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

I'm Joe Moore from Morgan Stanley. Very happy to have with us here today, Mark Papermaster, the Chief Technology Officer of AMD.

So maybe we could just get started with a couple of bigger picture questions. I mean, I guess, as you guys think about competing in both microprocessors and graphics with companies that spend more than you, but you have a better GPU than Intel has, then a better CPU than Nvidia has and you certainly succeeded in both of those markets. Can you just talk about the challenges of doing that and the opportunities of having both of those technologies going forward?

Absolutely, Joe. First of all, thanks for having me here. We are in an incredibly exciting market, compute, because it's just exploding. It's been exploding and it's doing nothing but going to a steeper and steeper hyperbolic growth. And what we said for years, we've said for the last decade that servicing this explosive demand in compute requires not just 1 engine but multiple engines. So that's been our strategy for over a decade at AMD. And so having both the CPU and GPU were quite fundamental to us. So it was prescient by leaders in AMD. It was driven first to accelerate with the graphics processing that is embedded with a CPU, which all of us use in most any mobile device or desktop device we have today, the PC and the mobile industry. But what we already knew back then, and has done nothing but moved faster than we originally anticipated, was that combination of CPU, GPU and other accelerators in the data center in the most demanding workloads. You saw that first in high-performance computing. That's -- high-performance computing, HPC is always the harbinger, Joe, of the most demanding workloads. But AI has turned to be the biggest consumer of these high-performance heterogeneous workloads. Because of the nature of the algorithms, it needs, CPU, GPU and specialized accelerators. That's what makes it an incredibly exciting time for the industry, and it makes us, at AMD, incredibly excited because this is a vision that we've been working for. This is a portfolio that we've steadily built up year after year, and we believe we're positioned very, very well for this hyperbolic growth.

Great. Maybe we could talk a little bit before we get into the products and markets, about the role of process technology for you guys. You haven't really been on the very bleeding edge. You didn't anticipate some of the Intel's struggles and you ended up having significant leadership. As you think about this going forward, do you need to push the envelope a little bit more on process technology? Do you assume that Intel will [ like to ] ship at some point? And how do you think you're going to be able to compete with them as they do?

Process technology, obviously -- think about it is foundational in our semiconductor industry. And you might say, is that changing is -- as technology evolves. I mean, Joe Moore, you don't ever slow down, but Moore's Law, it's slowing down. The cost of transistors is going up per node. The type of scaling you get, the type of circuitry that gets the benefit out of each new technology node, is getting less. Some of the circuit types don't scale. So what that means is a design and process technology have always been important, but now they have to be really designed in a partnership. And that's what we do extremely well. So we've led with the design approach that anticipated the slowing of Moore's Law. About a decade ago, we reengineered AMD engineering approach. We went to a modular approach. And we came up with what we called the Infinity Architecture. So we could partition out these different circuit types. A CPU engine, a GPU engine, separated from the circuitry that you use to connect it to the IO, to connect it to the memory around it. And we did that to give us scale. This allowed us to punch above our weight with the resources we had because then we could reuse circuitry across our portfolio, shared across product lines, and we did it because we anticipated this -- the change in rate and pace of search technology. That said, the new nodes remain vitally important, and they mostly remain important for the engine itself. Those transistors want to operate at the most efficiency you can. And that's what each technology node new brings. Remember, I said the transistors and the new node are getting more expensive, but they're still giving you efficiency gains, more performance at less watts of energy expended. So with our modular approach, Joe, we work very closely. We're partnered with TSMC, GLOBALFOUNDRIES. We're always looking at other foundry options. But with TSMC, is our primary partner and who we deeply partner in design technology co-optimization, DTCO. We are for high-performance, pretty much the lead intro of each new node. And as such, we've leveraged that modular technology to put the new node where it gives you the most benefit. And it remains important going forward, really as far as I can see in the process technology road maps.

Great. So maybe shift to some of the end markets, starting with server. If you could talk about the importance of Genoa, I know this is a really important new product for you guys. You've talked about it being kind of a bigger ramp -- more transitional ramp where Milan and Genoa kind of both do heavy lifting over the course of the year. Can you just talk about the importance of Genoa in your road map?

Yes. Genoa is our fourth-generation x86 EPYC server. And as the fourth generation, it incorporates all of the learning of the previous 3. And what we really wanted to achieve, and did achieve with Genoa was a huge step in total cost of ownership, TCO. But that's the buying factor for our enterprise and our data -- our bigger hyperscale accounts. And so as such, we designed it with a 96-core base, that's the base node, 96 cores with lots of IO, 128 IO PCIe Gen 5 lanes. Gen 5 was a step forward on the interconnect standard, has 12 channels of DDR5 memory. That's a new memory interface. And the CPU is in 5-nanometer. That's a new technology node, which has ramped exactly as expected with our partner, TSMC. So it was a big step forward. With that type of generational change and a new sockets, all that new memory, new IO goes into the new sockets, and that is a longer adoption cycle. It goes first in the hyperscale nodes, enterprise can then take advantage of it. But it's a more purposeful ramp for us. And the reason is that Milan, the third-generation EPYC, is still the TCO advantage out in the marketplace. It's 64 cores in a really efficient, cost efficient and performance efficient of 7-nanometer node. And so when you look at our stack up, Milan, third-generation EPYC, is a huge piece of our product stack along with Genoa on top. And that third generation still is a performance leader in a number of cases versus the just released competitor of x86 server. So it's a -- third generation is very, very strong. Fourth generation knocks it out of the park with a total cost of ownership and a consolidation play. So it is indeed a very, very, very important ramp for us. It's a very purposeful ramp, and it's going right on track.

Okay. Can you talk to -- I know there was a prominent blogger who talked about an issue with some kind of single channel DDR5 -- only having a single channel available for DDR5. I don't -- it seems like you guys have disagreed with some of those conclusions. But I guess, just bigger picture, you've said it's on track. You said you where you want it to be. Is there anything that changes with that platform in the second half that increases rate of adoption?

No. You're talking about a single dimm per channel, 1 DPC. That's the predominant implementation of Genoa. That's what the vast majority of our customers use. And so running that at 4800 speed, that was our primary plan for launch, and that's what we did. And the 2-dimm per channel, which is I think what you're referring to, is following. So that is for a targeted -- a much smaller targeted set of customers. Those speeds will be announced later this quarter, and that will ramp as well, but this number of customers for 2 dimms per channel is much smaller.

Okay. That's helpful. I wonder if we could talk about server. Your server progress really in 3 segments. It's enterprise server, enterprise-facing cloud, and then kind of more internal-facing cloud. Enterprise sever, I think, your share is the smallest, but I know you've made some inroads there. Can you talk to your progress?

You bet. First of all, let's talk about going to order of first internal properties at the hyperscale. That is the fastest adoption. You've seen us grow across our EPYC server lines most quickly there. And the reason is when you have a kind of performance and energy efficiency and TCO advantage like we've had with EPYC, that's an easy transition for hyperscale because they're massive buyers. They see the TCO advantage. They have a set of workloads that they can -- it's x86. So it's less than ships, and you're running, you're getting all of that TCO advantage. So historically, and with Genoa, that's just a very, very fluid ramp, and it's just very straightforward with the TCO benefits that we bring. With enterprise, that is generally a little bit of a lower core count because the VMs, the hyperscales have the virtual machines, they can take advantage of the full density that we have in EPYC. Some of the enterprise accounts still prefer a lower density of cores. We, of course, have that. We have a full stack up. And so what we've done is two things. We've increased the sales force that we have that's really educating our enterprise customers on the commanding TCO and the commanding sustainability gains that we have. We have such an energy efficiency for compute. And CIOs are asked by their Boards constantly, what are you doing for sustainability? Well, EPYC is a big piece of that answer because it brings such an energy efficiency to the compute. And so with that, we're increasing our awareness in the enterprise industry. We've made tremendous gains in our structure and basically our feet on the street and our focus for enterprise. We've also created, and you'll see we just put some news on that out today, a very easy way to do a virtual machine migration from if you're running our competitors x86 platform, to run a very, very straight button, push button, virtual machine, VM migration over to EPYC. And so we think that will be a big assist as well with enterprise. And there's a third category, and that's where people are running both in the cloud and on-prem. And they're moving workloads back and forth. We do that at our IT instantiation at AMD. And we're extremely well suited for that case. We have such a high presence in the cloud. You can mimic that in your on-prem and shift your workloads back seamlessly.

And when you think about those bigger cloud workloads, the more internal facing, can you talk about the role of the sort of more cloud native products like Bergamo that you're coming out with? And is that market going to be more competitive with other architectures? We've seen some arm encroach and things like that. Can you talk to that?

Yes, cloud native. Some of the -- you look at some of these cloud native applications, Java being one of them, and there's many others, Java workloads is what I referred to. They don't need necessarily the highest frequency. They need a lot of throughput. They need a lot of core, core density. They don't need to run at the highest performance, the highest frequency. And so what is the nature of the culture that we have at AMD is we listen. We listen to our customers. We see where workloads are going. In this case, we saw that. And we are quickly pivot our road map to add a swim lane, to add a product, I mean the dense cores you've mentioned the codenamed Bergamo. So Genoa -- instruction sets. You run the same code exactly as you optimize on Genoa, you can run it on Bergamo. It's optimized for cloud native. So instead of 96-core cluster, it's 128 cores. Doesn't run at the same peak frequency of Genoa and it's optimized for cloud native. So we think we're very, very excited about coming out first half of this year, right on track. And it really provides that compute density advantage for cloud native. But even more importantly, just like when we added our 3D stack cash that accelerated workloads in such as high-performance compute and seismic analysis, et cetera, it shows you that AMD is going to continue to listen to the customers, understand where the workloads are going, Joe, and make sure that we're there and we're there with our economy of scale and making microprocessors. And that is a key facet that we bring to our customers. Anyone can make their own silicon and optimize it for a specific workload. We have the advantage of making millions and millions of processors each year with finely-tuned process [ both ] development, test, manufacturing, reliability to serve these markets at the enterprise and data center, greater quality required.

I wonder if we could shift to within data centers still, the role of GPUs. You guys have been successful in cloud gaming, in areas like supercomputers. Where are you when it comes to more of the machine learning type workloads in terms of progress there? And where is that going to go, going forward?

Thanks for the question, Joe. We had a very thoughtful approach when we saw where AI was going because we knew the type of compute that it needed. It really gets accelerated by parallel processing. It's the nature of the algorithms when you do a -- both a forward propagation and a reverse propagation to complete that learning loop. It's very much needs parallel processing, vector processing. And so our strategy was, we knew we could develop the leading-edge hardware. And so we started on that path with our AMD Instinct line of GPU processors. And we focused first on HPC because HPC, given the deep heritage we had, we had a very, very good software stack already to build upon. It's called the ROCm stack. And so ROCm 4.0 was released a couple of years ago that was production level for HPC. And having this type of leading-edge hardware and a production HPC stack led to key wins across the HPC sector. While that entire development effort was going on in HPC, we were in parallel optimizing for AI workloads, adding support for PyTorch, adding support for TensorFlow, and other frameworks, and as well as optimizing within our GPU instruction set to speed the processing. Really excited with the advent of late last year of ROCm 5.0. The fifth generation of our software stack ROCm is production level for AI. Now if you go to PyTorch, you see only 2 software stacks rated at production level on Linux, and that is AMD and our GPU compute competitor NVIDIA. So we're really, really pleased with the progress that we're making. And so MI250, Instinct 250, is out in production today, winning in HPC. And now, at the early innings of growing on that now production level of AI software stack, Instinct is at Azure. You've heard announcements last year that Azure was standing up MI250 and starting -- and really, they've been tremendous partners with us to tune workloads needed to run on MI250. And so that's the first marker is we're out of the starting gate with MI250. We're incredibly excited about the performance with MI250. It takes on the A100 head on. And so our strategy is when you have a hardware that's completely competitive and the software, which is coming up rapidly, but not covering every market is to focus where we target that software stack. We're targeted, Joe, on the hyperscalers, rather than trying to hit every vertical market that may need tremendous amount of code. And so it is well underway with MI250. And MI300, our next generation, is on track. We'll be announcing that second half of this year. And that's a beast because it takes 4 GPU CPUs -- 4 Genoa CPUs, and embeds it with our graphics processing. So just like we shipped combined CPU and GPU for years in PC and embedded markets, we've now brought that approach to the data center with the Instinct 300. So we couldn't be more excited to launch that. Next year, it's going to bring commanding HPC, and it was equally optimized for AI. So again, coming to market and being announced second half of this year and ramping in 2024.

Great. And then last data center question, I guess, on inference with the focus now on generative AI and cost per query as people start to move to those types of workloads. How does AMD view that? Is that a CPU problem to solve? Is it a GPU problem to solve? And what do you think AMD's role is going to be?

Joe, great question. The answer is, it's all of that. Why is that? Inferencing is simply everywhere now. It's in -- it's going to be -- it's either there or going to be in almost every device that you interface with, whether it be a device on your home on the wall that's doing natural language recognition or processing of some of the telemetry it has, all the way to the biggest compute devices which are out there. And optimizing the way that you do your work, the way that you program and code, the way that you write your next speech show that you're giving, you might be tapping into ChatGPT and using the inference capabilities there. But every example I gave requires a different inferencing. Let's start with the latest example. You want to use ChatGPT to help you in that next speech, you're going to need an inferencing capability that mimics that large language model where the training was done in a GPU, and certainly with the larger memory space. It may not have to be a GPU, but it's going to have to really be able to process that large language model. And then as you step through the other applications, there's different requirements. So at AMD, what we're trying to bring is the right inferencing solution for the right inferencing problem. And so it is GPU for a very dense, large language model inferencing and typically done in hyperscale. It's leveraging the AI engine that we've successfully is shipping, that came to AMD in the Xilinx acquisition, a performance leading AI engine with a robust production level inference optimization stack. That's shipping in our Xilinx adaptive compute product line, and it's been permeated over time elsewhere across our product road map. Again, inferencing to run everywhere. We started already -- Lisa Su announced at CES, our Ryzen 7000 product line that has embedded in it an AI engine that accelerates inference workloads on PC or embedded applications. And then go back to the CPU. CPU is the workhorse of inferencing today because if it's a lower level inference that can run on the CPU that you actually have with you in your PC today or in your data center product you have today that's still where the bulk of inferencing is done. And to that end, we added vector and neural net instruction acceleration and AVX-512 into our fourth generation Zen, which is in Genoa and the EPYC line, and it's in our Ryzen 7000, which I mentioned just a moment ago for PCs and embedded.

Great. And I did want to ask just one question that I'm getting a lot in the last couple of days. Chinese server company, Inspur, was added to the Entity List last week. Can you talk to whether is it too early to say how much impact there is to AMD?

Well, AMD, like everyone in our industry, we, of course, follow all of the guidelines of export controls from the U.S. government, including, of course, the Entity List. And so we did see that news. But we're seeking clarification as I think the rest of the industry is because Inspur is a large holding company. It serves many, many markets. So we'll -- we're looking to get clarification on those guidelines. Of course, we'll abide by the resulting information.

Great. So shifting gears to Xilinx, which you guys have now owned for a while. Where have you seen the synergies of that acquisition? Is it primarily bringing AMD CPUs into the embedded space? Do you still see opportunities for Xilinx in the compute space? Just talk generally to the opportunity there.

Joe, we actually just hit on February 14, our 1-year anniversary of closing acquisition of Xilinx. And we could not be happier with how the integration has gone. The cultures of the company, we're very, very aligned. And so the engineers have just really enjoyed coming together. And you see that synergy across our product implementation and you see it in terms of how we're going to market with these products together. When we talked about the acquisition, we talked about what was very, very clear at the time, look at the embedded markets, that's where you started with. So you can look at the comm sector where with the advent of 5G, we now, the new AMD, has an end-to-end from the control plane with the presence we already have with our EPYC product line. By the way, it's getting -- that has expanded because the fourth generation EPYC adds Ciena, a telco optimized version of our fourth-generation EPYC coming out second half of this year. But that's now paired with the rich portfolio Xilinx had to attack the comms and 5G space with a large installed base. So that was sort of the obvious synergy that we expected. You couldn't have seen that more on display than Mobile World Congress just the prior week, where you can just read the press, it was well received the depth and breadth and competitiveness of the AMD portfolio that we have end-to-end to address telco. But beyond that, what we've seen is the synergies in other embedded markets like automotive. So traditional, the legacy AMD headwinds in automotive and you look at, for instance, the Tesla infotainment and others. And so many legacy Xilinx wins across the embedded space in LiDAR and in parking and image recognition and analysis, et cetera. So that's gone as expected. What's gone beyond our expectation is what we've done with AI engine. Already launching that AI engine that came from Xilinx and the -- as I said, the PC embedded space, as I mentioned earlier in my comments. And we see opportunities for it elsewhere in the portfolio. And then more importantly, the software stack that Victor Peng and the Xilinx team brought in on AI was very mature, and it pairs very, very well with ROCm, the existing AI software stack that AMD has. We announced recently that we're bringing all of that together under Victor Peng. So all of our AI efforts are centralized under Victor and a very, very concerted AI effort across AMD. Very excited about that.

Great. So I did want to ask you about client compute. The market share has gotten very noisy in a challenging market. There's been inventory kind of build from everywhere. Where do you think you are today from a market share standpoint? Intel seems more competitive in some of the desktop areas, but you guys are still getting significant wins in commercial and notebook. Can you just talk in general to how you see that shaking out?

We have a very competitive road map for client and a great road map coming forward. It is very synergistic to the AMD portfolio. Again, I talked to you about the modular approach that we adopted years ago. And what the PC market benefits from at AMD is all of the focus that we have on performance and performance for lot of energy flows right into the PC market synergistically with our data center and server compute market. So when you look at our road map, you look at the latest desktop announcement is the Ryzen 7900. And it has a commanding desktop leadership. It has a vertically stacked 3D cache on top of the CPU chip, which allows us to simply stream in the most demanding applications like gaming, which is one of the highest purchasers of high-performance desktops, but also work staging applications, et cetera, a benefit from this construct. And you see the portfolio synergy, that type of investment to 3D stack with hybrid bonding cache memory right on top of CPU could not happen without the modular approach in sharing that same technology with our server. Likewise, on notebook, you see that synergy with our graphics market. So the Ryzen 7000 series has not only that fourth generation Zen processor but RDNA 3, our latest generation of graphics processing. It is an absolute leader in notebook processing, absolute leader in battery life, and it's the first x86 processor to embed AI acceleration with the AI engine that I mentioned earlier. So a leadership road map. It's a very competitive market. The TAM is expected to be slightly down this year, about 260 million units is what we are projecting. But it's a great market for us, and it really leverages the synergy of our portfolio.

Okay. Great. Well, that brings us up to the end of our time. So Mark, thank you so much. Appreciate it.

Thank you very much for having me, Joe.

Thank you.