The Six Five On the Road with Victor Peng and Mark Papermaster at AMD Advancing AI

Thanks for having us.

Yeah, and I think if you look at it from an innovation perspective, that's the other thing, which is AI is just changing super rapidly. Like it is, believe it or not, still in its early stages, and so it's going to continue to evolve and in addition to proliferate. And that also means not only do you need multiple suppliers that are viable, but you really want multiple sources of innovation. And I think that's really critical. And I think that's what we pride ourselves on, innovating at multiple levels: technology level, architectural level, and software level.

Yeah, look, I think we not only strengthened our development teams and got them really focused in AI, right? Like we did a reorg early in the year where we put all the software resources all into one place and got them super focused in AI. But we also did some acquisitions, as I shared, Mipsology and Nod.ai. And we're working with the ecosystem. So I think each one of those elements is really important. From organic growth, just strengthening, bringing in more AI practitioners, doing some inorganic acquisitions of folks who really were contributing both to key technologies like ML but also have been maintainers in the ecosystem and working with the community. And then these standards bodies and these standard frameworks, they see AMD is going to be attracting a lot of developers, so they're also on board with supporting us as foundational partners. So I think each of those elements are really important. And by the way, again, it gets back to innovation too, right? We have deep engagements with our customers. They're telling us the problems they're solving, whether it's their first-party tools or when they're getting ready to stand up infrastructure, public instances. We have to be prepared to support all of that, and that has pushed us to innovate on the software side.

Yeah, the good thing about this is that not just within SaaS but in general, the industry is moving away from optimizing at a very low level. Why are they doing that? It's a combination of the fact that this is moving so fast. If you do that and then there's a new innovation, you have to redo all of that before you could actually move on. The other thing is that things like compiler technology, MLIR, is getting much more powerful. So the significance of, not everybody's familiar with OpenAI Triton, but the significance of that is that people can take the backend from frameworks like PyTorch and it'll still compile it and you still get excellent performance, right? So everybody's going to move to that because this is about productivity, right? And even the folks that still want the ninjas, that still want to get that last bit, they're being really careful about where they're going to do it. They don't want to do a lot of it, right? So I think what I would say is that that's just a general trend that we're just getting behind. So it's not even about just the competition, it's just that that's the way the industry wants to go, and we're aligning ourselves with what the customers want.

Yeah, I think, look, I don't have a crystal ball either. But what I do believe is going to happen is there's going to be innovation across the whole breadth of things. What gets attention is the people developing the really, really large foundational models, and we're working with folks that are working with them. But there's lots of innovation happening in more moderate size, like 100, 200 billion parameters, and even on device, like 10 billion. That's technology that trains in years. But I do think the other interesting thing to add to this is now we have AI to help with the AI. And as many people know, it's not just the infrastructure, it's the data. And people generating data through AI. So it gets back to the level of innovation that's going on is just phenomenal, and I don't think that's cornered in any one place. And I do think what makes AMD's position unique to capture that is because we are in clients, we are in GPUs and infrastructure, we have two different architectures: a spatial data flow architecture as well as a mainstream, really good, leading-edge GPU architecture. And stay tuned, you will definitely see us doing things in servers. And we're in all kinds of use cases. I think that's the other thing that doesn't always get appreciation. We talk about the breadth of our product portfolio. We have an incredibly broad market portfolio. We're in communications infrastructure, we're in autonomous vehicles, we're in healthcare, we're in factory automation, and traditional data centers, client, gaming, things like that. There are very few companies on the planet that can really say that they have both the platform architectures and the applications and use cases. And we work with the leaders in every one of those markets. So I don't know what the answer is, but I bet one of my customers in one of those markets has some of those answers. And it's just going to be exciting. It's a great time to be in technology.

Yeah, that's great. And actually, the last place you left off, maybe I'll lead off on, which is, look, at the end of the day, we've been in many different environments. Back when I was at Xilinx as well, we would supply to a key customer and they also have a group within them that competes. That's not really new. But I think it comes down to, are you delivering value? Are you delivering enough value that it really makes sense, even if they have some capability that they're going to work with you? Like we talk about open systems a lot. It's really in our DNA because we do think that we don't necessarily have all the best answers to everything. However, we are confident about what we do. We really are world-class at what we do, and we have to keep showing that and delivering that to our customers. So that's thing one. I think the other thing is that the pie is large, not only from a market size perspective. Every time you look at it, the TAM goes up. But also, back to the innovation, this is unlike anything I've seen in my career. I've been around for a while because I'm an old dude, right? But I mean, seriously, when do you have a new application come out and you're fundamentally changing data types? Like when social came out, you didn't have to change data types. And we're just doing quantization. We haven't really tapped into sparsity. There's just so much innovation. So I think the other thing that that diversity tells you is that not only is it a big market, but there's not one size-fits-all answer canonically across everything. There's innovation on the model side, there's innovation on algorithms, there's innovation on the software stack, and they all kind of interact. Innovation on the architectures too. So I think that's going to go on for a while. So the key thing is it's an 'and' function, it's not an 'or' function. Now, 10 years from now, will there be 15 different things? Probably not. But there's not going to be two. I honestly don't think there'll be two, right? Even for us, we have to have a different architecture for a client-based system versus what we're going to put in a massive supercomputer for training foundational models. You don't have an architecture that can span multiple orders of magnitude in TOPS per watt.

Yeah, and I think the other thing too is that you have to look at long lead time things. You opened up by saying some decisions we had to work on like a decade ago. So we're still working on deep technology things like really advanced packaging, really advanced integration technologies, more than two levels of stacking. We're working with TSMC and our entire supply chain, pushing the state of the art and getting it ready for high volume production at the right time. This is all about one thing: is that technology ready to go? Not too early, a lot of people crash and burn on that, and you certainly don't want to be too late. So I think at every level of technology, and one more thing, it keeps coming back to software, like you said, because these are complex systems. Systems are hardware and software. So the thing about this software is it's got to hide that complexity but let you take advantage of it. And that's actually a hard thing to do. But that's exactly because if you hide the complexity, people can move fast and you can be asynchronous with how you're lifting up the substrate and just giving a much more powerful engine. Back to your point, the software doesn't run on air, but at the same time, you don't want to have such tight coupling, and that's why you don't want to be doing assembly code or whatever, because you just can't move fast. So I think it's the whole stack: the software solution, the microarchitecture partitioning, chiplets, and all the way down to base technology. And we really can do that, and you could tell we're excited about that because we can keep talking about it.

Great.