So thank you, Forrest.
Thanks. Good to be here.
So, Forrest, I wanted to start. You've got the broadest portfolio of data center silicon. You've got some new networking technology. You have, you know, CPU which feeds GPU. Maybe you can start by broadly talking about your strategy around your entire data center portfolio.
Sure. Yeah. We've been embarked over the last decade. I joined AMD almost exactly 10 years ago now, to build out a portfolio of data center technology to really, eventually take leadership in providing silicon solutions for data centers. We began, of course, with the CPU. That's where we had some heritage. We began with the EPYC line of the CPU, and that's done very well. We're about 34% share right now, and then we built out the GPU, and then we acquired a networking team and began building that out. The purpose of having all of those pieces is that in the end, we do believe that to really optimize the solutions, particularly for the AI era, that having all of the elements in-house is very helpful.
And so our intent is to build out reference solutions for, you know, AI systems that incorporate world-class CPU, world-class GPU, world-class networking, and the sum of the parts is greater. The whole is greater than the sum of the parts. And we believe that we were well on the way to offering, you know, the best possible solutions, easy to adopt, limited for customers to adopt, and that they can source through a variety of partners.
Can we first talk about CPU? You've been gaining share on the server side fairly consistently for years now. Your share in cloud is in the 50% range, much lower in enterprise.
Right.
But you're beginning to see some success in enterprise. So, can you talk about your roadmap and why you're winning on the CPU side? And if you can particularly talk about enterprise and what sort of you've done differently there to catalyze some recent wins there.
Sure. And as I mentioned, we're up to, I think, the most recent quarter, just around 34% revenue share. And that is over 50% in cloud, and it's in the 20-ish% for enterprise. The reason we've had the success is that we have, you know, generally focused on producing higher performance parts with better power efficiency than our competitors. And with the cloud guys in particular, the data center is their factory. And so as they're making their decisions on what to buy, they're very focused on what's driving superior TCO, what drives the most efficiency in their data center. And so we've had relatively rapid growth there. On the enterprise side, the equation is slightly different. There are certainly some high-value workloads for which that performance and TCO benefit is critically important.
But the CIO also has to balance perceived risk and, you know, they're much more concerned about getting taken to task for a problem than, you know, getting kudos for having a slightly more efficient data center. And so there, it's been about retiring the perceived risk and so building out confidence in the ecosystem and our customers that AMD has not just a superior solution. I think most folks are convinced of that, but it's low risk. It's easy to adopt. And so our focus there has been over the last few years, really, working with the other partners in the ecosystem to ensure that we've got qualified solutions that we're demonstrating case studies that we're really retiring that perceived risk. I think that the inflection point really has come this year where people have enough familiarity with our solutions.
We've done enough POCs that they're recognizing, "Hey, this is easy." You know, there's really not a port. It's not a situation of, "I need to port my application." In fact, the funny thing is the instruction set that both Intel and AMD execute is actually the 64-bit instruction set that AMD created. So the application's actually written to conform to the AMD instruction set architecture. And so it's very easy to port. And I think that realization has really metastasized now in the market, coupled with, you know, some of the recent concerns around Intel and the press about them. I think that's also providing an impetus for folks to consider AMD.
Got it. I wanted to ask you about Turin and what Turin brings to the table, and how you see it, you know, stacking up with Sierra Forest for E-core and for, you know, Granite Rapids. How does it stack up? And sort of what's the early customer reaction to Turin?
Yeah. Turin, you know, continues our strategy of providing leadership CPUs, both per-performance as well as power performance. And, you know, I think the third-party measurements so far, there's very limited Granite Rapids out there, but there's been some third-party benchmarking that's showing, you know, like Phoronix showed across a very wide suite of, you know, 200 different benchmarks. Turin is 40% faster than the top end of the Granite Rapids stack. And so I think that sort of continues our clear performance leadership with Turin against Sierra Forest. There's no comparison. I mean, Sierra Forest is a great part. Don't get me wrong. But with the E-cores, it's much, much lower performance both on a per-thread as well as an overall socket level. And so it's really not competitive at all.
We see, you know, great success with Turin, both at the 128 as well as the 192 cores. The great success there, and I think that's gonna continue. I think, you know, we've got another generation of absolute double-digit performance leadership across every application.
There seems to be some perception that your success on the CPU side is almost all because of process. And I think it goes way beyond just process. And you've always said that you assume that the process gap will close. That is your.
That's right.
That is your base assumption. Can you just double-click on how much of it does relate to process?
You know, I think, certainly, we've got a great relationship with TSMC and, you know, access to their process is helpful. No question about it. And more and more, we're doing things to optimize the TSMC process and our circuit designs concurrently. We were the first, I think, as well to, you know, ship production-level silicon in their most recent process. But it goes way beyond that. I mean, our designs, I think, are exceptionally good. Our embrace of advanced packaging, chiplet technology, you know, early use of CoWoS. I think we've been at the forefront of things beyond the process that are gonna be increasingly important to developing high-performance chips going forward. So, you know, we were the first to embrace chiplets. We were the first to embrace 3D stacking.
We were the first to, you know, incorporate HBM in a major way in some of our data center parts, and by the way, on the Xilinx side as well, there's a rich heritage there on that technology, and I think that's served us in very good stead in continuing to push the performance, and we have a very mature process around retiring risk on new technology development, and I think that's super important because that means that our roadmaps have been very predictable. You know, I think the other thing we pride ourselves on is, you know, we put out the roadmap. We say what we're gonna do, and then, I think I hope you'll give us credit. We've.
Mm-hmm.
We've done exactly that for generation after generation.
Definitely. Let me shift over to GPU. You're shipping MI300. You're just starting to ship MI325X, really more in Q1 than in this quarter, so your roadmap is really beginning to evolve. You recently extended the TAM. You took it from 427 to 528, the accelerator TAM. There's a bunch of questions around that, and really, they are how do you see your position in that TAM and sort of what would define success? I've heard Lisa in the past say, "Well, look at our share in CPU." You know, we're, you know, not saying that that's what the goal is, but that we're in the market to be a major player, so what would define success?
I think I'm gonna keep myself out of trouble by agreeing with her. I would anyway. No. Look, I think we certainly aspire to be relevant to the market and, you know, to be relevant to the market, to be important to the market. I think you have to be, you know, strong double-digit percentages. You know, some people would say 20%. I think over time, that's what we're looking to, and I'm not giving you an exact number, but we're looking to make sure that we're relevant, we're important to the ecosystem.
You're, I mean, even today, if you look at the inference TAM, you're probably pretty close to 10% share this year just of the inference TAM. So you're not, you know, you're not saying that you have caught up to NVIDIA, but you're, give or take, 10% share of the inference TAM, which is not nothing in a very, very short amount of time.
Yeah. I mean, we've been super pleased with the progress on MI300. I, you know, I think we've said a couple of times it's the fastest-growing product we've ever had. You know, so from essentially $0- $5 billion in a year has been great. That's certainly a waypoint, not a destination. And, you know, it does offer demonstrably superior performance in TCO particularly for inference. And we think it's gotten competitive on training as well. Our intent is to continue pushing that very heavily with the roadmap. So you mentioned the 325, you know, beginning to ship this quarter, and then 355 in the second half of next year. Both of those parts are designed just as we did on the CPU side to be, you know, iteratively more performant, more competitive versus the competition, more differentiated.
And our intent is to continue leadership in inference across the board, which we think is more and more important as chain-of-thought models and similar approaches, inference TAM approaches become more important, and to continue developing ourselves on training as well. Long term, we don't aspire to be an inference solution. We aspire to be a provider of great AI solutions on the training side, on the inference side, on any new evolving models as well.
If I said that in a few years, if you achieve 20%, is that a failure or is that a success?
Look, again, I don't want to I don't want to, you know, validate a specific goal, specific share goal. But we, again, we aspire to be relevant. We aspire to be important in this market. And I think, you know, we certainly are, are looking in any, any similar number in such a large market is, is a huge revenue opportunity for us.
Got it. Yes.
For sure.
Sort of, what have you learned from experience so far with MI300? You've had a lot of success with a few large customers. You've been very public about Microsoft. Meta has been at, you know, some of your events.
Yeah. Yeah.
What have you learned that you take forward, and especially as you go to MI400 in 2026?
Sure. Well, look, we're the challenger in this market, right? Even though it's a new market, obviously, NVIDIA, great company, great products, and they are the de facto standard. And so as the challenger, you have to have a strategy that minimizes the friction of adoption, minimizes the effort that customers need to go through to adopt you because they're already using the alternative, right? And then you also have to have differentiation. You have to give them a reason to adopt you, notwithstanding the fact that nobody really wants a market in which there's only one player. Nevertheless, you know, tragedy of the commons, if you don't give a customer a specific reason to adopt you, they won't, no matter how much they would theoretically want an alternative.
So, you know, our approach with the 300 has been to, you know, first off, minimize the friction of adoption on the system side by making it very easy to drop MI300s into infrastructure that was candidly, you know, first conceived to host NVIDIA solutions, software solutions. And then secondly, on the software side, our real thrust has been to minimize, first off, the time to functionality so that if somebody's got a model, make sure they can immediately run it on MI. And by the time we got to introducing the MI300, we had accomplished that goal. Pretty much any model written for the standard framework, if you, if you're running on NVIDIA, you could run it on MI300 out of the box day one, and it would work. But it might not be performance.
So if you roll back 12 months ago, that random model, you might get 130% of the performance of the NVIDIA solution at the same time. You might get 50% of the performance. And that's friction. That's friction of adoption. If you don't know what you're going to get, that's a problem. And so we've really been focused over the last year on maturing the software ecosystem, the math libraries, the frameworks, working with the guys that are developing the foundational models to make sure their models are AMD-aware, and really working to minimize the friction of adoption such that somebody can pick up the, you know, just your random model and run it on an AMD solution today, and you're gonna get excellent performance. And so those two those, and you're gonna get better performance on inference.
You're gonna be differentiated with our greater memory capacity, our greater memory bandwidth, a few other things on caching. You're gonna get a differentiated solution that gives you a reason for adopting. As we've built out our roadmap going forward, we've tried to stay true to those learnings. How do we minimize the friction of somebody adopting us for inference or training? And how do we make sure that we have, you know, notwithstanding you can't be too different, how do we have some differentiation that provides that impetus, that prize for adopting AMD?
Can you talk about just the success you've had in software? One thing, I mean, NVIDIA's done a fantastic job with CUDA and with all their transformer models. And, you know, you've been playing catch-up and you've, you know, made the acquisitions and, and you've come a long way. But where do you assess where you are today versus where you need to be? Do you have all the transformer models you need? Do you have the capabilities you need?
You know, I think, again, it's a journey, not a destination because, you know, even as we're developing NVIDIA's great set of software resources, they're continuing to advance. But I think where we are today is that, as I mentioned a moment ago, for customers that already have a model or want to develop or, you know, or fine-tune a model, we're at a point where it's very easy to do so. We don't have the same level of vertical solutions that NVIDIA has for that. We're really, you know, focused on working with the other members of the ecosystem to develop those.
But I think we're at the point where it's, again, relatively easy to adopt, relatively easy to move over to AMD or to add AMD, really more realistically, add AMD as an alternative for your infrastructure and then place the workloads between NVIDIA and AMD where you get most benefit.
Can we talk about any supply constraints that you foresee in this coming year? You've gotten out, I think, pretty well in front of CoWoS. Of course, your allocation's not as much as NVIDIA's is, obviously, but you've done a pretty good job of getting out in front of that. Do you foresee memory being a potential bottleneck for you? I know, you know, every day I'm sure I have five emails from, you know, somebody asking me about the challenges that one of the large, you know, companies is having and whether that's a problem for you or, you know, or NVIDIA, frankly.
Yeah. Yeah.
So is that something that could constrain your growth, either of those two things or really any factor?
Yeah. I think we're in pretty good shape. We've got an excellent supply chain team and excellent operations team. And I think, more importantly, we've got outstanding relationships with all of our partners in the ecosystem. And it's not anybody's best interest, maybe one company's, but it's not anybody else's best interest to have one customer dominating the consumption of any particular component, be it CoWoS, be it memory, whatever. And so we have outstanding support and outstanding partnerships really from all of our partners, be it substrates, be it wafers, be it memory. And I think we've done a lot to build and develop those relationships, but it's also our interests are very congruent with their interests, which is customer diversity. Nobody wants to be locked down to one customer. It's way too dangerous.
And so, I think that has been a great set of relationships. And as we look forward into 2025 and beyond, we're very confident that we can get the support for the components that we need.
And you feel like, I'll just ask you directly. So you feel like if the challenges from that main or from that large supplier persist, that you still have other sources of supply?
Oh, sure. We work on the memory side, for example. We work with all of the major vendors, and we've got great relationships with all of them.
Great. Can we talk a little bit about beyond MI325X and talk about MI350? Some investors have wondered, just how that product is positioned in the marketplace at that time. Will there be liquid-cooled and air-cooled SKUs? Because the market will be significantly more shifted over to liquid cooling by then. And so, like, how do you intersect your roadmap with the infrastructure, which is also evolving?
Yeah. So look, we do see a rapid transition in the focus of particularly new data center builds, you know, predominantly shifting more and more to liquid-cooled builds to support the extremely high densities that people want to get. We're very comfortable in being able to support liquid-cooled environments. Really, even for the 300, we've got solutions available today. Surely that will be growing through the 325 and the 355. But you'll also be able to put it into an air-cooled infrastructure as well, which I think is a real advantage for us because it gives the customers flexibility in their deployment scenarios. We think 355 will compete very well in inference and training against the GB200 and follow-on solutions.
We think we will have a broader support at high density, at fairly high densities, for both liquid as well as air-cooled.
What about from a rack scale point of view? Because your first product that's fully rack scale's not until MI400. So will that aspect hold you back? Will the market want rack scale so much in the back half of 2025 that that's a factor that maybe could constrain 350? I'm just kind of wondering that's like another.
You know, I don't think it's gonna be a major issue. And that's certainly not what we're hearing from our customers, that the question is really, you know, how well does it fit in the data centers? What's the level of performance they can get? Are they stranding anything? Are they stranding power cooling or anything else? And I think we've been thoughtful about the design, and our customers are flexible enough in their data center deployments that that's not an issue, that won't preclude us. We will have full rack scale infrastructures available, of course, in the 400 timeframe.
You know, that's the point where, you know, you're starting to get up into the, you know, 200-plus kilowatt per rack regime, in which case you really do need to have a complete rack scale architecture.
Yes. Totally. Just on that front, can we talk about the ZT deal? Give us an update there and sort of review the rationale. It's my thinking that by the time that it really impacts your roadmap, it's probably more impactful for MI400 than for MI355.
Yeah. No, I think you got the timing right. So first off, we're progressing very well, and we're still very confident of closing the transaction in the first half of next year. We've already gotten regulatory approval in the US and a number of other geos. We're waiting for a few, but that all looks like it's on track. And we're very optimistic that we'll close the deal in the first half of next year. We have already started working through a set of contractual agreements. Of course, we're two different companies, so we can't operate as one yet, but we can put in place strong contractual agreements that allow us to engage the ZT resources on the forward-looking products. And we have already done so on 355, on 450, 400 series, and quite candidly, beyond.
And so that has already started. You will see some contribution from the ZT Systems resources in the MI350 series systems that our customers deploy. So I do think you'll start to see a little bit of contribution there, but certainly you'll see a major contribution from the ZT Systems engineering teams on MI400 and beyond.
Can we talk about the,
Oh, I'm sorry. You asked me about the rationale. This is worth touching on as well. The rationale for buying ZT really is twofold. The first is that as we're designing for these 200-kilowatt-plus racks, you really do need to comprehend the requirements at the rack and cluster level as you're designing the silicon. So being able to do that system and cluster level design very early on allows you to define and design a better piece of silicon to fit into it. So that's a big part of it. The second part is, look, we want to support the ecosystem in adding value to our solutions. So we don't want to take the approach that we have a one-size-fits-all.
We're not trying to take the Henry Ford Model T approach of you can have your hyperscale data center rack any way, any color you want as long as it's this color black. We're not taking that approach. So we're investing in enough systems engineering not only to produce a great set of base designs and elements, but also to allow others in the ecosystem to do variations to add their own value. That actually takes a little bit more engineering upfront to add the hooks, and design the components such that others can do that. But we think by doing so, we better harness the engineering talent really across the industry to accrete value to our ecosystem.
Great. Can we talk about the backend part of the network? I think on a recent earnings call, you said this is 30% of the cycle time, backend, backend part of the network. And we all know how strong NVLink is, to drive true cache coherency. You have UALink.
Yeah.
But that's for a pretty small cluster. Can you talk about how you're positioned for these larger clusters as we get to 100,000 and potentially a, you know, million cluster?
Yeah. So our focus, our focus has clearly been with, we think, the preponderance of the industry on how do we evolve Ethernet, to address those challenges. And we think there's no, there's no impediment for doing so. We think in the end, Ethernet is a superior solution to InfiniBand or any of the other alternatives that have been proposed in dealing with high performance, high scale, you know, scaling out behind some of the physical limit. One of the things people don't realize is there's physical limits on an InfiniBand network that, that limits you to a certain size. We've, we've helped form the Ultra Ethernet Consortium. That consortium's going to publish their first specification, we believe in Q1.
And we think we're well on track to be one of the first to offer an Ultra Ethernet networking solution that really will unlock, you know, million-plus node clusters with very high performance, very high resiliency, very high manageability, very high debuggability , which are all critically important aspects of not just how do you, you know, build a cluster, but how do you keep it running and keep it running efficiently. And those are key attributes we think of what Ultra Ethernet and Evolved Ethernet are going to provide.
So bottom line question that I'm asked all the time. When do you think, given what you know about your roadmap and given what you know about NVIDIA's roadmap, when do you think you're gonna have a product in the marketplace at the same time as them that is equivalent or better that you've caught up?
You know, from a day-one perspective, right? Since, you know, I would say right now the products in the market from NVIDIA are H100 and H200, and I think we're very well positioned with 300 and 325 against them. But you're saying when are we gonna introduce at the same time? Look, we're taking the same approach on the GPU side as we did on the CPU side, which is, you know, build a multi-generational roadmap, put in place the engineering discipline to retire technology risk during the development cycles in a predictable way and run them down. And so that's what we're doing. So we're doing the same general approach on the GPU that we did on the CPU.
I think that we're, you know, by the time you get to the middle of next year, GB200, I think really will be deployed in volume at that point. That's when it's really gonna be starting to ramp up in volume. I think we're gonna be there with MI355, and I think there's no questions, no asterisk on our MI400 generation. We aspire to be there with the leadership, training, and inference and chain of thought solution with MI400.
Great. And with the last minute we have for us, how do you give investors confidence that you can keep up with this annual product cadence? Because they really have accelerated the roadmap to this annual cadence. Really, it's a new platform every two years, not every year, with a you know memory upgrade every year. So do you feel like AMD's a big enough company to keep up with this annual cadence?
Yeah. I think we are, and we've already adjusted our resourcing to do so, you know, some time ago. So I think, you know, wait and see. But we're very confident of being able to hang to this annual cadence. And very importantly, we are extremely experienced in critical elements of technology that we think will be increasingly important around chiplets, 3D stacking, you know, very large body, and substrate, devices. And we know how to retire that risk. We know how to deliver those without surprises. I think that as others, you know, have to go down that path, they're going to potentially encounter problems, and I think they already have. And so, you know, we're all about execution fidelity and meeting our commitments.
I think we've shown that on the CPU side, and I'm confident we'll deliver on the GPU as well.
Thank you, Forrest.
Thanks a lot.
Appreciate it.
All right.