NVIDIA's GTC Keynote Breakdown in 30 Minutes

0:00In this special episode of the Denoise podcast, we're going to do a breakdown of NVIDIA's GTCKeynote. Let's get into it.

0:10All right. Welcome, new location, Addy. Yeah, this is great. Thanks for stopping by. All right. So we got a special episode. The NVIDIA keynote just happened this morning. We watched it. Let's do kind of a play-by-play summary. Coming in at the angle of M &E, media and entertainment. Yeah, sounds good. And we also saw the CES keynote together. Yeah. So we'll talk about sort of the differences between the two. Yeah, things we noticed that are different, things that are coming and kind of what that means, and a summary too, so you don't have to spend two hours watching the keynote, which also had a couple of live stream issues.

0:39Oh yeah, and also it cut out, the whole thing cut out at one point. Oh really? Yeah, the entire live stream. Oh my gosh. The first thing that stood out to me, and this was right in the intro, they had the garden area that they showed at the CES keynote that was talking about rendering and how it's being rendered in real time, real time ray tracing. And Jensen noted again that for every one pixel that was generated, 15 were rendered with AI. For every pixel that's rendered, Artificial intelligence predicts the other 15. And I feel like, I didn't check the notes, but I think in CES, he said for every one pixel, it was rendered, eight pixels were generated.

1:14So I don't know if that's just a increase or a misspeak, but an increase in improvements. I think he's talking about DLSS. Which they didn't talk about at all in this keynote. Okay, yeah. I mean, that was a big part of CES keynote. And we use DLSS a lot in virtual production because we're dealing with really large walls with large resolution, but we're limited by the amount of GPU and the power of Unreal Engine, if you will. So DLSS is a quick way for us to go from like a 2K render on a computer up to an 8K render or whatever that the wall needs. So in this case, there's - Without having to process every single pixel.

1:53Right. Like the computer figures out what the other pixel should be. It's crazy. It's generatively filling in all of the details. I know your friend James Blevins has also been a big fan. and of DLSS. Yeah, and hyping it up and talking about like, people need to pay attention to this. Yeah, absolutely. He's right. DLSS is supposed to be way more computationally efficient than just rendering those extra pixels out. It also pairs well with denoisers. So denoisers and DLSS together means you just have to do less in real time. Give it a noisy image, use AI to decipher and upscale what that final image should be.

2:25That's it. Which we have talked about here in past episodes. Yep. On denoised. On denoised. But that was the only kind of mention And that just stood out to me because I don't know if it was just a mix of the numbers or if there was an improvement that they didn't really highlight in DLSS or how they're able to do real time rendering. That was pretty much it for a lot of kind of M &E specific related stuff. Moving through sort of what the highlights were, he kind of did a recap of stuff he talked about at CES and sort of the development of AI and that we're progressing to agentic AI agents, which we've talked about here on the podcast.

2:54And the next step to that is physical AI and robotics AI, which we just talked about in the episode that went out today. Perfect timing. Yes. And DeepMind does come back to play a role at the very end. Yeah. Spoiler alert. Spoiler alert. The next thing that he moved on to, which stood out a lot, is AI factories. And this is a fancier word for data centers. Now, this is a very big idea. Whereas in the past, we wrote the software and we ran it on computers, in the future, the computer is going to generate the tokens for the software. And so the computer has become a generator of tokens, not a retrieval of files.

3:37From retrieval-based computing to generative-based computing, from the old way of doing data centers to a new way of building these infrastructure, and I call them AI factories. They're AI factories because it has one job and one job only, generating these incredible tokens that we then reconstitute into music, into words, into videos, into research, into chemicals or proteins. We reconstitute it into all kinds of information of different types. Yeah, I love the terminology here because Jensen and NVIDIA is really differentiating AI data center from the rest of the data centers by using a colloquial term that we all know and love and admire.

4:24Factories. Factories are the basis of revolutions. Industrial revolutions was all factories. So, you know, in the same sense, the next era of human achievement is going to be AI factories. And I feel like that's the way he puts it. Yeah. And one thing he did sort of frame the conversation at the beginning was how computing in the past was done from a retrieval computing model and that the future he's trying to go to is generative computing model. So instead of retrieving the files or finding the things you're looking for, generating the things you're looking for. Yeah. Generative AI fundamentally changed how computing is done from a retrieval computing model.

5:03We now have a generative computing model, whereas almost everything that we did in the past was about creating content in advance, storing multiple versions of it and fetching whatever version we think is appropriate at the moment of use. And that was sort of another theme that came up recurringly in the keynote. I mean, a lot of what Jensen said is so 30 ,000 feet up CEO speak, if you will, that I have a hard time following because I'm a concrete example kind of guy. And I always look for examples and cues and things like that. But again, this is in the realm of what do you mean you're going to generate the thing that I'm looking for?

5:43So is it going to remain there after I finish it? Or is it just like, I just make it every time I need it? Right. Yeah. It was like, what does that look like? What does that mean? What is that? Yeah. Are we getting rid of buckets and files? But it's his job to stay super ahead, you know, two, three, four steps ahead of what we are. And this keynote was a combo of like, here's things that are coming out that you probably already knew. And then here's our roadmap for the next few years. Yeah. And I feel like coming at this too, he's kind of got to be like, well, because this is coming into after DeepSeek comes out and it's like, well, hey, do we need AI factories?

6:14Do we need these massive things anymore to run these models or train these models? And just to put it into context, DeepSeek did a lot of damage, not a lot of damage, but a significant amount of damage to NVIDIA stock prices. Yeah. And we did see the stock price dip during GTC, and it never really recovered from that initial dip. Yeah, and I don't know what it's going to be like after people see it, but yeah, it's a bit of like, well, do we need these huge factories? which when people build factories, they need to buy a lot of your chips. A lot of hardware. So they need to buy as much of the chips.

6:47And I'd say there's still some cases of what he was talking about of where, yes, you would need that. Yeah, I mean, there's no doubt that the world is already building quote-unquote AI factories, right? Like if you look at any of the major tech companies like Meta or Google, they have data centers that are specifically doing AI computation at scale, quote-unquote hyperscalers. So these are all the new terminology, at least for us here in M &E that we're having to be accustomed to. Yeah. And one of the examples that he did give, and he sort of spent a lot of time explaining the challenge for an AI company in that they are trying to generate millions of tokens and tokens is sort of the backbone of generative AI of like the things you make that turn into text or images or speech and generating millions of tokens quickly.

7:36And sort of the battle - The rate of generating the millions of tokens. Right. And if that basically, it gave the example of the same issue that search has, where if you type a search query and you want like a response quickly, and so the longer it takes, maybe you get a better response, but it takes so long that people bounce from a user perspective. Absolutely. I love what NVIDIA is doing with terminology and language around all of this. They're really putting into structure what we'll be saying just regularly over the next few years. So tokens, from what it sounds like, is like the currency of inference.

8:09So for example, for JADGPT to generate one word of response, they need to generate 100 tokens in order to get to the end result of one word. So if you want to serve 100 customers and each customer needs like 1 ,000 tokens a second to be served, then you're making 100 ,000 tokens a second as a factory. Yeah, and the factory is the backbone. He said their AI factory is because it has one job and one job only, generating these incredible tokens that we then reconstitute into music, into words, into videos, into research, into chemicals or proteins. And the fact that you need the factories, you need their processors to make the foundation.

8:48But I love the fact that Jensen is generalizing all of the different segments of AI down to tokens, which is like the US dollar. Like you can buy so many things with a dollar and it all comes down to how much dollars you can earn. So you can buy those things. The solution to this was one of the big announcements, NVIDIA Dynamo. NVIDIA Dynamo does all that. It is essentially the operating system of an AI factory. Which is basically, it seemed like a new operating system for AI factories. Yeah. So the way Jensen put it was right now, a AI factory has to ride on top of many layers of technology. You have to run VMware.

9:30VMware will then build virtual machines on the cloud, and that cloud then runs on the data center. The data centers use GPUs. So there's all these types of layers of technology that are working together. how he puts Dynamo is like that goes directly to the NVIDIA infrastructure which encompasses their GPUs their connections from GPU to one GPU to the next one rack to the next and so basically like kind of turning your racks of GPUs into one single massive GPU yeah that this language is able to communicate with that quickly and also the other one of the other products announced integrated silicon photonics, which it's a mouthful.

10:08And I think you better understand this than I did, but it was a way to speed up networking all of the different racks. Yeah. So right now, if you put a data center together, you know, it's going to have, let's say a giant data center has somewhere like a hundred racks and within a hundred racks, you have 15 or 20 different servers, but they all have to interconnect. They all have to behave as one single organism. The way they do that is there is a fiber connection from one machine to the next, to the next, to the next. And then that rack is fiber connected to the next rack and down to a mainframe switch, if you will.

10:43What if you were to bypass all that and that fiber can connect directly to your GPU and then to the next GPU? So that's what this is. This is an optical interface right out of the silicon. I've never seen anything like this before. That's why. Yeah. That's why. Still trying to wrap my head around that. But basically, yeah, faster, better way you can make these look like one giant GPU. Yeah. You're no longer using copper or electricity to transport information. You're using light, which is going to have a higher throughput always. You can do terabytes of information per second on a fiber network.

11:25And then what's even cool about it is because it connects directly at a GPU level, stringing together tons of GPUs becomes easier than ever. And now we're still talking data center scale, massive scale. These are like enterprise. I just want to go back one step and just really look at the flavor and feel for this conference. CES felt more like NVIDIA is doing a lot of cool stuff. Obviously, consumer focus at CES. Yeah. So we got to see more Omniverse stuff. we got to see tons of like consumer like more autonomous car stuff digits they announced digits with a personal like actual computer put on your desk kind of thing yeah this one this is their own conference and because of that they get to drive the message and what they want to share with the world it felt like nvidia's like most of the energy here was spent on data centers yeah and looking at last year's as well this is a big enterprise focused data center conference yeah so clearly where they see their business or the bulk of their business is at the hands of hyperscalers who are going to build these massive AI factories.

12:33I mean, we're talking about 100 megawatt factories. That's like the output of a nuclear power plant. Yeah. Right. And that's why you need a nuclear power plant to power these. Yeah. Yeah. That was mentioned in, in the, right. That was a, yeah. A hundred megawatt factory using Blackwell chips, which can run or generate 12 billion tokens per second yeah yeah i believe currently if it was on hopper chips it was the comparison was like 300 million tokens a second or something which is still a lot yeah but gen the argument jens is making is as your need for ai becomes more sophisticated you'll need more tokens per user to generate that response or whatever a video or and especially if we're going to go with his vision of a generative computing future yeah where we need the factory instead of retrieving you're Generating, yes.

13:22So yeah, this all ties into his prediction and his worldview. How much do you think that will be the case? How much do you think? Like, I mean - Good question. Coming from our industry where we probably have some of the largest files, deal with the largest files of raw media, raw video, and kind of saying, oh, we're not gonna do a retrieval system anymore. We're gonna do a generative system. Yeah, I don't know about that. What does that even mean? Well, maybe not for media and entertainment. Maybe it doesn't apply, but certainly for you know social media user generated content a lot of that could be just generated on the fly i could see that yeah i mean or or if like maybe it's combo where you have like your source files yeah and but you're generating your output you're generating like your on the fly yeah social media ads based on a collection of photos and videos again i like retrieving both of ours both of us are super speculating here because yeah he didn't give actual concrete examples of that no and it was yeah this is more but what i will say this is like the entire keynote was based around one theme to me is that you can just keep throwing hardware at the problem as the problem gets bigger and more complex so yeah for a lot of stuff so not only gpus but now you're throwing factories at the problem or calling yeah calling them factories which we've had for a while, but yeah, AI factories.

14:43Sort of on the flip side, one of the things they did call out too in one of their little pie charts of new features or code bases was 6G, Edge 6G. But one of the areas that I'm super excited about is Edge. And today we announced, we announced today that Cisco, NVIDIA, T-Mobile, the largest telecommunications company the world, Cerberus ODC, are going to build a full stack for radio networks here in the United States. And that's going to be the second stack. So this current stack, this current stack we're announcing today, will put AI into the edge. You've worked in 5G. Yes. Where does 6G go? And what does this potentially mean for AI?

15:34I'm assuming this is some sort of like edge AI functioning on your phone locally. You don't need an AI factory to like run a model. Maybe we should ask ChatGPT, but it's when we were working on 5G, we were deploying 5G at Verizon. This was 2018 or so, a few years back. Verizon at the time was already moving on to 6G. So at that time, they were specifying what exactly 6G meant. Is it one terabyte a second download speed? You know, is it sub-millisecond latency? Is it an edge computer that's always connected to your phone that does all the heavy lifting and your phone is just a thin client? So all of these things.

16:17So it's not clear to me what 6G means, but clearly we're past 5G and moving into a world where the interconnectivity needs are so great that we have to create a whole new set of architecture to accommodate it. Interesting. The other thing announcement and kind of big focus was on autonomous vehicles, self-driving vehicles, Cosmo, which you talked about in the last episode of their data set for driving and understanding the real world to create autonomous vehicles. And they announced a partnership with GM where GM is going to be using the NVIDIA chips to power their self-driving fleet. Yeah, that's practical use case.

16:54Pretty dope. Yeah. First of all. And we talked about how this has to be something too that has to happen in real time. Yeah. Obviously. We literally just covered this last episode. Yeah. Self-driving vehicles is already here. I'm in Santa Monica, and I just literally saw 10 Waymo cars just driving around here. The question is, how do we get this to scale? How do we get the cars to be more autonomous? So if a car is offline, it could still do what it needs to do and just do inference locally on device. Right, yeah. Yeah, driving the canyon areas here, and you're like, you lose cell phone reception, you're like, I can't pull up my maps anymore, how do I get out of here?

17:32You're driving the car, not a big deal, but if your car is reliant on computing, AI to figure it out. And then, like you and I were saying, it has to compute in real time at hundreds of frames a second because a car is moving so fast, right? So even like a 30 frames per second camera, you know, whatever milliseconds at whatever speed, that's the chunk of distance you're covering, right? So the requirements are enormous. So I think NVIDIA is right to be in this business because it's also such a massive business. It's trillions and trillions of dollars in people reinvesting in their vehicles, buying a car that could drive itself over the next 10, 15 years.

18:12Yeah. Yeah. Talking about something that's revolutionized like transportation. Right. So I'm curious to see where that goes with GM and kind of how they release it in the cars. Yeah. I just quickly did a chat GPT on 6G. And yeah, I was right. So 6G is supposedly one terabytes per second, which is 100 times faster than 5G. Okay. Lower latency? Expected to be under millisecond, one millisecond. Yeah. So I was right about that too. So 5G is anywhere between, I believe it's a single digit. So two to eight. What does that even look like when they're like, we got to update to 6G? Is that more antennas?

18:45Is that different antennas? What does that even? Completely different infrastructure. So when we went from 4G to 5G, the 4G stuff could not be used. So they had to go and do they repurpose existing antennas? Do they build new ones? Do they have to have more? How does that work? The 4G antennas remain because your cell, your phone can interchange between, it has two antennas. But the way 5G coverage worked, and I think I went over this on one of the podcasts, is you just need to blanket more cell towers because it's connecting at different frequencies at the same time. Whereas 4G is like a single direct link.

19:21imagine 5g is like 10 15 different links at the same time so in an area like you know this area we're in one 4g tower will cover like 10 15 city blocks but you'll need one or two 5g towers per block or something insane to cover it correctly okay yeah so i would imagine 6g is even crazier even more towers for six maybe okay possible possible yeah i'm trying to scrape every M &E thing that was sort of covered. They always have these really cool animations flying around through their San Jose headquarters. And so they had these 3D animations moving around the office and then he did mention like these were all Gaussian splats.

20:02Gaussian splats just in case. So they scan the entire building and we're doing these animations with Gaussian splats that they turned into animations. It's cool. I'm scraping for M &E stuff here. Well, Omniverse is everywhere. Omniverse is now the primary engine to train all of the autonomous vehicle models. And if you combine Omniverse with Cosmos, which is their world model, you can pretty much simulate any robot in that environment. Or rather, you can have a robot train in that synthetic environment. Okay, right. Yeah, there's a ton of use. Omniverse is very omni. It's very big. Yeah, we've seen that.

20:43We've seen it with self-driving vehicles. We've seen a lot with like factory planning and building out your factory and running simulations. They also talked about building out your AI factory at a data center and running tasks and simulations, which is a new feature use case that they focused a good amount of time on. It's like, how do you actually design and build this hundred megawatt insane AI factory? Well, NVIDIA has a tool for that. Yeah, we have a tool. We have a tool for you. Basically making it frictionless for you to spend a billion dollars on a factory. Yeah. As far as new chips and updates, they did announce the new Blackwell Ultra NVL72, which is a bigger, better Blackwell chip coming out later this year.

21:23And then they laid out the framework for the next generation of chips, which they're going to call the Rubin Ultra, named after the person who discovered Dark Matter. Yeah. And that's, they said, would come out in the second half of 2027. Wow. So that's a big announcement for also like... That's hella early. Laying out, yeah, announcing this now and being like it's going to come out about two years from now. Right. Yeah, I think I just want to give people this roadmap of like there is bigger and better things coming. I think it's also like a stock price play. I mean, look, the biggest product for any public company is their share price.

21:57So the entire conference is just meant to build confidence in investors. Yeah, and especially when it's like they're riding so high, but everything's tied to a quarter and it's like, well, what's next? what's bigger and better and it's like a million chips everywhere and training AI stuff it's not like this stuff isn't happening fast enough we need bigger and better. The other hardware thing was DGX stations and so these are ranging from sort of like personal desktop size AI stations basically AI workstations for AI training and data science that they're going to partner up with all the major computer manufacturers and sell versions of of these DJX stations.

22:38I have no idea what the pricing is, but it seems like much more higher end computer to do. Yeah. Have you seen? Yeah, we almost bought a DJX station six, seven years ago when we were doing heavy Unreal work. At the time, I think it was like four or maybe eight GPUs in a single box. They were retailing for$100 ,000 or so. OK. Yeah, I think, I don't know if there's a max or just what the average is, 7 ,84 gigabytes unified memory. That's so important for AI stuff. the unified memory as i'm finding out doing a lot of local stuff on my computer and reading more into what we talked about the updates to the mac studio that was something that also 512 gigabytes unified yeah that um people are very impressed and we're able to run yeah when it when a memory is unified it's shared across the cpu and the gpu i think uh so there is no sort of transfer between the two what would you use the dgx i mean aside from like was that use case oh let's just get a souped up station so we could run on real is that better than just doing a build out yourself like what are the use cases for uh what i would do or what nvidia envisions both so for me it's just running a really fancy version of comfy ui and doing a lot of local inference local you know video generation style transfers things like that as a creative technologist just running ai locally I think how NVIDIA envisions it is any kind of professional work that you do, whether you're an accountant, whether you're a software programmer, whether you're a content creator, you're going to be using AI for it.

24:11And you need a different type of compute infrastructure for the next generation of tools. Than what you've traditionally had. Yes. Of a CPU, GPU. If you ask NVIDIA, they'll say you need DGX for every type of professional work. That's just my guess. Yeah, I'm curious to kind of see where the heavy VFX or where the kind of M &E applications are of getting a souped up computer like this. And then they touched on robotics, Omniverse, brought that up again. Physical AI and robotics are moving so fast. Everybody pay attention to this space. This could very well likely be the largest industry of all.

24:51Cosmos, which is their real world kind of data set of car driving, understanding the 3D world. It's the closest thing we have to a world simulation. So it takes into account things like air, water, physics-based collisions, materials of different types. I'm sure steel, bricks, wood, what have you. The more you put onto the Cosmos model, the heavier it gets and the more computationally expensive it is to simulate it. And lastly, big finale. And so today we're announcing something really, really special. It is a partnership of three companies. DeepMind, Disney Research, and NVIDIA, and we call it Newton.

25:40and then he brought out the uh bdx droid robot uh from star wars that we just talked about on the last podcast and how we'd seen it in some imagineering videos and the whole training that robot's gone viral dude yeah i mean it's extremely cute and then it's gonna go even more viral after this because it like kept talking and beeping to him i did notice the robot had a mic pack taped to its back that kind of why do you think that is so they could pick up the beeps and the noises and stuff yeah yeah so it was uh yeah it was very cute interaction with uh the robot and Jensen, it didn't go into much about what Newton is or what this partnership means.

26:13The only thing he did mention was that the Droid had two NVIDIA computers inside it. Just so you know, Blue has two computers, two NVIDIA computers inside. Didn't say anything else. It goes back to your question of why do you need DGX? Because when you're doing local inference, you absolutely need a ton of hardware to compute it in real time. And we covered this literally on the last episode. It's like the reasons robots are back into the spotlight now is because we're pretty close to having real-time AI systems, if not already, and stuff that can truly be autonomous. You don't have to program it.

26:53It'll just figure out the world. And so that plays right into this Newton partnership. Figure out the world. If it has the chips on it, they can also process it. It has Cosmos loaded on it. Real-time. Yeah. Yeah, right. And something that's small enough, it could fit in this pretty small robot. Yeah. And one of the words that you'll hear a lot is distilled model, distilled model. What that means is taking like a big heavyweight foundational model, essentially making a lightweight version of it, a travel pack, if you will, that can then run on limited hardware of a robot. Yeah, that was pretty much the highlights.