WEBVTT 00:00:00.000 --> 00:00:02.720 Okay, so we're back and ready to go deep on this 00:00:02.720 --> 00:00:08.879 one. It's NVIDIA's GTC March 2025 keynote. And, 00:00:08.879 --> 00:00:11.460 well, Jensen Wong did not disappoint, to say 00:00:11.460 --> 00:00:13.880 the least. Not even close. We've got the full 00:00:13.880 --> 00:00:16.420 keynote transcript, the slides, the demos. I 00:00:16.420 --> 00:00:18.640 mean, it was a firehose of information and some 00:00:18.640 --> 00:00:20.859 really, really big announcements in there. Absolutely. 00:00:21.440 --> 00:00:24.300 So the mission today is to cut through all that 00:00:24.300 --> 00:00:26.760 and really distill, you know, the key takeaways. 00:00:27.100 --> 00:00:29.320 What are the big moves NVIDIA is making? What 00:00:29.320 --> 00:00:31.839 does it mean for the industry as a whole? And 00:00:31.839 --> 00:00:33.780 most importantly, how does it connect back to 00:00:33.780 --> 00:00:35.590 what our listeners need to know? I think the 00:00:35.590 --> 00:00:37.850 biggest theme right from the start was just how 00:00:37.850 --> 00:00:40.450 intertwined everything's becoming. You know, 00:00:40.490 --> 00:00:42.670 AI obviously is at the center, but the way it's 00:00:42.670 --> 00:00:44.729 impacting graphics, data centers, cars, robotics, 00:00:45.090 --> 00:00:47.509 even basic scientific research. Yeah. It's all 00:00:47.509 --> 00:00:49.810 connected now. Right, right. And the speed at 00:00:49.810 --> 00:00:51.670 which it's all happening is, well, it's kind 00:00:51.670 --> 00:00:54.369 of mind blowing. So let's jump in. One of the 00:00:54.369 --> 00:00:55.710 first things that really caught my attention 00:00:55.710 --> 00:00:58.710 was this concept of AI as a new kind of factory, 00:00:58.869 --> 00:01:00.850 a token factory. Can you break that down a bit? 00:01:01.179 --> 00:01:03.359 Yeah, it's a really powerful analogy. I mean, 00:01:03.380 --> 00:01:05.659 think about it. Traditional factories, they take 00:01:05.659 --> 00:01:08.159 raw materials and produce physical goods. But 00:01:08.159 --> 00:01:10.459 these AI factories, they're producing something 00:01:10.459 --> 00:01:12.799 different, right? They're churning out tokens 00:01:12.799 --> 00:01:16.219 as their output. Tokens. So in simple terms, 00:01:16.359 --> 00:01:19.379 what exactly are these tokens? Essentially, think 00:01:19.379 --> 00:01:21.219 of them as the fundamental building blocks of 00:01:21.219 --> 00:01:24.239 information that AI models use. Like a token 00:01:24.239 --> 00:01:26.599 could represent a word. a piece of code, a pixel, 00:01:26.659 --> 00:01:29.760 and an image. There would allow AI to process 00:01:29.760 --> 00:01:31.879 and generate all these different kinds of data. 00:01:32.180 --> 00:01:34.219 And what's really amazing is just how versatile 00:01:34.219 --> 00:01:36.379 they are. Okay, that's starting to click. So 00:01:36.379 --> 00:01:39.519 the more complex the AI, the more tokens it's 00:01:39.519 --> 00:01:42.299 processing and generating. Exactly. And that 00:01:42.299 --> 00:01:44.379 connects to this whole evolution of AI that Jensen 00:01:44.379 --> 00:01:46.459 Huang talked about. We went from AI that was 00:01:46.459 --> 00:01:48.340 primarily about perception, like recognizing 00:01:48.340 --> 00:01:50.239 objects and images. Then we had generative AI, 00:01:50.459 --> 00:01:52.920 creating new content. Now we're moving into what 00:01:52.920 --> 00:01:55.730 he calls agentic AI. Agentic AI. What does that 00:01:55.730 --> 00:01:58.650 mean? Like AI with a sense of agency? Yeah, essentially. 00:01:58.790 --> 00:02:00.750 I mean, it goes beyond just generating outputs. 00:02:00.849 --> 00:02:04.890 It's AI that can reason, plan, make decisions, 00:02:05.049 --> 00:02:07.909 use tools that can actually take actions to achieve 00:02:07.909 --> 00:02:11.550 a goal. Wow. That's a big leap. And the next 00:02:11.550 --> 00:02:14.500 stage, he said, is physical AI. I mean, that 00:02:14.500 --> 00:02:16.180 sounds like something out of science fiction. 00:02:16.400 --> 00:02:19.240 Right. It's about bringing AI into the real world, 00:02:19.379 --> 00:02:21.780 you know, enabling it to truly understand and 00:02:21.780 --> 00:02:24.759 interact with the physical environment. Robotics 00:02:24.759 --> 00:02:27.060 is a huge part of that. OK, so we have this clear 00:02:27.060 --> 00:02:29.819 progression, but each of these steps comes with 00:02:29.819 --> 00:02:32.400 its own challenges, right? Like accuracy for 00:02:32.400 --> 00:02:36.039 perception, coherence for generative AI. What 00:02:36.039 --> 00:02:38.319 are the big hurdles for agentic and physical 00:02:38.319 --> 00:02:42.229 AI? For agentic AI, it's all about trust. I mean, 00:02:42.229 --> 00:02:44.030 if these AI agents are going to be making decisions, 00:02:44.229 --> 00:02:45.870 we need to be confident that they're making the 00:02:45.870 --> 00:02:47.389 right ones, that they're aligned with our values 00:02:47.389 --> 00:02:50.569 and that they won't go rogue. Right, right. That's 00:02:50.569 --> 00:02:52.389 a bit unsettling to think about. And for physical 00:02:52.389 --> 00:02:54.810 AI? Well, the real world is messy. I mean, it's 00:02:54.810 --> 00:02:57.330 unpredictable. So the challenge there is building 00:02:57.330 --> 00:02:59.650 AI that can handle all that complexity, that 00:02:59.650 --> 00:03:02.270 can adapt to changing conditions and still operate 00:03:02.270 --> 00:03:05.030 safely and effectively. And all of this, of course, 00:03:05.069 --> 00:03:07.729 requires massive amounts of computational power. 00:03:08.009 --> 00:03:10.789 I mean, Jensen Huang really hammered home the 00:03:10.789 --> 00:03:13.129 point that the need for compute is just exploding. 00:03:13.370 --> 00:03:15.590 Yeah, it's almost like a wake -up call. He said 00:03:15.590 --> 00:03:18.090 the computational demands for these new AI models, 00:03:18.389 --> 00:03:21.229 especially the agentic AI, it's something like 00:03:21.229 --> 00:03:23.430 100 times greater than what we were anticipating 00:03:23.430 --> 00:03:26.669 even a year ago. Wow. And what's driving that? 00:03:26.770 --> 00:03:29.669 Is it just the sheer size of these models? Well, 00:03:29.689 --> 00:03:31.430 that's part of it, but it's also the techniques 00:03:31.430 --> 00:03:33.129 they're using. Like he talked about chain of 00:03:33.129 --> 00:03:35.689 thought reasoning. Basically, it means the AI 00:03:35.689 --> 00:03:38.409 breaks down a problem into smaller steps, generating 00:03:38.409 --> 00:03:41.710 a ton more tokens in the process. So the smarter 00:03:41.710 --> 00:03:44.550 and more complex the AI gets, the more compute 00:03:44.550 --> 00:03:47.379 it needs. And that brings us to, well, NVIDIA's 00:03:47.379 --> 00:03:49.240 bread and butter, right? Accelerated computing, 00:03:49.680 --> 00:03:53.180 GPUs, the whole CUDA ecosystem. Absolutely. It's 00:03:53.180 --> 00:03:56.120 all about harnessing the power of parallel processing 00:03:56.120 --> 00:03:58.539 to tackle these massively complex computational 00:03:58.539 --> 00:04:02.080 tasks. And CUDA, that's NVIDIA's platform for 00:04:02.080 --> 00:04:03.939 making that happen. Okay. For those who might 00:04:03.939 --> 00:04:05.740 not be familiar with C2, can you give us the 00:04:05.740 --> 00:04:09.159 quick rundown? Sure. Think of CUDA as the language 00:04:09.159 --> 00:04:11.379 that lets developers write software that runs 00:04:11.379 --> 00:04:14.500 on NVIDIA GPUs. Unlike CPUs, which are good at 00:04:14.500 --> 00:04:17.560 doing a few things really fast, GPUs have thousands 00:04:17.560 --> 00:04:20.459 of smaller cores designed for massive parallel 00:04:20.459 --> 00:04:23.620 processing. So for tasks like AI training and 00:04:23.620 --> 00:04:26.680 inference, that's where GPUs really shine. And 00:04:26.680 --> 00:04:29.279 then there's all these CUDAx libraries. It felt 00:04:29.279 --> 00:04:31.079 like they were announcing a new one every few 00:04:31.079 --> 00:04:32.879 minutes during the keynote. Can you give us a 00:04:32.879 --> 00:04:34.819 sense of what they are and why they're so important? 00:04:35.300 --> 00:04:37.720 Yeah. Think of them as specialized toolkits, 00:04:37.720 --> 00:04:40.680 right? Like they're built on top of CUA and they're 00:04:40.680 --> 00:04:43.019 specifically optimized for accelerating different 00:04:43.019 --> 00:04:45.439 areas of science and industry. So instead of 00:04:45.439 --> 00:04:47.160 starting from scratch, developers can just plug 00:04:47.160 --> 00:04:49.509 in these libraries and boom. Instant performance 00:04:49.509 --> 00:04:52.290 boost. And there's a library for almost everything, 00:04:52.389 --> 00:04:54.569 right? I mean, we saw QNumeric for NumPy, Keyletho 00:04:54.569 --> 00:04:57.850 for chip manufacturing, Arial for 5G. Right. 00:04:57.910 --> 00:05:00.089 And each one is addressing a very specific need. 00:05:00.209 --> 00:05:03.170 Like QNumeric, that's a huge deal for data scientists. 00:05:03.670 --> 00:05:05.930 NumPy is like the foundation of scientific computing 00:05:05.930 --> 00:05:08.389 in Python. And QNumeric makes it run seamlessly 00:05:08.389 --> 00:05:11.920 on GPUs with massive speedups. In Kulitho, that's 00:05:11.920 --> 00:05:13.779 the one for computational lithography, right? 00:05:13.899 --> 00:05:16.720 The process of making chips. That seemed to tie 00:05:16.720 --> 00:05:19.139 in really closely with this idea of AI factories 00:05:19.139 --> 00:05:21.379 that Jensen Huang kept talking about. Totally. 00:05:21.439 --> 00:05:24.379 His point was that every company that has a physical 00:05:24.379 --> 00:05:27.100 factory, they're also going to need an AI factory 00:05:27.100 --> 00:05:30.699 to design and optimize the processes in that 00:05:30.699 --> 00:05:33.329 physical factory. And Caletho, it's a perfect 00:05:33.329 --> 00:05:35.949 example. It's accelerating the creation of those 00:05:35.949 --> 00:05:39.389 incredibly complex masks used to etch circuits 00:05:39.389 --> 00:05:41.970 onto silicon wafers. So it's almost like a factory 00:05:41.970 --> 00:05:44.509 within a factory, right? The AI factory powering 00:05:44.509 --> 00:05:46.810 the physical factory. Exactly. And he said it's 00:05:46.810 --> 00:05:48.209 not just chip manufacturing. It's going to be 00:05:48.209 --> 00:05:51.680 true for every industry. Aerospace. automotive, 00:05:51.680 --> 00:05:54.439 pharmaceuticals, AI is becoming fundamental to 00:05:54.439 --> 00:05:56.300 the design and creation of physical products. 00:05:56.519 --> 00:05:58.319 That's a big statement. And then there was Arial, 00:05:58.399 --> 00:06:01.220 which is enabling GPUs to basically function 00:06:01.220 --> 00:06:04.480 as 5G radios. How does that even work? It's pretty 00:06:04.480 --> 00:06:06.579 mind -blowing. Arial is a library that allows 00:06:06.579 --> 00:06:09.019 GPUs to handle all the complex signal processing 00:06:09.019 --> 00:06:12.100 that's needed for 5G communication. And then 00:06:12.100 --> 00:06:15.160 they layer AI on top of that with AI -RAN, which 00:06:15.160 --> 00:06:17.970 stands for AI and Radio Access Networks. AI, 00:06:17.970 --> 00:06:20.370 Ran. So AI is being used to optimize the network 00:06:20.370 --> 00:06:23.689 in real time. Yeah. It can dynamically adjust 00:06:23.689 --> 00:06:26.709 things like power allocation, beamforming all 00:06:26.709 --> 00:06:28.310 the things that affect the performance of the 00:06:28.310 --> 00:06:30.750 network. The goal is to make it more efficient, 00:06:30.930 --> 00:06:33.889 more responsive, and ultimately deliver a better 00:06:33.889 --> 00:06:36.420 user experience. That's fascinating. And it's 00:06:36.420 --> 00:06:38.800 another example of how AI is blurring the lines 00:06:38.800 --> 00:06:41.759 between hardware and software. Absolutely. And 00:06:41.759 --> 00:06:45.240 that trend continues with Qopt, the library for 00:06:45.240 --> 00:06:47.600 numerical and mathematical optimization. Right. 00:06:47.879 --> 00:06:49.959 Optimization problems are everywhere, right? 00:06:50.019 --> 00:06:52.339 From logistics and supply chains to financial 00:06:52.339 --> 00:06:54.800 modeling. Exactly. It's all about finding the 00:06:54.800 --> 00:06:56.980 best solution out of a huge number of possibilities. 00:06:57.660 --> 00:07:00.379 And traditionally, those calculations can be 00:07:00.379 --> 00:07:03.620 extremely computationally intensive. Qopt brings 00:07:03.620 --> 00:07:06.019 the power of GPUs to bear on those problems, 00:07:06.199 --> 00:07:08.699 significantly speeding up the process. And then 00:07:08.699 --> 00:07:10.879 there was the whole rapid fire section on all 00:07:10.879 --> 00:07:13.279 the other CDX libraries, Parabricks for gene 00:07:13.279 --> 00:07:16.500 sequencing, MONAI for medical imaging, Earth2 00:07:16.500 --> 00:07:18.399 for weather prediction. It was almost overwhelming, 00:07:18.579 --> 00:07:20.970 but in a good way. Yeah, it really showcased 00:07:20.970 --> 00:07:24.610 just how broad the impact of accelerated computing 00:07:24.610 --> 00:07:27.329 is. I mean, these libraries were enabling breakthroughs 00:07:27.329 --> 00:07:30.290 in critical areas like health care, climate science, 00:07:30.449 --> 00:07:32.990 drug discovery. It's incredible to see the potential. 00:07:33.189 --> 00:07:35.509 And they're even dipping their toes into quantum 00:07:35.509 --> 00:07:38.910 computing with KuQuantum and CuDay Quantum. They 00:07:38.910 --> 00:07:41.670 even had their first quantum day at GTC, which 00:07:41.670 --> 00:07:44.199 is. Pretty forward -looking. Definitely. Quantum 00:07:44.199 --> 00:07:47.019 computing is still early days, but NVIDIA is 00:07:47.019 --> 00:07:49.699 clearly taking it seriously. These libraries 00:07:49.699 --> 00:07:51.959 are providing tools for researchers to explore 00:07:51.959 --> 00:07:54.959 and develop quantum algorithms and to start integrating 00:07:54.959 --> 00:07:57.819 quantum processors with classical GPUs and hybrid 00:07:57.819 --> 00:08:00.779 systems. So NVIDIA is laying the groundwork for 00:08:00.779 --> 00:08:02.879 a future where quantum and classical computing 00:08:02.879 --> 00:08:05.399 work together. Exactly. And then there was CUT 00:08:05.399 --> 00:08:07.819 ESS, the big announcement for sparse solvers 00:08:07.819 --> 00:08:10.860 in CAE, which is computer -aided engineering. 00:08:11.279 --> 00:08:13.079 Right. That one seemed to generate a lot of buzz. 00:08:13.259 --> 00:08:15.920 What's so significant about it? Well, SCAR solvers, 00:08:15.980 --> 00:08:17.800 they're used in all sorts of engineering simulations, 00:08:17.800 --> 00:08:21.259 like how a bridge will withstand stress or how 00:08:21.259 --> 00:08:23.379 a new airplane design will behave in flight. 00:08:23.620 --> 00:08:26.959 And until recently, NVIDIA themselves were using 00:08:26.959 --> 00:08:29.180 general purpose computers to design their own 00:08:29.180 --> 00:08:31.519 accelerated computing hardware. Wow. So they 00:08:31.519 --> 00:08:34.399 were essentially using slower tools to design 00:08:34.399 --> 00:08:38.419 faster tools. Exactly. But with Cut ESS, they've 00:08:38.419 --> 00:08:41.600 optimized those solvers for CD. which means they 00:08:41.600 --> 00:08:44.820 can run on GPUs. So now the entire process of 00:08:44.820 --> 00:08:47.100 engineering design, it's getting supercharged. 00:08:47.440 --> 00:08:50.039 Faster simulations, more complex designs. It's 00:08:50.039 --> 00:08:52.039 a big leap forward. And they also touched on 00:08:52.039 --> 00:08:55.139 CutEF for data frames and Warp for physics simulations 00:08:55.139 --> 00:08:57.059 and Python. I mean, the list just goes on and 00:08:57.059 --> 00:08:58.919 on. And that's the key takeaway, right? They're 00:08:58.919 --> 00:09:01.759 building this incredibly rich ecosystem of software 00:09:01.759 --> 00:09:05.639 tools, all powered by CBA and GPUs. And because 00:09:05.639 --> 00:09:09.100 CDA is so widely adopted, these tools are immediately 00:09:09.100 --> 00:09:11.620 valuable to millions of developers across all 00:09:11.620 --> 00:09:14.000 these different fields. It's why NVIDIA believes 00:09:14.000 --> 00:09:15.820 we've reached this tipping point of accelerated 00:09:15.820 --> 00:09:18.500 computing. Before we move on from CDA, I wanted 00:09:18.500 --> 00:09:20.960 to touch on Halos, which is their framework for 00:09:20.960 --> 00:09:23.399 automotive safety. They didn't spend a ton of 00:09:23.399 --> 00:09:25.639 time on it, but given the growing importance 00:09:25.639 --> 00:09:28.080 of self -driving cars, it seems pretty crucial. 00:09:28.419 --> 00:09:31.120 Oh, absolutely. Halos, it's their way of making 00:09:31.120 --> 00:09:34.309 sure that self -driving cars are, well... safe. 00:09:34.570 --> 00:09:37.129 They're doing rigorous testing on every line 00:09:37.129 --> 00:09:39.629 of code, which millions of lines, by the way. 00:09:39.750 --> 00:09:42.370 And they're building in principles of diversity, 00:09:42.750 --> 00:09:46.070 transparency and explainability. They even dedicated 00:09:46.070 --> 00:09:48.429 a whole workshop at GTC to it. So they're not 00:09:48.429 --> 00:09:50.409 just focused on the performance of self -driving 00:09:50.409 --> 00:09:53.029 systems, but also on their safety and trustworthiness. 00:09:53.090 --> 00:09:55.090 Exactly. It's not enough for these systems to 00:09:55.090 --> 00:09:57.110 just work. They have to work safely and predictably. 00:09:57.230 --> 00:09:59.230 OK, so we've covered the foundational layer, 00:09:59.389 --> 00:10:01.710 the software tools. Now let's get into the hardware. 00:10:02.450 --> 00:10:04.889 NVIDIA announced their new Blackwell architecture, 00:10:05.269 --> 00:10:07.149 and the big news was that they're essentially 00:10:07.149 --> 00:10:10.610 putting two GPUs in a single package. What's 00:10:10.610 --> 00:10:12.330 the thinking behind that? Well, it's all about 00:10:12.330 --> 00:10:14.730 increasing compute density, right? Putting more 00:10:14.730 --> 00:10:17.730 processing power into a smaller space. But the 00:10:17.730 --> 00:10:19.889 really interesting part is the disaggregated 00:10:19.889 --> 00:10:22.570 NVLink switch, which they're calling MVLink Now. 00:10:22.929 --> 00:10:25.230 MVLink. So what does that mean and why is it 00:10:25.230 --> 00:10:28.600 important? So NVLink, that's NVIDIA's high -speed 00:10:28.600 --> 00:10:30.980 interconnect, that allows GPUs to communicate 00:10:30.980 --> 00:10:33.879 with each other really fast before it was integrated 00:10:33.879 --> 00:10:37.000 onto the motherboard. But now with NVLink, they've 00:10:37.000 --> 00:10:39.179 made it a separate switch. And that gives them 00:10:39.179 --> 00:10:41.419 a lot more flexibility. Flexibility in what way? 00:10:41.600 --> 00:10:44.200 Well, they can connect more GPUs together with 00:10:44.200 --> 00:10:46.779 higher bandwidth and with more simultaneous communication 00:10:46.779 --> 00:10:50.220 pathways. And that's crucial for scaling up these 00:10:50.220 --> 00:10:53.529 massive AI models. So MVLink is all about allowing 00:10:53.529 --> 00:10:55.610 the GPUs to talk to each other more efficiently. 00:10:55.830 --> 00:10:58.990 Exactly. And that, combined with the move towards 00:10:58.990 --> 00:11:01.590 liquid -cooled data centers, means they can pack 00:11:01.590 --> 00:11:03.929 a lot more compute into a single rack. I mean, 00:11:03.950 --> 00:11:05.809 they talked about achieving one exaflops in a 00:11:05.809 --> 00:11:08.669 single rack. One exaflops? That's mind -boggling. 00:11:08.669 --> 00:11:10.970 I mean, what does that even translate to in practical 00:11:10.970 --> 00:11:13.669 terms? It's an astronomical amount of computing 00:11:13.669 --> 00:11:16.529 power. And it's all made possible by liquid cooling. 00:11:16.919 --> 00:11:19.000 which is much more efficient at removing heat 00:11:19.000 --> 00:11:21.580 than traditional air cooling. So they're pushing 00:11:21.580 --> 00:11:24.340 the limits of both chip design and thermal management 00:11:24.340 --> 00:11:27.179 to achieve these massive scales. Absolutely. 00:11:27.399 --> 00:11:29.620 And it's not just about training these AI models. 00:11:29.940 --> 00:11:31.919 They also talked a lot about the importance of 00:11:31.919 --> 00:11:35.279 inference, which is actually using the trained 00:11:35.279 --> 00:11:37.820 models to generate outputs. Right. That's where 00:11:37.820 --> 00:11:39.879 the rubber meets the road, right? It's what actually 00:11:39.879 --> 00:11:42.720 delivers value to users. Exactly. And inference, 00:11:42.960 --> 00:11:45.559 especially for these new reasoning -based AI 00:11:45.559 --> 00:11:48.509 models. it can be incredibly computationally 00:11:48.509 --> 00:11:50.429 demanding. I mean, they gave this example of 00:11:50.429 --> 00:11:53.070 a wedding seating problem where a reasoning model 00:11:53.070 --> 00:11:56.509 used over 8 ,000 tokens compared to just under 00:11:56.509 --> 00:11:59.669 500 for a basic language model. Wow. So reasoning 00:11:59.669 --> 00:12:02.529 requires a lot more token generation, which translates 00:12:02.529 --> 00:12:05.669 to more compute. Exactly. And users expect those 00:12:05.669 --> 00:12:07.190 outputs quickly, right? They don't want to be 00:12:07.190 --> 00:12:09.350 waiting around. So you need a really powerful 00:12:09.350 --> 00:12:12.460 infrastructure to handle all that. And of course, 00:12:12.460 --> 00:12:14.500 these models are so big that they have to be 00:12:14.500 --> 00:12:18.220 split across multiple GPUs. They mention techniques 00:12:18.220 --> 00:12:21.440 like tensor parallelism, pipeline parallelism, 00:12:21.539 --> 00:12:24.840 expert parallelism. Can you give us a quick overview 00:12:24.840 --> 00:12:27.240 of what those are? Yeah, they're basically different 00:12:27.240 --> 00:12:29.600 ways of breaking down the model and its computations 00:12:29.600 --> 00:12:32.279 so that it can be distributed across multiple 00:12:32.279 --> 00:12:35.429 GPUs working together. Each technique has its 00:12:35.429 --> 00:12:38.490 own advantages and challenges, and the best approach 00:12:38.490 --> 00:12:41.950 often depends on the specific model and the hardware 00:12:41.950 --> 00:12:44.269 you're using. So it's a complex orchestration 00:12:44.269 --> 00:12:46.110 problem trying to manage all these different 00:12:46.110 --> 00:12:48.610 pieces. Totally. And that's where NVIDIA Dynavo 00:12:48.610 --> 00:12:50.649 comes in. Yeah. They're describing it as the 00:12:50.649 --> 00:12:53.470 operating system for these AI factories. An operating 00:12:53.470 --> 00:12:55.830 system for AI. What does that mean? Basically, 00:12:55.830 --> 00:12:58.029 it handles all the low level stuff like scheduling 00:12:58.029 --> 00:13:01.330 workloads, routing data between GPUs, optimizing 00:13:01.330 --> 00:13:04.149 the distribution of compute resources. It's like 00:13:04.149 --> 00:13:06.210 the conductor of the orchestra, making sure everything 00:13:06.210 --> 00:13:08.370 runs smoothly. And they're making it open source, 00:13:08.429 --> 00:13:10.649 right? So other companies can build on top of 00:13:10.649 --> 00:13:12.809 it. Exactly. And they're collaborating with companies 00:13:12.809 --> 00:13:15.250 like Perplexity to really build out this ecosystem 00:13:15.250 --> 00:13:18.000 around Dynamo. OK, so we've got the hardware, 00:13:18.159 --> 00:13:20.500 we've got the software infrastructure. Now, what 00:13:20.500 --> 00:13:22.980 about the actual performance games? I mean, they 00:13:22.980 --> 00:13:25.679 were throwing around some impressive numbers. 00:13:25.980 --> 00:13:28.220 Yeah, they showed Blackwell with MVLink 72 and 00:13:28.220 --> 00:13:30.740 Dynamo delivering up to 40 times faster token 00:13:30.740 --> 00:13:33.940 generation compared to Hopper, especially for 00:13:33.940 --> 00:13:36.659 those reasoning models. And the ISO power efficiency 00:13:36.659 --> 00:13:40.539 gains were around 25x. ISO power. Can you explain 00:13:40.539 --> 00:13:42.360 what that means? Sure. Basically, it means they're 00:13:42.360 --> 00:13:44.980 getting 25 times the performance for the same 00:13:44.980 --> 00:13:47.019 amount of power consumption. So it's not just 00:13:47.019 --> 00:13:49.159 about raw speed. It's about efficiency, too. 00:13:49.240 --> 00:13:51.240 Right. Which is really important for data centers 00:13:51.240 --> 00:13:53.799 where power is a major cost factor. Absolutely. 00:13:54.059 --> 00:13:55.679 And then there was this whole discussion about 00:13:55.679 --> 00:13:59.559 digital twins and the role of NVIDIA Omniverse 00:13:59.559 --> 00:14:02.559 in designing and managing these AI factories. 00:14:02.759 --> 00:14:05.220 Can you unpack that a bit? What are digital twins 00:14:05.220 --> 00:14:07.820 and why are they important in this context? A 00:14:07.820 --> 00:14:10.460 digital twin, it's basically a virtual replica 00:14:10.460 --> 00:14:13.149 of a physical system. So in this case, it would 00:14:13.149 --> 00:14:16.129 be a virtual representation of the entire AI 00:14:16.129 --> 00:14:19.789 infrastructure, the servers, the GPUs, the networking, 00:14:19.990 --> 00:14:22.370 the cooling systems. And the idea is that you 00:14:22.370 --> 00:14:24.769 can use this digital twin to simulate different 00:14:24.769 --> 00:14:27.990 scenarios, optimize performance, and even identify 00:14:27.990 --> 00:14:30.289 potential problems before they happen in the 00:14:30.289 --> 00:14:32.539 real world. So it's like having a virtual test 00:14:32.539 --> 00:14:35.379 bed for your entire AI factory. Exactly. And 00:14:35.379 --> 00:14:38.639 NVIDIA Omniverse, that's their platform for creating 00:14:38.639 --> 00:14:41.059 and running these digital twins. They're partnering 00:14:41.059 --> 00:14:43.299 with companies like Vertiv and Schneider Electric, 00:14:43.539 --> 00:14:45.759 who are experts in data center infrastructure, 00:14:46.059 --> 00:14:48.639 to make these digital twins as realistic and 00:14:48.639 --> 00:14:50.820 useful as possible. OK, so they're really thinking 00:14:50.820 --> 00:14:53.100 holistically about the entire AI infrastructure 00:14:53.100 --> 00:14:55.840 from the chips to the software to the physical 00:14:55.840 --> 00:14:58.120 data center itself. Totally. And then there's 00:14:58.120 --> 00:15:00.279 the networking piece, which is critical for scaling 00:15:00.279 --> 00:15:02.600 these systems. They talked about their Spectrum 00:15:02.600 --> 00:15:05.279 X Ethernet solution and how Cisco is adopting 00:15:05.279 --> 00:15:07.799 it for enterprise AI. Cisco is a big name in 00:15:07.799 --> 00:15:10.179 networking. What's significant about this partnership? 00:15:10.559 --> 00:15:13.200 Well, it means that high performance networking 00:15:13.200 --> 00:15:17.149 for AI. It's becoming more mainstream. Cisco's 00:15:17.149 --> 00:15:19.629 putting Spectrum X into their enterprise products, 00:15:19.929 --> 00:15:22.269 making it more accessible for companies that 00:15:22.269 --> 00:15:24.830 want to deploy AI at scale. And they also talked 00:15:24.830 --> 00:15:27.929 about the limitations of traditional copper interconnects 00:15:27.929 --> 00:15:30.110 at these massive scales. What's the alternative? 00:15:30.730 --> 00:15:33.309 They're betting on silicon photonics, which is 00:15:33.309 --> 00:15:36.669 essentially using light to transmit data instead 00:15:36.669 --> 00:15:38.889 of electrical signals. It offers much higher 00:15:38.889 --> 00:15:41.450 bandwidth, lower latency and lower power consumption. 00:15:41.750 --> 00:15:44.639 So it's like fiber optic cables on a chip. Yeah, 00:15:44.700 --> 00:15:47.600 sort of. They're co -packaging the optical transceivers 00:15:47.600 --> 00:15:49.720 directly with the processing chips, which is 00:15:49.720 --> 00:15:51.980 a pretty big engineering feat. And they highlighted 00:15:51.980 --> 00:15:54.500 this new type of modulator called a micro ring 00:15:54.500 --> 00:15:57.080 resonator or MRM. What's the advantage of that? 00:15:57.500 --> 00:16:00.629 MRMs are much smaller. and more power efficient 00:16:00.629 --> 00:16:03.230 than traditional modulators so they can pack 00:16:03.230 --> 00:16:05.370 more optical connections into the same space. 00:16:05.850 --> 00:16:08.330 It's a big step towards making silicon photonics 00:16:08.330 --> 00:16:11.090 practical for large scale data centers. OK, so 00:16:11.090 --> 00:16:12.909 they're laying the groundwork for the next generation 00:16:12.909 --> 00:16:15.850 of high performance networking. Now let's get 00:16:15.850 --> 00:16:18.509 to the roadmap. They talked about Blackwell Ultra 00:16:18.509 --> 00:16:21.389 coming in the second half of 2025, followed by 00:16:21.389 --> 00:16:24.529 Vera Rubin in the second half of 2026. What's 00:16:24.529 --> 00:16:27.149 the big deal about Rubin? Rubin is a complete 00:16:27.149 --> 00:16:30.139 architectural overhaul. They're introducing a 00:16:30.139 --> 00:16:34.980 new CPU, a new GPU, a new networking fabric called 00:16:34.980 --> 00:16:40.100 MVLink 144, and the latest HBM4 memory. It's 00:16:40.100 --> 00:16:42.279 like a whole new generation of their high -performance 00:16:42.279 --> 00:16:44.820 computing platform. And then there's RubinUltra 00:16:44.820 --> 00:16:47.840 in 2027, which promises even more extreme scale 00:16:47.840 --> 00:16:50.460 -up. They were talking about 15 exaflops of compute 00:16:50.460 --> 00:16:53.399 and 4 ,600 terabytes per second of bandwidth. 00:16:53.620 --> 00:16:55.799 Those are just insane number. Yeah, it's clear 00:16:55.799 --> 00:16:57.720 they're not slowing down. And they showed how 00:16:57.720 --> 00:16:59.659 the total cost of ownership is actually going 00:16:59.659 --> 00:17:02.740 down with each generation, even as the performance 00:17:02.740 --> 00:17:05.460 is skyrocketing. So these incredibly powerful 00:17:05.460 --> 00:17:07.539 capabilities are becoming more and more accessible. 00:17:07.960 --> 00:17:09.660 That's an important point. It's not just about 00:17:09.660 --> 00:17:11.400 pushing the limits of what's possible. It's about 00:17:11.400 --> 00:17:13.619 making those capabilities affordable and practical 00:17:13.619 --> 00:17:16.819 for a wider range of users. Exactly. And then 00:17:16.819 --> 00:17:19.220 they shifted gears to talk about AI for enterprise 00:17:19.220 --> 00:17:22.180 computing. They announced these new DGX systems, 00:17:22.440 --> 00:17:25.380 DGX Spark, which is like a personal AI development 00:17:25.380 --> 00:17:29.400 platform, and DGX Station, which is a super -powered 00:17:29.400 --> 00:17:31.220 workstation. What's the target market for those? 00:17:31.420 --> 00:17:32.980 I mean, who needs that much power at their desk? 00:17:33.160 --> 00:17:35.740 Well, DGX Stark. That's aimed at individual researchers 00:17:35.740 --> 00:17:38.160 and developers who want a dedicated AI development 00:17:38.160 --> 00:17:40.420 environment. It's powerful enough for serious 00:17:40.420 --> 00:17:43.079 work, but still relatively compact and affordable. 00:17:44.519 --> 00:17:46.920 That's for the real power users, people doing 00:17:46.920 --> 00:17:49.440 cutting -edge AI research or developing really 00:17:49.440 --> 00:17:51.920 demanding applications. It's like having a mini 00:17:51.920 --> 00:17:54.400 supercomputer on your desk. And they talked about 00:17:54.400 --> 00:17:57.160 reinventing storage for AI. What does that mean? 00:17:57.359 --> 00:17:59.460 It's all about moving away from traditional file 00:17:59.460 --> 00:18:01.599 -based storage to something that's more intelligent 00:18:01.599 --> 00:18:03.619 and aware of the meaning of the data. They're 00:18:03.619 --> 00:18:06.160 calling it semantics -based retrieval, and they're 00:18:06.160 --> 00:18:08.160 partnering with all the major storage vendors 00:18:08.160 --> 00:18:11.500 to make it happen. So it's about making the storage 00:18:11.500 --> 00:18:14.579 itself more AI -aware. Exactly. And then there's 00:18:14.579 --> 00:18:16.960 the open sourcing of their reasoning model as 00:18:16.960 --> 00:18:19.420 part of their NIMS initiative. What's the significance 00:18:19.420 --> 00:18:22.869 of that? It's about democratizing access. to 00:18:22.869 --> 00:18:26.089 these advanced AI capabilities. By making the 00:18:26.089 --> 00:18:29.069 model open source, they're allowing anyone to 00:18:29.069 --> 00:18:31.210 use it and build on top of it. So they're not 00:18:31.210 --> 00:18:32.849 just building the tools, they're also making 00:18:32.849 --> 00:18:35.190 the core technologies more accessible. Exactly. 00:18:35.410 --> 00:18:37.529 And finally, there was the big push in for robotics. 00:18:38.190 --> 00:18:41.069 Jensen Wong declared that the time has come for 00:18:41.069 --> 00:18:44.109 robots, and they laid out their vision for how 00:18:44.109 --> 00:18:47.220 AI is going to revolutionize this field. They 00:18:47.220 --> 00:18:49.680 talked about this continuous loop of simulation, 00:18:50.099 --> 00:18:53.039 training, testing and real world experience. 00:18:53.259 --> 00:18:55.220 Can you walk us through that? It's about using 00:18:55.220 --> 00:18:57.880 simulation to generate massive amounts of data 00:18:57.880 --> 00:19:01.990 for. training robot AI, then testing those AI 00:19:01.990 --> 00:19:04.569 policies in virtual environments before deploying 00:19:04.569 --> 00:19:07.670 them to real world robots. Right. And then the 00:19:07.670 --> 00:19:09.529 data from the real world feeds back into the 00:19:09.529 --> 00:19:11.910 simulation, creating this continuous cycle of 00:19:11.910 --> 00:19:13.789 improvement. So it's like a virtual proving ground 00:19:13.789 --> 00:19:16.609 for robot AI. Exactly. And they announced these 00:19:16.609 --> 00:19:19.809 new tools to make that happen. Cosmos for generating 00:19:19.809 --> 00:19:22.250 synthetic data. Newton for physics simulation. 00:19:23.210 --> 00:19:25.390 And then there was a big announcement of Isaac 00:19:25.390 --> 00:19:28.569 Groot N1, their generalist foundation model for 00:19:28.569 --> 00:19:30.829 humanoid robots. And they're open sourcing that 00:19:30.829 --> 00:19:33.150 as well, right? Yes, which is huge. It's like 00:19:33.150 --> 00:19:36.809 giving the entire robotics community access to 00:19:36.809 --> 00:19:39.890 this incredibly powerful starting point for building 00:19:39.890 --> 00:19:43.349 their own robots. So it's clear that NVIDIA is... 00:19:43.759 --> 00:19:46.160 taking a very comprehensive and forward -looking 00:19:46.160 --> 00:19:49.279 approach to robotics. Totally. And that's really 00:19:49.279 --> 00:19:51.259 the takeaway from the entire keynote, right? 00:19:51.420 --> 00:19:53.299 They're not just focused on one piece of the 00:19:53.299 --> 00:19:55.019 puzzle. They're thinking about the whole stack, 00:19:55.220 --> 00:19:57.619 from the hardware to the software to the applications, 00:19:57.900 --> 00:19:59.839 and they're pushing the boundaries in every direction. 00:20:00.269 --> 00:20:02.210 It's an incredibly exciting time to be following 00:20:02.210 --> 00:20:04.329 this field. I mean, the pace of innovation is 00:20:04.329 --> 00:20:06.970 just relentless. It is. And it's not just about 00:20:06.970 --> 00:20:08.930 the technology itself. It's about the impact 00:20:08.930 --> 00:20:11.609 it's going to have on our lives. I mean, AI is 00:20:11.609 --> 00:20:13.690 going to transform everything from the way we 00:20:13.690 --> 00:20:15.930 work to the way we interact with the world around 00:20:15.930 --> 00:20:19.289 us. So for our listeners, the key takeaway is 00:20:19.289 --> 00:20:21.990 that this stuff matters. Even if you're not a 00:20:21.990 --> 00:20:25.210 developer or a data scientist, AI and accelerated 00:20:25.210 --> 00:20:27.509 computing are going to shape the future. And 00:20:27.509 --> 00:20:29.410 understanding the trends and the key players. 00:20:29.609 --> 00:20:31.849 is going to be increasingly important. Absolutely. 00:20:32.289 --> 00:20:35.549 This keynote, it was a glimpse into that future, 00:20:35.670 --> 00:20:38.069 and it's clear that NVIDIA is playing a central 00:20:38.069 --> 00:20:40.130 role in shaping it. Well, on that note, we'll 00:20:40.130 --> 00:20:42.230 wrap things up here. As always, thanks for joining 00:20:42.230 --> 00:20:44.670 us for another deep dive. We'll be back soon 00:20:44.670 --> 00:20:47.410 to dissect another fascinating topic. Until then, 00:20:47.509 --> 00:20:50.369 keep exploring, keep learning, and stay curious. 00:20:50.430 --> 00:20:51.230 See you next time.