Great, we're gonna get started. Thank you. I'm, Tim Arcuri. I'm the semiconductor analyst, analyst here at UBS. Very pleased to have AMD, and we have Forrest Norrod, who's, EVP and GM of the Data Center Group. Thank you, Forrest.
Thanks, Tim. Good to be here, good to see you again.
Great. Well, I just wanted to start off and talk about what is on most people's minds, which is AI. You're hosting your AI day next week in-
That's right.
Thursday.
That's right.
I'm not asking you to front run what you're gonna say next week, but I'm just sort of wondering what things we could expect you to highlight. What are the things, you know, driving that business? Why should people be excited about your AI position and your AI opportunity?
Sure. Well, next week we're going to sort of set the next milestone on our journey to be a strong contender in the AI space. What we're going to unveil is we're gonna launch the MI300 formally in a couple of different variants. And do so while we, you know, report out on the progress of, I think maybe most importantly, of the software development of the software ecosystem to support MI300 and its use in AI, as well as, you know, we'll be there with quite a number of customers and partners who will relate their experience with and embrace of MI300 as well. So we expect to mostly be about MI300, and our position now as a very credible alternative for generative AI and large-scale AI systems.
We'll also talk a little bit about the rest of our pervasive AI strategy, including what we're doing on the PC side. We've already shipped millions of PC processors with AI accelerators built into them. We're gonna be introducing our next generation of that very, very soon, and we'll lay out a vision and roadmap for that as well. I think maybe that's something people aren't expecting. We do think that will also be very important in terms of the way that people interact with AI, and it augments their work and life experiences, won't simply be through the cloud, but it'll also be on the PC as well.
Great. And maybe with respect to the ultimate market opportunity for MI300X, you know, Lisa did offer this new $2 billion bogey for next year. But certainly, you know, the numbers we keep hearing is that you're roughly 10%, you know, unit-wise, of what NVIDIA is. That would suggest numbers that are way, way bigger than that. So can you just talk about what the TAM is maybe, or at least how to think about the TAM? Because if you look at, you know, their numbers, you can get very, very excited about what the TAM is for you.
Yeah, you know, I think. So certainly the market is moving very rapidly and, you know, we've seen explosive growth this year as really the promise of AI has caught everyone's attention. And the early indicators, the early signposts on the level of productivity enhancements that people are seeing with some of these POCs and initial deployments of generative AI are incredible. And we're—By the way, we're seeing the same. And so we think that in the long run, the TAM of the AI market is very difficult to precisely predict, but you can easily convince yourself it's very, very, very large.
If you really can deliver 20%, 30%, 80% productivity enhancements for some, you know, white-collar jobs that really have been resistant to any sort of substantive productivity enhancement in the past, you know, that justifies an enormous incremental investment in, in IT to allow companies to seize those, productivity enhancements. And we certainly do see indications that in many fields, you are gonna see very material productivity enhancements. And so what it is long term, it's a little bit hard to say. I've seen numbers up to, you know... I've seen some numbers up into the trillions, which, you know, I don't. I'm not claiming that by any stretch of the imagination, but I'm just using that to illustrate it's a little bit too early to tell how fast this develops.
That said, we do think that, you know, over the next 24-36 months, you know, the, the interest level is extraordinarily high. We do think it's gonna continue to grow materially from 2024 over 2023. We think it's gonna grow sequentially in each quarter. And so as we think about it, you know, we're- we've, we've made one comment about the, the, the size. We certainly expect to exceed $2 billion in revenue for MI300 next year, and we're, we're making sure, of course, that we have supply for, for more than that.
Can you talk about what the use case, or use cases are for the, you know, customers that are adopting it? Is it, is it mostly inference?
It's really a mix. MI300X has a distinct advantage on inference over our competitors' solutions by virtue of its memory bandwidth as well as memory capacity. You can do a lot more inferencing on each GPU. You can fit more, and you can fit larger models and have fewer numbers of GPUs to deliver the same, the same results, which translates into very material TCO advantages. On the training side, though, as well, we're seeing we have a very competitive part. I don't think we've got quite the same level of advantage that we do on inference, but we think MI300X is a very credible training part, and that's what our customers are telling us as well. And so we really think we're gonna see you know, large-scale deployments in both inference as well as training.
We'll see in both. Okay. And maybe we can pivot to CPU, because that's a big piece that, you know, we don't talk about probably enough. And we're sort of shifting to a market where server units might not go up very much, but core count's going up a lot.
Right.
And so can you talk about that dynamic, about, you know, how much core count is going up and how much, you know, pricing per core? Because, you know, we had seen pricing per core come down a lot, but now pricing per core seems to be, you know, flattening out a little bit. So we could begin to see some growth in a market that people are skeptical can grow.
Yeah, we do think that one of the dominant factors as you think about this market is exactly the one you just articulated, and it's something we've really been driving, which is really driving up the core counts within the CPU. You know, driving up the core counts of high performance cores, and let me come back to that in just a second. And so that's been driving the ASP trends you've already seen in the CPU market. ASPs have been going up on the server CPU side pretty dramatically, and the average core count has been really going up over the last couple of years, principally driven by AMD. We were the first to 32 cores, you know, 64 cores.
We have 96 and 128 cores now, where our competitor has really 56, they may claim 60. And so we don't see that slowing down. You know, the next generation, we think that the core counts go up again substantially. That translates into much more performance per system, and we do expect that to drive fewer number of systems, but still with an aggregate performance continuing to grow. We're gonna always give customers sort of more each generation in terms of more performance core per dollar. But I do think that you know, one way to think about the pricing is probably relatively flat in terms of per core going forward, substantial performance per core advantages each generation, translating into TCO advantages at the system level.
So we're so accustomed to talking about your share from a unit perspective, but it sounds like we should be measuring your share much more from a dollar point of view.
Yeah, I mean, at the end of the day, I actually don't really care that much about the unit, the unit share. You know, there's some, you know, long tail, very low calorie, very low value, components out there. We're much more interested from a revenue perspective, revenue share perspective. We're at 30% revenue share today, as we sit here today. And that's more weighted towards cloud than enterprise, but we do see enterprise starting to grow substantially. But our long-term ambition, just to be clear, is to continue to grow our revenue share, and we aspire, in the fullness of time, to be the market leader from a revenue share perspective.
It seems that your share, really, as I think about it, people just, I, I think, have this mental thought: Well, if you're—if you, if you have a performance advantage and if your, you know, competitor catches up, well, the share should ultimately be 50/50. So your share's, you know, 30%, sure, you can get to 50. But, but when you really deconstruct the source of your share, it's really pretty narrow. I mean, you have significant share at a couple of U.S. cloud vendors, but you don't have that much share at a couple of other large cloud vendors, nor do you have that much share in enterprise.
So your share. Is it fair to say that your share, while it is 30% from a dollar perspective, it's actually a bit narrow when you actually, you know, deconstruct it, and you know, what you're seeing now is you're seeing the broadening out of your share?
Yeah, I think that we've had made tremendous progress, you know, first with the hyperscalers, the cloud, the cloud guys. And that's... Quite candidly, that's where we first focused. You know, the... You know, if you think about an enterprise customer, there's many factors that go into selecting the CPU and, you know, the Intel guys have been buying me lunch, you know, once a quarter for the last 20 years. It is. I don't mean to be funny, but there are a number of factors that make, you know, adoption of new technology for many enterprises to be a slow thing, and we knew that.
And so for the cloud guys, however, the data centers are factories, and so having the best TCO, having the best efficiency, the highest performance, drops directly to their bottom line. And so we knew that was the first place to go, and I think we've been very successful there. You know, our aggregate share in North America hyperscalers is certainly well above 50%. In the enterprise, it's in the mid-teens right now. But when I look at the leading indicators, particularly as we've put more and more focus on enterprise over the last couple of years, and as our partners, HPE, Dell, Lenovo, Supermicro, et cetera, have broadened the portfolio of AMD solutions, we're seeing, you know, great leading indicators of share gain in enterprise as well.
That's a key focus for us going forward: hold on and continue to grow on, on the cloud, but really double down and focus on enterprise share growth going forward. I'm very confident, particularly when I see the engagement that we have, early engagement, with a large swath of customers, that, that we're gonna see substantial share gain growth in the next few years.
How dependent on process leadership is that picture? I think some people, you know, mentally ascribe your share gains pretty much entirely to having access to, you know, advanced process node, but it goes way beyond that. So even if things ended up where your competitor caught up from a, you know, process point of view, you still feel confident that you can still execute on those gains?
Yeah, absolutely. I mean, AMD has a tremendous history of design innovation and design excellence. By the way, if you look back at, say, the Opteron days, you know, back 20 years ago, AMD had a deficient process, you know, certainly in at least a node behind Intel, and yet was able to deliver substantially higher performance parts on the basis of superior design. I think that we've got both assets today. We've got a excellent design, and we've been driving innovation around, particularly things like chiplets, packaging, advanced integration. I mean, you see, you know, now Intel sort of progressing along that same road, but, you know, we've been there for many years and continue to push forward.
So when I look at the future roadmap, I'm very confident in it. We always assume that Intel is going to, from this moment forward, execute perfectly and do everything that they say they're gonna do. That's how we plan. And on that basis, with what they've articulated on the process side and what we know of their product development plans, I'm very confident that we're gonna maintain both performance as well as power efficiency leadership for the foreseeable future.
Great. Just to that point about chiplet, there was a question that came in that I think is a good one. And the question is: What advantage does your experience in with chiplets, you know, offer you, particularly in AI?
Well, I think that, you know, our principal competitor, NVIDIA, has, you know, been focused on delivering monolithic solutions for some time. So if you take a look at Hopper, for example, it's one large monolithic die on a CoWoS substrate surrounded by the HBM. That's as far as you can go. You know, to add additional capability, you're either just dependent on the transistor increase that you can get from process, or you have to embrace chiplets. And so we think we're substantially ahead of the learning curve of NVIDIA in terms of chiplet technology. We've been on this journey for many times, and MI300 is sort of the ultimate expression of advanced chiplet technology.
It's, it's you know, 13 different, 12 or 13, depends on the configuration, different chiplets, you know, interconnected via via 3D stacking, you know, on a very large CoWoS substrate. We've got a lot of experience doing this, and I think, you know, NVIDIA is about to have to go through that journey to continue driving performance, and we expect them to do so.
Can we talk just as it relates to the TAM, you know, longer term for AI, about the restrictions in China? It's not as much of a factor for you, given that you're earlier in your ramp. But how do you think about the impact that these restrictions have on the TAM over the longer term?
You know, I think it's a little. Again, it's a little bit hard to say. I think that, you know, one of the things that we do expect, we do expect to be able to offer China products that are fully compliant with all U.S. export regulations, and we expect NVIDIA to do so as well. But, you know, clearly and candidly, you know, certainly there will be other domestic producers that, or designers that will design, try very hard to design parts as well. I think that the promise of the capabilities of generative AI are so large that, and China is such a large market, that they will embrace generative AI. They will work with us, they will work with NVIDIA, they will work with other sources.
And so, you know, perhaps it's slightly delayed from what it might have otherwise been, but I do expect it to continue to develop.
And it also seems like if you actually read the document, it seems that there are some on-ramps that the government is actually providing, where you could take a chip that, from a performance perspective, is otherwise banned. But if you include some you know cryptographic aspects to it, where you can track how it's being used, that the government might work with you to allow you to ship chips that would otherwise, from a performance perspective, be banned.
Yeah, I don't want to get into that level of detail of discussion of roadmap options, but certainly we are, we are focused on, again, two things: always being fully compliant with the regulatory regime, and then, you know, within that set of guidelines, being very aggressive at supporting our customers worldwide.
Great. Can we just talk about your x86, you know, franchise? We did actually talk about that before, but there seems to be this perspective that the, you know, GPU, which, you know, for you is fine because you also make GPUs, but there, you know, seems to be this view that the x86 TAM is just gonna keep on getting cannibalized, and cannibalized, and cannibalized. So can you talk about the sort of interplay between, you know, GPU and, you know, any other, you know, custom accelerator, for that matter-
Yeah
... and GPU?
You know, our view is that certainly this year, and to some degree, next year, there is some cannibalization of the general compute TAM by AI. Nobody walked into 2023 expecting to see, you know, such a large imperative to embrace generative AI. And so we did see cannibalization of some of the TAM over to AI systems, no question about it. But in the long run, I—we don't view this as an either/or. We view AI as an additive workload. The need for the existing workloads generally does not go away.
Generally speaking, you're gonna still need the same systems of record, you're gonna need your transaction processing, you're gonna need your web front ends, you're gonna need the vast majority of the existing infrastructure to power the business, and the existing workloads, that need remains. What AI promises is an incremental set, a differentiated set of capabilities to build on top of that set of workloads, to take the data, in many cases, from them, generate insights and generate actions that we do think will be incredibly valuable. But we don't generally see the existing workloads going away. We think that this is going to be additive over the arc of time.
In the short term, there's gonna be some cannibalization, although again, you know, from our perspective, that may even be helpful because it places a premium on trying to get the most out of your compute infrastructure, CapEx dollars. And so the performance, TCO, and power efficiency leadership that we have, that we think we maintain for the foreseeable future, I think, you know, in some ways it will add—it is adding further weight to consideration and further weight to adoption of AMD's CPU technologies.
And just as a, you know, follow-up to that, with everyone doing more custom chips, does this limit your XAI TAM? I mean, everyone's talking about doing Arm PC and how do you view the competitive landscape with, you know, not only what's going on on the merchant side with Arm, but also what's going on on the custom side?
Yeah. So you know, there's certainly... The interest in custom chips is actually not new. It's been going on for some period of time, and we actually have a custom chip business within AMD that's been very successful. But when we think about AI in particular, you know, GPUs are very flexible, very powerful and very flexible, generally programmable devices that have broad applicability. In the early innings, and we think we're still in the early innings, when as algorithms continue to develop, be refined, are continuing to change, that general purpose GPU, we believe, not just we believe, but we're hearing from our customers as well, is very, very valuable.
You know, will there, over time, be some, you know, stabilization of the algorithms and some place where you see custom chips or purpose-built ASICs with lower programmability, you know, servicing part of the market? Probably. But, you know, we think GPUs have quite a way to run. And we also, you know, are very closely working with a number of large hyperscalers that as they build out their overall strategy for silicon, that encompasses semi-custom as well as custom components, that we can offer them solutions as well.
Great. When I asked you the first question about what you're gonna talk about next week, you mentioned software. And that's viewed by the investment community as, you know, you being significantly behind. However, you bought Mipsology, you bought Nod.ai last quarter. Can you talk about your journey in software and some of your efforts and, you know, why maybe the investment community is too skeptical of your- of the, you know, progress that you're-
Yeah, you know, software is, you know, absolutely the key element here. I would say, well, there's three key pillars. There's software, there's sort of the base compute in the GPU, and then there's networking. They're all absolutely essential. We've been very focused on the software journey for the last three or four years, not just the last couple of years. So Nod and Mipsology are great additions, are great teams, are great additions to AMD. But the central thrust of our software efforts really has been in close partnership with a number of large hyperscalers. We've had very large software teams working on ensuring that the ecosystem of AI frameworks, compilers, libraries, models, you know, are fully supportive of AMD and increasingly are optimized both on AMD as well as NVIDIA.
I mean, we. Look, we realize we're entering this market as a second player. You know, NVIDIA's got a long head start. So when you're facing a market like that, it's critical that you think about: How do I decrease the friction? How do I minimize the friction for adoption, for consideration and for adoption? And so, you know, we set out, you know, multiple years ago on that strategy to primarily focus on the frameworks, focus on the open source community, work with like-minded partners that are some of the largest customers for this technology, that very desperately wanted to have, you know, more competition, to, you know, drive, you know, better TCO, but also more innovation and more openness in the ecosystem. And so we've been in on this journey for a long, long time.
So things like PyTorch 2.0 last year, when it came out, or sorry, earlier this year, when it came out, you know, day zero, optimization and support was fully there for NVIDIA, and it was fully there for AMD, and that was the, those were the only two players with day zero support. Similarly, Open Triton, JAX, you know, if you look across the full spectrum of the ecosystem, we'll talk a lot more about this next week. Our perspective is partner with others, focus on open source, focus on where people are developing, which is around the, the frameworks, around the open source compilers, et cetera, and make sure that we absolutely minimize the friction for someone to adopt AMD.
How much of an impediment is a lack of a high-speed network or and a proprietary high-speed networking, you know, solution, as NVIDIA has with InfiniBand? How much of an impediment is that to you, you know, growing the business?
Yeah, and we'll talk more about this next week as well. Look, networking is critically important for these generative AI systems, particularly on the training side. You know, you have to orchestrate across tens of thousands or shortly hundreds of thousands of GPUs to train these large models, and so the performance of the backside network is critically important. But likewise, our customers are telling us in no uncertain terms that they want openness, they want Ethernet. And so we have been. We've embarked over the last several years to evolve Ethernet to address any deficiencies it has vis-a-vis InfiniBand, and actually solve some of the scale issues that even InfiniBand has. So we were the principal force behind the Ultra Ethernet Consortium, that now pretty much everybody in the industry has joined to evolve Ethernet forward.
We'll talk about the journey next week of where we are today, where many of the largest models today are trained on Ethernet backside networks, and how we and the rest of the industry are going to drive that forward.
Do you think that it's not going to be a?
I don't think it's going to be a long-term impediment, no.
And can you talk at all about, if we play out that Intel does catch up in process, can you envision a scenario where you would engage with Intel from a foundry perspective? Is that out there in your sort of mental roadmap where... Look, we're agnostic, you know, sure, we, you know, we're a, you know, major customer of TSMC, but if Intel's going to catch up and they're going to have a viable foundry business, even though they compete with you, if it was structured properly, you would consider to engage?
I think we would certainly listen. But I think TSMC has been a fantastic partner, you know, for us. And it is. It. By the way, it's not just about process, it's about the full ecosystem around it. You know, you can have a foundry with a great process that can be competitive, but if you don't have the rest of the ecosystem out there with the right other IP, analog or other IP that you might need, if you don't have the right design tool support, if you don't have the rest of the value chain that you really need on the foundry side, it's a tough road to hoe. So I would say never say never. It is a little bit difficult to see how we would embrace them quickly.
Yeah.
you know, I think that's okay in the interim because, you know, we've got such a strong partnership with TSMC, and they are so capable.
Just as a last question, op margins in your business were down about 1,000 basis points year-over-year last quarter on fairly comparable revenue. I mean, I think there's a lot of investments you've made that are yet to be monetized, you know, needless to say.
Yeah.
Should we still think about your business as being a 30%+ op margin business?
Yeah, absolutely. Our ambition for the data center business absolutely is 30%+ op margin. Over time, we are in a heavy investment cycle right now to make sure that we're well-positioned to ride the wave of growth that we think is coming in AI and other places. And so, yeah, it's really an OpEx-driven phenomena, you know, getting ready for the revenue ramp that we think will come.
Great. Well, we're out of time, but thank you, Forrest. Really appreciate it.
Thanks a lot.