All the pieces Introduced at Nvidia’s CES Occasion in 12 Minutes
At CES 2025, Nvidia CEO Jensen Huang kicks off CES, the world’s largest shopper electronics present, with a brand new RTX gaming chip, updates on its AI chip Grace Blackwell and its future plans to dig deeper into robotics and autonomous automobiles.
Right here it’s. Our model new GForce RTX 50 collection, Blackwell structure, the GPU is only a beast, 92 billion transistors, 4000 tops, 4 petaflops of AI, 3 occasions increased than the final era Ada, and we’d like all of it to generate these pixels that I confirmed you. 380 ray tracing teraflops in order that we may for the pixels that we’ve to compute, compute essentially the most stunning picture you presumably can and naturally 125 shader teraflops. There’s really a concurrent shader teraflops in addition to an integer unit of equal efficiency. So two twin shaders, one is for floating 0.1 is for integer. G7 reminiscence from micron 1.8 terabytes per second, twice the efficiency of our final era, and we now have the power to intermix AI workloads with laptop graphics workloads. And one of many wonderful issues about this era is the programmable shader can be in a position to now course of neural networks. So the shader is ready to carry these neural networks and consequently, we invented. Neuro texture compression and neural materials shading with the Blackwell household RTX 5070, 4090 efficiency at 5:49. Unattainable with out synthetic intelligence, unattainable with out the 4 tops, 4 tear ops of AI tensor cores. Unattainable with out the G7 recollections. OK, so 5070, 4090 efficiency, $549 and this is the entire household ranging from 5070 all the best way as much as $5090 5090 {dollars}, twice the efficiency of a 4090. Beginning In fact we’re producing a really massive scale availability beginning January. Properly, it’s unbelievable, however we managed to place these gigantic efficiency GPUs right into a laptop computer. It is a 5070 laptop computer for 1299. This 5070 laptop computer has a 4090 efficiency. And so the 5090, the 5090. Will match right into a laptop computer, a skinny laptop computer. That final laptop computer was 14, 4.9 millimeters. You bought a 5080, 5070 TI and 5070. However what we mainly have right here is 72 Blackwell GPUs or 144 dies. This one chip right here is 1.4 exaflops. The world’s largest supercomputer, quickest supercomputer, solely lately. This complete room supercomputer solely lately achieved an exaflop plus. That is 1.4 exaflops of AI floating level efficiency. It has 14 terabytes of reminiscence, however this is the wonderful factor the reminiscence bandwidth is 1.2 petabytes per second. That is mainly, mainly all the. Web site visitors that is occurring proper now. The complete world’s web site visitors is being processed throughout these chips, OK? And we’ve, um, 10 130 trillion transistors in complete, 2,592 CPU cores. Entire bunch of networking and so these I want I may do that. I do not assume I’ll so these are the blackwells. These are our ConnectX. Networking chips, these are the MV hyperlink and we’re attempting to fake concerning the the the MV hyperlink backbone, however that is not attainable, OK. And these are all the HBM recollections, 1214 terabytes of HBM reminiscence. That is what we’re attempting to do and that is the miracle, that is the miracle of the black wall system so we nice tune them utilizing our experience and our capabilities and we flip them into the Llama Nemotron suite of open fashions. There are small ones that work together in uh very very quick response time extraordinarily small uh they’re uh what we name tremendous llama Nemotron supers they’re mainly your mainstream variations of your fashions or your extremely mannequin, the extremely mannequin could possibly be used uh to be a trainer mannequin for a complete bunch of different fashions. It could possibly be a reward mannequin evaluator. Uh, a decide for different fashions to create solutions and resolve whether or not it is a good reply or not, mainly give suggestions to different fashions. It could possibly be distilled in loads of other ways, mainly a trainer mannequin, a information distillation, uh, uh, mannequin, very massive, very succesful, and so all of that is now out there on-line and Through Cosmos, the world’s first. World basis mannequin. It’s educated on 20 million hours of video. The 20 million hours of video focuses on bodily dynamic issues, so dynamic nature, nature themes themes, uh, people, uh, strolling, uh, palms transferring, uh, manipulating issues, uh, you recognize, issues which are, uh, quick digicam actions. It is actually about instructing the AI, not about producing artistic content material, however instructing the AI to know the bodily world and from this with this bodily AI. There are various downstream issues that we may uh do consequently we may do artificial knowledge era to coach uh fashions. We may distill it and switch it into successfully to see the beginnings of a robotics mannequin. You may have it generate a number of bodily based mostly, bodily believable, uh, eventualities of the long run, mainly do a Physician Unusual. Um, you possibly can, uh, as a result of, as a result of this mannequin understands the bodily world, in fact you noticed a complete bunch of pictures generated this mannequin understanding the bodily world, it additionally, uh, may do in fact captioning and so it may take movies, caption it extremely properly, and that captioning and the video could possibly be used to coach. Giant language fashions. Multimodality massive language fashions and uh so you possibly can use this know-how to uh use this basis mannequin to coach robotics robots in addition to massive language fashions and so that is the Nvidia cosmos. The platform has an auto regressive mannequin for real-time purposes as diffusion mannequin for a really prime quality picture era. It is unbelievable tokenizer mainly studying the vocabulary of uh actual world and an information pipeline in order that if you need to take all of this after which prepare it by yourself knowledge, this knowledge pipeline as a result of there’s a lot knowledge concerned we have accelerated all the things finish to finish for you and so that is the world’s first knowledge processing pipeline that is may accelerated in addition to AI accelerated all of that is a part of the Cosmos platform and at this time we’re saying. That Cosmos is open licensed. It is open out there on GitHub. Properly, at this time we’re saying that our subsequent era processor for the automobile, our subsequent era laptop for the automobile is known as Thor. I’ve one proper right here. Grasp on a second. OK, that is Thor. That is Thor That is It is a robotics laptop. It is a robotics laptop takes sensors and only a insanity quantity of sensor info, course of it, you recognize. Umpteen cameras, excessive decision radars, LIDARs, they’re all coming into this chip, and this chip has to course of all that sensor, flip them into tokens, put them right into a transformer, and predict the subsequent path. And this AV laptop is now in full manufacturing. Thor is 20 occasions. The processing functionality of our final era Orin, which is actually the usual of autonomous autos at this time. And so that is simply actually fairly, fairly unbelievable. Thor is in full manufacturing. This robotics processor, by the best way, additionally goes right into a full robotic and so it could possibly be an AMR, it could possibly be a a a human or robotic, uh, it could possibly be the mind, it could possibly be the, uh, manipulator, uh, this this processor mainly is a common robotics laptop. The chat GPT second. For basic robotics is simply across the nook. And actually, all the enabling applied sciences that I have been speaking about is. Going to make it attainable for us within the subsequent a number of years to see very speedy breakthroughs, shocking breakthroughs in generally robotics. Now the rationale why basic robotics is so essential is whereas robots with tracks and wheels require particular environments to accommodate them. There are 3 robots. 3 robots on this planet that we are able to make that require no inexperienced fields. Brown discipline adaptation is ideal. If we, if we may presumably construct these wonderful robots, we may deploy them in precisely the world that we have constructed for ourselves. These 3 robots are one agentic robots and agentic AI as a result of you recognize they’re info staff as long as they might accommodate uh the computer systems that we’ve in our workplaces, it is gonna be nice. Quantity 2, self-driving automobiles, and the rationale for that’s we spent 100+ years constructing roads and cities. After which quantity 3, human or robots. If we’ve the know-how to unravel these 3. This would be the largest know-how business the world’s ever seen. That is Nvidia’s newest AI supercomputer. And, and it is lastly known as Undertaking Digits proper now and if in case you have a superb title for it, uh, attain out to us. Um, uh, this this is the wonderful factor, that is an AI supercomputer. It runs all the Nvidia AI stack. All of Nvidia software program runs on this. DGX cloud runs on this. This sits Properly, someplace and it is wi-fi or you recognize linked to your laptop, it is even a workstation for those who prefer it to be and you possibly can entry it you possibly can you possibly can attain it like a like a cloud supercomputer and Nvidia’s AI works on it and um it is based mostly on a a brilliant secret chip that we have been engaged on known as GB 110, the smallest Grace Blackwell that we make, and that is the chip that is inside. It’s it’s in manufacturing. This prime secret chip, uh, we did in collaboration with the CPU, the grey CPU was a, uh, is constructed for Nvidia in collaboration with MediaTech. Uh, they’re the world’s main SOC firm, and so they labored with us to construct this CPU, the CPU SOC, and linked with chip to chip and the hyperlink to the Blackwell GPU, and, uh, this little, this little factor right here is in full manufacturing. Uh, we’re anticipating this laptop to uh be out there uh round Might time-frame.
