The connection, Silicon Valley and Wall Street. I'm John Furrier, host of the Cube here with Dave Vellante, my co-host.
Welcome to the Rave Summit coverage, brought to you by Solidigm, with support from Tentative D Matrix and Ardentum.
Welcome back to the Cube's live stream in Paris, France for the Rave Summit. I'm John Furrier, host of the Cube. Day one kicking off day zero side car event yesterday with Machine as well. Physical AI and robotics, which is going to lead into hot year in robotics. But that's all part of the power and the computing and the software efficiencies coming from all these new integrated systems. This next guest is a legend in the industry, the conscious engineering brain for AMD, 40-year history, Mark Papermaster, CTO and EVP with AMD. Thanks for coming on the Cube. Good to see you again.
Cube alumni back. I said you're the conscious brain of AMD because the engineering mindset of AMD has been one of the hallmarks of the company. And there's so much you guys are doing. It's so broad and in the spectrum of what capabilities, but performance has been a very huge deal. So thank you for what you do. We appreciate more horsepower, go faster. What's your current take right now because right now the demand is so high for performance, but it's not the way it was last decade, last generation. It's transitioning to a new architecture, new requirements. We're paving new roads, we're pouring new concrete in AI infrastructure. What's your take on all this?
Well, I love your analogy. We're paving new roads. John, so true. Because look, the need for high performance, that's actually not new. So I'm four decades plus in the industry, so that means four decades plus driving more and more performance and performance at efficiency across the computer infrastructure. So that's been my whole working career. But it is a new road now because generative AI is changing everything. It's such a diverse workload, John, that is driving us at AMD and certainly the whole industry to think, design, architect, and implement much more holistically than we ever have. And what I mean by that is the workloads are so complex because people are looking at what they do end to end. They're looking at whole processes, not just one bespoke task. And that means you need different computing engines, and they need to work together. And they need to work together at scale. I mean, we're talking across massive clusters of racks.
I love that analogy of like just focusing here, now we're looking at a larger scale because at the end of the day, it's all a system. So, if you have a system mindset, what's changed the most for you as you evaluate efficiency and some of the constraints and managing through those because when you look at the broader workloads that are running on a gigantic, there are use cases, there are new requirements. What are some of those system elements that need to be in place as first principles? That's different than building a motherboard and putting software on it, accessing it. Now we got and then putting it into a data center, then putting it in the cloud. What's different now from a system standpoint from a first principles?
You know, when you look at data center, the movement of the data needs to be so efficient because it's always been moving data from memory into that beast, that processor core, moving it to IO into that processor core as efficiently as you can. But, again, what I said a moment ago, it's heterogeneous now. So now you have to think about that data movement and optimizing that performance across CPU, across GPU, and across the network. And all of that to a memory complex, you need low latency to memory. You need to get out to storage. And so, what's changed is not the local optimization on each of the elements. Still have to do all of that. But you have to, as you said, at a system level, really optimize. And that's what we've changed at AMD. We've become system optimizers. Still optimizing every component, but then equally looking at how you optimize how each of the pieces come together. And in that vein, we've expanded our portfolio. So look at what we've done over the last few years, right? We added networking capability. We acquired Xilinx. We acquired Pensando. And then we acquired ZT Systems that brought a whole rack design capability. So for the data center design, we're optimizing at the rack level, hardware and software.
I have to ask you this question because it comes up a lot. Okay, I got this large-scale system. I want to put this AI factory or refinery into the system. Great. I plug it in. I put it into a data center. Great. But whoa, I got all this other stuff. So backwards compatibility is a big discussion that's come up a lot on the queue, but also backwards compatibility with existing workloads. So I have a lot of AMD, but I might want to run at the edge. I'm going to have distributed computing architecture. Now I'm looking at a cost envelope and a performance envelope. So I might look at like I might get some performance here, but if I look at the big picture, the performance cost ratio changes. A lot of these infrastructures have a lot of AMD, a lot of x86.
What's your vision on that? Because this has come up a lot.
Yes. It's coming up in you know, you say a lot. I'll say for me, every enterprise conversation, John, I'm having now, this comes up. Because enterprises have decades they've been running x86. They're not going to move that install base. And what we've done at AMD, you know, since we launched the new Zen processor back in 2017. We're now on our sixth generation. So at our Advancing AI event on July 22nd and 23rd, we're rolling out this new generation. You know, it continues the kind of leadership x86 CPU, but it's designed in such a way that it matches what I just described a moment ago. It's optimized for standalone x86 traditional workloads. But what about when you're running heterogeneous workloads that you need a GPU, but then you still have your traditional x86 apps? Helios has that sixth generation epic processor built in. And so the very control of the Helios rack with our new MI255 that we're again, we'll be sharing more details in a couple weeks. We'll be there, by the way. We're streaming live from your event.
So we'll be there. Let's get into the price performance because you're what you're getting into is obviously your efficiency your career does that. Performance is obviously great, but the cost also is important. Price performance has always been a benchmark. Let's map that to the current environment. So x86 factors into the cost. What about when you look at things like, okay, agentic workloads spanning to the edge and inference needs that are different in scope? So I might say, 'What's the weather like in Palo Alto?'
That's different than saying, 'Hey, solve my high-end math problem and write a like this new thesis and solve these big problems.' They have different levels of differentiated services. So I'd love to have that route in the right areas. How does that impact you guys because that's coming up a lot, especially policies policy-based resources is becoming a big
Yeah, no, you're spot-on. And that's what we've architected in our stack. So our AI stack's open. It's called ROCm. It's the same ROCm that we run on those massive clusters. When you're running AI at the edge, you're running AI locally on a PC like you have in front of you there, it's also ROCm. And so if you think about what we're doing, we're making it very easy to port. Now what you have to do is you have to orchestrate your biggest task, a massive inference model that's running on a foundation model, that's running in the cloud or a big data center. But most enterprises, that's very expensive if you run everything there. So they're exactly on the point you made, John. They're looking to run that more economically and often at the edge it has to be done locally because you need real-time response. And so we've done that for not only our CPU and GPU, but the embedded neural processors that we have on the PCs and also in the embedded edge.
All right. So I want to bring something else up. It's not related to the question I'm going to ask, but it's a way to tease out the question. So if you look at what's going on in China, they get constrained GPUs.
And so they're smart engineers too out there. And they do is they actually figure out how to run better, faster
They cluster it together.
All entrepreneurial engineers will figure out how to work around constraints. So their constraint I don't have the big GPUs. So what do they do? They engineer better solutions to run on non-GPUs.
I think that's going to be something that's
or even just lower capability GPUs.
So I think that's a tell sign to how the engineering entrepreneurial thinking's going to happen. The constraints are going to be in memory. We're going to see some constraints around cost and deployment savvy IQ. I think this is a topic. What's your thoughts on this engineering around constraints as they will come up because we're still at the beginning of a new era. What are some of those thoughts around what will come because an edge box might not have a big footprint than a big monster data center.
