Advancing AI 2026 | Andrew Dieckmann, AMD

Clips
More from Advancing AI 2026

Andrew Dieckmann

CVP & General Manager

AMD

play_circle_outline AMD's transformation from chipmaker into full rack-scale systems company

play_circle_outline Helios Rack and ZT Systems Acquisition Expand Liquid‑Cooled Frontier AI Across Racks, PCIe, Edge and Laptops

play_circle_outline Instinct GPU cards and EPYC CPUs integrated into system-level solutions

play_circle_outline Enterprise on-prem adoption, OEM go-to-market, hybrid cloud and POC strategies

play_circle_outline Reliable Liquid-Cooled Data Centers: Lessons from Rack Failures and Validated Design Collateral

play_circle_outline HBM memory cost increases versus performance-per-dollar design tradeoffs

Info

Andrew Dieckmann, AMD

Andrew Dieckmann

CVP & General Manager AMD

search

(INTRO)

Dave Vellante

>> Hi everybody, welcome back to Moscone West here in San Francisco. My name is Dave Vellante and here with my co-host for this segment, Bob O'Donnell, the principal and founder of TECHnalysis. Bob, great to be with you. Thanks Dave. Thanks for having me. Really appreciate you subbing in here. Super analyst, a lot of detail that we're going to go into. This is theCUBE's live coverage of AMD's Advancing AI 2026. Andrew Dieckmann is here, he's the Corporate Vice President and General Manager of the data center GPU business here at AMD, leading Instinct, the accelerator business, and the big baby, Helios. You gave birth, congratulations.

Dave Vellante

>> Thank you.

Dave Vellante

>> Love it.

Dave Vellante

>> We're very proud of the new baby.

Dave Vellante

>> Well this is a big deal, I wrote a piece this week, I've been following, as you have Bob, AMD for decades. Yeah, totally. And I did a deep dive for a week, I just went into the history and just to try to refresh my memory and try to come up with some conclusions, and it's very clear, AMD has transformed from a chip company going after Intel's x86 business to a systems company. And you have compressed 15, 20 years of ecosystem development by NVIDIA into five years through R &D, focused R &D and innovation, M &A to fill the gaps, obviously Xilinx huge, Pensando, et cetera, ZT Systems, and then ecosystem development, amazing. Amazing, in a very short time frame. So, pretty remarkable transformation. What's your perspective on that transformation? What are you guys today?

Andrew Dieckmann

>> I would agree with you, we're a hardware company, a chip company at our core, but we have transformed into a systems company, and it's really, it's required in order to deliver the rack scale infrastructure that we're doing with Helios. Whether it's the design, the co-design with our customers, all of the design complexity of the rack itself, but also the operational needs to make sure that the entire supply chain is planned for the large production ramp of these systems and that all of the right components exist in the right places with the right quantities.

Dave Vellante

>> Okay, so when we hear the term co-design, I said earlier when I was talking to Soni, when I first heard that from Jensen, I said, oh, they must be designing with the ecosystem. I don't think that's really the philosophy. It's designed in -house and then put it out to the ecosystem and plug it in. you have a different philosophy. You actually do co-design with the ecosystem. What benefits does that bring and what challenges does that bring?

Andrew Dieckmann

>> Yeah, I think...

Dave Vellante

>> More cooks in the kitchen.

Andrew Dieckmann

>> Yeah.

Andrew Dieckmann

>> So how do you deal with that?

Andrew Dieckmann

>> There's more cooks in the kitchen. I think we've tried to be very responsive to customer needs and requests as they've come along because the development cycles for these systems and chips are a couple of years long. And so, you learn so much during the development phase that we find keeping modularity and optionality as we go through the design is super important so that we can be responsive to new industry or customer feedback as it comes. And so that's really what we've tried to do is to enable that modularity both at a silicon level with our chiplet design, also at a rack scale level to enable different modularity at the system level. And then at the software, with our open source software stack, being able to incorporate learnings, pre and post production has been super helpful.

Dave Vellante

>> Okay, but so is there a trade off, and if not, how do you deal with the pace that you have to you're blowing away Moore's Law, it's not every 18 to 24 months, it's 12 to 15 months you're innovating, and you have to do that with your partners, there's multiple partners, so you have to be selective and figure, pick your spots, but what's your kind of methodology to make sure you can get to market as fast as possible for your customers?

Andrew Dieckmann

>> I think with Helios, we designed Helios to really enable us to do training and inferencing at the frontier scale. So we've been partnering with some of the leaders in frontier model building. And so, yeah, it's not hundreds of folks that we're doing co-design with simultaneously. It's a handful of folks. And honestly, we look for commonality between the needs of those people. And I think we are also very pragmatic in different tradeoffs as we go through the design process. We're not trying to boil the ocean, we're trying to be very selective in the market that we're addressing and making sure that we're getting the key elements delivered to deliver the best value to the customers.

Andrew Dieckmann

>> Bob, you know this,

Andrew Dieckmann

>> a lot of ways to solve problems in tech.

Andrew Dieckmann

>> No, but what's fascinating is,

Andrew Dieckmann

>> and I think this is reflective, because I, like Dave, I've followed AMD for literally decades, the company has always been focused on trying to hear what the customers want, right? And this feels like this co-design effort really feels a lot like, hey, let's find out what these folks need, let's tweak and tune, and being receptive to that idea. Lisa inherently is, she's just like that. She's got that personality of being like, hey, I want to take whatever feedback you can give me and incorporate that into our designs, and I think that's really critical. the other interesting thing for me is, I feel like the ZT Systems purchase was a way bigger deal than I think a lot of people gave it credit for. Because I feel like that was like, oh, yeah, that's a rack company. But then when you see how incredibly complex these things are, you have to think, wow, there was a lot of expertise that was acquired in that and setting you up, again, not just for this first Helios, but then, frankly, for that annual cadence that Dave mentioned as well. Because you had to be thinking way in advance on how to do all this stuff. So I'd love to hear a little bit more about how that piece came in. Yeah, you're right. That was a critical piece of our ability to execute on Helios. we had been building over the years that capability internally, hiring lots of mechanical, thermal engineers, et cetera. But the ZT acquisition supercharged that. We added north of 1 ,000 new engineers, specifically on rack design, thermals, mechanicals, and all of the hardware components of the system that have to come together. So that was a big part of our vision. And we knew that we needed to supercharge that effort in order to meet our time -to -market goals. And the ZT acquisition was a big part of that.

Andrew Dieckmann

>> Yeah. And one of the things,

Andrew Dieckmann

>> and you and I were chatting about this beforehand, we're seeing, obviously, the Helios interest initially from the big guys, the hyperscalers, what have you. One of the big themes, however, from Advancing AI overall, I feel like AMD was talking about was how to distribute AI workloads. And that would include doing things with enterprises. And so a couple questions come to mind. First of all, I know you have the standalone Instinct cards that you could plug into traditional air -cooled servers. That's certainly a step, but would we ever see an AMD Helios? is there a sort of the Helios (AMD Helios rack-scale AI system) version that maybe an enterprise would want to tackle?

Dave Vellante

>> AMD HeliosWell, there's that too, but actually my point is

Andrew Dieckmann

>> other types of form factors for Helios moving forward. Yeah, I think there will be a variety of form factors. Lisa has, for the last couple years, talked about the need for compute diversity. And as the models have continued to evolve and the capabilities, Jack talked a little bit about it, the capabilities of a nine billion parameter model now, exceeding that of a much larger model just last year. So that is coming to pass now, and so I think AMD's vision of everything from the client, the edge, all the way to the rack scale solutions for doing frontier model training or inferencing at scale in the cloud, and everything in between is required. And the different form factors, we're just pragmatic about it. Helios is a beast. it requires specialized data centers that are liquid cooled data centers. And those are being built in abundance right now, and so there's many nice homes for Helios to be deployed into, but there's a lot of legacy data centers that are air cooled, that PCIe cards or four-way servers fit nicely in, and can do a lot of good work. And then right in the hands of the consumer, the laptop capabilities continue to grow, and the model capabilities continue to enhance. and so we really see a very broad spectrum of compute being needed to serve the very large compute needs of the world moving forward.

Andrew Dieckmann

>> Yeah, well, and to me,

Andrew Dieckmann

>> one last point I was just going to make is it's been fascinating to watch this whole thing evolve because the truth of the matter is, the news around AI infrastructure has been utterly dominated, of course, by the other guys, right? It's crazy, and yet people have been saying, oh, we want alternatives, we want alternatives, and it feels like it takes the actual release of the rack for people to finally go, okay, now we've got some real competition here, let's make this happen. That's why it feels like, philosophically, it's even a bigger moment today than the actual product release, in a way. Yeah, we're obviously proud of the product. We think it's a leadership accelerator and a leadership AI rack on the market today. And as much as we're saying that, we think the specs support that claim. The fact that we have large partners like Anthropic, OpenAI, Meta endorsing it and saying, hey, we're going to buy gigawatts of AMD rack-scale infrastructure, I think is a good testament to the value and the capabilities of the system. And so we're very proud of those partnerships and excited about what the future holds there.

Dave Vellante

>> Well, I think Bob's question, too, speaks to the way that infrastructure requirements are changing. You remember VCE? VCE took, I think Intel servers and Cisco switches and EMC storage and it hooked them together, like bolted them together and put a software layer on top of them. I think EPYC Instinct, you've got Pensando, you've got ROCm, but they're not bolted together. they're designed as a system. So, can you speak to how infrastructure requirements have changed, but also the design of infrastructure has changed?

Andrew Dieckmann

>> Yeah, our customers don't want component solutions, they want system solutions and really cluster solutions, because beyond the system, we give complete design collateral, this is how you lay out your data center, and people want things that are validated and easy to deploy. And so, we have modularity in our design in which we can support different partners. We also need to have something that is reliable, quick to market, because we're turning these things. new GPUs come out every year.

Andrew Dieckmann

>> Right.

Andrew Dieckmann

>> And we want to deploy gigawatts of them and then turn to the next thing because the next thing is going to be that much better. And so I think it's really enabling our customers with both choice, but also with some playbooks to get things to market rapidly so that they can start to monetize those solutions.

Dave Vellante

>> I wasn't able to sit in on the analyst sessions because I was doing theCUBE, but boy, Lisa's keynote didn't pull any punches on the benchmarks specifically. So I want to ask you about that. You put forth some significant, bold statements about the performance, not only of your CPUs, your GPUs, but also Helios relative to the Vera Rubin racks and also other chips. Talk about those benchmarks. Where do they come from? I don't know if this came up in the analyst session, you guys must have been very inquisitive. So what can you share with us?

Andrew Dieckmann

>> Yeah, I think we are comparing our benchmarks and specs to the published specs of our competitor, and we have a capable competitor who we respect greatly, but we're very proud of what we've been able to produce here. With our chiplet approach, we are able to provide significantly more compute per GPU within the same power envelope as our competitor is doing. That results in, I think we're at 20 % more peak flops in some cases, and then the HBM bandwidth, we have a leadership position in HBM bandwidth, and we provide up to 50 % more HBM capacity per GPU. What that means is you can distribute your model over fewer number of GPUs, because a lot of these are simply memory capacity limited. And so you're getting a lot more efficient use of the hardware, and so we're simply comparing our specs to the published specs of our competitor there, and we think that we show very well.

Andrew Dieckmann

>> It was impressive, to be clear.

Andrew Dieckmann

>> And I think it was important to do that, because people are saying, as I said, people are like, yeah, we want an alternative, but we don't want any alternative, we want a solid alternative. And I think what came out today was, hey, this is a really solid alternative. And again, by laying out the roadmap for the future, it also points to, all right, this is not a one-time shot. this is something that's going to keep going. And it's a bit of the keeping up with the Joneses kind of a thing, but it's important to make it clear to people that you're here to play and you're serious about it. And I think that's going to be a big step.

Andrew Dieckmann

>> If we could,

Andrew Dieckmann

>> I'd love to revisit the enterprise question a bit longer because one of the challenges, of course, I kind of was joking about the Helios (AMD Helios rack-scale AI system), but most enterprises don't have liquid-cooled data centers. A few do, but most don't, so those plug-in cards are an interesting solution. But in general, the other thing is, it seems to me that most enterprise data centers have been used to deploying either Intel or AMD infrastructure, right? Yeah. Not necessarily NVIDIA. And so, in my mind, that theoretically gives you guys sort of a philosophical lead in that enterprise market because companies are used to dealing with AMD for their enterprise. Is that something that you're seeing? Do you think you could take advantage of that in that market? Yeah, and importantly in the enterprise, because it is a, there's Fortune 500 companies, we don't touch all of those customers as AMD directly. The large AI customers, we interface directly with and have very strong engagements. We actually leverage our partners, our OEM partners are a key go-to-market route for us for the broad enterprise, super important, because those customers also require more handholding, more help, more assistance to get the solutions actually deployed. And so I think, yeah, that existing infrastructure that we have through our CPU go-to-market motion has also proven to be helpful for the on-prem AI deployments of the PCIe cards and other things. Yeah, because I've talked recently with Dell, with HPE, with Lenovo, all of those guys are talking about all the GPU equipped servers they're selling. Those are the more traditional two -way, four-way, or eight GPU kind of systems. Or again, now they have the options of the plug-in cards as well. So it's going to be interesting. But one question is how do these companies start to integrate this stuff? Because it's hard, this is not Greenfield, like they already have their existing stuff, now okay, now how will they put this new stuff in there and really leverage it? Yeah, it's a good question and I think it's, Lisa said, we're in the early innings of AI, we're in the super early innings of a lot of this on-prem AI adoption, right? And I think people are experimenting, doing some POCs, okay, how can I actually match the right hardware to the workloads that my company is using? And as those tools evolve, I think they're just looking, where do I get the best economics? And we think that it's going to be a hybrid approach where there'll be some well-suited localized workloads for on-prem deployments, and there'll be a lot of cloud infrastructure that they're going to need to take advantage of as well from CSP partners. So I think the market is figuring that out, and they're figuring it out actually rather rapidly. So we'll see how that evolves over the next 12 months here. And guys like TensorWave, we just had Darrick on. At TensorWave, yeah.

Dave Vellante

>> And TensorWave, what did I say, TensorWave? CoreWeave. You know why I said it, because I was pushing him on, how do you differentiate? He said, well, what was the mindset of going all AMD? And he said, well, we watched CoreWeave, and, they're all NVIDIA. We didn't want to be second to them. We wanted to be number one in a market, and we thought that by leaning into AMD, and obviously they had to trust that you were going to deliver what you said, this is 2023.

Dave Vellante

>> Yeah.

Dave Vellante

>> They said, we can focus and we can be the best. And so my point being, his claim was that differentiation is really the simplicity and obviously the reliability, which is kind of like what CoreWeave is carving out, but they're doing that with an all AMD approach. Point being, that simplifies it for enterprises. They can get their feet wet, they can practice, get things going, do their POCs, actually go into production, and then bring in, and maybe it's plug-in cards, maybe it's, I don't know, maybe it's doing their own Helios at some point in time. The economics might be so compelling down the road, maybe five years down the road, that the cost per token of the liquid cooled may be an incentive to refactor your data centers, who knows? But you're in a bridge mode, right? You're going from dominant x86 competitor to now going into a new market with a completely different playbook.

Dave Vellante

>> Yes.

Dave Vellante

>> And you don't want to give that existing market up.

Andrew Dieckmann

>> No, no, and that's why now we have, even at this point in our go-to-market motion, we are engaging with enterprises, and we're engaging with them. It's a much smaller market than what the large model builders are doing, but we think it's an important market to get that learning in conjunction with our customers so that we're not figuring this out, hey, we're going to start to figure this out in five years. No, it's important to figure that out now and really understand where those pain points are, and you get that learning through those customer experiences so that we can roll it into our software stack and also into our next product iterations as we get that learning together.

Dave Vellante

>> And it really is a nascent market. The enterprise AI market is just really starting to get going. you're hearing some numbers, but let's face it, the fact that people are so concerned about token costs is because they're not scaling without labor like they're hoping to. They have to be careful. They can't just open up the checkbook and fucking max out, right?

Andrew Dieckmann

>> I mentioned to Bob, at the end of the day, economics does prevail. It is an important factor in people's decisions and how they're going to deploy this stuff.

Dave Vellante

>> So, I got to ask you, what you're doing with Helios is complicated. It's not like I said with VCE bolting together some compute storage and networking and calling it a day with some software veneer. Very complicated stuff. When GB200 racks started to ship, they were too heavy for the floor. GPUs were melting, the failure rates were like 50 % on those things. So my question is, have you learned from that? Are you getting similar concerns? What gives you confidence that you're not going to repeat those mistakes?

Andrew Dieckmann

>> We kind of touched on it earlier, we're a learning company, so we take all the learning that we can, too, and the industry is learning, right? the industry has learned so much in rack scale deployments over the past couple years. I think we've tapped into that, we've gotten a lot of counsel from our customers and where are the pitfalls, what do we avoid, and how do we really make these deployments resilient, reliable, serviceable, and all those good things. So there are many challenges, and it's also why we go right to data center providers themselves, work with them right at the shovel-ready stage, make sure that those data centers are spec'd appropriately, work with our Neo Cloud partners really early on, and then of course deliver the system. It's really a broad spectrum of things you need to look at to make sure the deployments are successful.

Dave Vellante

>> I want to ask you a question. Do you guys tout the increased capacity that you have with high bandwidth memory?

Dave Vellante

>> Yes.

Dave Vellante

>> My understanding is that costs have gone up slightly in HBM. And so how is it that you're able to keep your token costs so attractive and yet accommodate all that extra HBM?

Andrew Dieckmann

>> We're putting the extra HBM on the module because our customers find value in it, right? And so at the end of the day, they're able to get significantly more performance out of the system because of the memory. So, yes, unfortunately, memory costs are on the rise. We're working with our key partners to kind of plan for that as we go into the future, make sure we're making the right design point decisions now for what's coming down the pipe in the next year or two in that area. but we know that for Helios, with the design point we've chosen, it is the optimum perf per dollar design point for that system, including the memory capacity. And so we've tested this, we've retested it, and we've tested it again. Although the memory is more expensive now than it used to be, it is still of great value to our customers.

Dave Vellante

>> Okay, so eat the margin and keep the customer loyalty and that'll come back and benefit you.

Andrew Dieckmann

>> It's not just about eating the margin, customers care about performance per dollar. The easiest way for us to deliver value is to get performance up, right? If performance goes up by X percent, they're willing to pay more.

Andrew Dieckmann

>> And I was going to say,

Andrew Dieckmann

>> looking forward, the other thing I'm curious about, different side of the equation though, is interconnect and optical. You guys talked about that. Can you give us a little hint on what we're going to see? Because that's been a very hot topic in rack design recently. Yeah, so with Helios, it's an all -electrical interconnect for the scale -up. We're using an Ethernet transport protocol for that. In the future, we are looking to expand our scale -up domain as we go to larger pod sizes. Again, customers find value in that. We want to deliver on that value. As you increase the pod radix, we are going to be integrating some optical technologies into future racks. We haven't revealed a lot of detail about that, but that's certainly an area of future investment and a direction that we're going.

Andrew Dieckmann

>> Yeah, fair enough.

Andrew Dieckmann

>> Going to be interesting to watch, that's for sure.

Dave Vellante

>> Okay, so you announced, I think you announced general availability.

Dave Vellante

>> Yes.

Dave Vellante

>> So ramping up in the second half of 2026, and then really steep ramp in 2027, right?

Dave Vellante

>> That's correct.

Dave Vellante

>> And Darrick was saying they've actually got a Helios in house. I asked him the same question about GPUs melting. He said, look, we use this stuff and we're very careful and he's super excited.

Andrew Dieckmann

>> We're highly against melting GPUs.

Dave Vellante

>> Yeah, we all are. And Meta, we had Meta on as well and very impressive what you guys are doing. Of course, Meta, in the early days, we were there, Bob, with theCUBE at the inaugural OCP event. Oh yeah. And Meta has given such a huge contribution to the marketplace with the open source infrastructure. So it's amazing.

Dave Vellante

>> Yeah, they've been a great partner and co-design partner.

Dave Vellante

>> And this gives you your open source card, it's a good marketing play.

Dave Vellante

>> It is.

Dave Vellante

>> And I think it adds some ecosystem value. As you say, you've mentioned many times, competition is a good thing. Customers want an alternative.

Andrew Dieckmann

>> Absolutely, and to play up on the open source thing, I think it drives more partnerships, and I think one of the most impressive things for me, and one of the most surprising things from the event, was the number of partners that you guys brought up on stage, and the level of engagement they clearly had. It was not just, it was obviously the OpenAIs, and the Anthropics, and what have you, but then to have Cisco up there, to have AT &T up there, these were not companies I'm expecting to see at this event. It was impressive to see how you're working across a whole tech ecosystem to make this come to life. That's right. Obviously, AMD puts a lot of energy into our software stack and other things, but the community contributing to that software stack and the open source software stack is much, much more than a marketing thing. with the advent of agentic tools and our customers' ability to apply agents as well as their own know-how, and then contribute those to the open source software stack. It's not a closed ecosystem. Even on stage, our partners, Anthropic and OpenAI talked about that, too, contributing their know-how and improvements to our software stack. Those are benefiting our entire ecosystem and really propelling us forward.

Dave Vellante

>> So hey, if we have time, I have one more question. Are we good? Quickly, please, yeah. We're eating into the staff break. Sorry. We had six hours of programming, one 30-minute break today.

Andrew Dieckmann

>> Oh, sorry.

Dave Vellante

>> That's all right, go.Cerebras. Cerebras.

Andrew Dieckmann

>> That was the other big surprise, as I was saying, surprises. Yeah. Are we going to see that integrated into the rack at some point?

Andrew Dieckmann

>> How do we think about that?

Andrew Dieckmann

>> So what we announced today is a data center-level partnership where their racks and our racks are going to be combined basically to deliver the benefits of their low latency solution combined with the benefits of the effective token throughput of Helios and that really allows us to deliver that benefit as well as drive the economics of that solution to a much better place for the ad.

Andrew Dieckmann

>> So it's rack to rack, not integration within the rack.

Andrew Dieckmann

>> Correct.

Dave Vellante

>> But in the big picture, that's for customers who need that ultra, ultra low latency and can monetize it on their end. It's an emerging segment of the inferencing market. How big do you think that's going to be? Is that like a 5 % to 10 % type of thing? Is that eventually potentially bigger or smaller?

Andrew Dieckmann

>> It could be bigger. We're watching that closely. There are some interesting use cases there. It's still nascent and sort of emerging. But I think these heterogeneous solutions like we announced today will be one way of expanding the total addressable market.

Dave Vellante

>> Got it. Cool. All right, Andrew, we got to jump.Thank you so much for that.

Andrew Dieckmann

>> Thank you.

Dave Vellante

>> Really appreciate it. And thank you for watching Dave Vellante for Bob O'Donnell. We'll be right back right after this short break. We're here at Moscone West. Stop by and see us. We're at the top of the escalator. You're watching theCUBE.

Andrew Dieckmann, AMD

search

(INTRO)

Dave Vellante

>> Hi everybody, welcome back to Moscone West here in San Francisco. My name is Dave Vellante and here with my co-host for this segment, Bob O'Donnell, the principal and founder of TECHnalysis. Bob, great to be with you. Thanks Dave. Thanks for having me. Really appreciate you subbing in here. Super analyst, a lot of detail that we're going to go into. This is theCUBE's live coverage of AMD's Advancing AI 2026. Andrew Dieckmann is here, he's the Corporate Vice President and General Manager of the data center GPU business here at AMD, leading Instinct, the accelerator business, and the big baby, Helios. You gave birth, congratulations.

Dave Vellante

>> Thank you.

Dave Vellante

>> Love it.

Dave Vellante

>> We're very proud of the new baby.

Dave Vellante

>> Well this is a big deal, I wrote a piece this week, I've been following, as you have Bob, AMD for decades. Yeah, totally. And I did a deep dive for a week, I just went into the history and just to try to refresh my memory and try to come up with some conclusions, and it's very clear, AMD has transformed from a chip company going after Intel's x86 business to a systems company. And you have compressed 15, 20 years of ecosystem development by NVIDIA into five years through R &D, focused R &D and innovation, M &A to fill the gaps, obviously Xilinx huge, Pensando, et cetera, ZT Systems, and then ecosystem development, amazing. Amazing, in a very short time frame. So, pretty remarkable transformation. What's your perspective on that transformation? What are you guys today?

Andrew Dieckmann

>> I would agree with you, we're a hardware company, a chip company at our core, but we have transformed into a systems company, and it's really, it's required in order to deliver the rack scale infrastructure that we're doing with Helios. Whether it's the design, the co-design with our customers, all of the design complexity of the rack itself, but also the operational needs to make sure that the entire supply chain is planned for the large production ramp of these systems and that all of the right components exist in the right places with the right quantities.

Dave Vellante

>> Okay, so when we hear the term co-design, I said earlier when I was talking to Soni, when I first heard that from Jensen, I said, oh, they must be designing with the ecosystem. I don't think that's really the philosophy. It's designed in -house and then put it out to the ecosystem and plug it in. you have a different philosophy. You actually do co-design with the ecosystem. What benefits does that bring and what challenges does that bring?

Andrew Dieckmann

>> Yeah, I think...

Dave Vellante

>> More cooks in the kitchen.

Andrew Dieckmann

>> Yeah.

Andrew Dieckmann

>> So how do you deal with that?

Andrew Dieckmann

>> There's more cooks in the kitchen. I think we've tried to be very responsive to customer needs and requests as they've come along because the development cycles for these systems and chips are a couple of years long. And so, you learn so much during the development phase that we find keeping modularity and optionality as we go through the design is super important so that we can be responsive to new industry or customer feedback as it comes. And so that's really what we've tried to do is to enable that modularity both at a silicon level with our chiplet design, also at a rack scale level to enable different modularity at the system level. And then at the software, with our open source software stack, being able to incorporate learnings, pre and post production has been super helpful.

Dave Vellante

>> Okay, but so is there a trade off, and if not, how do you deal with the pace that you have to you're blowing away Moore's Law, it's not every 18 to 24 months, it's 12 to 15 months you're innovating, and you have to do that with your partners, there's multiple partners, so you have to be selective and figure, pick your spots, but what's your kind of methodology to make sure you can get to market as fast as possible for your customers?

Andrew Dieckmann

>> I think with Helios, we designed Helios to really enable us to do training and inferencing at the frontier scale. So we've been partnering with some of the leaders in frontier model building. And so, yeah, it's not hundreds of folks that we're doing co-design with simultaneously. It's a handful of folks. And honestly, we look for commonality between the needs of those people. And I think we are also very pragmatic in different tradeoffs as we go through the design process. We're not trying to boil the ocean, we're trying to be very selective in the market that we're addressing and making sure that we're getting the key elements delivered to deliver the best value to the customers.

Andrew Dieckmann

>> Bob, you know this,

Andrew Dieckmann

>> a lot of ways to solve problems in tech.

Andrew Dieckmann

>> No, but what's fascinating is,

Andrew Dieckmann

>> and I think this is reflective, because I, like Dave, I've followed AMD for literally decades, the company has always been focused on trying to hear what the customers want, right? And this feels like this co-design effort really feels a lot like, hey, let's find out what these folks need, let's tweak and tune, and being receptive to that idea. Lisa inherently is, she's just like that. She's got that personality of being like, hey, I want to take whatever feedback you can give me and incorporate that into our designs, and I think that's really critical. the other interesting thing for me is, I feel like the ZT Systems purchase was a way bigger deal than I think a lot of people gave it credit for. Because I feel like that was like, oh, yeah, that's a rack company. But then when you see how incredibly complex these things are, you have to think, wow, there was a lot of expertise that was acquired in that and setting you up, again, not just for this first Helios, but then, frankly, for that annual cadence that Dave mentioned as well. Because you had to be thinking way in advance on how to do all this stuff. So I'd love to hear a little bit more about how that piece came in. Yeah, you're right. That was a critical piece of our ability to execute on Helios. we had been building over the years that capability internally, hiring lots of mechanical, thermal engineers, et cetera. But the ZT acquisition supercharged that. We added north of 1 ,000 new engineers, specifically on rack design, thermals, mechanicals, and all of the hardware components of the system that have to come together. So that was a big part of our vision. And we knew that we needed to supercharge that effort in order to meet our time -to -market goals. And the ZT acquisition was a big part of that.

Andrew Dieckmann

>> Yeah. And one of the things,

Andrew Dieckmann

>> and you and I were chatting about this beforehand, we're seeing, obviously, the Helios interest initially from the big guys, the hyperscalers, what have you. One of the big themes, however, from Advancing AI overall, I feel like AMD was talking about was how to distribute AI workloads. And that would include doing things with enterprises. And so a couple questions come to mind. First of all, I know you have the standalone Instinct cards that you could plug into traditional air -cooled servers. That's certainly a step, but would we ever see an AMD Helios? is there a sort of the Helios (AMD Helios rack-scale AI system) version that maybe an enterprise would want to tackle?

Dave Vellante

>> AMD HeliosWell, there's that too, but actually my point is

Andrew Dieckmann

>> other types of form factors for Helios moving forward. Yeah, I think there will be a variety of form factors. Lisa has, for the last couple years, talked about the need for compute diversity. And as the models have continued to evolve and the capabilities, Jack talked a little bit about it, the capabilities of a nine billion parameter model now, exceeding that of a much larger model just last year. So that is coming to pass now, and so I think AMD's vision of everything from the client, the edge, all the way to the rack scale solutions for doing frontier model training or inferencing at scale in the cloud, and everything in between is required. And the different form factors, we're just pragmatic about it. Helios is a beast. it requires specialized data centers that are liquid cooled data centers. And those are being built in abundance right now, and so there's many nice homes for Helios to be deployed into, but there's a lot of legacy data centers that are air cooled, that PCIe cards or four-way servers fit nicely in, and can do a lot of good work. And then right in the hands of the consumer, the laptop capabilities continue to grow, and the model capabilities continue to enhance. and so we really see a very broad spectrum of compute being needed to serve the very large compute needs of the world moving forward.

Andrew Dieckmann

>> Yeah, well, and to me,

Andrew Dieckmann

>> one last point I was just going to make is it's been fascinating to watch this whole thing evolve because the truth of the matter is, the news around AI infrastructure has been utterly dominated, of course, by the other guys, right? It's crazy, and yet people have been saying, oh, we want alternatives, we want alternatives, and it feels like it takes the actual release of the rack for people to finally go, okay, now we've got some real competition here, let's make this happen. That's why it feels like, philosophically, it's even a bigger moment today than the actual product release, in a way. Yeah, we're obviously proud of the product. We think it's a leadership accelerator and a leadership AI rack on the market today. And as much as we're saying that, we think the specs support that claim. The fact that we have large partners like Anthropic, OpenAI, Meta endorsing it and saying, hey, we're going to buy gigawatts of AMD rack-scale infrastructure, I think is a good testament to the value and the capabilities of the system. And so we're very proud of those partnerships and excited about what the future holds there.

Dave Vellante

>> Well, I think Bob's question, too, speaks to the way that infrastructure requirements are changing. You remember VCE? VCE took, I think Intel servers and Cisco switches and EMC storage and it hooked them together, like bolted them together and put a software layer on top of them. I think EPYC Instinct, you've got Pensando, you've got ROCm, but they're not bolted together. they're designed as a system. So, can you speak to how infrastructure requirements have changed, but also the design of infrastructure has changed?

Andrew Dieckmann

>> Yeah, our customers don't want component solutions, they want system solutions and really cluster solutions, because beyond the system, we give complete design collateral, this is how you lay out your data center, and people want things that are validated and easy to deploy. And so, we have modularity in our design in which we can support different partners. We also need to have something that is reliable, quick to market, because we're turning these things. new GPUs come out every year.

Andrew Dieckmann

>> Right.

Andrew Dieckmann

>> And we want to deploy gigawatts of them and then turn to the next thing because the next thing is going to be that much better. And so I think it's really enabling our customers with both choice, but also with some playbooks to get things to market rapidly so that they can start to monetize those solutions.

Dave Vellante

>> I wasn't able to sit in on the analyst sessions because I was doing theCUBE, but boy, Lisa's keynote didn't pull any punches on the benchmarks specifically. So I want to ask you about that. You put forth some significant, bold statements about the performance, not only of your CPUs, your GPUs, but also Helios relative to the Vera Rubin racks and also other chips. Talk about those benchmarks. Where do they come from? I don't know if this came up in the analyst session, you guys must have been very inquisitive. So what can you share with us?

Andrew Dieckmann

>> Yeah, I think we are comparing our benchmarks and specs to the published specs of our competitor, and we have a capable competitor who we respect greatly, but we're very proud of what we've been able to produce here. With our chiplet approach, we are able to provide significantly more compute per GPU within the same power envelope as our competitor is doing. That results in, I think we're at 20 % more peak flops in some cases, and then the HBM bandwidth, we have a leadership position in HBM bandwidth, and we provide up to 50 % more HBM capacity per GPU. What that means is you can distribute your model over fewer number of GPUs, because a lot of these are simply memory capacity limited. And so you're getting a lot more efficient use of the hardware, and so we're simply comparing our specs to the published specs of our competitor there, and we think that we show very well.

Andrew Dieckmann

>> It was impressive, to be clear.

Andrew Dieckmann

>> And I think it was important to do that, because people are saying, as I said, people are like, yeah, we want an alternative, but we don't want any alternative, we want a solid alternative. And I think what came out today was, hey, this is a really solid alternative. And again, by laying out the roadmap for the future, it also points to, all right, this is not a one-time shot. this is something that's going to keep going. And it's a bit of the keeping up with the Joneses kind of a thing, but it's important to make it clear to people that you're here to play and you're serious about it. And I think that's going to be a big step.

Andrew Dieckmann

>> If we could,

Andrew Dieckmann

>> I'd love to revisit the enterprise question a bit longer because one of the challenges, of course, I kind of was joking about the Helios (AMD Helios rack-scale AI system), but most enterprises don't have liquid-cooled data centers. A few do, but most don't, so those plug-in cards are an interesting solution. But in general, the other thing is, it seems to me that most enterprise data centers have been used to deploying either Intel or AMD infrastructure, right? Yeah. Not necessarily NVIDIA. And so, in my mind, that theoretically gives you guys sort of a philosophical lead in that enterprise market because companies are used to dealing with AMD for their enterprise. Is that something that you're seeing? Do you think you could take advantage of that in that market? Yeah, and importantly in the enterprise, because it is a, there's Fortune 500 companies, we don't touch all of those customers as AMD directly. The large AI customers, we interface directly with and have very strong engagements. We actually leverage our partners, our OEM partners are a key go-to-market route for us for the broad enterprise, super important, because those customers also require more handholding, more help, more assistance to get the solutions actually deployed. And so I think, yeah, that existing infrastructure that we have through our CPU go-to-market motion has also proven to be helpful for the on-prem AI deployments of the PCIe cards and other things. Yeah, because I've talked recently with Dell, with HPE, with Lenovo, all of those guys are talking about all the GPU equipped servers they're selling. Those are the more traditional two -way, four-way, or eight GPU kind of systems. Or again, now they have the options of the plug-in cards as well. So it's going to be interesting. But one question is how do these companies start to integrate this stuff? Because it's hard, this is not Greenfield, like they already have their existing stuff, now okay, now how will they put this new stuff in there and really leverage it? Yeah, it's a good question and I think it's, Lisa said, we're in the early innings of AI, we're in the super early innings of a lot of this on-prem AI adoption, right? And I think people are experimenting, doing some POCs, okay, how can I actually match the right hardware to the workloads that my company is using? And as those tools evolve, I think they're just looking, where do I get the best economics? And we think that it's going to be a hybrid approach where there'll be some well-suited localized workloads for on-prem deployments, and there'll be a lot of cloud infrastructure that they're going to need to take advantage of as well from CSP partners. So I think the market is figuring that out, and they're figuring it out actually rather rapidly. So we'll see how that evolves over the next 12 months here. And guys like TensorWave, we just had Darrick on. At TensorWave, yeah.

Dave Vellante

>> And TensorWave, what did I say, TensorWave? CoreWeave. You know why I said it, because I was pushing him on, how do you differentiate? He said, well, what was the mindset of going all AMD? And he said, well, we watched CoreWeave, and, they're all NVIDIA. We didn't want to be second to them. We wanted to be number one in a market, and we thought that by leaning into AMD, and obviously they had to trust that you were going to deliver what you said, this is 2023.

Dave Vellante

>> Yeah.

Dave Vellante

>> They said, we can focus and we can be the best. And so my point being, his claim was that differentiation is really the simplicity and obviously the reliability, which is kind of like what CoreWeave is carving out, but they're doing that with an all AMD approach. Point being, that simplifies it for enterprises. They can get their feet wet, they can practice, get things going, do their POCs, actually go into production, and then bring in, and maybe it's plug-in cards, maybe it's, I don't know, maybe it's doing their own Helios at some point in time. The economics might be so compelling down the road, maybe five years down the road, that the cost per token of the liquid cooled may be an incentive to refactor your data centers, who knows? But you're in a bridge mode, right? You're going from dominant x86 competitor to now going into a new market with a completely different playbook.

Dave Vellante

>> Yes.

Dave Vellante

>> And you don't want to give that existing market up.

Andrew Dieckmann

>> No, no, and that's why now we have, even at this point in our go-to-market motion, we are engaging with enterprises, and we're engaging with them. It's a much smaller market than what the large model builders are doing, but we think it's an important market to get that learning in conjunction with our customers so that we're not figuring this out, hey, we're going to start to figure this out in five years. No, it's important to figure that out now and really understand where those pain points are, and you get that learning through those customer experiences so that we can roll it into our software stack and also into our next product iterations as we get that learning together.

Dave Vellante

>> And it really is a nascent market. The enterprise AI market is just really starting to get going. you're hearing some numbers, but let's face it, the fact that people are so concerned about token costs is because they're not scaling without labor like they're hoping to. They have to be careful. They can't just open up the checkbook and fucking max out, right?

Andrew Dieckmann

>> I mentioned to Bob, at the end of the day, economics does prevail. It is an important factor in people's decisions and how they're going to deploy this stuff.

Dave Vellante

>> So, I got to ask you, what you're doing with Helios is complicated. It's not like I said with VCE bolting together some compute storage and networking and calling it a day with some software veneer. Very complicated stuff. When GB200 racks started to ship, they were too heavy for the floor. GPUs were melting, the failure rates were like 50 % on those things. So my question is, have you learned from that? Are you getting similar concerns? What gives you confidence that you're not going to repeat those mistakes?

Andrew Dieckmann

>> We kind of touched on it earlier, we're a learning company, so we take all the learning that we can, too, and the industry is learning, right? the industry has learned so much in rack scale deployments over the past couple years. I think we've tapped into that, we've gotten a lot of counsel from our customers and where are the pitfalls, what do we avoid, and how do we really make these deployments resilient, reliable, serviceable, and all those good things. So there are many challenges, and it's also why we go right to data center providers themselves, work with them right at the shovel-ready stage, make sure that those data centers are spec'd appropriately, work with our Neo Cloud partners really early on, and then of course deliver the system. It's really a broad spectrum of things you need to look at to make sure the deployments are successful.

Dave Vellante

>> I want to ask you a question. Do you guys tout the increased capacity that you have with high bandwidth memory?

Dave Vellante

>> Yes.

Dave Vellante

>> My understanding is that costs have gone up slightly in HBM. And so how is it that you're able to keep your token costs so attractive and yet accommodate all that extra HBM?

Andrew Dieckmann

>> We're putting the extra HBM on the module because our customers find value in it, right? And so at the end of the day, they're able to get significantly more performance out of the system because of the memory. So, yes, unfortunately, memory costs are on the rise. We're working with our key partners to kind of plan for that as we go into the future, make sure we're making the right design point decisions now for what's coming down the pipe in the next year or two in that area. but we know that for Helios, with the design point we've chosen, it is the optimum perf per dollar design point for that system, including the memory capacity. And so we've tested this, we've retested it, and we've tested it again. Although the memory is more expensive now than it used to be, it is still of great value to our customers.

Dave Vellante

>> Okay, so eat the margin and keep the customer loyalty and that'll come back and benefit you.

Andrew Dieckmann

>> It's not just about eating the margin, customers care about performance per dollar. The easiest way for us to deliver value is to get performance up, right? If performance goes up by X percent, they're willing to pay more.

Andrew Dieckmann

>> And I was going to say,

Andrew Dieckmann

>> looking forward, the other thing I'm curious about, different side of the equation though, is interconnect and optical. You guys talked about that. Can you give us a little hint on what we're going to see? Because that's been a very hot topic in rack design recently. Yeah, so with Helios, it's an all -electrical interconnect for the scale -up. We're using an Ethernet transport protocol for that. In the future, we are looking to expand our scale -up domain as we go to larger pod sizes. Again, customers find value in that. We want to deliver on that value. As you increase the pod radix, we are going to be integrating some optical technologies into future racks. We haven't revealed a lot of detail about that, but that's certainly an area of future investment and a direction that we're going.

Andrew Dieckmann

>> Yeah, fair enough.

Andrew Dieckmann

>> Going to be interesting to watch, that's for sure.

Dave Vellante

>> Okay, so you announced, I think you announced general availability.

Dave Vellante

>> Yes.

Dave Vellante

>> So ramping up in the second half of 2026, and then really steep ramp in 2027, right?

Dave Vellante

>> That's correct.

Dave Vellante

>> And Darrick was saying they've actually got a Helios in house. I asked him the same question about GPUs melting. He said, look, we use this stuff and we're very careful and he's super excited.

Andrew Dieckmann

>> We're highly against melting GPUs.

Dave Vellante

>> Yeah, we all are. And Meta, we had Meta on as well and very impressive what you guys are doing. Of course, Meta, in the early days, we were there, Bob, with theCUBE at the inaugural OCP event. Oh yeah. And Meta has given such a huge contribution to the marketplace with the open source infrastructure. So it's amazing.

Dave Vellante

>> Yeah, they've been a great partner and co-design partner.

Dave Vellante

>> And this gives you your open source card, it's a good marketing play.

Dave Vellante

>> It is.

Dave Vellante

>> And I think it adds some ecosystem value. As you say, you've mentioned many times, competition is a good thing. Customers want an alternative.

Andrew Dieckmann

>> Absolutely, and to play up on the open source thing, I think it drives more partnerships, and I think one of the most impressive things for me, and one of the most surprising things from the event, was the number of partners that you guys brought up on stage, and the level of engagement they clearly had. It was not just, it was obviously the OpenAIs, and the Anthropics, and what have you, but then to have Cisco up there, to have AT &T up there, these were not companies I'm expecting to see at this event. It was impressive to see how you're working across a whole tech ecosystem to make this come to life. That's right. Obviously, AMD puts a lot of energy into our software stack and other things, but the community contributing to that software stack and the open source software stack is much, much more than a marketing thing. with the advent of agentic tools and our customers' ability to apply agents as well as their own know-how, and then contribute those to the open source software stack. It's not a closed ecosystem. Even on stage, our partners, Anthropic and OpenAI talked about that, too, contributing their know-how and improvements to our software stack. Those are benefiting our entire ecosystem and really propelling us forward.

Dave Vellante

>> So hey, if we have time, I have one more question. Are we good? Quickly, please, yeah. We're eating into the staff break. Sorry. We had six hours of programming, one 30-minute break today.

Andrew Dieckmann

>> Oh, sorry.

Dave Vellante

>> That's all right, go.Cerebras. Cerebras.

Andrew Dieckmann

>> That was the other big surprise, as I was saying, surprises. Yeah. Are we going to see that integrated into the rack at some point?

Andrew Dieckmann

>> How do we think about that?

Andrew Dieckmann

>> So what we announced today is a data center-level partnership where their racks and our racks are going to be combined basically to deliver the benefits of their low latency solution combined with the benefits of the effective token throughput of Helios and that really allows us to deliver that benefit as well as drive the economics of that solution to a much better place for the ad.

Andrew Dieckmann

>> So it's rack to rack, not integration within the rack.

Andrew Dieckmann

>> Correct.

Dave Vellante

>> But in the big picture, that's for customers who need that ultra, ultra low latency and can monetize it on their end. It's an emerging segment of the inferencing market. How big do you think that's going to be? Is that like a 5 % to 10 % type of thing? Is that eventually potentially bigger or smaller?

Andrew Dieckmann

>> It could be bigger. We're watching that closely. There are some interesting use cases there. It's still nascent and sort of emerging. But I think these heterogeneous solutions like we announced today will be one way of expanding the total addressable market.

Dave Vellante

>> Got it. Cool. All right, Andrew, we got to jump.Thank you so much for that.

Andrew Dieckmann

>> Thank you.

Dave Vellante

>> Really appreciate it. And thank you for watching Dave Vellante for Bob O'Donnell. We'll be right back right after this short break. We're here at Moscone West. Stop by and see us. We're at the top of the escalator. You're watching theCUBE.