Mark Papermaster on Advancing AI 2025 and the AMD Open-Source AI Software Stack

Hi, Bryan. Thanks. Great to have you here. Yes.

Well, what I think the key takeaway from today was the progress we've made. You refer back to the last time we chatted and, you know, so let's say go back to the last episode and where were we at that time? We had a few key partners or hyperscalers that had taken our Instinct products, had applied ROCm, and were actually running production inference. So it was a huge milestone and a lot of questions people had. Will AMD be able to actually keep up with the roadmap? Will they be able to be, in fact, on an annual cadence, to keep performance moving up into the right? Which is the that and at the torrid pace that our customers demand. And where will the ROCm software stack go? And will AMD in fact be able to make good on their commitment to create an open ecosystem and an open software stack? Will there be traction? And so what I love about today is emphatic statements on all three. We showed our roadmap that we, in fact, pulled our next generation MI350 series in several months, and we're into production start of Q3. Fantastic inferencing products where you can get up to 35x throughput improvement. But it also is designed to support right through mid scale training in addition to inference, and as well starting to pull the veil back a little bit about the MI400 series the following year and how that is being readied for mass scale for both training and inference. So I think, you know, one was showing what has been a hallmark at AMD and that's our actual execution excellence, really doing what we said, and executing at quality. Secondly, from a software standpoint, really hearing from customers, hearing from their customers, you know? Yes. Those hyperscalers that we heard from, Meta and the progress that they have made and much expanded deployment. But we heard from many others who are running on the ROCm software stack and could attest to the progress that we've made. And then thirdly, the resounding statement and commitment of AMD to open standards. We made announcements of where we actually have specs out to support interoperability with others in the industry. And in fact, where we are in AMD in our roadmap to support them. So those are the several key takeaways from today.

Well, I think you have to think about the history of computing. It doesn't do well when there's just one player. And we've had this several times in history. There was, you know, I grew up at IBM, so I well know what a complete vertical walled garden solution looks like. And if you look at the history of IBM, it kept that mainframe solution. There's still mainframes today, but the industry demanded open solutions, and IBM went that way. Open source at the end always wins out. So you see it the same in phones. There's now open source alternatives with Android, many different Android models out there. This is the moment for the AI ecosystem to stand up and create an actually open alternative. And we're doing that with open standards. So on connectivity at the network level, Ultra Ethernet promises excellent performance, as well as features that have been only available in InfiniBand. There's really only one company offering InfiniBand today, but with Ultra Ethernet, it's going to be a broad consortium. It's going to have all of the congestion management and the features that people need to really build out AI networks. Secondly, Ultra Acceleration Link. And that's not the scale out like UEC, Ultra Ethernet. It's a scale up. How do you build big AI nodes? And we're seeing a strong adoption of UALink. And what we showed today was the commitment we have and how we are deploying UALink in our roadmap. And we have partners coming on stage with us to be a part of that. So that along with our open source ROCm stack is a very clear message of choice that we're giving to the industry.