Advanced Micro Devices, Inc. (AMD) Earnings Call Transcript & Summary

Okay. Okay. Here we go. I apologize for the technical issues. Thank you, everybody, for attending today's fireside chat. We're with Victor Peng, AMD President -- AMD's President of Adaptive & Embedded Computing Group. Victor, welcome. How are you?

Good. Thanks. How are you, Hans. Thanks for having me.

Yes. Good, good. Well, we saw you for nearly 15 years at Xilinx. You now are at AMD. What are your early impressions? And by the way, I understand you have a new role. You're an executive sponsor for AI. So how's that and what's that all about?

Yes. Look, first of all, joining AMD has been really exciting. I think we knew how great combinations would be. Lisa and I, it's been quite a journey, of course, to get to the point where we are today. But since we've closed for about 0.5 year now, I would say that it's been even more amazing in terms of the opportunities we see and kind of the really positive synergies. In terms of my role being the executive sponsor for our Pervasive AI initiative Yes, I think that when the deal was announced, I was going to continue to be responsible for the Xilinx business plus some other embedded business that AMD had. But there were some strategic initiatives that I was going to help drive. I think once we close and we were able to talk more deeply and openly about things, we centered in AI, I suppose, in one sense, that shouldn't be a surprise. So yes, I'm super excited about that. There's a lot of -- tremendous amount of opportunity, a lot of work to be done, but a tremendous amount of opportunity. So it's all good.

Fantastic. Before we dive in, I just want to tell the audience that Victor is not going to be too prominent with street events. And so we're delighted to have him here and because it's AI-centric, we're going to try to keep it that way. And for those of you that want to ask questions, I'll be monitoring the Q&A, monitor here on my end. So feel free to ask or type it out. I'll be checking periodically, and we'll go from there. So let me ask you, Victor, from a corporate culture perspective, I think that AMD and Xilinx make a fantastic combination. That's kind of like my view. Is that indeed the case? And what can we see from 2 organizations that are addressing very different markets and how the synergies could play out over time?

Yes. Look, I think -- first of all, I think everybody knows as we spoke about at the Financial Analyst Day that you look at the combined portfolio, and it's an incredible product portfolio of leadership CPUs, GPUs, adaptive SoCs and FPGAs. And then we at Xilinx had adaptive SmartNIC, and then we doubled down with Pensando. Now we have these DPUs, right? So if you look at the breadth of our product portfolio, that's -- I think that's quite unmatched. We also had significant overlap, obviously, in terms of the markets we are servicing and data center. But there was also complementary as well as in the marketplace, right? So the breadth that we have in the embedded markets. And the thing about AI, right, is it encompasses all of that, right? As big and as exciting it is, we're still really in the early innings. And so I think that broad product portfolio, the common customers that are all the thought leaders and drivers of AI in the compute world but also in the embedded space, I think that's really unique. Now we'll -- I'm sure we'll discuss this a little bit later, but there's also, as, again, we shared in the Financial Analyst Day that we also have a vision not only on the hardware side but what we're going to do on software. And I think once we get both of those things really going, the opportunity is just tremendous. And our customers are really excited about it, right?

It's interesting that you bring up software. Your competitor on the GPU front has made it famous now, this move to CUDA and this platform that they've developed. Is that what you are saying when you say software kind of on top of the hardware to making these platform-specific solutions for customers in many markets?

Yes. I would say at a high level, there are similarities, but I think what we're trying to do is goes beyond just the CUDA part. And I think, again, it is important to think about both the hardware and the software, right? It's -- you need really both to sort of make this work. And again, with the breadth of, I guess, what would I say, targeted platforms that we have. What our software stack can to enable you to work from a common environment to hit, whether it's CPUs or the right execution engine or GPUs or FPGAs or adaptive SoCs, we're going to enable our customers to leverage all of that. I think the -- AI is so -- such an evolving technology. And really, there's not always one right answer. In fact, one thing that's been clearly accelerated is heterogeneous computing, right? And you really need all those targets. And not only do you have that need from a compute perspective because computing -- the demand for computing is virtually insatiable. But really, you can have other bottlenecks in the data center, like memory storage and then the network that brings that all together. And since we really have a tremendous offering also on the networking side, we also enable things like smart storage. What you're really going to see this is going to be all about how the software enables that entire heterogeneous environment and how you get the most throughput through the entire infrastructure, not just a node, right? I think that's a thing to keep in mind. Final thing in terms of some of the -- you brought CUDA, but I think more and more people are working in standard AI frameworks like PyTorch and TensorFlow and so forth. So in that sense, that's the environment people are familiar with. And our development environment is all interface to that seamlessly. So I think that certainly levels the playing field a lot. And again, we'll be building out the rest of that software stack over time.

Great. There is -- the street has -- we kind of struggled to understand what does it mean when an AI model or the road map for these models, the complexity of the parameters is doubling every few months in Moore's Law. We're lucky to get doubling every 2 or 3 years. What does that mean exactly? What does that mean to designs and engagements with customers and so on? If you can give us something in lay terms that we can understand as investors.

Yes. Hans, that's a great point. In fact, before I talk about some of the things I think AMD can bring the table to address that trend, let me add to the problem statement a little bit more. I alluded to it a minute ago. I mean these models are getting so large that they don't fit on a single node. In fact, they require racks and racks of whether it's GPUs or other kinds of accelerators. And then once again, because the parameters, I mean, we're talking about hundreds of billions of trillions of parameters, so there's a lot of memory. And bringing that all together, you could be limited again by the networking and your memory and even storage in some instances. So you really have to look at an entire problem. And once again, AMD actually has that product portfolio. It's a support all of that. The other thing is that the complexity is growing so much that the amount of computes that they need to even take these large models and take upwards of a month to train, they're also burning power on an exponential level, right? And these exponentials really are not sustainable. And what that's driving the industry to do is we must innovate. We must innovate on multiple different levels, right? I talked about how that's driving heterogeneous compute environments because different types of workloads work better, are more effective or efficient on different types of compute platforms. But the other thing is that we have to innovate, and I'll try not to make this too complex, but there's a lot of calculations that are being done. There needs to be more innovation in the data types, in other words, the representation of the numbers, right, to reduce the compute load but also reduce the power and reduce the memory. So there's innovation that has to happen on these data types. There's also a characteristic of these deep neural networks called sparse, right? You can think of it as your brain -- neural networks that are inspired with brain like just in terms of connectivity, how often things fire are relatively speaking sparse. It's not all kind of firing way at the same time. The same thing happens with a lot of these networks. And so we have to take advantage of that sparsity. And what that will do, again, is reduce the compute demands, like you said, so that it's something that's tractable. It also reduces the power. So AMD, really, we have product offerings that enable people to innovate in those areas, right, and also eliminate bottlenecks, as I mentioned, like on the networking side and others. And the important thing to remember is that as big as AI is already today, we're still in the very early innings of that. And so I think a lot of innovations ahead of us. And AMD, I think, really enables our customers and partners and ISVs to sort of do that.

That's excellent, that sparsity issue. And to go down there a little bit more, so you can enable customers or help them along in getting that more efficient data type or reducing the amount of data that is kind of irrelevant to what the model is trying to do? Is that...

Yes. That's right. I mean -- so broadly speaking, that optimization people refer to is quantization, right? So you can have different number representations that are less math -- compute-intensive and you can get the same level of accuracy, right? And if you recall, in the early days, everyone said, "Oh, you need a full floating point." And then it kind of got reduced sort of what they call a half-precision flowing point. Now people keep reducing the number of representation once again so that you don't need quite as much silicon real estate to do computation so you can effectively do more computations. And it also has fewer memory requirements. So as models get bigger and there are more parameters and so forth. So you've seen some of that in GPUs but also some of these proprietary architectures. Now our adaptive SoCs and FPGAs really can help you do that in spades, right? So that's one aspect. And then on the sparsity side, there's also things that are just emerging in some of the more traditional architectures. But once again, our AIE for AI engine IP that was developed at Xilinx and we're going -- to refresh people, we're going to leverage that more broadly across the AMD portfolio. That really lets you innovate quite more extensively both in data types and in sparsity. And you can optimize things in terms of the memory and how the data flows, and all that is both good for performance and good for power efficiency. So yes, that's why once again, our GPUs or CPUs and our adaptive FPGAs, we bring that together, like that breadth is really hard to match from other players.

Yes. That's interesting. That may be the first time I heard that the wording of an AI engine, and it was developed by Xilinx. It's fascinating. Along those topics but just moving along, so AMD innovated with the EPYC processors by deploying tiles, these little chiplets. And maybe they're not tiles. Your competitor calls them tiles. But these chiplets that have scale and you can deploy them in other market segments like desktops, is there going to be a road map that incorporates chiplets that are FPGA chiplets or GPU chiplets? Is that a way to look at that? And would it be put together with an Infinity Fabric version of the future?

Yes. Just -- I know you've followed us, Xilinx, for quite some time, of course, AMD. But just to refresh everyone, both companies, I'd say, had anticipated the slowing of Moore's Law more than a decade ago, right? So both companies, I think, have a lot of leadership in terms of chiplets 2.5D and also now moving to 3D technologies, right? So together, I think we really are clearly one of the leaders in that capability. So that gives you some context of how we got here. What I would say is that absolutely, we view that we're going to proliferate some of the IP. The leadership IP from both sides across the combined product portfolio over time. We've already talked about shared the fact that we've integrated that AI engine, that inference engine into a client CPU, which will be announced down the road. And you'll see more proliferation. That's our strategy to proliferate that IP in both directions. That could be done monolithically integrated or it can be done through chiplets. Down the road, we could also look at how we stack things vertically. I mean we have that capability. We have that track record. And by the way, a number of our customers are quite excited about not only the breadth of the portfolio but exactly because we have those capabilities. And in some cases, some of our top customers have certainties in which we could offer that capability to help tailor things to their needs. So yes, we're definitely going to be leaning in, in terms of our overall product offering. And I think we have some capabilities that a lot of our customers are quite excited about and how they might leverage for their -- to solve their problems.

That's fascinating. We do have a question from an investor. It doesn't have to do with AI. If GLOBALFOUNDRIES cannot migrate to more advanced nodes, what kind of options does AMD have due to political tensions in Taiwan? Will AMD seek foundry support from Intel? I think he's talking about TSMC or -- anyway, so that's the question.

So I guess, look, I don't want to speculate about political tensions. However, what I would just say is that, as everyone knows, TSMC as well as Samsung as well as Intel, a number of folks in the foundry industry are going to be building foundries. They're in the process of that now on U.S. soil. So clearly, we used TSMC and GLOBAL. And frankly, [ ACG ], we also use Samsung. We've used a number of foundries, and the fact that a number of them have plans to sort of put foundries onshore is good. But I guess, again, I don't want to be speculating too much about future geopolitical situations.

Okay. I think you did a good job with that one, Victor. Now I might be bouncing around here, but I do get a lot of questions on this whole notion of data processing units, DPUs, custom DPUs, FPGAs that act like DPUs, Pensando DPUs. What is the use case? Or what -- how did that emerge? And how does -- obviously, it seems to be a legitimate category. And how does AMD play there with Pensando, with the Xilinx portfolio and so on?

Yes. Look, I kind of alluded to it already. I kind of mentioned how these days -- I mentioned the context of AI, but this is true for many workloads, right? Increasingly, things performance in actual TCO through your computing infrastructure can be limited by the network. So it's really important that you do that also when you can offload CPU cores from some of those networking functions. Then you could obviously free up those compute resources to monetize doing real applications that customers are paying money for. So there's a whole host of reasons why that's really important. The second thing in terms of DPU is in whether it's the Pensando, where things are highly programmable, but kind of from a software perspective versus what was AMD, the Xilinx-originated SmartNIC that's also very adaptable but can be done at a lower level in terms of the hardware. The common thing you should take away from this is that customers need things customized for the workloads and what they're running in their data center infrastructure. That's really important both from a performance and throughput but also come from things like security, right? And being able to have security from the cloud to the edge, that's really all important. And there are different levels of program building and different levels of optimizations that you can make. And AMD, we have doubled down on this, so covered the broadest range of needs from our customers, right? Whether they want to take advantage of very, very ultimate performance and verifying our control but requires working at a more detailed level or a customer that doesn't need that but really wants to -- and doesn't have the resources to sort of do things at a more detailed hard level, we could support all of that. And also just the ease of deployment of that software kind of approach. So I think the general trend is the network is going to be increasingly important. And now we have incredible assets and capabilities that we could offer our customers wherever they are in that kind of range of what they're trying to solve for. So yes, we're really excited about it. And I think it really goes hand-in-hand with the tremendous compute leadership that we have across the CPUs and the GPUs and adaptive FPGAs and SoCs.

That's great. Just one off, just I was just thinking here in real time. Do you foresee having at some point the need to have a switch, like a top-and-rack switch, that kind of thing to communicate?

I guess what I would say, we are always thinking about where things are going from an architecture perspective. We listen very closely to what our customer needs are. Again, I would just generalize that to that whole thing of the architecture of the data center is getting -- is being rethought and innovated again because of things like AI, but not only AI. And you're going to see increasingly computation happening with data, both at rest in your storage or memory, close to where it's being used by some applications; and potentially in flight, both in the network side and potentially in the switch side. And we have tremendous technology in all those areas. So if that's the right thing for our customers, we have the capability of doing it.