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

Thank you very much for everybody coming out to the keynote session here of TD Cowen's TMT Conference. My name is Matt Ramsay. I work in the semiconductor research team. We are really, really excited to have Mark Papermaster from AMD, who's been the CTO of the company for a long time after a distinguished career at IBM, at Apple, at Cisco and has been CTO for the last 10 or 12 years, and the products that his team have engineered in the market have changed the AMD market cap by 150x, something like that. So I mean, maybe that's good. So we're -- Mark's been a friend for a long time, and we're delighted to have him here. So thank you very much, Mark, for coming.

Thanks for having me.

People would ask me, why we chose to invite AMD to be the keynote here? And I think it's pretty simple. AI is affecting computing at literally every level from the data center to the PC, embedded wireless infrastructure, gaming, gaming consoles, and AMD is the one company that has either a #1 or #2 position in literally every one of those markets, and so I think Mark is a great person to give us all some perspective on what's going on high-performance computing and AI. So on behalf of everyone here before we get started into the fun stuff, Mark, just thank you for your partnership.

Where I wanted to start this conversation is a little bit backward-looking actually. It's been interesting to watch your company grow from a couple of billion in market cap to $270 billion or whatever it is this morning. And it would be interesting if you could just kind of talk about how when yourself and Lisa came to AMD, where you started from, what you learned along the way, big pivotal moments in the company's history and maybe some things that, man, if I had that one back, I'd do something a little different?

Sure. Well, it's been an incredible journey, and we have a long way to go on the journey. But if we take a moment and look back, I came into this role as Matt said, just over 12 years ago. And at that point, I was a 30-year veteran in the industry, having had the opportunity to work on some great products, all around high performance, all about really making a difference in computation. And I came to AMD knowing that it had a rich history of IP building blocks, of having had success in CPU and GPU through the ATI acquisition and great technologists. I had worked with a number of technical leaders at AMD over my career prior to joining, and so it was really a thrill to have that opportunity. And so if you look back on the journey, the opportunity taken was to trust the 30 years of experience and to know that as my intuition told me that workloads were very much moving to needing more and more high-performance. Think about it, it was, again, 12 years ago, AI was just starting. So acceleration was really coming into its own. And I looked at where the trends of putting that technology together and the building blocks AMD had, and I knew that we had all the building blocks to be the right provider in that direction. And so when you think about what are the tough decisions made, it was to trust that intuition. So you have to have the courage to trust that and to lean in on the high-performance road map as we did. First, rebuilding Zen and our CPU franchise because that's the fastest path to revenue. We had well established ourselves in the x86 ecosystem and so rebuilding that Zen CPU prowess. But we didn't trail far behind in the GPU. What most people don't realize is that we immediately went after bids and won bids with the Department of Energy that built up our GPU capability and the software, a ROCm stack, starting with HPC and then, of course, expanding that to AI. So we've been at it over a decade, Matt, of high performance, and we made a number of tough decisions under Lisa's leadership, my role in CTO, that just leaned in on our strengths and not getting distracted by all the other things that are going on in the industry. And if you ask me, the biggest thing that surprised me most, it was simply how fast that the uptick of AI demand has gone since the advent of Generative AI because it didn't change the secular trend that we had been marching to on the high-performance compute capability, but Generative AI fundamentally plummeted the barrier to leverage and get access to high-performance devices, because now it's a spoken interface. It's into that prompt the U.S., then it's tapping vast computing and vast data. So that's probably been the biggest surprise that I had, and I think many of us here in the room.

Yes, I think that's right. I mean we've -- in my team, anyway, we've had the thesis that for semiconductors, right? If you look at basically every single vertical of the economy, more CapEx and more OpEx go into computing. And I think from my point of view, Generative AI is just a massive accelerant of that, and that's what we're seeing. So it's interesting, right? The last 18 months, 24 months have been the fastest change I've seen in the data center market. I think that's pretty much consensus. From your seat as CTO, what big changes are you making in allocation of R&D dollars of hiring a focus just across the organization that you manage. There has to be some -- I mean products now for AMD, like the MI300 and the successors to that, are the most important products at the company now. And if I had investor meetings 36 months ago, no one would ask about them. And I think that's true of your big competitor as well. It's true of everyone in high-performance computing. So what are the big changes you've made in the last 12 months?

Well, we run a very well-oiled capital allocation process. So we look every year as we're planning ahead to where we're going to spend our R&D. And, clearly, if you look at where we're going to resource the last 2 years, it's been vastly dominant for AI. We've added hundreds of engineers, actually, I'll have to say, thousands of engineers across AI because we have AI enabled our entire portfolio. We're shipping AI enablement across our entire portfolio. And the last couple of years have leaned in even further on the data center AI and really bringing to market our MI300, which expanded us from production level with dominant on HPC and really shifted us over to having 2 versions, one, which is HPC-optimized, MI300A, which is powering the Lawrence Livermore Supercomputer, is being stood up right now. And MI300X, which is what Microsoft announced just last week at their build, which is in production in Azure. And so one of the things, though, Matt, that's really helped us is because of the modular design approach that we adopted to field the turnaround at AMD, it makes us more agile and allows us to have high reuse. So the fact that we're leaning in to the extent we are in AI, it doesn't mean that we stop investment on other core businesses that we have. They may not have the explosive growth rate that we're seeing in data center AI, they have strong secular growth each year, and they share so many of the building blocks that we have when you think about how the high compute devices that we have, how do they connect to memory, how do they connect to IO, how we're handling, for instance, the neural net engine that we reuse, it came into the Xilinx application. It's already in our Ryzen PCs and yet it's also in our embedded devices. So a very thoughtful process and one that really leverages modular hardware and modular software, we're going to have at the end of this year, a unified software stack across every AI implementation we have, whether you're running on the GPU, our CPU, embedded or PC, one unified software stack.

That makes a lot of sense. So I wanted to shift the conversation a little bit because as you and I talked leading into this, the audience here is a little bit wider than just us knuckle heads in semiconductors. So one of the things that Lisa and yourself orchestrated at the company is the AMD's internal IT organization actually reports into you as CTO. And I think what a lot of us are wondering just broadly is, how are broader enterprises actually trying to tackle AI? Things are moving so rapidly at the hyperscale that -- and those are the most well resourced smartest companies in the world, they're doing their thing. But broader enterprise, every boardroom meeting has to be about what are you doing in AI, and how are you adopting it. So I'd be interested to hear your perspective as maybe a customer of your AI business in the AMD IT organization, and what the challenges are, what you're excited about, et cetera?

Yes. It was definitely a smart move by Lisa. I think it was at this point, 9 years ago, where I had IT fall under the CTO and Head of Technology and Engineering role that I play because at every meeting that we have as we're thinking about that next-generation CPU or our road map that we're tackling, we're also thinking about the very problems that we have as a Fortune 250 company in tackling all of the huge compute demands that we face, which grow at every product turn, and we have over 150 billion transistors on the MI300. So you can imagine the compute required. And now the AI capability. So our Board, like every CIO has the same question to us as every other CIO. And that is, what are you doing with AI? How are you improving your processes, improving your productivity? And what we're -- all of us are doing is we're looking at the data we have and the processes we have and are we truly leveraging the data, harnessing the data, are we creating models from that data that can solve problems faster? And that's across, in our case, engineering, it's across our supply chain, it's across our financial and our HR processes, and we're having advantage because we're a technology company. So what do we do? We immediately started optimizing our own models using open source models, training them on the data, now fine-tuning them to get a higher accuracy. But that helps me and my CIO and as well as the rest of the leadership in AMD, as we talk to other CIOs, they're going through the same journey. They're wanting to run hybrid. Sometimes they're running their most demanding task in the cloud, but they also want to have a local capability, some of the models and the data are so precious, they actually don't want it leaving on-prem. And we are able to walk them through based on our own experience, different options that they have. And of course, explain to them how AMD has brought competition and can actually bring tremendous AI solutions to it with our product portfolio to bear.

So I guess, I made this joke last night at dinner, but it went well, so we'll try it again. My first question on the products is going to be on gaming. I'm sorry, we're not going to do that. We're going to do it on AI, bad crowd here. I just wanted you to -- AMD -- this has been -- since yourself and Lisa came to the company, the most important business as you restarted it was going to be data center. And that's had different instantiations from turning the server business around completely from a standstill to leadership position pretty much unquestioned across most metrics to evolving in folding Xilinx into the portfolio and then eventually Pensando acquisition and some more work you're doing on networking and now obviously, Generative AI has gone vertical, and we're watching the innovations there. But I just want to start -- like first order of tenets of your data center strategy, like is it just high-performance leadership across the board, no matter what application? I just want to understand how you think about the goals of the data center franchise?

Yes. Data center, I grew up in data center. Those who don't know my history, I had over 2 decades at IBM providing technology to data centers across the world. And what you learn when you grow up in that environment, that first and foremost, when you serve the data center, you're serving customers that are entrusting their business to you. So yes, they want the economics. So it's a given that you have to bring total cost of ownership advantage. That's a tenant. That is certainly very clear to me as I took this role and we started driving our road map to absolutely bring that total cost of ownership. But equally, it has to be reliable. You have to build in the fact that it can be basically hit the kind of five 9s, the reliability levels, the data centers need to know that you're going to be up and running and meet your serviceability. Beyond that, you have to be a trusted partner. You have to execute. And so in our data center strategy, it's about not only providing a leadership technology, not only providing one that's reliable and has the standards that it needs to meet the diversified product SKUs that our customers may need, but that we'll hit every product cycle when we commit it on time. That's what it means to service a data center, and that's what we've been focused on. And I mentioned that we understood where technology was going at AMD. So we saw that Moore's Law was slowing. And so for the data center, to provide that TCO gain, generation in and generation out, you have to, in your strategy encompass the technology trends. And we developed a modular approach, one that allowed us to be first to the industry of chiplets. And that allowed us to have our compute engines on the cutting-edge node and then our connections to memory I/O and support circuitry on prior generation nodes. And we adopted, we were the first to adopt this chiplet approach in both lateral 2.5D and vertical 3D stacking. And so Matt, I think the story is, if you want to play in data center, you've got to have a broad approach that's customer-centric, you're listening to where the workloads are going, and yet you have a portfolio of technology that can bring the best solution, the best semiconductor node to the problem at hand.

Definitely makes sense. It's funny that you talk about the history of the data center business like that. I remember you and I being on stage 8 or 9 years ago, one of these things and being asked if you guys were just taping desktop parts together to try to make servers? And I was like, well, we've come a long way since then, with leadership basically across the board in the server industry. So I wanted to ask, and we're kind of at a similar point now with Generative AI, right? There's a market leader that's fairly obvious to the investors in the room. And there's a very deep set of relationships AMD has built with all of those key customers, both on hardware, on software, on design, and also a need, I think, a recognition by all of those customers that they would like to have some diversity in their supply, some customization in their partnership et cetera. So maybe you could just give us a little bit of a status report on where you are now? Is it similar to what happened in server 7, 8 years ago? Is it not? What the relationships are and what the road map looks like right now?

Our history at AMD of what we've done in the data center with server is a huge boost to the efforts that we now have as we expand into GP, GPU in the data center. The reason for that is straightforward. It's what I said a moment ago, you have to establish yourself as a trusted supplier. And you need to establish the relationships from everything from the product itself, are you listening? Are you incorporating the customer requirements? And are you establishing the ties with all the way through the field organization on hardware and software. And so when you think about the data center and what we did with X86, it was not only to provide our generation after generation, we're shipping our fourth gen of EPYC and soon in the second half of this year, to ship churn, our fifth generation of EPYC servers. And it's been done in such a way where we've listened to customers, understand the requirements, hardware and software unfolded in. That communication, that trust is directly applicable. What the learning is for data center GPU for us, is not the hardware elements. We were tried and true in terms of the hardware elements. But where we had to race and putting together the software structure is building out in a rapidly changing AI software and ISV environment. As soon as you think you have the complete set of software support, a model changes. And so you have to in that relationship with the largest hyperscalers driving the most demanding AI models, you have to develop a set of partnership and listening skills and agility in the road map that is beyond what we had ever done before. Our time to market of the MI300 was vastly faster than any data center product that we had ever done in the history of AMD. And then the resulting revenue ramp has been the fastest that we have ever experienced at AMD. The rate and pace of that AI is moving is really astounding. And it's required us to almost put the company on its head in terms of reinventing ourselves and ensuring that we can react at that kind of work speed. But that's exactly what we've done with the MI300 program and the road map, which follows it.

Yes, a couple of follow-ups there. I think one is on that pace and cadence. Your large competitor recently very publicly changed the cadence of their product introductions for once every 2 years, give or take, to about once every year. And there's 2 sides to that coin. One, do you guys think you can keep up with that cadence? And the other side is, do you think the customers can assimilate technology at that cadence? So that's kind of the first part. And then the second part is, you mentioned on the forward road map, maybe this isn't the right forum, but I wouldn't be doing my job if I didn't ask. So anything -- any comments on the forward road map? I know there's a lot of folks that are kind of interested in that, let's say.

Well, first, let's talk about our commitment to road map at AMD. Again, what did you see us do as we entered the data center CPU market. We set an aggressive road map. We committed to it, and we executed on the regular drumbeat. That drumbeat for the x86 server market is about every 18 months. You'll see a new product introduction. And honestly, that is what I think that the rate of absorption is of our customers' industry. It typically is 18 to actually 24 months. But we are at our hearts, we are a nimble competitor. And what the market demands, we will deliver. We were on an annual cadence. So as we refresh our Ryzen line of PC chips every year, where the annual 12-month cadence and what we're doing for our GPU products. And so we'll let the market decide. We modified our cadence, and we'll be sharing more as the year progresses with our road map. But we've accelerated our road map as well. And if our customers can consume productivity enhancements in that GP road map at a 12-month cadence we're certainly prepared to do so. Again, we've been doubling down on our R&D investments. And we also leverage the chiplet methodology that I mentioned earlier. I mean you think about what we did with MI300, no, I'll just hold up an MI300 module here. The version I'm showing you here is MI300x. This is what Microsoft announced last week, it is in production in Azure. It's a virtualized instance. So it's leveraging that same virtualization capability, hardware-based virtualization. It's very, very efficient. We've had on CPU. That's in multiple generations of instinct, including MI300. It's been in production for video survey, now in production for AI, and Azure. And with that chiplet methodology, we pivoted to offer both an HPC optimized version of MI300 and an AI optimized. So going forward, as you think about how do we handle changes and increase cadence, we will leverage our modularity as you pick and choose which lever are you tapping across memory enhancements, I/O enhancements, new GPU cores, other accelerators. So new math formats that you have. So the name of the game going forward in this astounding pace that we are all on, in terms of AI compute, is holistic design, is what I call it. You have to bring all of those elements together. You have to work across the stack. So Matt, that's how we're going to respond to the cadence, and we're going to do it with an ecosystem. We're committed to our development leveraging our open source in terms of our software stack and open standards in terms of how we put solutions together right through the rack level.

It's interesting, what you just mentioned there was actually my next question. So maybe your eyes are really good. The one of the things that has happened through generative AI over the last 24 months is, NVIDIA has changed itself from a GPU company to a card company to a server company to a rack scale company. And I think, when you look at what they've done, I ask myself as an analyst, okay, well, what companies actually have all the building blocks and that technology kick actually can fight back against that? And the list gets really short, really fast. And so I'd like to hear what you and your team are thinking about in terms of AMD's expansion to the system level and the rack level and the data center level, frankly, as you partner with customers.

Yes. Again, I grew up for over 2 decades at IBM. And so I'm well versed in what it takes to put together a system solutions and optimize system solutions. I think our competitor is almost taking the mainframe approach, putting the whole solution together and you can very finally optimize that, and there's certainly the merits to that approach. But the other side of the coin and where AMD will very much differentiate is you bring more people with you and you can offer more options when you commit yourself to not only an open software that brings a whole development community with you, but open standards is how the solutions are put together. Open standards don't mean that everything gets tied up in committee. It's typically done best with a small focused consortia that decide the pinch points of what is the best way to put these system solutions together. Matt, we've been very actively engaged in the open compute platform, OCP. If you look at our platform that we shipped the MI300 today, it's an OCP platform. And so is our competitors today with the H-100. So you can take out a competitive slide, you can drop in an AMD MI300 sled. And so if you change the differentiation point to be the rack, our strategy is going to hold. We're going to continue to differentiate all the way through the rack level. We have engineers skilled at each of the areas of optimizing around rack level implementation, but we're going to do it in such a way where we partner and we create an ecosystem. And what that's going to create is a broader set of SKUs of offerings, and it's going to create a broader ecosystem. So we're going to -- our strategy is to take the ecosystem with us as we move forward and as we optimize at a stack level. It is not new in the data center to optimize at system level. We've been doing it for generations in x86 servers. And it's not that dissimilar as you think through what is the workload? How is it running? How are you scaling out to achieve it? And are you partnering with ISVs thoroughly to optimize that is a constant across data center applications, and that is the equation and strategy that we're bringing to data center GPU.

Makes a lot of sense. There's only going to be a couple of instances where I'm going to be able to take a question or 2 from the audience. So if anybody has something they want to try to bounce off of Mark, we'll get you the microphone. And Erica is right in the front here and caught my eye. So...

First of all, thanks, Mark, for being here with us and spending time with us, so we can learn from you. I was wondering you kind of alluded to this earlier in response to a question from Matt, but how important has it been your initial success with the hyperscalers and enterprises with whom you've partnered with on MI300 on the hardware side. But importantly, on the software side, how has that been critical to that success with MI300, what have you learned? And how is that helping you win new customers beyond the Microsoft Meta, Oracle that I think we're aware of? And like can you take those learnings and use them with other customers, I guess, it's more of an IP question as well.

Yes. It's a great question. We are a collaboration company at AMD. I mean that's what we are famous for, and we listen extremely well. And that's what we've done in each of the markets that we serve. But for data center GPU, we had to earn the seat at the table to be that collaborator. So we had to demonstrate that we had brought our software stack, the ROCm stack and AI up to a proficiency and that we have the compute capability up to efficiency. And that's what we did in the instinct road map building up to the MI300. And so the deep partnership we've had with several hyperscaler players has been absolutely essential because these are the companies that are working on the absolute cutting edge of technology of the next generation large language models. And so when you service these largest kind of opportunities and you have that seat at the table, you know the problems you need to solve. And it's invaluable of not running down the rabbit trail of something that is solving a problem, but not the most important problem at the hand, and so what the collaboration has done with us with Microsoft, Meta and others has really helped us prioritize what matters most to them, which, therefore, is what matters to the many internal customers and thousands of external customers that run on their third-party platforms. So that part has been invaluable. But also, I think we can plainly state that it was the pull to have competition in the market that we needed. So one, we had to be credible, and we had to have established that trust that we did through the -- our buildup of the EPYC x86 server line. But we're not confused. We have to deliver here to very, very aggressive performance. We've made tremendous strides in listening very well and folding improvements into our software and hardware stack. So we announced ROCm 6.0, is in production level at the end of last year, yet we just announced a 6.1. And what does 6.1 do? It's folding in the learnings as we've come up to that production ramp that you now saw announced by Microsoft last week and the same with the feedback we have from the other partners. So we'll continue to collaborate. We'll continue to move very, very quickly and iterate our hardware and software in this fast-paced environment. But we're really pleased to have earned the opportunity to be the only other production level GP, GPU in the industry today with the kind of a competitive level of training and leadership level we have today. And again, you asked about our road map, I won't be going through details here today, but we have increased the cadence of our road map, and we'll be sharing more as the year progresses, give specific details.