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

Hi. Welcome every, everybody, to our fireside with AMD and Dan McNamara, Head of Server. Dan, maybe if you could talk a little bit about yourself, your experience, your background as head of Server at AMD. And this kind of very important role, as I was saying to you before, all eyes are on server here and particularly with this Xilinx acquisition as well. Can you give us all an idea of why you're here, what you're going to do for AMD, what you're going to do for this division and how you're going to change the industry here?

Sure, Chris. And I just wanted to say thanks for having me. It's a great pleasure to join, and thank you for everyone who joined the webcast also. So Chris, it's -- my background is very relevant to the combined company today. And let me give you a little feel for my background and then talk a little bit about joining AMD and the thought process there and then how I see it going forward. So for about 15 years, previous to AMD, I was in the FPGA business. So I held leadership roles across applications engineering and business units at Altera. And then when the Altera acquisition from Intel happened, I actually ran that group called Programmable Solutions Group at Intel, and then added network compute and ASIC groups over the course of 4 years at Intel. So bringing Xilinx in at this point is not only exciting to me and kind of fits a sweet spot, it actually creates what I would call the best and most complete portfolio in the marketplace at a time when it's so important. We're in the compute mega cycle and our products really align. If you think about CPUs, GPUs, traditional FPGAs but also adaptive SoCs, which Victor and team really drove and are ahead of the market on, obviously, integrated software ASIC capability. So we really bring to bear an incredible portfolio here and it's quite exciting to have them on board. And then the other thing I would say about bringing on Xilinx, although I competed with them for 15 years, having them on board, they have an extremely similar culture to AMD and they are also executing at the same cadence as AMD. And so let me talk a little bit about joining AMD because 2 years ago, it was a really important decision for me. And I recall that there were 3 main elements of why I was drawn and ultimately joined. And the first one was really their vision for -- to change the landscape of compute in the data center. It was built on the highest performance and the best TCO. And also, they had a vision of how the heterogeneous compute world would evolve. So I thought that was really compelling. Secondly, the technology and the execution. So starting with IP and chiplets to advanced process technology with TSMC, to advanced packaging and the software development, it was clear that they were ahead of the industry, and also, they're on an execution cadence that I know the industry saw. But most importantly, it was building trust with the customer base, which was really, really important. And then lastly and very importantly is really what I just mentioned, which is culture. So every single executive at AMD has a singular mindset about listening to the customer and delivering what we say we'll do, meeting their requirements, and that is really a critical piece, too. So as you look at this and you kind of fast forward back to today, I couldn't be more excited about leading the Server business at this juncture. As I mentioned, the megatrends are attuned to our products at a time when our products are in probably the best shape in the industry and we really feel good about that. And also, the complete software set across these products is really an impressive portfolio that we're going to bring to market. So when you look back again on that strategy and vision, I would say that after 2 years, it's not only coming to fruition but it's actually accelerating, as you saw from our Q4 results and full year 2021 results. So we're pretty excited about where we are today and the promise of where we're going within this compute mega cycle.

Yes. Thanks for that, Dan. That was a great intro. AMD has recently announced, I believe, a larger data center TAM when it closed Xilinx. How should we think about this bigger TAM? And what are your market share goals more broadly, not just for the industry but also as we think about cloud, as we think about enterprise, as we think about HPC?

So we see the data center market, and we just raised this to about $50 billion from our previous number of about $35 billion. And we see it across really 3 areas. One is just expanding compute and the trends in the industry, so there's an increase in demand. Our expanded portfolio that we're expanding here from both 7-nanometer and going into 5s to address a wider swath of workloads. And then the addition, obviously, of Xilinx with both FPGAs and adaptive SoCs, so we see that increasing to about $50 billion. Xilinx brings added opportunity across all of the segments. But let me give you a little bit of feel for where we are on the segment. So from a cloud standpoint, clearly, our strongest position today, we see very, very solid growth in public instances. In Q4 alone, we grew it quite dramatically and I think we exited the year with about 400 public instances, which was pretty dramatic growth. So we're seeing that accelerate based on Milan getting out there. And actually, we're -- there's new releases of instances almost every day here. But more importantly, the internal properties. And the internal properties, think about internal properties as more of the Platform-as-a-Service or Software-as-a-Service offering from a traditional cloud service provider. We are very, very strong there and that was really the majority of our share. And the reason being there is the value we bring in terms of performance and overall TCO typically will go to the bottom line of those separate internal property businesses. So the leaders of those businesses see, whether it was Rome in the beginning and now Milan, see it as a huge win for their business. So it's -- we see continued growth in internal properties, and we also see that the public instance market will continue to grow. When you look at HPC, we're doing extremely well. And with our top-of-the-stack Milan at 64-core and sort of the road map we have, we feel very, very good about continued progress there. But just in Q4 alone, if you think back in November, we were -- we tripled our place in the top 500 supercomputer list to 73. And even more importantly, we are 8 of the top 10 in the supercomputing 500 Green List, which speaks to not only our performance but our performance per watt, which is a really important piece for not only supercomputers but in the enterprise on sustainability goals. So we see traction continuing there quite aggressively. And then enterprise. So it's probably the place where we're most underrepresented but we did hit an inflection point last year with the Milan launch. And we tuned that launch to true enterprise type workloads. We brought out a stronger ecosystem, almost 2.5x the solutions around. We brought out over 100 platforms across all of our top OEM partners. And we delivered a per-core performance advantage that they did not have with Rome. So we really feel like that traction will continue. And we have, currently today, over 250 world records in performance with Milan, and a lot of those are tuned to enterprise-type workloads. So we feel really good about enterprise going forward and continuing that traction. When you take a step back with Xilinx, as I mentioned, they add to all of these in terms of the opportunity, but they also add a nice opportunity for us to enhance our opportunity in 5G and network function virtualization in the edge. They have strong relationships there. They've been in that business for quite a while, and we really think that the combined company around 5G, both core and edge and then just the broader tools of edge compute, we feel very, very strongly about. So the net is we're really operating in an expanding TAM at a time when we really believe our portfolio is delivering the maximum customer value. So it's a pretty exciting time for us.

Great. Something you guys really pioneered here was multichip modules, I think, with your first EPYC. And maybe there were some growing pains there but you quickly fixed that. And you were a pioneer here, and now it looks like Intel is copying that approach with Sapphire Rapids. I guess, first of all, is this the future of where we're going with server compute? And what is it that hyperscalers like about this approach? And does it give them an ability to customize for different environments if you're doing this multichip approach?

Yes, Chris, it's been an amazing innovative journey for AMD, in my opinion. And it all started back in 2017 with Naples. And that was a 14-nanometer die and it was the first sort of introduction of chiplets. So it was kind of the reentry into server for AMD. And then we doubled down in August of 2019 of Rome, which was basically a 2x gen-on-gen performance, right? Went to 7-nanometer, delivered 64-cores top of stack solution, PCI Express Gen 4, 32 megahertz DDR, and we really delivered a strong, strong product to the market with Rome. And then last year, March, as I mentioned, the Milan launch. Milan, we came back at 7-nanometer and the same stack pretty much, where we delivered 25% performance per watt improvement and 19% IPC uplift. We have tremendous momentum with Milan today, and we actually believe Milan stacks up extremely well against Sapphire Rapids today. But I think the question you had was, is this the future? We certainly believe it is. And the reason being is if you tie it back to the hyperscalers, what they see are 2 things from us. First is what I talked about on internal properties. The high density, the socket level throughput, the performance per watt and most of all the TCO. So if you can deliver more VMs in a rack, it's a win for cloud and hyperscalers. So that's one of the main reasons. And this chiplet architecture is really the way we got to that high density and we'll continue to go there. But secondly, the point you made is really key. The customization capability that we have based on our chiplet architecture and our advanced packaging is really a complete win. Most of our SKUs for cloud are custom and a couple of different examples. One is you can -- we can design a part for maximum cores and maximum power to get absolutely the most performance out of that CPU, or we can take a power ceiling and deliver the maximum performance within that ceiling and optimize on core count and memory bandwidth. So we bring a really nice flexibility to the cloud provider. And as we engage with customers and we continue to engage with our road map, we see the customers are engaging with us over 3 to 5 years in terms of what are we bringing to market. And we believe that certainly, chiplets will stay, certainly increased core counts and we will continue to deliver this customization. And the interesting thing about customization is as we expand our portfolio with the likes of Xilinx and GPUs, it gives us a lot of flexibility to bring to the hyperscaler going forward. So we feel really, really good about our road map and our traction here along with this journey that we're on, and we're going to continue this journey very aggressively going forward.

Yes. It's been -- it has been quite a quick and fast journey, too. You've had some announcements recently to the server portfolio. Milan-X, that's the HBM version. Maybe talk about that and whether you think that's -- HBM is going to gain quite a bit of traction. But of course, Genoa and you recently announced Bergamo, I believe the higher core count version there in 2023, Bergamo in 2023, but Genoa before the end of the year here. Maybe you can talk about that and how you think this is going to allow you to compete against Sapphire Rapids and also anything that might be coming from ARM as well. And I'll have a quick follow-up after that.

So I mentioned our road map a couple of times now. We are very, very excited about our road map, both at 7-nanometer today and going to 5-nanometer. And what happened, Chris, is while we were talking to our customers, it became very, very clear that workloads were diversifying or even fragmenting across the data center. And it became very, very clear that a singular view of general-purpose compute was not going to deliver the performance and optimizations that these diversified workloads needs. So we really took a step back and said, all right, how do we go -- really address this. And we looked at process technology. We looked at core design. We looked at advanced packaging. And really, we came up with this road map, which is quite exciting, and it starts with Milan-X. And if you think about Milan-X, Milan-X is really another extension of our 3D chiplet and stacking technology. We basically are stacking cache onto our Zen 3 cores. And it's a partnership with TSMC with driving hybrid bond approach, which is 200x the interconnect density of 2D and 15x the interconnect density of similar 3D packaging technologies. It's going to deliver up to a 50% performance increase across technical computing workloads. Workloads like computation fluid dynamics, EDA and any sort of advanced physics modeling are going to see a tremendous uplift here. So we're really excited about this family. It's going to triple the amount of cache offered from Milan. And the uplift we're seeing, as I mentioned, is pretty dramatic. So we're sampling that -- that product is in GA today and we're going to launch that product actually in the end of this month. So we're very, very excited about that. As you move to Genoa, Genoa is obviously our fourth gen, driven on 5-nanometer TSMC process. This is a step function in performance. 96 cores, 192 threads, 12 channels of DDR5, PCI Express Gen 5, the new CXL coherent interconnect. So this will deliver extreme performance advantage across anything in the market across core enterprise, high-performance computing and public cloud. And that we are sampling today and we're going to bring that out in the launch later this year. And then Bergamo, as you mentioned, very, very exciting product. It is 128 high-performance cores, extremely power-optimized and it's focused on cloud-native and open source applications. So we really feel that, that is a socket level, socket throughput leader in the market. That will be available in the first half of '23, but we believe that will stack up to most of the ARM offerings in the market today. And we believe that Genoa stacks up extremely well with Sapphire Rapids as well as Milan and Milan-X. So we feel, Chris, very, very good about where we are from both products today that are leadership and then moving to more leadership going forward over the next 4 to 5 quarters. And then we always expect competition but we do want to stay focused on our game. If we're staying close to our customers and we're driving and listening to them and executing, we believe we're in very, very good shape going forward.

And then, Dan, just a quick follow-up there on the topic. Intel announced Granite Rapids. Actually pushed it out a little bit to 2024. We've kind of been saying that for a while now, but that was going to be on Intel 3. Maybe talk about the timing of that, how you might stack up that offering versus your offerings in that time frame. And also, does this mean that you have to pull some of your future products into 3-nanometer or 4-nanometer or something like that? Is that an opportunity for you or do you not feel that you need that?

Chris, we -- so we haven't given out a lot of details on the follow-on beyond Zen 4. And we did see the Granite communication. We do feel good about where we are on next generation. We have many engagements across both hyperscale and the enterprise and HPC markets. We feel very, very good. But we haven't given out really the detailed plans there, probably at a later date. But overall, it hasn't changed our plans. We're on our cadence and we believe we'll be very well set up in that time frame.

Yes. And is this an opportunity for you, the fact that Granite has pushed? Are you viewing that as an opportunity?

Look, I believe -- back to the road map cadence. If we stay on our cadence, it's certainly an opportunity any time a product comes out later. But when you look at how we plan the business and the road map, nothing's changing. We're going to stay on our cadence. We're going to deliver Zen 4. And then as I mentioned, we haven't really talked a lot publicly about Zen 5 but we feel very, very good about where that product sits today and we're in execution mode on it.

Got it. Switching gears. It sounds like hyperscale is just a fantastic opportunity for you. Enterprise, I think, is an area that will take a little bit more time to develop here and I think you touched on that. Where are we now in terms of innings or however you want to describe it? What needs to be done in terms of the ecosystem to kind of develop this? And what's the timing of this ramp? Is it going to be a hockey stick kind of inflection at some point or is this going to be a longer process? How should we think about it?

So Chris, cloud does remain the majority of our revenue today. However -- and we do believe that cloud will continue to grow, right, as we talked about with hyperscalers. As I mentioned, we saw that inflection point on Milan launch, and we actually really accelerated enterprise in the second half of last year. And I think in the Q4 results that Lisa just reviewed, we showed that we doubled that business year-on-year in Q4. We expect that type of growth to continue and here's the reasons why. we feel good about it. And I'll talk about a couple of different things. But first is the ecosystem, right, like you mentioned. We're delivering more platforms and we're going to deliver more platforms on Milan-X going forward also and then with Genoa. But over 100 platforms and the solution sets that we put together were across multiple verticals and very specific IT workloads. So we feel good about the traction with the ecosystem. Secondly, the modernization of the on-prem environment is accelerating and it's accelerating for us in the right direction. So the move is to hyper-converged infrastructure. That enables multi-cloud, hybrid cloud environments. And for us, that feeds right into our strength. So HCI solutions for us with our density and our TCO actually is a really strong solution set to really help enterprises modernize and create this multi-cloud environment [indiscernible] market. So we really feel good about that. And lastly, we will continue to drive per-core performance. So not so much top of stack but middle of the stack performance such that, that sweet spot in the core enterprise, we will compete very aggressively. And we also bring a very, very strong performance per watt. There's been numerous announcements recently about customers seeing either better performance or equal performance at 20% to 40% lower power consumption. Sustainability is a critical piece for most CIOs today, and we're bringing that capability very, very strongly as part of our solution set. And then lastly, we are very much engaging with the broader channel like the VAR network and the NSP network to engage them and really enable that. So it's still a work in progress. But our goal, and we're on track to do this, is get it aligned with the overall server business growth.

Okay, that certainly makes sense. HPC, we haven't talked about that yet. A smaller portion of your TAM but it seems like you guys have some really kind of strategic success stories here. CPU but also CPU and GPU combos, very important for HPC. And then finally, you have a new weapon here with FPGA as well. And maybe talk about this. And Intel might be co-packaging some of their offerings. Are you planning to do the same?

HPC is definitely a smaller portion of the market but it is a very important part of the market because they -- those customers drive and push the envelope of performance more than anyone. So usually, what we'll see is the architectures that they develop actually get more mainstreamed into other places, even in cloud architecture. So it's a really, really key segment. And most importantly, for us, it plays to our strengths of high performance. So we've had some -- a bunch of wins, as you mentioned, some of the most notable ones on the supercomputers of Frontier and El Capitan. And with our solution, CPU plus GPU, and by the way, we launched our Radeon Instinct MI200 last year in November and it's seeing very, very strong uptick. But the CPU plus GPU advantage is the Infinity Architecture and the cache coherency between the 2 devices. Both devices are delivering extreme performance on their own, but tying them together cache coherence is really, really critical to add to that performance. And we're driving a complete software stack, open software stack with a number of libraries for high-performance computing and machine learning, training and inference. So we have a lot of momentum in this space. When you look at Xilinx, it's very exciting because Xilinx brings a tremendous amount of flexibility, as you know, with FPGAs. And we look at Xilinx to come in and look at those next-level checkpoint -- choke points, I should say. So whether it's a compute workload that really needs to be accelerated or the network, right, network offload is a big area. So the combination now of CPU, GPU and adaptive SoC with Xilinx, I think, is world-class. And when you look at it, we can do it any number of ways. We can put those 3 on a board. We can put them all in a package. As long as they're under an overarching software stack, we'll provide the solution that the customer desires. So overall, the addition of Xilinx in this space really helps us. And as I mentioned, our GPU business with the MI200, is really picking up steam in terms of performance across not just HPC but even across the cloud also.

Interesting. EPYC just even right off the bat, did really well in part because of memory and you were addressing memory in a way that maybe Intel wasn't. Can you discuss the importance of this memory controllers, more channels? You guys seem to be pushing the edge there. DDR5, maybe there's a little bit of consternation out there in the market today, but obviously, this is the future. And talk about all of that, whether you're seeing sort of complications around stuff like DDR5 but also just the importance of memory and how you're going to compete there against Intel as you continue to push the envelope.

Yes, Chris. Memory is a critical and, I would say, growing part of the performance equation. And if you think about why, it's because we and the industry are driving higher and higher core counts. And the real important piece is keeping those extra cores fed with data consistently, right, to optimize the performance. You don't want idle CPU cycles. So it's really, really critical. So we're working with -- across the ecosystem and drive the optimal points around bandwidth, latency and efficiency to help our customers really try and solve their memory-intensive workloads. If you look at it one way, you can increase the data rate, the clock rate and the density through channels and multi-DIMMs. And DDR is a really, really important next step for that. So we are very much engaged in the DDR5 ecosystem. And we feel that, that will be a big boost to the next generation for us with Genoa. And secondly with the CXL consortium. We're working with a number of partners there to look at ways to also enhance memory performance and capacity. So there's an area there. And then lastly, if you look at it, we are constantly looking at on-chip versus off-chip. And Milan-X that I just talked about is a good example of stacked cache, so we're going to continuously look at that. But memory is going to be even more critical as we go up in the core counts. So we're going to maintain a very tight tie to our industry partners and try and be an innovator continuously here in the memory space.

Great. And you guys have forecasted really nice growth again in '22. And supply seems to be a constraint, obviously, for most of the industry. What improvements are you expecting as we look through '22? I think you guys have talked a little about this. But even into '23, does that mean -- if there are improvements, does that mean even higher growth for you guys? And server overall, I imagine high gross margin. You guys are probably a beneficiary there. But there are a lot of guys who want their allocation, but maybe put all the pieces together there.

So supply really has been a dynamic conversation now for probably 5 or 6 quarters. And I would say that we're pretty pleased where we are. And I would -- I'd need to call out our supply chain group. They have done just a tremendous job helping the company grow in '21. And it's not just one business, right? It's several businesses across AMD that grew pretty aggressively in '21. So they've just done a tremendous job. And it's through their tight partnerships across the entire ecosystem with TSMC on wafers and all the way through the OSAT network. They've really done a very, very good job of setting us up and investing for the future. Secondly, we have tremendous visibility. I talked about our customer focus. And the customers are giving us a lot of visibility in trying to drive our share up with them. And this is not a current quarter, current year kind of conversation. It's a multiyear journey with them. And I believe we have that visibility to go drive and invest in the back-end capacity to really support our both near-term and long-term growth goals. So we feel very, very good. Server, in general, the demand is still outpacing the supply a bit, but we do see with these investments through the course -- we have enough to support our growth targets. And as we go later into 2022 and into '23, we feel very, very good about where we are to really drive the long-term growth and share gains that we're after.

Excellent. And I know we had a half hour but I'm going to go a little bit over here, Dan. We've seen some pretty good cloud CapEx numbers out there. Facebook was one of them that I think caught a lot of people's attention. But assuming the industry solves some of these supply issues that we're having right now, how big of a growth opportunity are some of these CapEx opportunities for AMD on the hyperscale side? And are you better positioned now with these offerings that you have versus the past few years?

So I'll go back to what I talked about at the beginning, Chris, the excitement that I have for leading server at such a time in the industry and the CapEx increases are so exciting to see. We really believe that we have the best portfolio. But what's really happening is this focus on what is the right solution for the right workload. It truly is evolving into a heterogeneous world. And we feel that we're set up extremely well here, right? The portfolio alone, CPUs, GPUs, FPGAs, adaptive SoCs, ASIC capability and the investments in software, we feel well really good about where we are. The key point here, though, is each one of those has a very good road map on its own. But when you take a step back and put these together, whether, as I mentioned earlier, if it's all discrete components on a board or we look at integrating within package, all integrated within one software stack, an overarching software stack, we feel like we can bring the solutions to the hyperscalers, and for that matter, from client all the way to cloud. We feel that we can deliver the best optimized solutions for the right workload for the foreseeable future. So we're investing heavily in all of those areas. And we really feel like we're in a really good position. We're going to continue to execute. And we've made up a lot of ground. But for me, I really feel like we're just getting started. And with the growth that we see in the compute mega cycle, this is going to be a really fun ride for us because if we stay focused on the customer base and continue delivering those optimized solutions, we should be really in good shape to drive share gains and revenue growth.

Well, I would agree, Dan. I think you're just getting started, and I also agree you're in a very good position here, just a wonderful road map. I look forward to you personally executing on it and I look forward to the results.

Well, thank you, Chris. It's been a pleasure to join you and really appreciate the time.

Great. Thanks so much. And operator, with that, if you could...