process neural networks. So the shader can bring these neural networks and as a result, we invented. Neuro texture compression and neural material shading in the Blackwell family RTX 5070, 4090 performance at 5:49. It is impossible without artificial intelligence, it is impossible without the four ends, 4 tear ops of AI tensor cores. It would be impossible without the memories of the G7. OK, so 5070, 4090 performance, $549 and here’s the whole family from 5070 to $5090 5090 dollars, double the performance of a 4090. Starting Of course we’re making a very large scale that will be available starting in January. Well, it’s unbelievable, but we managed to put these great GPU performances in a laptop. This is a 5070 laptop for 1299. This 5070 laptop has 4090 performance. And so the 5090, the 5090. Fits a laptop, a thin laptop. The last laptop is 14, 4.9 millimeters. You get the 5080, 5070 TI and 5070. But what we have here is 72 Blackwell GPUs or 144 dies. This one chip here is 1.4 exaflops. The world’s largest supercomputer, fastest supercomputer, recently. This whole room supercomputer has just achieved exaflop plus. This is 1.4 exaflops of AI floating point performance. It has 14 terabytes of memory, but here is the amazing thing that the memory bandwidth is 1.2 petabytes per second. That’s basically, basically the whole thing. The internet traffic that is happening now. The world’s internet traffic is processed by these chips, OK? And we have, um, 10 130 trillion transistors in total, 2,592 CPU cores. It’s a whole bunch of networking and that’s why I wish I could do it. I don’t think I do it’s the blackwells. This is our ConnectX. Networking chips, this is the MV link and we’re trying to pretend it’s the MV link spine, but that’s not possible, OK. And this is all in HBM memories, 1214 terabytes of HBM memory. This is what we are trying to do and this is the miracle, this is the miracle of the black wall system so we adapted it well with our expertise and our capabilities and we made it the Llama Nemotoron suite of open models. There are small ones that interact with uh very fast response time very small uh they are uh what we call super llama Nemotron supers they are the main versions of your models or your ultra model, the ultra model will be can be used uh to be a teacher model for a whole bunch of other models. This can be a reward model evaluator. Uh, a judge for other models to make answers and decide if it’s a good answer or not, basically give feedback to other models. It can be distilled in many different ways, basically a teacher model, a distillation of knowledge, uh, uh, model, very big, very capable, and so it’s all now online and Via Cosmos, the first in the world. Model of the foundation of the world. It was trained on 20 million hours of video. 20 million hours of video focused on physical dynamic objects, highly dynamic nature, nature themes, uh, people, uh, walking, uh, moving hands, uh, manipulating objects, uh, knowing You know, things that, uh, move fast in the camera. It’s really about teaching AI, not about creating creative content, but teaching AI to understand the physical world and from it to have this physical AI. There are many things below that we can do as a result of which we can create synthetic data to train models of uh. We can distill it and make it effective to see the start of a robotics model. You can do this to create a lot of physically based, physically believable, uh, future scenarios, usually a Doctor Strange. Um, you can, uh, because, because this model understands the physical world, of course you see a whole bunch of images that generate this model that understands the physical world, it also, uh, can of course be captioned and to be able to capture the videos, write them very well, and that captioning and the video can be used in training. Large language models. Multimodality large language models and uh so you can use this technology to uh use this foundational model to train robotics robots as well as large language models and this is the Nvidia cosmos. The platform has an auto regressive model for real-time applications as a diffusion model for very high quality image generation. Amazing tokenizer that basically learns vocabulary in the uh real world and a data pipeline so if you want to take all of this and then train it on your own data, this data pipeline because there’s a lot of data involved we’re making it all end to end for you and that’s why it’s the first data processing pipeline in the world that can be accelerated as well as AI accelerated it’s all part of the Cosmos platform and today we announce. Cosmos is openly licensed. It’s open access on GitHub. Today, we’re announcing that our next-generation processor for the car, our next-generation computer for the car is called Thor. I have one here. Hang on for a second. OK, this is Thor. This is Thor This is a robotics computer. It’s a robotics computer that takes sensors and an insane amount of sensor information, processes it, you know. Umpteen cameras, high-resolution radars, LIDAR, they all come on this chip, and this chip has to process all the sensors, turn them into tokens, put this transformer, and predict the next way And this AV computer is now in full production. Thor 20 times. The processing capability of our last generation Orin, which is really the basis of today’s autonomous cars. And so it’s actually quite, unbelievable. Thor is in full production. This robotics processor, on the other hand, also goes to a full robot and so it can be an AMR, it can be aaa human or robot, uh, it can be the brain, it can be the, uh, manipulator, uh, this processor is a universal robotics computer. The chat GPT instance. Because general robotics is just around the corner. And in fact, all the enabling technologies I mentioned are. Let’s make it possible in the next few years to see very rapid advances, surprising advances in general robotics. Now the reason why general robotics is so important is that robots with tracks and wheels need special environments to accommodate them. There are 3 robots. 3 robots in the world we can make without needing green fields. The adaptation of the brown field is perfect. If we, if we can make these amazing robots, we can deploy them in the exact world we are building for ourselves. These 3 robots are an agent robot and agent AI because you know they are information workers as long as they can accommodate uh the computers that we have in our offices, it’s good. Number 2, self-driving cars, and the reason for that is because we’ve spent 100+ years building roads and cities. And then number 3, human or robot. If we have the technology to solve these 3. This is the biggest technology industry the world has ever seen. This is Nvidia’s latest AI supercomputer. And, and it’s finally called Project Digits now and if you have a good name for it, uh, contact us. Um, uh, here’s the amazing thing, it’s an AI supercomputer. It runs the full Nvidia AI stack. All Nvidia software runs on it. DGX cloud runs on it. It sits Well, anywhere and it’s wireless or you know connected to your computer, it’s a workstation if you want and you can access it you can reach it like a cloud supercomputer and Nvidia’s AI works on it and um it’s based on a super secret chip that we’re working on called the GB 110, the smallest Grace Blackwell that we’ve made, and this is the chip that’s inside. It is in production. This top secret chip, uh, we created in collaboration with CPU, the gray CPU is a, uh, built for Nvidia in collaboration with MediaTech. Uh, they’re the leading SOC company in the world, and they’re working with us to build this CPU, the CPU SOC, and connect the chip to the chip and the Blackwell GPU link, and, uh, this little, this little thing here. is in full production. Uh, we expect this computer to be available uh around May.