AMD Advancing AI 2026: Lisa Su Full Keynote

[00:00:00] Good morning. [00:00:05] That's a pretty good good morning. [00:00:07] How about let's try one more time? [00:00:07] Good morning. [00:00:09] Good morning. [00:00:10] And welcome to Advancing AI 2026. [00:00:14] It's so great to be back here in San Francisco [00:00:16] and to see so many friends and partners and customers [00:00:19] and especially all the developers that are here with us today. [00:00:24] And I want to say a big welcome to everyone [00:00:26] who's joining us online from around the world. [00:00:29] This is my absolute favorite event of the year. [00:00:33] It's where we bring the entire AI ecosystem together [00:00:36] to show what we've been building and where we're going next. [00:00:39] And this year, this is our biggest show ever [00:00:42] because we have so much to tell you. [00:00:45] So let's go ahead and get started. [00:00:48] At AMD, our mission is to push the boundaries [00:00:51] of high performance in AI computing [00:00:53] to help solve the world's most important challenges. [00:00:56] And I often say AI is the most important technology [00:01:00] of the last 50 years. [00:01:02] And frankly, the progress that we're seeing in the industry [00:01:05] is just incredible. [00:01:07] Like every month, every few months, [00:01:09] we see something new that was far beyond our imagination. [00:01:13] And you can already see the impact across every industry. [00:01:17] In healthcare, AI is helping researchers [00:01:20] identify new drug candidates faster. [00:01:22] In science, we're solving problems [00:01:25] that we didn't think were possible a few years ago. [00:01:28] And across every industry, AI is changing the way we get work done. [00:01:33] And the most important thing is [00:01:35] we're still in the very, very early innings of what's possible. [00:01:40] Now, just take a look at some of these charts. [00:01:43] You know, we look at these every year [00:01:44] and you kind of see the rate and pace of AI growth. [00:01:49] A couple of years ago, we were just starting. [00:01:51] People were experimenting with chatbots [00:01:53] and, you know, it was really cool. [00:01:55] But if you look at today, [00:01:57] we're really seeing more than 35 quadrillion tokens [00:02:01] consumed every single month. [00:02:04] That's an increase of nearly 160 times in just two years. [00:02:09] And that curve is actually just getting steeper. [00:02:12] You're going to hear a lot about that today. [00:02:14] So just looking at, you know, where is all that demand? [00:02:17] I mean, obviously, training remains incredibly important [00:02:20] and foundational. [00:02:21] And over the last four or five years, [00:02:23] you know, we see the compute that's used for training, [00:02:26] advanced models, continuing to increase by, [00:02:28] let's call it roughly 5x every year. [00:02:31] And those models are getting better. [00:02:32] Like, we're all experiencing that. [00:02:34] We're seeing better reasoning. [00:02:35] We're seeing new capabilities. [00:02:37] We're now seeing a growing number of specialized models. [00:02:40] And one of the things that I really believe in [00:02:42] is there is no one perfect model. [00:02:45] I think we're all going to use a slew of different models, [00:02:48] depending on what task you're trying to solve [00:02:51] and which industry you're in. [00:02:53] But the bigger shift is actually in inference. [00:02:56] And we expected this. [00:02:58] We expected that inference would grow faster than training. [00:03:01] And we can say that in 2026, for the first time, [00:03:05] the world is using more AI compute to run models than to train them. [00:03:10] And we've seen that to a point where we think this year, [00:03:14] roughly 60% of the global AI compute capacity [00:03:17] will be used for inference. [00:03:19] And the reason for that is actually pretty simple. [00:03:23] As billions of people every day use AI, [00:03:27] the workloads shifts from building models to putting them to work. [00:03:30] And actually, we're going to talk a lot today about agentic AI. [00:03:35] And that's accelerating the shift even faster. [00:03:37] So if you think about all of this, [00:03:40] you know, we see agents as the next big step for AI. [00:03:44] And the most interesting thing is, [00:03:46] like, this has really just started accelerating [00:03:49] over the last, I would say, five or six months this year. [00:03:52] So when we started with LLMs, [00:03:54] they were great to answer questions. [00:03:56] But we found that to really get the power of AI, [00:04:00] you want agents to be able to really answer full questions. [00:04:05] So you give it a goal, [00:04:07] it keeps working until it's solved the problem. [00:04:10] And what that means for us in the compute industry, [00:04:13] it means that compute demand is growing at an incredible pace. [00:04:19] We're actually seeing a step change in compute demand. [00:04:22] Because when you ask the agent to do something, [00:04:25] it actually has dozens of steps. [00:04:28] And it has to reason, and it has to call tools, [00:04:30] and it has to access data, [00:04:31] and it has to keep doing it over and over [00:04:34] until it solved the problem. [00:04:35] And so you need lots of GPUs to do all that reasoning. [00:04:39] But importantly, you also need a lot of CPUs [00:04:42] to orchestrate every step around it. [00:04:45] So when you look at, you know, [00:04:47] just what that means from a market standpoint, [00:04:49] it's really hard to put this market data up [00:04:52] because I feel like every few months, [00:04:54] we're changing, you know, sort of the perspective [00:04:56] based on what our customers are saying. [00:04:58] But what we're seeing is that the shift to agentic AI [00:05:02] is certainly growing the AI accelerator market [00:05:05] very significantly. [00:05:07] It was actually last year at this event [00:05:09] that we called the AI accelerator market [00:05:12] at about 500 billion by 2028. [00:05:15] And at that time, [00:05:17] it actually felt like a very big number. [00:05:20] Didn't you think it was a big number? [00:05:22] Oh, yeah. [00:05:23] But honestly, you look at it today [00:05:25] and it looks conservative [00:05:26] because the AI demand [00:05:28] is just continuing to accelerate. [00:05:31] And you can see it, right? [00:05:32] Better models, more usage, [00:05:33] more usage requires more compute, [00:05:36] and more compute actually builds better models. [00:05:39] And so we're now expecting [00:05:40] that by 2030, [00:05:42] the AI accelerator market [00:05:44] is going to reach about 1.4 trillion by 2030. [00:05:48] So what do you think about that? [00:05:49] Is that a big number? [00:05:55] I mean, what that means is [00:05:57] by the end of the decade, [00:05:58] the AI accelerator market [00:06:00] is going to approach the size [00:06:01] of the entire semiconductor market today. [00:06:05] And although there will be [00:06:06] many different types of accelerators, [00:06:08] you know, I'm a big believer [00:06:09] in there's no one-size-fits-all [00:06:11] as it comes to chips, [00:06:13] we do expect that GPUs [00:06:15] are going to make up [00:06:16] the vast majority of that market [00:06:17] because the algorithms [00:06:19] are still very much in their infancy [00:06:21] and we're still continuing [00:06:23] to see the workloads change. [00:06:24] And that favors programmability [00:06:26] in the overall silicon ecosystem. [00:06:31] Now, perhaps the most interesting part [00:06:34] of the last six months [00:06:35] has been, you know, [00:06:36] GPUs are only part of this story. [00:06:39] What we're actually seeing [00:06:41] is Agentic AI is creating [00:06:43] an entirely new growth vector [00:06:45] for server CPUs. [00:06:47] Now, I've updated this number a lot also [00:06:49] over the last, you know, [00:06:51] six to eight months. [00:06:53] But look, we call it like we see it. [00:06:56] And, you know, we were early. [00:06:57] I mean, we saw from our largest [00:06:59] hyperscale customers, [00:07:00] and you're going to see some of them [00:07:01] here today, [00:07:02] who said, look, [00:07:03] as AI inference is going up, [00:07:06] we need more server CPUs. [00:07:07] And so we saw growth [00:07:08] in the server CPU market. [00:07:11] And we said, you know, [00:07:12] last year we thought it might grow, [00:07:13] let's call it 18 to 20% CAGR [00:07:15] to an overall market size [00:07:18] of 60 billion. [00:07:19] But frankly, [00:07:20] what we're seeing is that [00:07:21] the rate and pace [00:07:22] of agentic AI adoption [00:07:23] is much, much faster [00:07:25] than any of us thought. [00:07:27] So we're talking about agents [00:07:29] going from millions to billions. [00:07:32] Like we're able to increase [00:07:33] all of our productivity. [00:07:35] And that requires a tremendous amount [00:07:37] of CPU infrastructure. [00:07:39] And every customer conversation [00:07:41] is telling us that this buildout [00:07:43] is just beginning. [00:07:44] So based on what we're seeing today, [00:07:46] and we're going to talk more about [00:07:48] how agentic AI is evolving [00:07:49] with CPUs, [00:07:51] we now expect the server CPU market [00:07:53] to grow by over 50% [00:07:55] to over 200 billion by 2030. [00:07:59] So that's starting from today, [00:08:01] a $25 billion market [00:08:03] going to over 200 billion. [00:08:04] So there's a lot of excitement [00:08:06] about CPUs as well. [00:08:10] Now, probably one of the things [00:08:12] that is differentiating AMD [00:08:15] in how we think about the market [00:08:17] is this is not just [00:08:19] a data center opportunity. [00:08:21] As AI becomes a larger part [00:08:23] of our daily lives, [00:08:25] we want the intelligence to run [00:08:27] right where the work happens. [00:08:29] And that means cloud [00:08:31] is super important, [00:08:32] but the devices that we use every day [00:08:34] are going to be extremely important. [00:08:36] And that means that we're going to do [00:08:37] a lot more at the edge as well [00:08:39] with autonomous machines [00:08:40] that sense and act in real time. [00:08:43] And we really believe that you need AI [00:08:45] to be infused everywhere. [00:08:47] And that's exactly what we're focused on at AMD. [00:08:50] So putting all of that together, [00:08:52] we actually expect that the market [00:08:54] for our high-performance [00:08:55] and adaptive computing products [00:08:57] to grow at roughly a 40% tagger [00:08:59] over the next several years, [00:09:01] approaching a $2 trillion market [00:09:04] by 2030. [00:09:05] And frankly, [00:09:06] the only way you're going to be able [00:09:08] to service a market like that [00:09:09] is for us to work together [00:09:11] as an ecosystem. [00:09:13] There's no one company [00:09:14] that can solve it all, [00:09:15] but this is the opportunity [00:09:17] to bring the best [00:09:19] and brightest together. [00:09:20] And we have never been [00:09:22] in a better position to lead. [00:09:24] We have the broadest product portfolio. [00:09:27] We have the strongest roadmaps [00:09:28] we've ever had. [00:09:30] And the thing that I'm most proud of [00:09:31] is we have the deepest partnerships [00:09:33] with the companies [00:09:35] that are building this future. [00:09:38] So a little bit about our strategy. [00:09:40] Look, I think we've been very consistent. [00:09:43] You know, we've talked about [00:09:44] our multi-year strategy, [00:09:45] and it's really built [00:09:46] around three priorities. [00:09:48] The first is just compute leadership. [00:09:50] We're building the broadest set [00:09:51] of compute engines in the industry [00:09:53] so that we have the right compute [00:09:55] for the right workload. [00:09:57] Second, it's going to be [00:09:58] about open platforms. [00:10:00] We want everyone to come together [00:10:02] in an open ecosystem. [00:10:04] We believe an open ecosystem [00:10:05] is essential to the future of AI, [00:10:07] and that's how we get [00:10:08] the force multiplier [00:10:10] of everyone coming together. [00:10:12] And that's open on hardware [00:10:13] in terms of hardware standards, [00:10:15] as well as open on software. [00:10:17] And that's why we're investing [00:10:19] so heavily in our Rockham software. [00:10:21] And you're going to see [00:10:22] that AI has been [00:10:23] a tremendous multiplier [00:10:25] in this rate and pace [00:10:27] of progress that we're making [00:10:28] in Rockham and in software. [00:10:31] And it really means that [00:10:32] if we put all of this together, [00:10:35] we can have our customers [00:10:36] deploy on AMD hardware [00:10:39] faster than ever. [00:10:41] And the third piece is [00:10:43] we're going to get a chance [00:10:43] to talk to you [00:10:44] about powering AI everywhere. [00:10:46] So that's AI in enterprise, [00:10:48] that's AI at the PC, [00:10:50] that's AI in the physical world. [00:10:53] And that's adding new capabilities [00:10:54] across all of our products. [00:10:56] So today, you're going to see [00:10:57] all of that in action. [00:10:59] We're going to start with compute [00:11:00] for the agentic error. [00:11:02] I'm going to show you [00:11:02] a lot of hardware. [00:11:04] And then we're going to talk [00:11:05] about some software [00:11:06] and then our overall products [00:11:08] from a platform standpoint. [00:11:09] And we're honored to have [00:11:11] some really special guests [00:11:13] who are going to join us [00:11:14] to really help bring [00:11:15] the technology to life. [00:11:17] So let's start [00:11:18] with the data center. [00:11:20] Today, AMD EPYC runs [00:11:22] on the most important workloads [00:11:23] in the world. [00:11:25] We power the largest cloud providers, [00:11:27] we power the digital platforms [00:11:28] that billions of people use [00:11:29] every day and the most important [00:11:32] critical systems that are used [00:11:34] by the largest businesses, [00:11:36] including more than 60% [00:11:37] of the Fortune 100 [00:11:38] are using AMD EPYC. [00:11:40] And that momentum [00:11:41] is just building. [00:11:43] Last quarter, [00:11:44] we reached a record 46% revenue share [00:11:47] in the server market. [00:11:49] And we're continuing to see [00:11:51] every major customer [00:11:52] move more workloads to EPYC. [00:11:56] And on the GPU side, [00:11:58] our instinct adoption [00:11:59] has also accelerated. [00:12:01] We have a broad set of customers [00:12:02] across the largest AI labs, [00:12:05] cloud providers, [00:12:06] leading AI startups, [00:12:08] as well as some of the national labs [00:12:10] and sovereign AI opportunities [00:12:11] are being built on AMD instinct. [00:12:14] And we really appreciate [00:12:16] that opportunity. [00:12:19] Now, AI computes [00:12:20] becoming a lot more complicated. [00:12:22] And what we're seeing [00:12:24] with Frontier AI [00:12:24] is it's really raising the bar [00:12:27] for what infrastructure needs to do. [00:12:29] It takes more than a single chip [00:12:32] or a single server. [00:12:34] You actually have to design [00:12:36] the entire rack as one system. [00:12:39] And that requires leading CPUs, [00:12:42] that requires leading GPUs, [00:12:44] that requires high-speed networking [00:12:46] that connects everything inside the rack [00:12:48] and across the data center. [00:12:50] And just as importantly, [00:12:51] this is about making these systems [00:12:54] super easy to use. [00:12:56] And so they need to be easy to deploy, [00:12:58] easy to service, [00:12:59] and very reliable. [00:13:01] And that is exactly what we built [00:13:03] with Helios. [00:13:05] So today, [00:13:06] I'm super excited [00:13:07] to launch Helios, [00:13:08] the industry's highest-performance AI ref. [00:13:12] Now, Helios is built [00:13:19] to train and run [00:13:20] the most demanding Frontier models [00:13:22] in the world [00:13:22] at massive scale. [00:13:24] And I have a lot of show-and-tell for you. [00:13:27] So for those of you who know me, [00:13:28] you know that I love holding up chips. [00:13:33] I call myself sometimes [00:13:34] the Vanna White of chips. [00:13:37] But it turns out with Helios, [00:13:39] there are lots and lots of components. [00:13:41] And so you can see, [00:13:43] these are the chips [00:13:43] inside the Helios rack. [00:13:46] These are the Instinct. [00:13:49] Thank you. [00:13:53] So this is Instinct, [00:13:56] Epic, [00:13:57] and then the Pensando chips [00:13:59] for networking [00:13:59] that are inside each rack. [00:14:01] And every one of these chips [00:14:03] is part of the system. [00:14:05] So MI-455 is the engine. [00:14:08] It's the highest-performance GPU [00:14:09] in the industry. [00:14:11] Venice, [00:14:12] which drives it, [00:14:13] is the world's fastest CPU cores. [00:14:16] And Pensando DPUs, [00:14:20] actually, [00:14:20] and NICs, [00:14:21] actually connect them [00:14:22] with leadership programmability [00:14:23] and scale-out bandwidth. [00:14:25] So, [00:14:26] I have a few more props [00:14:27] to show you. [00:14:29] Let's kind of take a look [00:14:30] at how these things fit [00:14:31] into the overall system. [00:14:33] So first, [00:14:33] this is now [00:14:34] the MI-455 Accelerator, [00:14:37] which is mounted [00:14:38] on what we call [00:14:40] an Enhanced Accelerator Module, [00:14:42] or EAM. [00:14:43] It's a production module [00:14:45] that goes into Helios, [00:14:46] and it is much more [00:14:48] than the GPU package. [00:14:50] The EAM [00:14:50] actually integrates [00:14:52] the GPU, [00:14:53] memory, [00:14:54] power delivery, [00:14:55] high-speed interfaces, [00:14:57] system management, [00:14:59] cold plates [00:14:59] for liquid cooling, [00:15:01] all into one [00:15:02] compact, [00:15:03] serviceable module. [00:15:05] Pretty cool, right? [00:15:07] Yeah. [00:15:12] Now, [00:15:13] if I just give you [00:15:13] some of the specs, [00:15:14] we're talking about [00:15:15] 320 billion transistors. [00:15:18] This is built [00:15:19] with TSMC's [00:15:20] leading 2-nanometer [00:15:22] and 3-nanometer [00:15:22] process technology. [00:15:24] And most importantly, [00:15:25] it brings together [00:15:26] nearly a decade [00:15:27] of AMD chiplet innovation. [00:15:29] So it allows us [00:15:30] to combine [00:15:30] 12 compute [00:15:32] and I-O chiplets [00:15:33] with 432 gigabytes [00:15:35] of memory, [00:15:36] all connected [00:15:37] by leading industry [00:15:38] to 3D chip stack packaging. [00:15:41] And I can say for sure, [00:15:42] it's the highest performance [00:15:44] AI accelerator [00:15:45] in the industry. [00:15:48] OK, so this is [00:15:54] the CPU board [00:15:55] that's inside [00:15:56] the Helios compute tray. [00:15:57] And I'm going to talk [00:15:57] a lot about CPUs later. [00:15:59] But just to give you [00:16:00] a high-level view, [00:16:03] it's high-speed, [00:16:04] 96-core epic processor, [00:16:06] DDR5 memory, [00:16:07] all the I-O [00:16:08] that's needed [00:16:08] to feed the GPUs, [00:16:10] all in a single motherboard. [00:16:12] And we also have [00:16:14] our Selina DPU. [00:16:16] This is a critical part [00:16:18] of the networking [00:16:18] infrastructure. [00:16:20] And it really allows us [00:16:21] to deliver both [00:16:22] front-end [00:16:23] as well as [00:16:23] the scale-out [00:16:24] and scale-across [00:16:25] bandwidth overall. [00:16:27] And this is our [00:16:28] Vulcano AI NIC. [00:16:30] And you can see, [00:16:31] this actually allows us [00:16:33] to really have [00:16:34] up to six Vulcano NICs [00:16:36] on a board. [00:16:37] And it uses [00:16:38] the open [00:16:38] Ultra-Ethernet standard. [00:16:40] So each Helios compute tray [00:16:41] includes two of these [00:16:43] Vulcano boards. [00:16:44] And that allows us [00:16:45] to have the scale-out [00:16:47] and scale-up [00:16:48] of the scale-up [00:16:49] networking overall. [00:16:51] So you put all that [00:16:52] inside the rack [00:16:53] and we have [00:16:54] an additional [00:16:55] six dedicated [00:16:56] networking trays [00:16:57] that handle [00:16:57] the scale-up networking. [00:16:59] And those are all [00:17:00] connecting 75 GPUs [00:17:02] together with [00:17:02] UA-Link over Ethernet [00:17:04] with Silicon [00:17:05] from our partners. [00:17:06] And that's what we mean [00:17:07] by an open ecosystem. [00:17:09] So when you bring [00:17:10] this all together, [00:17:11] I want to say, [00:17:13] Helios is simply [00:17:14] the best AI rack [00:17:15] in the world. [00:17:24] Now, [00:17:25] let's take a look [00:17:25] at some numbers. [00:17:26] Clearly, [00:17:27] it's a very competitive [00:17:28] world out there. [00:17:29] So when you compare [00:17:30] Helios to the competition, [00:17:32] we're delivering [00:17:33] 15% more compute, [00:17:36] 50% more HBM4 memory [00:17:38] capacity and memory [00:17:39] bandwidth, [00:17:40] and 50% more [00:17:41] scale-out bandwidth. [00:17:43] And what that means [00:17:44] is that every Helios [00:17:46] can deliver more performance [00:17:47] for the largest models, [00:17:49] more capacity [00:17:50] for longer context, [00:17:52] and the bandwidth [00:17:52] to scale [00:17:53] across thousands [00:17:54] of racks. [00:17:56] And now, [00:17:57] we can kind of [00:17:57] come over here [00:17:58] and take a look [00:17:58] at another [00:17:59] very large piece [00:18:00] of hardware. [00:18:02] This is now [00:18:03] the production hardware [00:18:04] that customers [00:18:04] are deploying, [00:18:05] and you can see [00:18:05] more of that [00:18:06] as you go through [00:18:07] the exhibit center. [00:18:10] And each one of these [00:18:12] weighs more than [00:18:13] 160 pounds [00:18:14] and stands less [00:18:16] than 2 inches tall. [00:18:18] That's just [00:18:18] an extraordinary [00:18:19] amount of technology [00:18:21] when you look [00:18:22] at what's in [00:18:22] one of these trays. [00:18:23] And when you bring [00:18:24] these things together, [00:18:25] the numbers [00:18:26] are just incredible. [00:18:28] More than 18,000 [00:18:29] cDNA 5GPU [00:18:30] compute units, [00:18:32] over 4,600 [00:18:33] Zen6 CPU cores, [00:18:35] and 31 terabytes [00:18:37] of HBM4 memory, [00:18:38] all in a single rack. [00:18:41] That's what it takes [00:18:42] to run [00:18:43] agentic AI at scale. [00:18:46] So, [00:18:47] today, [00:18:47] I'm excited [00:18:48] to announce [00:18:48] that Helios [00:18:49] is in full production. [00:18:55] We have [00:18:59] shipments on track [00:19:00] to start [00:19:01] at the end [00:19:02] of the third quarter [00:19:02] and ramping [00:19:03] into the fourth quarter [00:19:04] and the second half [00:19:05] of the year. [00:19:06] And I can tell you, [00:19:07] customer demand [00:19:08] for Helios [00:19:08] is extremely strong. [00:19:10] And we're extremely proud [00:19:12] of the work [00:19:12] that we have done [00:19:13] across the leading [00:19:14] AI labs [00:19:15] to adopt Helios. [00:19:16] And so, [00:19:17] let's start [00:19:18] with some of our guests. [00:19:19] Today, [00:19:19] I'm excited to be joined [00:19:21] by our newest [00:19:22] Instinct partner [00:19:22] and one of the leaders [00:19:24] in Frontier AI. [00:19:25] To share more [00:19:26] about what we're [00:19:26] doing together, [00:19:27] please welcome [00:19:28] Anthropic co-founder [00:19:29] and Chief Compute Officer, [00:19:31] Tom Brown. [00:19:41] Hello, Tom. [00:19:43] It is so great [00:19:44] to have you here [00:19:45] and it is really [00:19:47] such an honor [00:19:47] to be working [00:19:48] with Anthropic. [00:19:49] We just, [00:19:49] we had a big week [00:19:50] this week. [00:19:51] We did, yeah. [00:19:52] So, look, [00:19:54] Anthropic has just [00:19:54] had incredible success [00:19:56] over the last few years. [00:19:58] You know, [00:19:58] Claude has made [00:19:59] such a big impact [00:19:59] on the overall industry. [00:20:01] Compute is such [00:20:02] an important factor [00:20:03] in the progress. [00:20:05] Can you just talk [00:20:05] a little bit [00:20:06] about your compute strategy [00:20:08] and just how things [00:20:09] have been evolving? [00:20:10] Yeah. [00:20:11] Yeah, and first, [00:20:11] thank you so much. [00:20:12] A hard act to follow [00:20:13] with Helios. [00:20:14] Amazing. [00:20:15] So, with Anthropic, [00:20:17] as you mentioned, [00:20:20] the scale [00:20:21] that we're growing [00:20:23] just as an industry [00:20:24] is enormous [00:20:25] and so we have been [00:20:27] working to make sure [00:20:29] that we have [00:20:30] the best, [00:20:31] like, different chips [00:20:33] and we can use [00:20:34] the best chips [00:20:35] for the best workloads. [00:20:37] Yeah, and look, [00:20:38] I think you've been [00:20:39] really leading in that [00:20:40] area and, you know, [00:20:42] I'm happy to say [00:20:43] that I've always [00:20:44] wanted to be [00:20:45] as part of your [00:20:46] infrastructure. [00:20:47] I remember the early [00:20:48] times we talked [00:20:48] as you were starting [00:20:49] Anthropic, [00:20:50] but we had a big [00:20:51] announcement this week. [00:20:52] We're honored [00:20:52] that you will be [00:20:53] deploying up to [00:20:54] two gigawatts [00:20:55] of Helios. [00:20:56] Why now? [00:20:57] What was the key reason? [00:20:59] Yeah, so I think [00:21:01] that, like, [00:21:02] the first thing [00:21:03] was just Helios [00:21:04] is an amazing machine. [00:21:05] It's, like, [00:21:06] it's absolutely [00:21:07] fantastic. [00:21:08] Do we like that? [00:21:09] Does that work for us? [00:21:14] And then I think [00:21:15] the thing that we [00:21:16] were thinking about [00:21:16] originally was [00:21:17] whenever we're [00:21:18] bringing up [00:21:18] a new hardware platform, [00:21:20] it's a big effort. [00:21:21] It's, like, [00:21:22] a huge thing. [00:21:23] And so as part [00:21:24] of the, [00:21:25] as we were thinking [00:21:26] about this, [00:21:27] we started doing [00:21:28] our own evaluation [00:21:30] of MI-355. [00:21:32] You guys generously [00:21:33] got us a rack [00:21:33] to start working. [00:21:35] And we expected [00:21:36] this to be [00:21:36] kind of a big process. [00:21:38] Our actual experience [00:21:39] was we had [00:21:41] one engineer [00:21:42] who started doing it. [00:21:43] They spun up, [00:21:44] they connected it [00:21:45] to Claude, [00:21:46] asked it, [00:21:47] hey, [00:21:48] bring up this machine, [00:21:49] left it going [00:21:50] over the weekend, [00:21:51] and we ended up [00:21:52] with a graph [00:21:53] of the actual [00:21:53] performance [00:21:54] of our leading [00:21:55] model on it [00:21:56] just going up [00:21:57] and up and up [00:21:58] over the weekend. [00:21:59] And so, [00:22:00] yeah. [00:22:03] And I think [00:22:04] that that's a testament [00:22:05] to the open platform [00:22:06] that you guys have made [00:22:07] where, like, [00:22:08] anyone, like, [00:22:09] human or AI [00:22:10] can now build [00:22:11] real models [00:22:12] on your platform. [00:22:15] Tom, [00:22:15] I kid you not, [00:22:16] when my team [00:22:17] told me that [00:22:17] you had an engineer [00:22:18] working on MI-355, [00:22:20] I said, [00:22:21] hey, [00:22:21] like, [00:22:22] are we, like, [00:22:22] making sure we help them? [00:22:24] And they're like, [00:22:24] they say they don't need [00:22:25] any help. [00:22:26] They got it all covered. [00:22:27] And I was like, [00:22:28] hey, [00:22:28] that's a wonderful story. [00:22:30] Thank you for that. [00:22:31] Now, [00:22:32] we're also super excited [00:22:33] about the work [00:22:34] we're doing together [00:22:35] because, [00:22:35] you know, [00:22:35] one of the most valuable [00:22:36] parts of this partnership [00:22:37] is, [00:22:38] you know, [00:22:38] Claude [00:22:39] and using Claude [00:22:40] within AMD. [00:22:41] Like, [00:22:41] we're big believers [00:22:42] in the fact that, [00:22:43] you know, [00:22:43] AI is this force multiplier [00:22:45] and the more we're able [00:22:46] to have the leading [00:22:48] foundational models [00:22:49] really familiar [00:22:50] with the AMD architecture, [00:22:51] you know, [00:22:51] the better we're going [00:22:53] to be able to service [00:22:53] the overall market. [00:22:54] So, [00:22:54] can you just, [00:22:55] you know, [00:22:55] talk a little bit [00:22:56] about how you're seeing [00:22:57] Claude really evolve [00:22:58] in, you know, [00:22:59] these types of, [00:23:00] you know, [00:23:00] highly engineering [00:23:01] specific workloads? [00:23:02] Yeah. [00:23:03] So, [00:23:03] I think this is a great place [00:23:05] for collaboration [00:23:06] where, [00:23:06] in turn, [00:23:07] like, [00:23:08] the bread and butter [00:23:08] of Claude [00:23:09] is doing software engineering [00:23:10] and now more and more [00:23:11] we're seeing it [00:23:12] help out with [00:23:13] the type of workloads [00:23:14] that you guys are doing also, [00:23:15] like design, [00:23:17] layout, [00:23:18] which is not quite [00:23:19] the normal software engineering [00:23:20] but is adjacent [00:23:21] and I do think [00:23:23] that that is the place [00:23:25] where we can work together [00:23:26] to make the next generation [00:23:28] of chips even better. [00:23:30] Yeah, [00:23:30] I think, [00:23:31] and also on the software side, [00:23:33] I think the work [00:23:33] that we've seen [00:23:34] with Claude [00:23:34] and the kernel development [00:23:36] has been just incredible. [00:23:39] So, [00:23:39] you know, [00:23:40] Tom, [00:23:40] one of the things [00:23:41] for our audience here, [00:23:42] we like to talk [00:23:43] about the present [00:23:43] but it's actually, [00:23:44] most of us [00:23:45] are working on the future. [00:23:46] so this is really [00:23:48] the beginning of, [00:23:49] you know, [00:23:49] what we believe [00:23:50] is a very strong [00:23:52] multi-year, [00:23:53] you know, [00:23:53] partnership. [00:23:54] Can you talk [00:23:55] a little bit about, [00:23:56] number one, [00:23:56] what are you most excited [00:23:57] about in the industry [00:23:58] and then, [00:24:00] number two, [00:24:00] like, [00:24:00] what can we do more together [00:24:01] to really make sure [00:24:05] that we're satisfying [00:24:06] those, [00:24:07] you know, [00:24:07] key opportunities? [00:24:08] Yeah, [00:24:09] that's a great question. [00:24:10] So, [00:24:12] I think the working together [00:24:14] for the scale-up [00:24:16] seems like [00:24:17] it's probably [00:24:18] the biggest thing [00:24:19] that I'm most excited for [00:24:20] where it does seem [00:24:21] like we see consistently [00:24:23] using more, [00:24:25] better compute [00:24:26] results in better models [00:24:28] that then add more value [00:24:30] to all the folks [00:24:30] using it downstream. [00:24:31] So, [00:24:32] I think that's [00:24:32] the biggest thing [00:24:33] and then I think [00:24:34] one thing now [00:24:35] that we're investing [00:24:36] more and more [00:24:37] with you guys too [00:24:37] and I think [00:24:38] we can continue to [00:24:40] is security [00:24:41] where that's a place [00:24:42] where now [00:24:43] as the machines [00:24:44] are getting more complex, [00:24:46] we can work together [00:24:47] to secure things [00:24:48] at the chip layer, [00:24:49] the server level, [00:24:50] the rack, [00:24:51] the whole network [00:24:52] and that'll help out [00:24:53] like, [00:24:54] not just us [00:24:55] but the entire industry [00:24:56] stay safe. [00:24:56] No, [00:24:57] I think you're [00:24:57] completely right. [00:24:58] We are super focused [00:24:59] on the idea of, [00:25:01] you know, [00:25:01] how can we use AI [00:25:02] to improve every aspect [00:25:03] of our product [00:25:04] whether it's, [00:25:05] you know, [00:25:06] performance, [00:25:06] power, [00:25:07] software, [00:25:07] and security [00:25:08] as you said. [00:25:09] So, [00:25:10] really, [00:25:10] Tom, [00:25:11] huge thank you [00:25:12] for the, [00:25:13] you know, [00:25:14] the real, [00:25:14] the opportunity [00:25:15] to work very closely [00:25:16] with you and your team. [00:25:16] Our team loves [00:25:17] working with you guys [00:25:18] and we look forward [00:25:19] to everything [00:25:20] that we're going [00:25:20] to do together. [00:25:21] Thank you so much. [00:25:22] Thank you. [00:25:23] Thank you. [00:25:23] Thank you. [00:25:23] All right. [00:25:30] So, [00:25:31] look, [00:25:31] Tom talked about [00:25:32] how important [00:25:33] performance is [00:25:34] and we have been [00:25:36] laser focused [00:25:37] on ensuring [00:25:38] that the real world [00:25:39] performance [00:25:39] of MI455 [00:25:41] and Helios [00:25:41] really comes true [00:25:43] when we look [00:25:44] at, [00:25:44] you know, [00:25:45] system results. [00:25:46] So, [00:25:46] let's take a look [00:25:47] at some of those results. [00:25:49] So, [00:25:50] here's an example. [00:25:51] We're running [00:25:51] DeepSeq V4 Flash [00:25:53] one of the latest [00:25:54] reasoning models [00:25:55] and we are actually [00:25:56] comparing MI455 [00:25:58] to MI355 [00:25:59] and, you know, [00:26:00] the key thing [00:26:01] is with every generation [00:26:02] we want to make [00:26:03] huge leaps [00:26:04] in performance. [00:26:05] So, [00:26:05] we're seeing that [00:26:06] at lower concurrency [00:26:07] that MI455 [00:26:09] delivers 4x [00:26:10] the throughput [00:26:11] of MI355 [00:26:12] but, [00:26:13] clearly, [00:26:14] what you see [00:26:14] with, [00:26:15] you know, [00:26:15] large users [00:26:16] with, [00:26:17] across the board [00:26:18] that you want [00:26:19] to increase [00:26:20] the number [00:26:20] of simultaneous users, [00:26:21] you want to increase [00:26:22] the amount of work [00:26:22] that you can do [00:26:23] at the same time. [00:26:24] And so, [00:26:25] at the highest concurrency, [00:26:26] we're seeing Helios [00:26:27] deliver up to 34 times [00:26:30] more throughput [00:26:31] than our prior generation. [00:26:33] And what that means [00:26:34] is more users, [00:26:36] much faster response, [00:26:37] and much better efficiency [00:26:39] at scale. [00:26:40] And performance [00:26:41] is just one aspect [00:26:43] of it, right? [00:26:43] The other aspect [00:26:44] and, you know, [00:26:45] what I should have mentioned [00:26:46] when Tom was on stage [00:26:47] is every one of us [00:26:48] in the enterprise [00:26:49] is seeing our AI budgets [00:26:50] go up every single month. [00:26:52] Is that right? [00:26:53] Are you seeing some of that? [00:26:55] And so, [00:26:56] customers really care [00:26:57] about cost per token. [00:26:59] And with every generation [00:27:00] of Instinct, [00:27:01] we're driving [00:27:02] the cost per token down. [00:27:04] And that's [00:27:04] with more memory, [00:27:06] more bandwidth, [00:27:06] and much more compute. [00:27:08] And we're taking [00:27:09] another major step [00:27:10] with MI455, [00:27:12] delivering up to 18 times [00:27:14] more tokens per dollar [00:27:16] so customers can serve [00:27:17] far more users [00:27:18] with the same investment. [00:27:20] And as you think about [00:27:22] how that comes together [00:27:23] in a rack, [00:27:25] many of our data centers [00:27:26] right now [00:27:27] are really limited by power. [00:27:28] So power is kind of [00:27:29] the maximum limiter. [00:27:31] And so we tested Helios [00:27:32] over a full range [00:27:34] of workloads [00:27:35] where you have [00:27:36] a fixed rack power [00:27:37] compared to the competition. [00:27:39] And what we're seeing [00:27:40] is across the highest [00:27:42] throughput workloads, [00:27:43] the most interactive applications, [00:27:44] all of these leading [00:27:45] inference workloads, [00:27:47] we're seeing Helios [00:27:48] an average of 10% to 15% [00:27:50] more performance [00:27:51] than the competition. [00:27:53] And that comes [00:27:54] with all of the system work [00:27:55] that we've done [00:27:56] in the overall system. [00:27:59] And how that translates [00:28:00] to a customer [00:28:00] is cost. [00:28:03] What we expect [00:28:03] is the Helios rack [00:28:05] delivers more performance [00:28:07] and it also delivers [00:28:08] up to 30% more tokens [00:28:11] per dollar [00:28:11] than the competition. [00:28:13] So I think that's [00:28:14] a pretty good [00:28:15] overall value proposition. [00:28:24] Now, I'm happy to say [00:28:26] that the customer demand [00:28:27] for Helios [00:28:28] is extremely strong. [00:28:29] From the largest AI labs [00:28:31] to hyperscale [00:28:32] and neocloud providers, [00:28:34] we are working [00:28:34] across the entire ecosystem [00:28:36] to enable Helios [00:28:37] together with our OEM [00:28:39] and ODM partners. [00:28:41] And one of our deepest [00:28:43] and earliest partners [00:28:44] deploying Helios [00:28:45] is OpenAI. [00:28:47] To talk about [00:28:48] where AI is headed [00:28:49] and the work [00:28:49] we're doing together, [00:28:51] please welcome [00:28:51] to the stage [00:28:52] OpenAI's [00:28:53] Head of Infrastructure, [00:28:54] Sachin Khati. [00:29:01] Thank you, Sachin Khaki. [00:29:03] Thank you, Sachin. [00:29:04] It's so great [00:29:05] to have you here [00:29:06] and most importantly, [00:29:08] we are so thankful [00:29:09] for the partnership [00:29:09] with OpenAI. [00:29:11] We've been through [00:29:11] so much [00:29:12] on this journey together. [00:29:13] So you're really [00:29:15] operating on the frontier [00:29:16] of AI. [00:29:17] Can you just tell us [00:29:18] a little bit [00:29:19] of what you're seeing [00:29:20] and just what compute [00:29:22] looks like for you? [00:29:23] Thanks. [00:29:24] Great to be here. [00:29:25] I was thinking [00:29:26] about this event [00:29:27] and over the last [00:29:28] two years, [00:29:29] it feels like [00:29:29] the attendance [00:29:30] at this event [00:29:31] is following [00:29:32] scaling loss [00:29:33] and it's been doubling [00:29:34] every year [00:29:35] over the last [00:29:36] two years. [00:29:37] But more seriously [00:29:38] to your question, [00:29:40] we've always [00:29:41] predicted that [00:29:42] models will evolve [00:29:44] where they will start [00:29:45] as chatbots, [00:29:46] then they become [00:29:47] reasoners, [00:29:48] then they become [00:29:48] agents, [00:29:49] then they become [00:29:49] interns. [00:29:50] and the trajectory [00:29:52] has kept up [00:29:53] with that. [00:29:53] So we are all [00:29:54] seeing that models [00:29:55] are becoming [00:29:56] a lot more capable, [00:29:57] more agentic [00:29:58] and able to do [00:29:59] very long running [00:30:00] tasks, [00:30:01] let them go [00:30:02] and they just [00:30:02] take care of it. [00:30:03] And so all that [00:30:04] really means is [00:30:06] we need to keep [00:30:07] scaling compute. [00:30:08] The more compute [00:30:09] we scale [00:30:10] and throw at training, [00:30:12] the more capabilities [00:30:13] emerge. [00:30:14] The more compute [00:30:14] we scale [00:30:15] and provide it [00:30:17] to distributing [00:30:18] intelligence to the world, [00:30:19] the more people [00:30:19] want to do with it. [00:30:20] and this is what [00:30:22] we see with [00:30:22] exploding token [00:30:23] budgets. [00:30:24] We've seen [00:30:25] internally, [00:30:26] for example, [00:30:26] in OpenAI [00:30:27] that not just [00:30:28] engineering, [00:30:29] but every aspect [00:30:30] of the enterprise [00:30:31] is beginning [00:30:32] to use agents [00:30:33] for every part [00:30:33] of the work. [00:30:34] This is why [00:30:35] we need platforms [00:30:36] like yours [00:30:37] that can scale [00:30:39] with how quickly [00:30:40] the capabilities [00:30:41] of these models [00:30:41] are scaling [00:30:42] as well as [00:30:43] how quickly [00:30:43] the whole world [00:30:44] is beginning [00:30:45] to embrace [00:30:46] what these things [00:30:46] can do. [00:30:47] Well, Sachin, [00:30:48] I don't think [00:30:48] I've ever [00:30:49] spoken to you [00:30:51] where you haven't [00:30:51] asked me [00:30:52] for more compute. [00:30:53] So he's actually [00:30:54] very, very consistent. [00:30:56] Look, our teams [00:30:57] have been working [00:30:58] really closely together. [00:30:59] I think it's been [00:31:00] a true journey [00:31:01] when we think [00:31:01] about just all [00:31:02] of the technology [00:31:03] that we've been doing. [00:31:04] You know, [00:31:05] last year, [00:31:05] we announced [00:31:06] a really, [00:31:07] you know, [00:31:07] landmark partnership [00:31:09] where you're going [00:31:10] to deploy [00:31:11] up to six gigawatts [00:31:13] of AMD infrastructure. [00:31:14] You're one of the first, [00:31:15] I think you're the first [00:31:16] to actually have [00:31:17] MI-455 racks [00:31:19] a few months ago. [00:31:20] So can you talk [00:31:21] a little bit [00:31:21] about the journey? [00:31:23] No, it's been [00:31:23] a phenomenal partnership. [00:31:25] As you mentioned, [00:31:27] we were first [00:31:28] to bet on AMD [00:31:29] and we are really thrilled [00:31:30] with how that bet [00:31:31] is turning out for us. [00:31:33] Got started on MI-300 [00:31:35] and expanded [00:31:36] very quickly [00:31:37] to 355. [00:31:38] And it's been [00:31:39] really fun [00:31:41] for our team [00:31:42] to see how [00:31:42] it was possible [00:31:44] to optimize [00:31:45] the software stack, [00:31:46] the networking [00:31:47] and everything [00:31:47] and deploy [00:31:48] our models [00:31:49] running on AMD [00:31:51] infrastructure. [00:31:52] So as Lisa mentioned, [00:31:54] so we just got [00:31:55] hands on the Helios [00:31:56] tracks three months ago [00:31:57] and the collaboration [00:31:58] has gone [00:31:59] to the next level. [00:32:00] Our engineers [00:32:00] are working [00:32:01] side by side [00:32:02] with AMD engineers [00:32:04] to optimize [00:32:05] the software stack [00:32:06] and run GVT class [00:32:07] workloads [00:32:08] on Helios [00:32:09] already. [00:32:10] Phil will probably [00:32:11] on stage later [00:32:12] talking about [00:32:13] how productive [00:32:14] and fast [00:32:16] that collaboration is. [00:32:17] So we are really [00:32:18] excited about [00:32:19] the capabilities [00:32:20] Helios is already [00:32:21] showing us [00:32:22] and we expect [00:32:24] that we'll be [00:32:24] deploying Helios [00:32:25] at massive scale [00:32:26] starting towards [00:32:27] the end of this year [00:32:28] and then accelerating [00:32:29] throughout 2027. [00:32:31] Like I said, [00:32:31] I know, [00:32:32] more faster. [00:32:33] We need it earlier. [00:32:36] So, [00:32:37] Sachin, [00:32:37] the other thing [00:32:38] is, you know, [00:32:39] one of the most [00:32:39] exciting parts [00:32:40] of our collaboration [00:32:41] is actually [00:32:41] some of the research [00:32:42] work. [00:32:43] So, you know, [00:32:43] very thankful [00:32:44] for the opportunity [00:32:45] to spend, [00:32:47] you know, [00:32:47] really bringing [00:32:47] our best engineers [00:32:48] together with your [00:32:49] researchers [00:32:50] to talk about [00:32:51] how AI [00:32:53] should be, [00:32:53] you know, [00:32:53] really implemented [00:32:54] and used [00:32:55] for, you know, [00:32:56] future both silicon [00:32:57] and software. [00:32:58] So can you talk [00:32:59] a little bit [00:32:59] about that work? [00:33:01] Yeah, [00:33:01] I think one of [00:33:03] our key bets [00:33:04] and I think [00:33:05] for a lot of us [00:33:06] in AI [00:33:06] is about recursion, [00:33:08] right? [00:33:09] AI recursively [00:33:11] improving the systems [00:33:12] it needs to run on [00:33:13] and eventually, [00:33:15] obviously, [00:33:15] AI doing the research [00:33:16] itself, right? [00:33:18] Today, [00:33:18] expert engineers [00:33:19] spend an enormous [00:33:20] amount of time [00:33:20] optimizing kernels, [00:33:23] figuring out [00:33:23] how to translate [00:33:24] new models [00:33:25] to run them [00:33:26] efficiently [00:33:26] on different hardware, [00:33:28] dealing with [00:33:28] all the usual things [00:33:29] you think about, [00:33:30] compilers, [00:33:30] communications, [00:33:32] libraries, [00:33:32] system configurations. [00:33:33] What we've been [00:33:35] super excited [00:33:36] and encouraged [00:33:37] by the work [00:33:38] we are doing [00:33:39] with you [00:33:39] is now [00:33:41] AI showing [00:33:42] the potential [00:33:43] and demonstrating [00:33:44] it in production [00:33:45] to automate [00:33:46] a significant portion [00:33:47] of the work [00:33:48] when anyone programs [00:33:49] to an AMD GPU. [00:33:50] So it dramatically [00:33:52] has shortened [00:33:53] the path [00:33:53] from a new model [00:33:55] coming up [00:33:55] to an efficient [00:33:57] production deployment [00:33:58] and something [00:33:59] that we can [00:34:00] much more easily [00:34:01] tune to changing [00:34:02] workload requirements, [00:34:03] right? [00:34:03] And that's [00:34:04] a big deal. [00:34:05] So one of the other [00:34:06] things that's [00:34:07] super exciting for us [00:34:08] is because AMD [00:34:09] is built around [00:34:10] an open software [00:34:10] ecosystem, [00:34:11] all these things [00:34:12] that we are now [00:34:14] inventing internally [00:34:15] for our own use, [00:34:17] we believe [00:34:18] we can also bring [00:34:19] it to the whole world [00:34:20] and allow the whole world [00:34:21] to use it [00:34:22] for all of the AI [00:34:23] models out there, [00:34:24] not just our models. [00:34:26] I think super excited [00:34:27] about that. [00:34:28] I think the idea [00:34:28] is that it's really [00:34:30] the tide [00:34:30] that lifts all boats, [00:34:31] right? [00:34:32] As we work, [00:34:33] it helps our work together, [00:34:34] but it really helps [00:34:34] the overall ecosystem. [00:34:36] So a little bit [00:34:37] of your crystal ball. [00:34:39] Looking forward, [00:34:40] I think we all know [00:34:41] that compute [00:34:41] is super critical. [00:34:43] What do you need [00:34:44] from the next generation [00:34:45] AI infrastructure [00:34:46] and what can we do [00:34:48] as really a key partner [00:34:51] to help you [00:34:52] accomplish that mission? [00:34:53] If you haven't [00:34:54] heard already, [00:34:55] I need more compute [00:34:56] more quickly. [00:34:58] At least he's consistent. [00:35:01] I think the thing [00:35:05] that the whole industry [00:35:06] is realizing [00:35:07] is it's a systems problem [00:35:08] at a data center scale. [00:35:10] So I think we obviously [00:35:12] very quickly internalized [00:35:13] that AI was a rack level [00:35:14] problem, [00:35:15] but now it is very clear [00:35:17] that it's not just [00:35:18] a rack level problem, [00:35:19] it's a rack in a data center. [00:35:20] So the whole data center [00:35:21] is a system. [00:35:23] And we have to think [00:35:24] about how we design [00:35:25] the future based on [00:35:26] what we know [00:35:27] about our workloads, [00:35:28] what's coming down the line, [00:35:29] but co-design it with you [00:35:31] from CPUs to GPUs [00:35:33] to memory, [00:35:34] networking, [00:35:34] storage, [00:35:35] power distribution [00:35:36] in a data center [00:35:37] and the cooling systems [00:35:38] that go with it. [00:35:39] These are all intertwined [00:35:40] and they all have [00:35:42] to be co-designed. [00:35:43] And one of the really fun [00:35:46] parts about what we do [00:35:47] with you and your team [00:35:49] is how easy [00:35:50] and productive it is [00:35:52] for us to work with you [00:35:53] and give you insight [00:35:54] into how we expect [00:35:56] the world to change [00:35:56] in the future [00:35:57] from what our workload [00:35:58] needs are, [00:35:59] what our models are, [00:36:00] and how quickly [00:36:01] your team responds [00:36:02] in turning those insights [00:36:04] into better chips, [00:36:06] better systems, [00:36:06] better software. [00:36:08] So that's worked well, [00:36:09] really well with MI400 [00:36:10] and left shifted [00:36:11] a lot of the engineering [00:36:14] work that we needed to do [00:36:15] to bring this up [00:36:16] online quickly. [00:36:17] And that's why [00:36:18] we are very confident [00:36:19] we can productionize [00:36:20] those systems [00:36:21] very quickly. [00:36:22] And so really looking forward [00:36:24] to accelerating that [00:36:25] with MI500 and beyond [00:36:26] and increasingly using AI [00:36:29] to actually do that work [00:36:30] for us rather than [00:36:32] being limited by humans. [00:36:34] That's fantastic, Sachin. [00:36:36] And look, [00:36:36] we really do recognize [00:36:38] that input is so, [00:36:39] so valuable. [00:36:40] So I have to say [00:36:41] thank you again [00:36:42] for just the extraordinary [00:36:43] partnership, [00:36:44] the amount of work [00:36:45] that our joint teams [00:36:46] are doing together. [00:36:47] And we couldn't be more [00:36:48] excited about the opportunities [00:36:50] in front of us. [00:36:51] Thank you, Sachin. [00:36:52] Thank you so much. [00:36:52] Yes. [00:36:58] All right. [00:36:59] So now let's turn [00:37:01] to the world of CPUs. [00:37:03] There's been a lot of talk [00:37:04] about CPUs recently, [00:37:06] and I can say [00:37:07] that this has been [00:37:07] the foundation [00:37:08] of our AMD data center [00:37:10] strategy for the longest time. [00:37:12] We launched Epic in 2017, [00:37:14] and we've been [00:37:15] on a very, very clear mission. [00:37:18] Every generation, [00:37:19] more performance, [00:37:20] more efficiency, [00:37:21] more capability [00:37:23] for a wider range [00:37:24] of workloads. [00:37:26] Now, Naples [00:37:26] was our first generation. [00:37:28] It established [00:37:28] the foundation, [00:37:30] but Rome and Milan [00:37:31] really changed [00:37:32] the economics [00:37:32] in the data center. [00:37:33] We did more [00:37:34] than what people expected [00:37:36] in terms of adding [00:37:37] core count [00:37:38] and adding throughput. [00:37:39] And then Turin set, [00:37:41] Genoa expanded [00:37:42] our performance, [00:37:43] and then Turin set [00:37:44] a new bar [00:37:45] for density, [00:37:46] throughput, [00:37:46] and total cost [00:37:47] of ownership. [00:37:48] That's really been [00:37:49] the Epic formula, [00:37:50] a clear roadmap [00:37:51] and very consistent execution [00:37:54] and leadership [00:37:55] that actually grows [00:37:57] with every generation. [00:37:58] This is why, [00:37:59] if you look today, [00:38:01] 5th Gen Epic Turin [00:38:02] is the best server CPU [00:38:03] in the world. [00:38:04] With up to 192 cores [00:38:07] and 384 threads, [00:38:09] Turin delivers [00:38:10] leadership performance [00:38:11] across a broad range [00:38:13] of cloud, [00:38:13] enterprise, [00:38:14] and HPC workloads. [00:38:16] But the thing [00:38:17] about the data center [00:38:18] is compute [00:38:19] never stands still. [00:38:21] With agentic AI, [00:38:23] it's actually creating [00:38:24] a whole new class [00:38:25] of infrastructure, [00:38:26] and the CPU matters [00:38:27] now more than ever. [00:38:29] And that's exactly [00:38:30] what we built Venice for. [00:38:33] So there's been [00:38:34] a lot of talk [00:38:35] about agentic AI, [00:38:36] but it's actually [00:38:36] a very new field. [00:38:38] And what we see [00:38:39] is that server computing [00:38:41] is actually splitting [00:38:42] across a couple [00:38:43] of different workloads. [00:38:45] So the first section, [00:38:47] you know, [00:38:47] is GPU servers, [00:38:48] and that's fairly well known. [00:38:50] That's where the GPU's job, [00:38:53] the CPU's job, [00:38:53] is basically to drive [00:38:54] the GPUs. [00:38:56] And so the job there [00:38:57] is all about speed. [00:38:59] We want the highest [00:39:00] frequency cores, [00:39:02] we want the fastest I.O. [00:39:03] and we want to keep [00:39:04] the GPUs fully fed. [00:39:06] Now the biggest middle part, [00:39:08] this is the largest growth, [00:39:09] is actually in agent servers, [00:39:11] or what we call [00:39:11] agent sandboxes. [00:39:13] And this is actually [00:39:14] a whole new class [00:39:16] of workload. [00:39:16] Here you execute code, [00:39:18] you actually call [00:39:19] a bunch of tools, [00:39:20] you actually query data [00:39:22] that's outside the model, [00:39:23] you have a whole bunch [00:39:24] of things around it. [00:39:26] And in this place, [00:39:28] the priority [00:39:28] is actually density. [00:39:30] So we want [00:39:31] the highest performing [00:39:32] cores per watt [00:39:33] to run thousands [00:39:35] of agents at once. [00:39:37] And then we have [00:39:38] our traditional [00:39:39] general purpose servers [00:39:40] that run the enterprise. [00:39:42] And here, [00:39:43] you're going to have [00:39:43] a diversity of workloads. [00:39:45] It's all about efficiency, [00:39:46] it's running the applications, [00:39:47] it's running databases, [00:39:48] it's running data services, [00:39:50] and all of that [00:39:51] is, you know, [00:39:52] what we would traditionally say [00:39:54] is, you know, [00:39:55] general purpose server. [00:39:56] And so with these [00:39:57] different workloads, [00:39:58] you actually need [00:39:59] the right CPU [00:40:00] for the right workload. [00:40:02] And Epic [00:40:03] is the only CPU portfolio [00:40:05] that leads [00:40:06] across all three. [00:40:09] So Venice [00:40:10] is our newest [00:40:11] and most advanced [00:40:12] server CPU family, [00:40:13] and it's actually designed [00:40:14] for the agentic era. [00:40:16] We're extending [00:40:17] TURN's leadership [00:40:18] across every single metric. [00:40:20] That is performance, [00:40:22] that's efficiency, [00:40:23] that's TCO for cloud, [00:40:25] that's enterprise efficiency, [00:40:27] that's HPC workloads. [00:40:29] And Venice starts [00:40:30] with our all-new [00:40:31] Zen 6 core. [00:40:33] We have higher IPC, [00:40:35] higher frequency, [00:40:36] and that delivers [00:40:37] up to 1.8 times [00:40:39] more performance [00:40:40] than TURN. [00:40:41] That is one of the largest [00:40:43] generational gains. [00:40:44] I mean, [00:40:44] we have been doing this [00:40:45] for six generations, [00:40:47] but this is one [00:40:48] of the largest [00:40:48] generational gains [00:40:49] in the history of Epic. [00:40:51] It's 203 billion [00:40:53] transistors built [00:40:54] on TSMC's newest [00:40:55] 2 nanometer process, [00:40:56] and it uses [00:40:58] a next-generation [00:40:59] chiplet that supports [00:41:00] up to 512 threads [00:41:02] per socket. [00:41:03] And this is actually [00:41:04] a really important point. [00:41:05] I'm going to show you [00:41:05] why in a few minutes. [00:41:08] This gives us [00:41:08] the highest compute density, [00:41:10] but it also allows us [00:41:11] to double both the memory [00:41:13] and the I-O bandwidth. [00:41:15] So, I have a few more [00:41:16] chips to show you, [00:41:17] if that's all right. [00:41:19] And what's clear [00:41:21] is Venice is not one chip. [00:41:25] Venice is actually [00:41:26] an entire family [00:41:27] of chips. [00:41:28] And this is where [00:41:29] our chiplet architecture [00:41:30] becomes a huge, [00:41:32] huge advantage. [00:41:33] It lets us take [00:41:35] the Zen 6 architecture [00:41:36] and build a whole portfolio [00:41:38] around it. [00:41:39] So, let's start [00:41:40] with this guy. [00:41:43] This is Venice HF. [00:41:45] This is the highest [00:41:47] performing CPU [00:41:49] that drives [00:41:50] AI host nodes. [00:41:52] It has [00:41:53] eight compute chiplets, [00:41:54] each with 12 cores, [00:41:56] running at up to 5 gigahertz [00:41:58] with the I-O [00:41:59] and memory bandwidth [00:41:59] to keep the GPUs [00:42:01] completely fed. [00:42:02] And this is the CPU [00:42:03] that I showed you [00:42:04] before shipping [00:42:05] inside Helios. [00:42:13] Now, [00:42:14] this is its [00:42:17] bigger brother. [00:42:19] It's Venice 256 core. [00:42:21] This is the chip [00:42:22] that's built [00:42:23] for EGentex sandboxes. [00:42:25] And it also has [00:42:26] eight compute chiplets, [00:42:28] but each of the chiplets [00:42:29] now has 32 cores. [00:42:31] So, it scales [00:42:32] all the way [00:42:32] to 512 threads. [00:42:34] And it is [00:42:35] the highest compute density [00:42:36] in the industry. [00:42:43] Customers get [00:42:44] more agent capacity [00:42:45] per watt, [00:42:46] per dollar, [00:42:47] per rack, [00:42:48] bar none. [00:42:52] Okay, [00:42:53] and then [00:42:53] one more member [00:42:55] of the family. [00:42:56] This one is [00:42:57] Venice 128 core. [00:42:59] And this guy [00:43:00] is optimized [00:43:00] for enterprise [00:43:01] and general purpose servers. [00:43:03] This is the same [00:43:04] leadership performance [00:43:05] per core, [00:43:06] but it's in a lower cost, [00:43:07] lower power design [00:43:08] for enterprise applications. [00:43:11] And this one [00:43:12] is really tuned [00:43:12] to give customers [00:43:13] the best performance [00:43:14] per dollar [00:43:15] across a wide range [00:43:16] of configurations. [00:43:17] So, you can see, [00:43:18] really, [00:43:19] the power of the family [00:43:21] as they come together. [00:43:23] And in addition, [00:43:24] with these guys, [00:43:25] I can tell you [00:43:25] that next year, [00:43:26] we're adding [00:43:27] a few other chips [00:43:28] to the family. [00:43:29] We're adding [00:43:29] Verano, [00:43:30] which is the next generation [00:43:31] for AI host nodes. [00:43:33] And what that adds [00:43:34] is a more power-efficient, [00:43:36] low-power memory, [00:43:37] as well as [00:43:38] faster interconnect [00:43:39] between the CPU [00:43:40] and the GPU. [00:43:41] And then we're also adding [00:43:42] for our high-performance computing, [00:43:44] our technical computing, [00:43:46] Venice X, [00:43:47] which is using [00:43:48] our 3D V-cache [00:43:49] stacking technology. [00:43:50] And we're really [00:43:51] the first [00:43:51] to stack [00:43:52] memory chiplets [00:43:53] right below [00:43:54] the compute chiplets [00:43:55] to accelerate performance [00:43:56] even for those [00:43:58] types of workloads. [00:43:59] So, you can see [00:44:00] one architecture, [00:44:01] a CPU [00:44:02] is not a CPU. [00:44:04] There are many [00:44:04] different types [00:44:05] of CPUs. [00:44:06] And for us, [00:44:07] one architecture [00:44:08] spans dozens [00:44:09] of different chips. [00:44:10] And that breadth [00:44:11] is someone [00:44:12] that no other [00:44:13] CPU vendor [00:44:14] can match. [00:44:15] And that's [00:44:15] what makes [00:44:16] 6th Gen Epic [00:44:17] the best CPU [00:44:19] in the data center. [00:44:19] Now, you're [00:44:27] going to bear [00:44:27] with me [00:44:28] because I'm [00:44:28] going to take [00:44:29] you through [00:44:29] some performance [00:44:29] because the numbers [00:44:30] are just incredible. [00:44:32] So, looking [00:44:33] at AI workloads, [00:44:34] right? [00:44:35] Again, you have [00:44:35] different use cases [00:44:37] and different workloads. [00:44:38] You know, [00:44:39] starting first [00:44:40] with GPU servers. [00:44:41] Many of the [00:44:42] largest frontier [00:44:43] models require [00:44:44] you to move [00:44:44] data between [00:44:46] the CPU [00:44:46] and the GPU. [00:44:47] And in that [00:44:48] scenario, [00:44:48] Venice moves [00:44:49] the data [00:44:49] faster than [00:44:50] the best [00:44:51] competitive [00:44:52] x86 processor [00:44:53] delivering up [00:44:54] to 1.8 times [00:44:56] more tokens [00:44:56] per second. [00:44:58] And when you [00:44:58] look at [00:44:59] the agentic [00:45:00] CPU servers [00:45:00] and sandboxes, [00:45:02] because of our [00:45:02] density, [00:45:03] Venice delivers [00:45:03] more than twice [00:45:04] the agents [00:45:05] per watt [00:45:05] for agent [00:45:06] sandboxes. [00:45:08] And for [00:45:08] general purpose [00:45:09] applications, [00:45:09] in this case, [00:45:10] we're standardizing [00:45:11] on a 100-kilowatt [00:45:12] rack, [00:45:13] Venice delivers [00:45:14] more than twice [00:45:15] the performance [00:45:16] per watt. [00:45:17] So, pretty [00:45:18] incredible. [00:45:18] results. [00:45:21] And the gap [00:45:22] is even wider [00:45:23] if you look [00:45:24] at ARM processors. [00:45:26] So, if you look [00:45:26] at the leading [00:45:27] ARM processors, [00:45:28] when we're [00:45:28] comparing at the [00:45:30] chip level, [00:45:31] Venice supports [00:45:32] up to 2.8 times [00:45:33] more agents [00:45:34] per watt. [00:45:35] And when you [00:45:36] look at the [00:45:36] rack level, [00:45:37] Venice is delivering [00:45:38] up to 3.3 times [00:45:40] more performance [00:45:41] per watt. [00:45:42] And what that [00:45:43] means is, [00:45:44] at the data [00:45:45] center scale, [00:45:45] you can just [00:45:47] get a lot [00:45:47] more agents [00:45:48] in the same [00:45:50] power envelope. [00:45:57] Now, [00:45:58] there's been a lot [00:45:59] of talk [00:46:00] about what matters [00:46:01] most for [00:46:01] agentic AI, [00:46:03] with some saying [00:46:04] that per core [00:46:05] performance under [00:46:05] load is the only [00:46:06] thing that counts. [00:46:08] that's really [00:46:09] only part of the [00:46:10] story. [00:46:11] Because agentic AI [00:46:12] actually runs as a [00:46:13] distributed platform, [00:46:15] and that is across [00:46:16] databases and vector [00:46:17] stores and [00:46:18] orchestration. [00:46:19] And all of that [00:46:20] happens has to happen [00:46:21] at once. [00:46:22] And that requires [00:46:23] not just raw speed, [00:46:25] but you absolutely [00:46:26] need density, [00:46:28] efficiency, [00:46:29] and per core [00:46:30] performance all [00:46:31] together. [00:46:32] And what I'm [00:46:33] happy to say [00:46:33] is that Venice [00:46:34] leads across [00:46:36] every one of [00:46:36] them. [00:46:37] So when you [00:46:38] compare Venice [00:46:39] against the [00:46:40] highest performing [00:46:41] ARM CPU from [00:46:42] our competition, [00:46:43] Epic delivers [00:46:44] 20% higher [00:46:46] per core [00:46:47] performance. [00:46:48] And when we... [00:46:49] I like that [00:46:55] number. [00:46:56] I like that [00:46:57] number. [00:46:58] And when we [00:46:58] pair that with [00:46:59] leadership core [00:47:00] density, [00:47:00] we're delivering [00:47:01] 2.2 times [00:47:02] more performance [00:47:03] at the socket [00:47:03] level. [00:47:04] And so the main [00:47:05] point is whether [00:47:07] a customer needs [00:47:08] per core or [00:47:09] per socket [00:47:10] performance, [00:47:11] Venice is the [00:47:12] leader. [00:47:13] And there's one [00:47:13] more point that [00:47:14] a benchmark [00:47:15] doesn't really [00:47:16] capture. [00:47:17] Because when you [00:47:18] think about what [00:47:18] agentic AI is, [00:47:20] it's really like [00:47:20] you're adding [00:47:21] thousands of [00:47:22] employees to your [00:47:23] enterprise, to [00:47:24] your sort of [00:47:24] traditional [00:47:25] workflow. [00:47:26] And that [00:47:27] software enterprises [00:47:28] have been running [00:47:29] on for years. [00:47:30] The truth [00:47:31] is, that [00:47:32] software runs [00:47:33] on x86. [00:47:35] So Venice [00:47:35] runs all of it. [00:47:37] And this is where [00:47:37] x86 has an [00:47:38] advantage. [00:47:39] You must have [00:47:40] the best [00:47:40] performance, you [00:47:41] must have the [00:47:41] best density, but [00:47:42] having that [00:47:43] software compatibility [00:47:44] is a big, big [00:47:45] plus. [00:47:46] So customers [00:47:47] scale up their [00:47:48] agents on the [00:47:49] stack that they [00:47:49] already have. [00:47:51] And now, let me [00:47:52] just show you what [00:47:53] that means at the [00:47:54] rack level. [00:47:55] So at the rack [00:47:55] level, what we [00:47:56] find is, you know, [00:47:57] density is super [00:47:58] important. [00:47:59] Everyone's trying [00:47:59] to optimize their [00:48:00] data center. [00:48:01] Everyone's trying [00:48:01] to optimize their [00:48:02] power envelope. [00:48:03] And with Venice, [00:48:05] every major [00:48:06] server OEM is [00:48:07] offering a broad [00:48:08] spectrum of racks [00:48:09] for agentic AI. [00:48:10] So what that [00:48:11] means is, we [00:48:12] have the full [00:48:12] spectrum from, [00:48:14] you know, 25,000 [00:48:15] cores to 50,000 [00:48:17] cores, and that's [00:48:18] more agents per [00:48:19] rack, and that [00:48:20] gives customers the [00:48:21] choice. [00:48:22] The key is, every [00:48:23] data center is [00:48:24] different. [00:48:25] And with that, the [00:48:26] flexibility allows [00:48:28] you to pick the [00:48:29] right cooling, the [00:48:30] right space [00:48:31] requirements for what [00:48:32] you're trying to do [00:48:33] in your data center. [00:48:34] So there you have [00:48:35] it. [00:48:36] Venice, the best [00:48:37] server CPU in [00:48:39] the industry. [00:48:46] And I'm very happy [00:48:47] to say also today [00:48:48] that Venice is in [00:48:49] full production. [00:48:51] Customer demand [00:48:51] is incredible. [00:48:53] It's the strongest [00:48:53] we've ever seen [00:48:54] for a new [00:48:55] epic generation. [00:49:01] We're seeing [00:49:02] every major [00:49:03] server OEM, [00:49:04] every major [00:49:05] cloud provider [00:49:05] on track to [00:49:07] begin rolling out [00:49:08] in the fourth [00:49:08] quarter as we [00:49:10] start with making [00:49:11] the broadest [00:49:11] epic launch [00:49:12] we've ever had. [00:49:14] So with that, [00:49:15] I want to turn [00:49:16] to my next guest [00:49:17] who runs some [00:49:18] of the largest [00:49:19] and most advanced [00:49:19] compute infrastructure [00:49:20] in the world. [00:49:21] And it's really [00:49:22] been our privilege [00:49:23] to partner with [00:49:23] them as they've [00:49:24] deployed multiple [00:49:26] generations of both [00:49:27] epic and instinct. [00:49:28] To share more [00:49:29] about our work [00:49:30] together, please [00:49:31] welcome to the [00:49:31] stage, Meta's [00:49:32] head of [00:49:33] infrastructure, [00:49:34] Santosh [00:49:34] Dinardin. [00:49:43] Santosh is a true [00:49:50] friend. [00:49:51] I have to say [00:49:52] it's been a [00:49:53] tremendous opportunity [00:49:54] over the last few [00:49:55] years. [00:49:56] My story about [00:49:57] Santosh is the first [00:49:58] time I sat in his [00:49:59] office. [00:50:00] He said to me, [00:50:00] Lisa, we're going to [00:50:02] deploy a lot of [00:50:03] CPUs. [00:50:05] please make [00:50:05] sure they [00:50:05] work. [00:50:08] And I said, [00:50:08] I will. [00:50:09] I will. [00:50:10] And you [00:50:10] lifted it up. [00:50:11] I understand. [00:50:13] I try. [00:50:13] But look, [00:50:14] Santosh, [00:50:14] Meta operates, [00:50:16] you operate some [00:50:17] of the most [00:50:18] advanced [00:50:18] infrastructure [00:50:18] in the world. [00:50:20] What's happening [00:50:21] in the data [00:50:22] center world? [00:50:22] What's happening [00:50:23] in your world? [00:50:23] Tell us a little [00:50:24] bit about [00:50:25] the architecture [00:50:26] and what you're [00:50:27] working on. [00:50:28] Sure. [00:50:28] First of all, [00:50:29] thank you. [00:50:29] It's awesome. [00:50:30] This is like a [00:50:30] hardware geek's [00:50:31] paradise, right? [00:50:33] You're talking [00:50:34] about transistors, [00:50:35] you're talking [00:50:35] about hardware, [00:50:36] you're talking [00:50:36] about cooling. [00:50:37] This is sort [00:50:38] of the gem. [00:50:38] I like it. [00:50:40] And I think [00:50:41] this is an audience [00:50:41] that's receptive [00:50:42] to it. [00:50:42] I usually talk [00:50:44] to audiences, [00:50:44] they have usually [00:50:45] either not been [00:50:46] to a data center [00:50:47] or not really [00:50:47] seen a chip ever. [00:50:49] So this is [00:50:50] refreshing, [00:50:50] I have to say. [00:50:52] The thing I'll [00:50:52] say about, [00:50:53] see, listen, [00:50:54] we are seeing [00:50:54] demand go [00:50:55] through the roof. [00:50:57] When you look [00:50:57] at inference, [00:50:58] training, [00:50:59] recommendation [00:51:00] system, [00:51:00] that's how [00:51:01] a newsfeed works. [00:51:02] content creation, [00:51:04] the demand for that [00:51:05] is just exponential [00:51:06] right now, right? [00:51:07] And when I think [00:51:08] about sort of [00:51:09] the overall, [00:51:12] Mark has this vision [00:51:13] of delivering [00:51:14] personal super [00:51:15] intelligence [00:51:15] to billions of people. [00:51:17] So if you go [00:51:18] to sort of Facebook [00:51:18] or Instagram [00:51:19] or WhatsApp [00:51:20] or whatever [00:51:20] surface you go to, [00:51:22] the idea is [00:51:22] that we'll meet [00:51:23] you there [00:51:23] and we'll deliver [00:51:24] super personalized [00:51:26] intelligence [00:51:27] right wherever you are. [00:51:29] And that's why [00:51:29] we established [00:51:30] this meta-compute [00:51:31] initiative [00:51:31] and because [00:51:32] it's a top level [00:51:33] that Mark himself [00:51:34] oversees. [00:51:35] Now, [00:51:35] what that means [00:51:36] is that [00:51:36] we are now [00:51:37] moving away [00:51:38] from an area [00:51:38] where we used [00:51:39] to take servers, [00:51:40] optimize it, [00:51:41] right? [00:51:42] This is what [00:51:42] the conversations [00:51:42] we were having [00:51:43] many years ago [00:51:44] that, hey, [00:51:44] I'm deploying a CPU [00:51:45] I really need [00:51:46] to maximize [00:51:46] performance out of it. [00:51:48] It's different now. [00:51:49] You talked about [00:51:50] how we should be [00:51:52] thinking about [00:51:52] the whole system [00:51:53] end to end. [00:51:54] We are at a point [00:51:55] where we have to [00:51:56] look about [00:51:56] the whole data center [00:51:58] and think about [00:51:58] it as one [00:51:59] integrated system. [00:52:00] Servers, [00:52:01] hardware, [00:52:01] networking, [00:52:02] cooling, [00:52:03] power, [00:52:03] all of that [00:52:04] is not optional. [00:52:05] All of that [00:52:06] has to work [00:52:06] together, [00:52:07] right? [00:52:07] And this is where [00:52:08] I think [00:52:08] there's a huge [00:52:09] opportunity [00:52:09] for collaboration [00:52:10] with partners [00:52:11] like AMD [00:52:11] because you have [00:52:13] to now co-design, [00:52:14] co-create systems, [00:52:16] not just deploy [00:52:17] sort of something [00:52:18] that comes [00:52:18] off the shelf, [00:52:19] right? [00:52:20] The other thing [00:52:20] is, by the way, [00:52:21] flexibility. [00:52:22] The thing I really [00:52:23] like about what you [00:52:23] said is that [00:52:24] this is such [00:52:25] a big opportunity [00:52:27] in the industry [00:52:28] right now. [00:52:29] All of us [00:52:29] need to lean in [00:52:30] and sort of [00:52:31] work on this [00:52:31] together. [00:52:32] So it needs [00:52:34] to be open, [00:52:34] it needs [00:52:35] to be heterogeneous [00:52:35] and I think [00:52:37] that's no one [00:52:38] company, [00:52:39] one partner [00:52:39] that works [00:52:40] for any one [00:52:41] of us. [00:52:42] We need to [00:52:42] sort of work [00:52:43] on this together [00:52:43] and that's why [00:52:44] I think [00:52:44] you're one of [00:52:45] our most important [00:52:46] strategic partners. [00:52:47] Thank you, [00:52:48] Santosh. [00:52:48] And look, [00:52:49] we completely [00:52:50] agree with [00:52:50] your philosophy. [00:52:52] Now, [00:52:52] you are actually [00:52:53] one of our [00:52:54] broadest partners [00:52:54] because if you [00:52:55] think about [00:52:55] our work [00:52:56] on CPUs, [00:52:57] GPUs, [00:52:58] you know, [00:52:58] we developed [00:52:59] the RackSkill [00:53:00] OCP systems [00:53:02] together. [00:53:03] You've deployed [00:53:04] lots and lots [00:53:04] of CPUs, [00:53:05] so millions [00:53:05] of CPUs [00:53:06] and, you know, [00:53:07] you're a lead [00:53:08] partner on [00:53:09] Venice. [00:53:09] Can you just [00:53:10] talk a little bit [00:53:10] about sort of [00:53:11] your CPU [00:53:11] and infrastructure [00:53:12] and some [00:53:12] of our work [00:53:13] together? [00:53:14] Sure. [00:53:14] It's been many [00:53:15] years and the [00:53:16] story that Lisa [00:53:17] was saying [00:53:18] was many years [00:53:18] ago. [00:53:18] I think we went [00:53:19] from Milan [00:53:20] to Bergamo [00:53:21] to Turin [00:53:22] now to Venice. [00:53:23] So at least [00:53:24] the fourth [00:53:25] generation, [00:53:26] so long-time [00:53:27] partnership, [00:53:27] obviously. [00:53:28] I actually think [00:53:30] the world is [00:53:30] changing in the [00:53:31] sense that [00:53:32] it used, [00:53:33] the world is [00:53:34] changing, [00:53:34] every year is a [00:53:34] new world these days. [00:53:35] It's like every [00:53:36] month. [00:53:36] Month, exactly. [00:53:38] So what ends up [00:53:38] happening now is [00:53:39] that it's not just [00:53:40] a GPU game [00:53:41] anymore, [00:53:41] it's CPUs [00:53:42] and GPUs. [00:53:43] If anything, [00:53:44] I think CPUs [00:53:44] are becoming [00:53:45] at least as [00:53:45] important, [00:53:46] if not more, [00:53:47] you're looking [00:53:48] at a world [00:53:48] where sort [00:53:49] of the [00:53:49] workforce, [00:53:50] the workload [00:53:51] is changing [00:53:52] because while [00:53:53] you have your [00:53:54] workhorses [00:53:55] and GPUs, [00:53:56] at the end [00:53:56] of the day, [00:53:57] there's [00:53:58] agentic workloads, [00:53:59] you have to run [00:53:59] your tools, [00:54:00] you have to run [00:54:00] your systems, [00:54:01] and all the code [00:54:02] that's being generated [00:54:03] still needs to run [00:54:04] somewhere, right? [00:54:05] So I'm super [00:54:06] excited to work [00:54:07] on Venice. [00:54:07] I think there's a lot [00:54:08] to come there. [00:54:10] And you have [00:54:12] to think again, [00:54:13] this is a theme [00:54:14] that I'm assuming [00:54:15] will go throughout [00:54:16] the presentation [00:54:16] that you have [00:54:17] to start thinking [00:54:18] about CPUs [00:54:18] and GPUs [00:54:19] as conjoint things. [00:54:21] You hand off [00:54:21] workloads, [00:54:22] depending on the workload [00:54:23] you employ [00:54:23] the right hardware. [00:54:24] Yes, [00:54:25] no, [00:54:25] absolutely right. [00:54:26] And we've gotten [00:54:27] a lot of feedback [00:54:28] from your technical team. [00:54:29] I think that's [00:54:29] what I really appreciate [00:54:30] about the partnership [00:54:32] with Meta. [00:54:34] Now, [00:54:35] clearly, [00:54:35] you're deploying [00:54:35] lots of GPUs too, [00:54:37] a lot of accelerators, [00:54:38] and you're also [00:54:39] one of our [00:54:40] deepest partners [00:54:41] on the accelerator side [00:54:42] starting with MI300 [00:54:43] and then we had [00:54:44] a very large [00:54:46] strategic partnership [00:54:48] announced around [00:54:49] up to 6 gigawatts [00:54:50] starting with MI450. [00:54:52] And what we're doing [00:54:53] is quite unique [00:54:53] with Meta. [00:54:54] So can you talk [00:54:55] a little bit about [00:54:55] sort of the evolution [00:54:57] and the MI450 plans? [00:55:01] Again, [00:55:01] multi-year collaboration. [00:55:03] I think we started [00:55:03] with MI300. [00:55:05] We did a bunch. [00:55:06] It had really nice [00:55:07] memory capacity. [00:55:07] I remember talking [00:55:08] to you about that, [00:55:09] right? [00:55:09] and then [00:55:10] there's a bunch [00:55:11] of pipe cleaning [00:55:11] with it. [00:55:12] It was important [00:55:13] to deploy that [00:55:14] at production scale, [00:55:16] sort of get hands-on [00:55:17] experience on both sides. [00:55:19] 350 is the first time [00:55:20] we deployed this [00:55:21] on our ranking [00:55:21] and recommendation systems. [00:55:23] So that's now [00:55:24] getting into [00:55:25] sort of some good numbers. [00:55:27] And 450 is, [00:55:28] I think, [00:55:28] I'm super excited about it. [00:55:29] We just got some [00:55:30] of the racks. [00:55:31] And we are going [00:55:32] to deploy it [00:55:33] across the board [00:55:33] because it gives us [00:55:35] the opportunity [00:55:35] to go and sort [00:55:37] really collaborate. [00:55:38] I think the difference, [00:55:39] big difference [00:55:40] between sort of [00:55:40] the 300, [00:55:41] 350 and 450 [00:55:42] is that we have had [00:55:43] an engineer sitting [00:55:44] in the same room [00:55:45] talking to each other, [00:55:46] co-designing, [00:55:47] co-creating, [00:55:48] like power cooling. [00:55:50] Like I was saying, [00:55:50] it's not just a single thing [00:55:52] and we are not shy [00:55:54] in feedback. [00:55:55] They're not shy, [00:55:56] for sure. [00:55:57] But you're very receptive. [00:55:59] So this is true partnership, [00:56:00] right? [00:56:00] This is the point [00:56:01] about co-designing [00:56:02] and co-creating [00:56:03] that I think. [00:56:04] And like I was saying, [00:56:06] that there's a huge [00:56:07] sort of deployment [00:56:08] coming in. [00:56:09] The more we deploy, [00:56:10] the earlier we co-design, [00:56:12] the better we are. [00:56:13] Yeah, absolutely. [00:56:14] Really, really appreciate [00:56:15] the effort together [00:56:16] on MI450. [00:56:18] Now, I'm going to ask you [00:56:19] also about your crystal ball. [00:56:21] So you look out [00:56:21] into the future [00:56:22] and you see what the next [00:56:24] few years really means [00:56:26] as you push forward [00:56:27] in the AI industry. [00:56:29] What are the biggest challenges [00:56:30] and really opportunities [00:56:32] for us? [00:56:32] This is the industry [00:56:33] ecosystem here. [00:56:34] So what should we be [00:56:35] focused on as an industry? [00:56:37] Listen, all of you [00:56:38] know about the bottlenecks. [00:56:40] You know about the power, [00:56:41] the data centers, [00:56:42] the silicon. [00:56:43] All of these, I think, [00:56:44] are choke points [00:56:45] that I think the industry [00:56:46] is sort of waiting on. [00:56:48] But those are things [00:56:49] I'm pretty confident [00:56:50] will sort of work [00:56:51] our way through. [00:56:51] At the end of the day, [00:56:53] I think that demand [00:56:54] is immense. [00:56:54] People are responding. [00:56:56] But the thing [00:56:57] that I really want [00:56:58] to make sure [00:56:59] all of us realize [00:56:59] is that we should not [00:57:01] think about systems [00:57:02] in isolations. [00:57:04] Like I was saying, [00:57:05] CPUs and GPUs [00:57:06] are one conjoint system. [00:57:07] You should start [00:57:07] thinking about it that way. [00:57:09] We really need [00:57:10] to co-design things early. [00:57:13] See, data centers [00:57:14] takes years to build. [00:57:15] One of my favorite stories [00:57:17] is Mark comes to me [00:57:18] and says, [00:57:18] hey, I want a gig [00:57:19] worth of data centers. [00:57:21] Well, you should have [00:57:21] talked to me two years ago. [00:57:23] It takes time. [00:57:24] It's the same with silicon. [00:57:26] It just takes time. [00:57:27] So we need to be starting [00:57:29] and sitting down [00:57:30] in a room, [00:57:31] co-designing today [00:57:32] for what we need [00:57:33] to deploy in 27 and 28. [00:57:35] That, I think, [00:57:36] is when we truly unlock [00:57:37] the powers of the system. [00:57:39] The systems are amazing. [00:57:40] The specs are amazing, [00:57:41] right? [00:57:41] But think about [00:57:42] how much more [00:57:43] you can get out of it [00:57:44] if you sit in a room [00:57:45] and just co-design it today. [00:57:48] And then in 28, [00:57:49] we'll be having [00:57:49] much better graphs out there. [00:57:52] Fantastic. [00:57:52] Santoj, [00:57:53] I completely agree with you. [00:57:54] I want to say again, [00:57:55] thank you. [00:57:56] We are so, so happy [00:57:57] with the amazing partnership [00:57:59] that we have, [00:58:00] you know, [00:58:00] across the board [00:58:01] and most importantly, [00:58:02] you know, [00:58:02] with your engineering teams. [00:58:03] And we truly are excited [00:58:04] about what we're going [00:58:05] to do in the future. [00:58:06] Thank you so much. [00:58:07] Thanks a lot. [00:58:07] Cheers. [00:58:07] Thank you. [00:58:08] Thank you. [00:58:10] So, you heard Santoj [00:58:15] talk about just [00:58:16] the diversity of workloads [00:58:17] and really the massive scale [00:58:20] that you require with AI. [00:58:22] Now, I want to turn [00:58:23] to another part [00:58:24] of the inference market [00:58:25] where the requirements [00:58:26] are actually very different. [00:58:27] So, lots and lots [00:58:29] of workloads [00:58:29] and as inference is moving [00:58:31] into more products [00:58:32] and services, [00:58:33] we're actually seeing [00:58:34] it segment [00:58:34] into different workloads [00:58:35] depending on what [00:58:36] you're trying to do. [00:58:37] Some of these are, [00:58:39] let's call it less [00:58:39] interactive, [00:58:40] so they're really tuned [00:58:41] for, let's call it [00:58:42] maximum throughput [00:58:43] or the lowest cost. [00:58:45] You know, [00:58:45] other of these applications [00:58:46] require more balanced throughput, [00:58:48] so you have to balance [00:58:49] throughput and responsiveness. [00:58:52] And then there's [00:58:52] this new class [00:58:53] of applications [00:58:54] that actually want [00:58:55] very, very fast results. [00:58:57] You have extremely [00:58:58] smart engineers [00:58:59] and they don't like to wait. [00:59:01] And that is where [00:59:02] ultra-low latency [00:59:03] or every millisecond [00:59:04] actually matters. [00:59:06] Each one of these [00:59:07] types of workloads [00:59:08] requires a different [00:59:09] type of compute. [00:59:11] And one of the best ways [00:59:12] to reach ultra-low latency [00:59:14] today is disaggregated inference. [00:59:17] So, if you think about, [00:59:18] you know, [00:59:18] the different pieces, [00:59:19] both pre-fill [00:59:20] and decode [00:59:21] are two different jobs. [00:59:23] So, pre-fill [00:59:24] tends to need more compute. [00:59:26] Decode tends to need [00:59:27] more memory bandwidth. [00:59:29] And with Instinct, [00:59:30] we have a very balanced machine [00:59:31] so that we deliver [00:59:32] both great compute [00:59:33] and great memory bandwidth. [00:59:35] But you can actually [00:59:36] take this a step further. [00:59:37] If you know what workloads [00:59:38] you're trying to run, [00:59:40] you can actually let customers [00:59:41] tune each of these pieces [00:59:43] independently. [00:59:44] And to address [00:59:45] this market opportunity, [00:59:47] we've been working [00:59:48] with Cerebrus. [00:59:49] So, to talk more [00:59:50] about what we're [00:59:50] building together, [00:59:51] please welcome to the stage [00:59:53] Cerebrus co-founder [00:59:54] and CEO, [00:59:55] Andrew Feldman. [00:59:56] Andrew, it's great [01:00:06] to have you here. [01:00:06] It's been an exciting [01:00:07] few months for you. [01:00:09] I know that [01:00:09] for those who don't know [01:00:11] exactly what you've [01:00:13] been working on, [01:00:13] can you talk a little bit [01:00:14] about Cerebrus [01:00:15] and what problem [01:00:16] you've been trying to solve? [01:00:17] Sure. [01:00:18] Great to be here [01:00:19] and great to be [01:00:20] sort of among people [01:00:22] who love hardware. [01:00:24] It's nice. [01:00:25] Look, I'm through [01:00:26] and thrilled to be [01:00:26] here with you [01:00:27] and to announce [01:00:29] our cool new partnership. [01:00:32] At Cerebrus, [01:00:33] we build the world's [01:00:35] largest and fastest chip. [01:00:37] It's a full wafer. [01:00:39] And we package this [01:00:40] into a system [01:00:41] and we package [01:00:42] two systems, right, [01:00:44] into racks [01:00:45] and then racks [01:00:46] into clusters. [01:00:48] And we deliver them [01:00:49] both on-premise [01:00:50] and via the cloud. [01:00:52] And you've had [01:00:53] some of our customers [01:00:54] up here already today [01:00:55] on the frontier labs. [01:00:57] We serve customers [01:00:58] like OpenAI [01:00:59] and in the hyperscalers [01:01:01] like AWS [01:01:02] and the small agentic [01:01:04] and the coding space, [01:01:06] leaders like Cognition. [01:01:08] And they do this [01:01:10] because we're blisteringly fast. [01:01:14] There's no question, [01:01:15] there's no question, [01:01:15] there's no question, [01:01:15] Andrew, [01:01:16] that you have [01:01:16] some tremendous innovation [01:01:17] with what you've been working on. [01:01:20] And our teams [01:01:21] have been collaborating [01:01:21] really closely [01:01:22] over the last few years [01:01:24] to really have Helios [01:01:26] together with the wafer scale engine [01:01:27] for ultra-low latency inference. [01:01:29] Can you just kind of just [01:01:31] educate the audience [01:01:32] a little bit? [01:01:32] What are we trying to do [01:01:33] and what does that mean [01:01:34] for customers? [01:01:35] Sure. [01:01:35] I think what's happened [01:01:37] and you described it previously [01:01:39] is that AI has moved [01:01:41] from being a novelty [01:01:42] to being useful [01:01:44] and then in some domains [01:01:45] a necessity. [01:01:47] And when something's [01:01:48] a necessity, [01:01:49] people want to use it [01:01:50] and they want to use it quickly. [01:01:52] And to serve this market, [01:01:54] this segment [01:01:55] of ultra-low latency, [01:01:57] we sought a partnership [01:01:58] that could extend [01:02:00] our capabilities [01:02:02] and our footprint. [01:02:04] And there was no better answer [01:02:05] than AMD. [01:02:06] We were already [01:02:07] an AMD customer [01:02:08] as we use AMD CPUs [01:02:10] to surround our systems. [01:02:13] And we were so excited [01:02:14] at the opportunity [01:02:15] to build a disaggregated solution [01:02:17] that combined AMD CPUs, [01:02:21] the cool new Helios rack [01:02:23] and the Cerebrus [01:02:25] wafer scale engine. [01:02:28] Yeah, look, [01:02:29] the technology [01:02:30] is pretty cool. [01:02:31] So just talk a little bit [01:02:32] about how it comes together. [01:02:33] Sure. [01:02:34] So as Lisa described, [01:02:36] what you have [01:02:37] with Instinct [01:02:38] and the Helios rack [01:02:39] is you have the leader [01:02:41] in performance [01:02:42] and memory capacity. [01:02:45] And you marry that [01:02:46] with our wafer scale engine, [01:02:48] which is the leader [01:02:49] in sort of in SRAM [01:02:51] and in memory bandwidth. [01:02:53] And that combination [01:02:55] allows us to deliver [01:02:56] a solution [01:02:57] that is unmatched [01:02:58] in the industry. [01:02:59] So very, very exciting. [01:03:00] I know that customers [01:03:01] are excited [01:03:02] about what we're going [01:03:02] to do together as well. [01:03:03] So when are we going [01:03:05] to have it in market? [01:03:07] Later this year. [01:03:09] And it will be first [01:03:10] in the Cerebrus cloud. [01:03:14] and it will be later [01:03:16] at a store near you. [01:03:18] I think the key thing [01:03:21] is we're going to give [01:03:22] customers a choice [01:03:23] to really put together [01:03:25] what is the solution [01:03:26] that they want. [01:03:27] And I think we're very, [01:03:28] very happy to be able [01:03:29] to do this together with you. [01:03:30] I think it's a huge step up [01:03:32] in terms of what we can do [01:03:33] for this ultra low latency segment. [01:03:35] I think that's right. [01:03:36] I think until recently [01:03:37] customers sort of, [01:03:39] they could have [01:03:40] sort of high throughput [01:03:42] or they could have [01:03:44] extraordinary speed. [01:03:45] And by bringing together [01:03:47] the Helios rack [01:03:49] with the Cerebrus [01:03:50] wafer scale engine, [01:03:51] we give you five times [01:03:53] the throughput [01:03:54] while continuing [01:03:56] to deliver [01:03:57] this extraordinary speed. [01:03:58] it's really something amazing. [01:04:02] Fantastic. [01:04:02] Well, look, Andrew, [01:04:03] huge congratulations [01:04:04] to what you [01:04:05] and the Cerebrus team [01:04:06] have done. [01:04:07] I mean, I think [01:04:07] it's been incredible. [01:04:08] I think you've been clear [01:04:10] on what you're trying [01:04:11] to accomplish. [01:04:12] And it's really nice [01:04:13] to see not only it [01:04:14] come together, [01:04:15] but us come together [01:04:16] to offer a solution [01:04:17] that will be very, [01:04:19] very compelling [01:04:19] to customers. [01:04:20] So we can't wait [01:04:20] to get this [01:04:21] into customers' hands. [01:04:22] Soon. [01:04:23] Thank you. [01:04:24] Good to see you. [01:04:24] Thank you so much. [01:04:25] Thank you, Andrew. [01:04:28] All right. [01:04:31] So look, [01:04:32] I think we've shown you [01:04:33] a lot of hardware [01:04:34] this morning. [01:04:35] But as we all know, [01:04:37] hardware is only part [01:04:38] of the story. [01:04:38] It's actually software [01:04:40] that turns all of this [01:04:41] into a platform [01:04:41] that developers [01:04:43] can really build on. [01:04:44] So to share the progress [01:04:45] we're making with Rockham, [01:04:47] please welcome [01:04:47] AMD's Senior Vice President [01:04:49] of AI, [01:04:50] Vamsi Bopana, [01:04:50] to the stage. [01:04:58] Thank you, Risa. [01:04:59] Good morning, everyone. [01:05:01] Morning. [01:05:02] It's great to be back [01:05:03] to talk about software. [01:05:05] Just a few years ago, [01:05:06] programming AMD GPUs [01:05:08] required deep expertise [01:05:09] and significant [01:05:10] engineering effort. [01:05:12] We've come a long, long way [01:05:13] in a remarkably short [01:05:15] period of time. [01:05:16] Through our open source [01:05:18] community collaboration [01:05:19] and sustained investment [01:05:21] in Rockham, [01:05:21] our software stack, [01:05:23] AMD Instinct GPUs [01:05:24] are powering some [01:05:25] of the most important [01:05:26] and consequential [01:05:28] workloads on the planet. [01:05:30] But the biggest change [01:05:32] is still ahead of us. [01:05:34] AI is transforming [01:05:35] how software is getting built [01:05:37] and we are bringing [01:05:38] that transformation [01:05:39] to Rockham. [01:05:41] And that's what I'm excited [01:05:42] to share with you guys today. [01:05:44] Our software teams [01:05:45] have been moving fast [01:05:46] with relentless focus [01:05:47] on developers. [01:05:49] We've invested [01:05:50] at every level of the stack [01:05:52] in build and test infrastructure [01:05:53] so we ship faster. [01:05:56] Rockham releases now go out. [01:05:57] But every six weeks, [01:05:59] not every four months, [01:06:01] we've continued [01:06:02] to expand the work [01:06:03] we do with our AI ecosystem [01:06:04] partners. [01:06:05] And the result? [01:06:07] More features, [01:06:09] faster performance, [01:06:10] and a significantly better [01:06:11] out-of-the-box experience. [01:06:14] Our strategy [01:06:15] that has gotten us here [01:06:16] has been remarkably consistent. [01:06:19] It's built [01:06:19] on two core principles. [01:06:21] Partner deeply [01:06:22] with the open source ecosystem [01:06:24] and build [01:06:26] with the right layers [01:06:27] of abstraction [01:06:27] to enable developer productivity. [01:06:30] Open source gives us velocity [01:06:32] and scale. [01:06:34] And abstraction [01:06:34] makes developers [01:06:36] more productive. [01:06:39] Our deep, deep commitment [01:06:40] to open source [01:06:41] has resonated [01:06:42] with the community [01:06:43] that has truly embraced AMD. [01:06:45] From frameworks [01:06:46] and compiler stacks [01:06:48] to inference engines [01:06:49] and model hubs, [01:06:50] AMD is now becoming [01:06:51] part of default enablement [01:06:53] for the most important [01:06:54] AI communities, [01:06:56] such as Hugging Face, [01:06:57] PyTorch, [01:06:59] Jax, [01:07:00] VLM, [01:07:01] and SGLang. [01:07:02] This is why new models [01:07:03] get day zero support [01:07:04] and why the world's [01:07:06] most important AI workloads [01:07:07] run on AMD today. [01:07:09] That is a big shift. [01:07:12] Inspired by engineers [01:07:13] seeking productivity, [01:07:16] abstractions have evolved [01:07:17] to enable work [01:07:18] at the right level of detail, [01:07:20] hiding complexity [01:07:21] when they want speed [01:07:22] and exposing control [01:07:24] when they need performance. [01:07:26] From low-level programming [01:07:27] in assembly and C [01:07:29] to block programming abstractions [01:07:30] like Triton [01:07:31] to Pythonic frameworks, [01:07:33] we've invested [01:07:34] in key abstractions. [01:07:36] Ater, [01:07:37] our kernel library, [01:07:38] and Atom, [01:07:39] our serving engine, [01:07:41] get you peak performance [01:07:42] without writing [01:07:43] the kernels yourself. [01:07:45] Fly DSL [01:07:46] is a brand new [01:07:48] Pythonic domain-specific language [01:07:49] that gives you [01:07:50] low-level control [01:07:51] with the performance [01:07:52] of hand-tuned C++. [01:07:55] And Mori, [01:07:56] our communications library, [01:07:57] helps deliver [01:07:58] leadership performance [01:08:00] on the latest models [01:08:01] like minimax. [01:08:04] But look, [01:08:05] something very, [01:08:06] very exciting [01:08:07] is happening. [01:08:08] Over the past year, [01:08:09] I've seen something [01:08:10] remarkable inside AMD. [01:08:12] our engineers [01:08:12] are using AI models [01:08:13] to generate GPU kernels, [01:08:16] optimize code, [01:08:17] debug issues, [01:08:18] and improve performance. [01:08:20] And in some cases, [01:08:21] these AI-generated kernels [01:08:22] are shockingly good, [01:08:24] better than what we expected, [01:08:26] sometimes better than [01:08:27] the most manually-tuned versions. [01:08:30] The first time you see this, [01:08:32] you're a little bit skeptical. [01:08:34] you run more tests, [01:08:35] you try to break it, [01:08:36] you look for what went wrong, [01:08:38] and then you realize [01:08:39] this is real. [01:08:41] What has happened [01:08:42] in general software development [01:08:44] is coming [01:08:45] to significantly transform [01:08:47] GPU programming. [01:08:49] It's going to reduce [01:08:50] the time it takes [01:08:51] to bring up workloads, [01:08:52] make optimization automated, [01:08:55] and remove any remaining barriers [01:08:57] to broad adoption. [01:08:59] We want to put [01:09:00] that capability [01:09:01] in the hands [01:09:03] of every developer. [01:09:05] So today, [01:09:06] I am excited [01:09:07] to introduce [01:09:08] Rockham.ai. [01:09:09] It's an agentic AI platform [01:09:19] that brings the capabilities [01:09:20] of AI-assisted GPU programming [01:09:22] to developers. [01:09:23] It starts with the simple idea [01:09:25] of giving developers access [01:09:27] to the power of an AMD ecosystem [01:09:29] through the AI coding agents [01:09:31] they already use. [01:09:33] Whether you're using cursor, [01:09:35] cloud, codex, [01:09:36] Rockham.ai helps those agents [01:09:38] understand AMD platforms, [01:09:40] understand Rockham, [01:09:41] and helps you build [01:09:43] and optimize your workloads. [01:09:45] In other words, [01:09:47] we are making [01:09:47] those popular coding agents [01:09:49] into Rockham superusers. [01:09:52] You should be able [01:09:52] to describe the workload [01:09:53] you want to run, [01:09:54] the performance target [01:09:55] you want to hit, [01:09:56] and then let the agents [01:09:57] help you get there. [01:10:00] For years, [01:10:01] we have been building [01:10:01] the open software foundation. [01:10:03] Now, [01:10:04] we are adding [01:10:05] an AI-assisted layer [01:10:06] that helps developers [01:10:08] use that foundation [01:10:09] faster and more effectively. [01:10:11] So let me show you [01:10:12] what is inside Rockham.ai. [01:10:15] It's built [01:10:15] on the foundational capabilities [01:10:17] of Rockham, [01:10:18] our core software stack. [01:10:19] Runtime, [01:10:21] libraries, [01:10:22] compilers, [01:10:23] tools, [01:10:23] framework integration, [01:10:25] all of the core software [01:10:26] that makes the stack work. [01:10:28] On top of that, [01:10:29] we built a layer [01:10:30] of AI-assisted optimization [01:10:32] called Hyperloom. [01:10:35] Using Hyperloom, [01:10:36] the system can analyze [01:10:37] the workload, [01:10:38] tune configurations, [01:10:40] select and tune kernels, [01:10:41] adjust parallelism strategies, [01:10:43] and iterate [01:10:44] towards performance goals. [01:10:45] to give you an example [01:10:47] of the power of Hyperloom. [01:10:49] I was talking [01:10:50] to the lead developer [01:10:51] last week. [01:10:52] He told me [01:10:53] they just pushed through [01:10:54] a suite of 14,000 models [01:10:56] through Hyperloom, [01:10:58] optimizing them [01:10:59] and creating valuable insights. [01:11:01] This would have been [01:11:02] impossible to imagine, [01:11:04] even with a large team [01:11:06] of engineers before. [01:11:07] That's the power [01:11:08] of Hyperloom. [01:11:08] And finally, [01:11:10] we provide an AI-native interface [01:11:12] for our developers, [01:11:14] AI skills that allow coding agents [01:11:16] to understand Rokam natively. [01:11:18] I'm incredibly excited [01:11:19] by what we have built [01:11:20] with Rokam.ai. [01:11:22] It's going to profoundly transform [01:11:24] the way developers interact [01:11:25] with our platforms. [01:11:27] So let me make this concrete [01:11:28] for you with an example. [01:11:30] Imagine you are a developer [01:11:31] who wants to optimize [01:11:33] Minimax M3 [01:11:34] with VLLM [01:11:35] on MI-355s. [01:11:37] Today, [01:11:38] there are many things [01:11:39] you need to get right. [01:11:40] You need the right [01:11:41] VLLM optimizations, [01:11:42] the right model recipes, [01:11:44] the right environment variables. [01:11:46] You may need to write [01:11:47] new kernels [01:11:47] or optimize kernels. [01:11:48] And then you still need [01:11:49] to tune all of this. [01:11:52] With Rokam.ai, [01:11:53] this interaction [01:11:54] becomes much simpler. [01:11:57] So let's see it. [01:11:58] You can say, [01:11:59] optimize Minimax M3 [01:12:01] with Hyperloom. [01:12:03] Enter. [01:12:05] And behind the scenes, [01:12:07] the agent does [01:12:08] what an expert would do. [01:12:10] Pulls a known good recipe, [01:12:12] runs and profiles the workload, [01:12:13] it creates and tests [01:12:14] a bunch of kernel [01:12:15] and runtime configurations, [01:12:17] and it keeps iterating [01:12:18] towards target. [01:12:20] You can actually watch [01:12:22] on screen here, [01:12:23] Rokam.ai write a GPU kernel. [01:12:25] It sees an opportunity [01:12:27] write a more optimized [01:12:28] MOE group gem. [01:12:30] And the result, [01:12:32] Rokam.ai delivers [01:12:33] a 38% more tokens [01:12:36] per second improvement [01:12:37] on this specific example. [01:12:40] This is the experience [01:12:41] we want for developers. [01:12:43] Simple to start. [01:12:47] That's pretty cool, huh? [01:12:52] Simple, simple to start [01:12:53] and powerful to optimize. [01:12:55] We've continued [01:12:56] our relentless focus [01:12:57] and performance. [01:12:58] With every release, [01:13:00] we've delivered [01:13:00] significant performance gains. [01:13:02] On leading models [01:13:03] like DeepSeq, [01:13:04] Rokam.ai delivers [01:13:05] 3.3 times speed up [01:13:07] over Rokam 7. [01:13:09] This has been made possible [01:13:10] by innovations [01:13:11] at every level [01:13:12] of the stack. [01:13:13] Techniques, [01:13:14] such as the use [01:13:15] of block-scaled, [01:13:16] fused MOE kernels, [01:13:18] KB cache quantization, [01:13:19] KB cache quantization, [01:13:20] expert parallelism. [01:13:21] But look, [01:13:22] the bigger story [01:13:23] is how AI agents [01:13:24] actually profiled the workload, [01:13:26] proposed new kernels, [01:13:27] tested configurations, [01:13:28] and actually validated [01:13:29] all of the results. [01:13:32] That exact same focus [01:13:33] is also on delivering [01:13:35] performance gains [01:13:36] for training. [01:13:37] On the same hardware, [01:13:39] Rokam.ai improves [01:13:40] training performance [01:13:41] by an average [01:13:41] of 2.4 times. [01:13:44] Through optimized kernels, [01:13:45] such as fused flash attention, [01:13:47] more efficient checkpointing, [01:13:49] and advanced parallelism strategies. [01:13:51] Our engineers [01:13:52] have done amazing work [01:13:54] to unlock the capabilities [01:13:55] of MI455 [01:13:56] through Rokam.ai. [01:13:58] So I'm delighted [01:13:59] to share some incredible results [01:14:01] on hardware. [01:14:02] On running real workloads, [01:14:04] we are able to see [01:14:05] the system deliver [01:14:06] 20 terabytes [01:14:08] of memory bandwidth [01:14:09] and 20 petaflops [01:14:11] of delivered compute. [01:14:13] These are the highest [01:14:14] demonstrated [01:14:15] compute plus memory [01:14:16] capabilities [01:14:17] of any accelerated platform [01:14:19] in the industry today. [01:14:21] It's not theoretical, [01:14:23] there's measured. [01:14:30] Super exciting [01:14:31] that we were able [01:14:31] to share this on day one. [01:14:33] I'm also thrilled [01:14:34] to share [01:14:34] that the leading [01:14:35] AI ecosystem partners [01:14:37] who have had their hands [01:14:38] on MI455 [01:14:39] have had a delightful experience. [01:14:41] You can see here [01:14:42] what they've had to say [01:14:43] from end-to-end [01:14:45] PyTorch testing [01:14:46] to ensuring [01:14:47] hugging face models [01:14:48] are getting validated [01:14:49] to VLLM [01:14:50] and SGLang running [01:14:51] well on MI455. [01:14:53] We are incredibly encouraged [01:14:55] and excited [01:14:56] by what the community [01:14:57] is about to unlock [01:14:58] with MI455. [01:14:59] This matters [01:15:00] because day zero readiness [01:15:02] is one of the most [01:15:03] important objectives [01:15:04] we strive for [01:15:05] and that's what [01:15:06] rockm.ai enables. [01:15:09] So now let's look [01:15:10] at rockm.ai in action [01:15:11] on Helios. [01:15:13] This time [01:15:13] we are going to use [01:15:14] a codex front-end. [01:15:16] Let's deploy [01:15:16] DeepSeq V4 Pro [01:15:18] a 1.6 trillion parameter [01:15:20] frontier model [01:15:21] on Helios. [01:15:23] Now watch what happens. [01:15:26] rockm.ai pulls [01:15:27] the right recipes, [01:15:29] it configures [01:15:30] the runtime, [01:15:30] it implements [01:15:32] the model graph [01:15:32] onto hardware [01:15:33] and finally [01:15:35] serves up the model. [01:15:38] There you go. [01:15:39] The model is actually [01:15:40] up and running. [01:15:42] And let's see the model [01:15:43] write a simple poem now [01:15:44] about advancing AI. [01:15:47] Let's hit enter. [01:15:50] Isn't that a pretty cool poem? [01:15:57] This is the experience [01:15:58] we want to give our users. [01:15:59] The developers [01:16:00] should not have [01:16:01] to manually reason [01:16:02] through every layer [01:16:03] of model architecture, [01:16:04] kernel selection, [01:16:06] comms strategy, [01:16:07] and rack-level deployment. [01:16:09] AI is transforming [01:16:10] this experience [01:16:11] and that is a very, [01:16:13] very big shift [01:16:13] for GPU hardware [01:16:14] and software. [01:16:17] Now, [01:16:17] there are few organizations [01:16:19] shaping the frontier of AI [01:16:20] as much as open AI. [01:16:22] You heard Lisa [01:16:23] and Sachin talk [01:16:24] about our partnership. [01:16:25] It's been truly wonderful [01:16:26] collaborating closely [01:16:28] with the open AI [01:16:29] technical team [01:16:30] to make rapid progress [01:16:31] with Rockham. [01:16:32] To talk about how [01:16:34] they are advancing [01:16:35] the frontier of software [01:16:36] development [01:16:36] and our collaboration, [01:16:37] I'm delighted to welcome [01:16:39] to stage [01:16:40] Philippe Tillet [01:16:41] from OpenAI. [01:16:50] Thank you for joining us [01:16:52] here, Philippe. [01:16:53] So just so you all know, [01:16:55] Philippe is the creator [01:16:56] of Triton, [01:16:57] one of the most important [01:16:58] software innovations [01:16:59] in modern AI infrastructure. [01:17:01] It's a huge privilege [01:17:01] having you here. [01:17:03] Thank you. [01:17:07] OpenAI has always pushed [01:17:09] the limits of infrastructure. [01:17:11] So, you know, [01:17:12] can you tell the team [01:17:12] a little bit here today [01:17:14] as to how your team [01:17:15] actually develops [01:17:15] GPU software [01:17:16] and how have technologies [01:17:17] like Triton evolved? [01:17:19] Yeah, for sure. [01:17:20] Well, first of all, [01:17:21] thanks so much [01:17:21] for having me. [01:17:22] I was here two years ago. [01:17:23] It's my great pleasure [01:17:24] to be back now. [01:17:27] A way of looking at things [01:17:28] is there's no one way [01:17:30] to fit all applications [01:17:32] and we really try [01:17:33] to meet users [01:17:34] where they are. [01:17:35] So for that reason, [01:17:36] we've developed [01:17:36] multiple solutions. [01:17:38] So for workflows [01:17:40] where speed of iterations [01:17:41] is more important, [01:17:42] we've developed Triton, [01:17:44] as you mentioned. [01:17:45] So this is typically [01:17:45] what researchers will use. [01:17:48] But more and more, [01:17:49] we've been in a case [01:17:50] where we've had to [01:17:51] hyper-optimize performance [01:17:52] and Triton didn't give us [01:17:53] the level of control [01:17:54] that we wanted. [01:17:55] So we've developed [01:17:56] Gluon for that, [01:17:57] which is a lower-level language. [01:18:00] And this way, [01:18:01] we've been really able [01:18:02] to optimize our kernels [01:18:03] for AMD, [01:18:04] among other things. [01:18:06] And, yeah, [01:18:08] like our agents [01:18:09] have become extremely good [01:18:10] at both of them [01:18:11] and we've worked with AMD [01:18:13] on both of them as well. [01:18:16] Yeah, and, you know, [01:18:17] more and more [01:18:17] what we're seeing [01:18:18] is agents [01:18:19] slowly taking over, [01:18:21] like, yeah, [01:18:23] they're getting just very good, [01:18:24] as you just mentioned, [01:18:25] at writing this kind of code. [01:18:26] Amazing. [01:18:27] So, look, [01:18:28] I think the work [01:18:28] you have done with Triton [01:18:29] and now Gluon, right, [01:18:30] it's been really, really [01:18:31] moving the ball forward [01:18:33] for the entire industry. [01:18:34] Now, we began our work [01:18:35] together with 300 [01:18:36] and, you know, [01:18:37] extended at 350, [01:18:38] but now we're collaborating [01:18:40] at a different scale [01:18:41] with 450 and Helios. [01:18:42] So I would love [01:18:44] to hear your thoughts [01:18:45] on how that collaboration [01:18:46] has been going [01:18:46] and how has that experience been? [01:18:48] Yeah, super well. [01:18:49] It feels like ages. [01:18:51] I think we started years ago [01:18:53] on hardware co-design, [01:18:55] you know, [01:18:55] when MI450 [01:18:56] was, like, [01:18:57] in the very early [01:18:58] prototyping stages. [01:19:00] I think we're able [01:19:01] to work really well together [01:19:02] and it's so nice [01:19:03] to see the outcome [01:19:04] of that [01:19:04] and the chip. [01:19:07] And more recently, [01:19:08] we've been working [01:19:09] very, very deeply [01:19:10] on software enablement [01:19:12] for MI450, [01:19:13] both for our models [01:19:14] but also [01:19:15] for the broader ecosystem. [01:19:16] So a lot of this development [01:19:17] has been done [01:19:18] open source. [01:19:20] And, yeah, [01:19:21] we're collaborating [01:19:21] on the full stack now [01:19:23] down to the very low-level [01:19:25] LVM co-generation, [01:19:27] you know, [01:19:28] so that, like, [01:19:28] all the nice [01:19:30] hardware advances [01:19:31] that came with MI450 [01:19:32] can be properly leveraged, [01:19:34] you know, [01:19:34] in end-to-end applications. [01:19:36] Yeah. [01:19:37] And, you know, [01:19:38] you got your Helios racks, [01:19:40] you know, [01:19:40] so tell us a little bit [01:19:40] about that, yeah. [01:19:41] Yeah, it worked. [01:19:42] Like, [01:19:43] we got the chip [01:19:44] and... [01:19:45] Yeah, [01:19:50] we got the chip [01:19:51] and, you know, [01:19:52] very quickly, [01:19:53] like, [01:19:53] a single engineer [01:19:54] in a few days [01:19:55] was able to confirm [01:19:56] that, you know, [01:19:57] everything worked. [01:19:58] Obviously, like, [01:19:58] we have to optimize [01:19:59] performance more. [01:20:01] But we're very confident [01:20:02] we can get in a very, [01:20:03] very good place. [01:20:04] Awesome. [01:20:05] That's awesome to hear. [01:20:06] So, look, [01:20:07] one of the more amazing things [01:20:08] that we've done recently, [01:20:09] right, [01:20:10] is our teams [01:20:10] have collaborated [01:20:11] even more closely [01:20:12] in terms of, [01:20:13] you know, [01:20:14] using AI [01:20:15] to program AMD GPUs. [01:20:17] So, you know, [01:20:17] would love to hear [01:20:18] your thoughts [01:20:19] and anything you can share [01:20:20] with the audience. [01:20:21] Yeah, no, [01:20:22] I mean, [01:20:23] everyone in this room [01:20:24] knows how much [01:20:25] agents are taking over [01:20:26] our jobs, [01:20:29] in a sense, [01:20:29] as kernel engineers. [01:20:31] And, [01:20:33] yeah, [01:20:34] and it's been amazing [01:20:35] just to see, like, [01:20:36] how much better [01:20:37] they've got over [01:20:38] just the past [01:20:39] six months. [01:20:40] And just now [01:20:40] they're extremely good [01:20:41] at generating [01:20:42] just high-quality [01:20:43] GPU kernels [01:20:44] in a way [01:20:44] that were not before. [01:20:46] And it's been [01:20:47] very great collaborating [01:20:48] with AMD [01:20:50] on just, like, [01:20:52] making sure [01:20:52] that agents [01:20:53] not only are capable [01:20:54] enough, [01:20:55] you know, [01:20:56] but also have [01:20:57] all the right context [01:20:58] they need [01:20:58] to be able [01:20:59] to make good decisions [01:21:00] and good optimizations. [01:21:02] That's great to hear. [01:21:03] Yeah, I think, [01:21:03] you know, [01:21:04] the quality of code [01:21:05] that's getting produced [01:21:06] by AI-assisted kernel generation [01:21:09] has been truly remarkable [01:21:10] through our collaboration. [01:21:12] Very excited. [01:21:13] Now, [01:21:13] as you look ahead, [01:21:15] right, [01:21:15] you know, [01:21:16] what crystal ball [01:21:17] do you have for us [01:21:18] for the future [01:21:19] of AI software ecosystem [01:21:20] and collaborations [01:21:21] like ours? [01:21:22] Yeah, [01:21:23] so there's two things [01:21:24] that really come to mind. [01:21:26] One, [01:21:26] you know, [01:21:27] as we keep talking, [01:21:28] is just how good [01:21:29] agents are getting. [01:21:31] And the other one is, [01:21:32] I think, [01:21:33] how much the open approach [01:21:35] to software of AMD [01:21:36] is enabling, [01:21:37] not only for OpenAI, [01:21:39] but I think [01:21:39] for the industry [01:21:40] as a whole. [01:21:41] You know, [01:21:42] I talked a little bit earlier [01:21:43] how we've been collaborating [01:21:44] on LVM code generation. [01:21:46] LVM is an open source project. [01:21:48] The entire compiler stack [01:21:49] from AMD is open source. [01:21:51] And that has allowed us [01:21:52] to make our agents [01:21:54] really good at very, [01:21:55] very low level code generation [01:21:56] down to, you know, [01:21:58] scheduling instructions [01:21:59] in assembly. [01:22:00] And this has led to, like, [01:22:02] very, very significant [01:22:04] performance gain [01:22:05] that I don't think [01:22:05] we have been able [01:22:06] to achieve [01:22:08] in a fully closed source stack. [01:22:10] Yeah, [01:22:10] that's one of the real [01:22:12] sort of big value propositions [01:22:14] that we have to offer, right? [01:22:15] Because everything that we do [01:22:17] is in the open [01:22:17] and the fact that AI [01:22:19] can actually pick that up [01:22:20] is a big one. [01:22:21] Thank you, Philippe. [01:22:22] We truly appreciate [01:22:23] the partnership [01:22:24] and we're really excited [01:22:25] about what we're going [01:22:26] to build together. [01:22:27] Yeah, my pleasure. [01:22:27] Thank you so much, Raheem. [01:22:34] What you just heard [01:22:35] is extremely important. [01:22:37] When we combine [01:22:38] better abstractions [01:22:39] with AI-assisted development [01:22:41] and closed hardware-software [01:22:43] co-design, [01:22:43] we can create [01:22:44] some incredibly [01:22:45] powerful capabilities. [01:22:46] And that's exactly [01:22:47] aligned with our strategy. [01:22:50] Now, everything [01:22:51] I've shown you so far [01:22:52] has been about AI. [01:22:53] But Raheem is also [01:22:55] a great stack [01:22:56] for developing [01:22:56] high-performance [01:22:57] and scientific computing [01:22:58] applications. [01:22:59] From frontier models [01:23:00] to climate simulation, [01:23:02] we use the same Raheem. [01:23:04] The same library, [01:23:04] same tools, [01:23:05] same AI-assisted development. [01:23:07] One open stack [01:23:08] for every single workload. [01:23:11] And in fact, [01:23:11] when you look at our roadmap, [01:23:13] we have retained [01:23:13] that very strong focus [01:23:15] on HPC [01:23:15] and scientific computing. [01:23:17] Just like MI-455 [01:23:19] leads in AI compute [01:23:20] with FP4 [01:23:21] and memory performance, [01:23:22] its MI-430X [01:23:24] leads HPC compute [01:23:25] with FP64 [01:23:27] and memory performance, [01:23:29] two platforms [01:23:29] purpose-built [01:23:30] for two very different [01:23:31] classes of workload. [01:23:34] The MI-430X [01:23:35] is built [01:23:36] by leveraging [01:23:37] our modular [01:23:37] chiplet architecture [01:23:38] to deliver [01:23:39] a purpose-built variant [01:23:40] with native [01:23:42] FP64 hardware, [01:23:44] giving customers [01:23:45] leadership performance [01:23:46] across AI and HPC [01:23:48] in a single platform. [01:23:50] And this is [01:23:51] hardware double precision. [01:23:52] not emulation. [01:23:54] For the scientific community, [01:23:56] that distinction [01:23:56] is everything. [01:23:58] This product delivers [01:23:59] 288 teraflops [01:24:01] of FP64 compute, [01:24:03] nearly nine times [01:24:04] better than competition. [01:24:06] And the same leadership, [01:24:07] memory capacity, [01:24:08] and bandwidth [01:24:09] of MI-455. [01:24:11] The MI-430X [01:24:12] ships [01:24:13] in the first half [01:24:15] of 2027. [01:24:16] moment. [01:24:16] This is why [01:24:19] leaders in sovereign AI [01:24:21] love this product. [01:24:23] The first sovereign [01:24:23] exascale AI factories [01:24:25] in both US and Europe [01:24:26] are being built [01:24:27] on the MI-430X. [01:24:29] Discovery [01:24:30] at Oak Ridge National Labs [01:24:31] and Alistair Koch [01:24:33] with CNES [01:24:33] and Gen-C [01:24:34] in Europe. [01:24:36] So look, [01:24:37] as we look ahead, [01:24:39] let me leave you [01:24:39] with this thought. [01:24:41] Every so often, [01:24:42] our industry goes [01:24:43] through inflection points. [01:24:45] We've all felt it [01:24:46] when neural networks [01:24:47] first started recognizing [01:24:49] cats and dogs, [01:24:50] when transformers changed [01:24:51] what was possible [01:24:52] with AI. [01:24:53] I believe we are experiencing [01:24:55] another one [01:24:55] right now, [01:24:57] a real inflection [01:24:58] in the ability [01:24:59] of AI [01:25:00] to program [01:25:00] complex hardware systems. [01:25:02] Building on the surface area [01:25:04] that's been exposed [01:25:05] from open source software [01:25:07] and leveraging the power [01:25:08] of abstracted interfaces, [01:25:10] AI is making it [01:25:11] dramatically easier [01:25:13] to program AI systems. [01:25:15] We cannot wait [01:25:16] what you will build with it. [01:25:19] Thank you. [01:25:28] Now, AI is getting [01:25:29] pervasively infused [01:25:31] into all forms of computing. [01:25:32] Enterprises are experiencing [01:25:34] the same shift [01:25:35] with their own [01:25:35] unique requirements. [01:25:37] So to tell us more [01:25:38] about that transformation, [01:25:40] please welcome to stage [01:25:41] Senior Vice President [01:25:42] and General Manager [01:25:43] of Compute [01:25:44] and Enterprise AI, [01:25:46] Dan McNamara. [01:25:55] Thank you, Vansi, [01:25:56] and good morning, everyone. [01:25:58] It is a great pleasure [01:26:00] for me to bring to you [01:26:02] the third pillar [01:26:02] of our strategy, [01:26:03] which is powering [01:26:04] AI everywhere. [01:26:06] And, you know, [01:26:06] the vision behind this [01:26:08] has remained the same [01:26:09] for several years [01:26:10] for us, [01:26:11] you know, [01:26:12] to deliver the right [01:26:13] compute engine [01:26:14] to the right workload [01:26:15] and then continuously [01:26:17] drive optimization points [01:26:19] across our entire portfolio [01:26:22] from CPUs, [01:26:24] GPUs, [01:26:25] networking, [01:26:25] and FPGAs. [01:26:27] So building on [01:26:29] what we've already shared, [01:26:30] I'll cover how our strategy [01:26:32] comes together [01:26:33] across our enterprise portfolio [01:26:34] and how the foundation [01:26:36] we built [01:26:37] with the CPU franchise [01:26:38] has truly positioned us [01:26:40] for the next era of AI [01:26:42] across multiple deployment models. [01:26:46] So Lisa talked about [01:26:47] the frontier model [01:26:48] and hyperscale landscape [01:26:49] and just the rapid [01:26:51] and massive advances [01:26:52] we've seen [01:26:53] over the last [01:26:54] six to 12 months. [01:26:56] And those advances [01:26:57] are driving [01:26:57] the adoption of AI [01:26:59] well beyond the cloud. [01:27:01] And as AI continues to grow, [01:27:03] it will demand compute [01:27:04] beyond the massive [01:27:05] purpose-built data centers [01:27:07] that we all know and love, [01:27:09] extending into enterprise, [01:27:10] personal, [01:27:11] and physical AI. [01:27:12] And each of these [01:27:13] brings a new [01:27:14] or unique set [01:27:16] of requirements. [01:27:17] Enterprises are usually [01:27:19] constrained by power and cooling. [01:27:20] Personal AI usually [01:27:22] must operate [01:27:23] within a fairly limited [01:27:25] compute and memory footprint. [01:27:26] And physical AI [01:27:28] must perform reliably [01:27:29] in real-world demanding environments. [01:27:33] And AMD [01:27:34] is the only company [01:27:35] delivering a complete portfolio [01:27:37] that spans [01:27:37] all of these [01:27:38] development models. [01:27:41] So I just mentioned [01:27:42] that our journey began [01:27:44] in the data center [01:27:46] with our CPU franchise [01:27:47] and I want to touch [01:27:48] on it a bit. [01:27:48] With each generation [01:27:50] of Epic, [01:27:50] our strategy has been [01:27:51] to listen to our customers, [01:27:53] address their evolving needs, [01:27:54] and deliver [01:27:55] the highest performance, [01:27:57] lowest total cost [01:27:58] of ownership, [01:27:59] and the fastest time [01:28:00] to value. [01:28:02] As Lisa mentioned earlier, [01:28:03] we've introduced [01:28:04] many industry firsts [01:28:06] across our generations [01:28:07] that our enterprise [01:28:09] customers [01:28:09] has actually asked for. [01:28:12] And that customer focus [01:28:13] has brought us [01:28:14] to where we are today [01:28:15] with Epic [01:28:16] as the leader [01:28:18] in enterprise computing. [01:28:20] And over the last three years, [01:28:21] there's been [01:28:21] very, very strong adoption [01:28:23] across the enterprise [01:28:24] with the world's largest companies [01:28:25] moving more and more [01:28:26] workloads to AMD. [01:28:28] And it's reflected [01:28:29] in a number of places. [01:28:31] First, public cloud, [01:28:32] where our VM consumption [01:28:34] has grown at 76% CAG [01:28:36] over the last three years. [01:28:38] And on-prem deployments [01:28:40] with our OEM [01:28:41] and ODM partners [01:28:42] is growing quite aggressively [01:28:44] across key industries. [01:28:46] But just as important, [01:28:49] this position allows us [01:28:51] to understand truly [01:28:52] the diversity [01:28:52] of the workloads [01:28:53] that our enterprise [01:28:54] customers run every day. [01:28:57] And each workload [01:28:59] places new demands [01:29:02] on the CPU. [01:29:03] Some require [01:29:03] maximum thread count, [01:29:05] some depend on [01:29:06] per-core performance, [01:29:07] while others depend [01:29:08] on memory band [01:29:09] with cache efficiency [01:29:10] or I.O. performance. [01:29:12] The really important point [01:29:14] is that there is [01:29:15] no single SKU [01:29:17] that solves [01:29:18] every enterprise application. [01:29:20] And that's why [01:29:21] our EPIC portfolio [01:29:22] with VENIS [01:29:23] spans 8 to 256 cores [01:29:26] with a range of power, [01:29:27] frequencies, [01:29:28] and I.O. capabilities, [01:29:29] giving customers [01:29:31] the flexibility [01:29:31] to choose [01:29:32] the right solution [01:29:33] for their workload [01:29:34] and most importantly, [01:29:36] at the most optimized cost. [01:29:39] So, Lisa mentioned [01:29:41] this earlier [01:29:42] and now we all know [01:29:43] we covered VENIS [01:29:45] and she showed [01:29:46] how we have [01:29:48] strong leadership [01:29:49] across AI host nodes [01:29:51] and the agentic workflows. [01:29:53] But I wanted to show you [01:29:54] how VENIS [01:29:55] actually leads [01:29:56] across the general purpose [01:29:58] workloads [01:29:58] of the enterprise [01:29:59] and these workloads [01:30:01] that are running, [01:30:01] our customers [01:30:03] rely on every day. [01:30:04] As you can see, [01:30:05] we have a minimum [01:30:06] of 2.5x [01:30:08] to performance [01:30:09] across our competition, [01:30:10] across these very, [01:30:11] very important [01:30:12] enterprise workloads. [01:30:15] So, now, [01:30:16] as we all know, [01:30:17] agentic AI [01:30:18] is really [01:30:20] the next major transition [01:30:21] for the enterprise [01:30:22] and it truly [01:30:23] is a force multiplier [01:30:25] for enterprises [01:30:26] and adoption [01:30:27] is accelerating. [01:30:28] In fact, [01:30:29] the latest data here [01:30:29] shows [01:30:30] is that nearly [01:30:31] every enterprise [01:30:31] next year [01:30:32] will have some form [01:30:34] of agentic deployments. [01:30:37] And as enterprises [01:30:38] move from, [01:30:39] you know, [01:30:40] pilots to production [01:30:41] deployments, [01:30:42] we've talked to CIOs [01:30:44] all the time [01:30:44] and they are looking [01:30:46] at three major things. [01:30:47] First, [01:30:48] predictable infrastructure [01:30:50] cost and token costs. [01:30:52] Secondly, [01:30:53] making sure [01:30:54] the data is secure [01:30:55] and then lastly, [01:30:57] and sometimes [01:30:58] this is overlooked, [01:30:59] right, [01:31:00] integrating AI [01:31:01] into their existing [01:31:02] infrastructure. [01:31:03] It's not all [01:31:04] greenfield out there [01:31:05] for the enterprise [01:31:06] customer. [01:31:08] So, [01:31:08] to solve [01:31:09] these problems, [01:31:10] we truly believe [01:31:11] that this is going [01:31:11] to be a distributed [01:31:12] deployment model [01:31:13] that leverages [01:31:14] frontier models, [01:31:16] cloud services, [01:31:17] on-prem infrastructure, [01:31:17] and of course, [01:31:18] AI clients. [01:31:20] And as AI [01:31:22] has moved [01:31:22] from chatbots [01:31:23] responding to prompts [01:31:24] to, you know, [01:31:25] agents doing real work, [01:31:26] writing code, [01:31:28] summarizing documents, [01:31:29] and running complex flows, [01:31:32] the requirements [01:31:33] have evolved also. [01:31:35] And those are exactly [01:31:36] the requirements [01:31:37] we've been designing [01:31:38] for for multiple years. [01:31:40] So, [01:31:41] while AI will be deployed [01:31:43] across all these [01:31:44] environments, [01:31:45] the on-prem deployments [01:31:46] in our enterprise [01:31:46] customers have [01:31:47] fairly unique constraints. [01:31:49] First and foremost, [01:31:50] power cooling. [01:31:52] Really big challenge [01:31:53] in space. [01:31:54] And of course, [01:31:56] costs remain [01:31:57] top of the list [01:31:58] for CIOs [01:31:59] that are trying to, [01:32:00] you know, [01:32:01] drive things forward [01:32:02] in the AI world. [01:32:03] And then lastly, [01:32:05] the full solutions [01:32:06] they need [01:32:07] to be easy [01:32:09] to deploy. [01:32:10] That's been common [01:32:12] for many, many years [01:32:13] for the enterprise. [01:32:14] But it has to, [01:32:15] that's something [01:32:16] we're very, very focused on. [01:32:17] So, [01:32:18] to address these constraints, [01:32:20] thank you, Laura. [01:32:22] I'm very happy [01:32:24] to announce [01:32:26] the launch [01:32:26] of the Instinct [01:32:27] MI350P. [01:32:34] So, today, [01:32:35] every enterprise [01:32:36] wants AI [01:32:37] in the data center. [01:32:38] But the challenge is, [01:32:39] how do you add [01:32:40] those capabilities [01:32:41] without starting over? [01:32:43] That's exactly [01:32:44] what the 350P [01:32:45] was designed to solve. [01:32:46] It's an air-cooled GPU [01:32:48] designed to fit [01:32:49] within the power [01:32:50] and cooling [01:32:50] of enterprise servers [01:32:52] today [01:32:52] without a facilities upgrade, [01:32:55] making it easy [01:32:56] to bring LLM scale [01:32:57] inference [01:32:58] into today's [01:32:59] enterprise data center. [01:33:00] And at the same time, [01:33:01] there's no compromise [01:33:02] in capability. [01:33:04] A single MI350P [01:33:06] can support [01:33:06] up to 260 billion parameters, [01:33:09] allowing customers [01:33:09] to run [01:33:10] the majority [01:33:11] of enterprise AI workloads [01:33:12] on a single GPU. [01:33:15] And the economics [01:33:16] are actually even better [01:33:18] and more compelling. [01:33:19] The 350P delivers [01:33:21] more than four times [01:33:22] the tokens per second [01:33:23] per dollar [01:33:24] than the competition. [01:33:26] It's pretty cool. [01:33:28] And it actually [01:33:32] turns the customer's [01:33:35] existing data center [01:33:36] into an AI data center. [01:33:38] Now, [01:33:39] we all know [01:33:41] that benchmarks [01:33:41] are one part [01:33:42] of the equation, right? [01:33:44] The other part [01:33:44] is really evaluating [01:33:46] does the advantage [01:33:47] hold up [01:33:48] across the actual workload [01:33:49] you run [01:33:49] and how your business runs. [01:33:51] So we tested [01:33:52] a set of workloads, [01:33:54] enterprise use cases [01:33:55] using production models. [01:33:57] And the 350P [01:33:58] delivers much higher [01:33:59] productivity [01:34:00] than the competition, [01:34:02] delivering up to [01:34:02] two to five times [01:34:04] tokens per, [01:34:05] faster tokens per second [01:34:07] than the competition. [01:34:08] So if you think [01:34:09] about that, [01:34:10] that's more tokens [01:34:11] per second, [01:34:12] more work completed, [01:34:14] more users supported, [01:34:16] all at a lower cost [01:34:18] for the business. [01:34:20] So another area [01:34:22] that I wanted [01:34:23] to talk about [01:34:23] is our earliest [01:34:24] customers, [01:34:25] our own AMD IT team. [01:34:28] And they're [01:34:30] a tough group [01:34:31] to work with [01:34:31] even for us. [01:34:32] So, [01:34:33] but they put us [01:34:35] through our paces [01:34:36] and, you know, [01:34:37] really this has been [01:34:38] our model [01:34:39] from day one [01:34:39] is we want [01:34:40] to put our technology [01:34:41] into our own [01:34:42] data center first. [01:34:43] So like many companies, [01:34:45] many enterprises, [01:34:46] we're looking at [01:34:46] how to deliver AI service [01:34:47] to our employee base [01:34:48] but also manage [01:34:50] costs [01:34:50] and protect our data [01:34:52] and manage [01:34:53] and control the data. [01:34:54] So we focused [01:34:56] on two use cases [01:34:57] that we wanted [01:34:58] to try out [01:34:59] and first was [01:35:00] autonomous threat detection [01:35:02] and, you know, [01:35:02] that is probably [01:35:04] one of the faster [01:35:04] growing agentic [01:35:05] applications [01:35:06] that we're seeing today. [01:35:07] And the second was, [01:35:09] you know, [01:35:09] a personalized AI assistant [01:35:10] running on OpenClaw. [01:35:12] So both of these [01:35:13] run route requests [01:35:15] through the router [01:35:16] and the gateway [01:35:18] sort of tests [01:35:20] the complexity, [01:35:21] the sensitivity [01:35:22] of the information [01:35:22] and the service level [01:35:24] requirements. [01:35:25] So some requests [01:35:26] are sent [01:35:27] to the Frontier model [01:35:28] while there's a router [01:35:29] to open weights [01:35:30] models running [01:35:31] on Epic [01:35:31] and an MI350P. [01:35:34] So with intelligent routing, [01:35:36] we reduced our token [01:35:37] cost by 43% [01:35:39] while delivering [01:35:41] up to 3x [01:35:42] faster response times [01:35:43] for the workloads [01:35:44] running locally. [01:35:46] So we believe [01:35:47] this is how [01:35:48] enterprise AI [01:35:49] will be deployed. [01:35:52] But we also know [01:35:54] that from our history [01:35:56] here with Epic [01:35:56] that enabling enterprise [01:35:58] takes a whole lot more [01:36:00] than great silicon. [01:36:02] It requires a complete [01:36:03] ecosystem of software [01:36:04] and solutions. [01:36:05] And over the last [01:36:06] several years, [01:36:06] we've truly grown [01:36:08] significantly [01:36:09] and expanded [01:36:09] our enterprise [01:36:10] AI ecosystem. [01:36:11] Today, [01:36:12] we support [01:36:12] numerous AI ISVs, [01:36:15] provide zero-day support [01:36:16] for all of the industry's [01:36:18] leading open models, [01:36:19] and deliver robust [01:36:20] frameworks [01:36:21] that help customers [01:36:22] bring AI [01:36:23] into production [01:36:24] as fast as possible. [01:36:26] And it's all built [01:36:27] on trusted platform [01:36:29] software and OEM [01:36:30] partnerships [01:36:30] that enterprises [01:36:32] rely on [01:36:32] every single day. [01:36:34] Now, [01:36:35] everything I just shared [01:36:38] comes down [01:36:39] comes down [01:36:39] to one goal [01:36:39] for us, [01:36:40] and that's [01:36:41] helping customers [01:36:42] deploy AI [01:36:42] where it creates [01:36:44] the most value [01:36:45] for them, [01:36:46] whether it's [01:36:47] in the cloud, [01:36:48] on-premise, [01:36:48] or at the edge. [01:36:50] So to bring this [01:36:51] to life, [01:36:52] I'd like to welcome [01:36:53] Jeremy Legg, [01:36:55] Chief Technology Officer [01:36:56] at AT&T. [01:37:04] Hey, crew, [01:37:05] good to see you. [01:37:06] Yeah, thanks for having me. [01:37:08] Yeah, this is great. [01:37:09] It's kind of a homecoming. [01:37:10] I grew up in the East Bay, [01:37:12] so it's good to be back. [01:37:13] Yeah, fantastic. [01:37:13] We really appreciate [01:37:14] you joining us today. [01:37:15] So, look, [01:37:16] I think everyone knows [01:37:18] AT&T is the largest [01:37:20] communications company, [01:37:21] one of the largest [01:37:22] in the world, [01:37:23] and I'm sure [01:37:23] all these people, [01:37:25] half of them out there [01:37:26] probably have AT&T [01:37:27] network right now, [01:37:29] but, you know, [01:37:29] hopefully more than half. [01:37:32] But, you know, [01:37:34] AI is touching [01:37:35] way more than the network, [01:37:36] and you and your team [01:37:37] are really doing [01:37:38] a fair amount of work [01:37:39] driving it. [01:37:41] So can you share [01:37:41] with us sort of [01:37:42] the opportunities [01:37:42] you approach, [01:37:44] you're working on, [01:37:44] and your experience to date? [01:37:46] Sure, happy to. [01:37:48] At AT&T, [01:37:49] we're now burning [01:37:50] about a trillion-plus [01:37:52] tokens per month, [01:37:54] and that number [01:37:55] has gone up [01:37:56] pretty dramatically. [01:37:57] It's moving up [01:37:58] double digits, [01:37:59] and so this is now [01:38:00] becoming very pervasive [01:38:02] inside of the enterprise. [01:38:04] We've got over [01:38:05] 100 Gen A models [01:38:06] in production [01:38:07] across our enterprise [01:38:10] as well, [01:38:11] and those are stretching [01:38:12] from, you know, [01:38:13] things that I think [01:38:14] a lot of folks [01:38:14] are doing in customer care, [01:38:16] things that we're doing [01:38:17] in fraud, [01:38:18] but also things [01:38:19] that you wouldn't [01:38:19] necessarily think about, [01:38:21] like where's the best place [01:38:22] to place a cell tower [01:38:24] or a RAN [01:38:25] that is the most effective [01:38:26] use of it [01:38:27] inside of a network. [01:38:29] So all of these things [01:38:30] are happening [01:38:30] across our enterprise, [01:38:32] but they're happening [01:38:32] at a scale [01:38:33] that is not necessarily normal. [01:38:36] We get over 300,000 calls a day. [01:38:39] The transcription [01:38:40] of those calls [01:38:41] is an enormous workload, [01:38:42] and then driving [01:38:43] the insights [01:38:44] out of those calls [01:38:45] that we then bring back [01:38:46] to the business [01:38:46] is another set of insights. [01:38:49] So it's really [01:38:50] an interesting thing, [01:38:51] and now we're beginning [01:38:52] to actually rebuild [01:38:54] entire workflows [01:38:55] inside of the company [01:38:56] beyond just the point [01:38:58] use of a use case. [01:39:00] So as we think about HR, [01:39:01] as we think about finance, [01:39:03] and we think about [01:39:03] different kinds of things [01:39:05] that each of those [01:39:05] organizations do, [01:39:07] cash forecasting, [01:39:09] things that we do [01:39:09] on onboarding employees [01:39:10] as examples, [01:39:11] we're now agentifying [01:39:13] those entire workloads. [01:39:14] Yeah, so it is, [01:39:17] you know, [01:39:17] we work closely [01:39:18] with your team, [01:39:19] and we've had [01:39:19] many conversations. [01:39:21] It's truly amazing [01:39:22] the number of actual, [01:39:23] you know, [01:39:25] opportunities [01:39:26] you're addressing [01:39:27] with AI. [01:39:28] So can, you know, [01:39:28] a super broad range of use. [01:39:30] So can you just explain [01:39:32] sort of how, [01:39:32] what are the key learnings [01:39:34] in moving these [01:39:36] to AI [01:39:37] at enterprise scale, [01:39:38] and then maybe a bit [01:39:39] on what we've done together? [01:39:41] Sure. [01:39:41] I mean, [01:39:42] I'll start at a place [01:39:43] that I don't think [01:39:44] everybody starts from here, [01:39:46] which is the human. [01:39:48] You need world-class teams [01:39:50] in order to do this, [01:39:51] and you need people [01:39:52] that think in workflows. [01:39:54] There's an enormous amount [01:39:55] of what I call [01:39:56] racing from stoplight [01:39:57] to stoplight. [01:39:59] Folks are going [01:40:00] from zero to 100, [01:40:01] and then they hit [01:40:01] the next stage [01:40:02] of the workflow, [01:40:03] and then there's [01:40:03] a red light, [01:40:04] and they haven't worked [01:40:04] their way all the way [01:40:05] through that. [01:40:07] So I would start [01:40:08] with people [01:40:09] and the quality [01:40:09] of the people [01:40:10] that you really have [01:40:13] across your enterprise [01:40:14] in order to do [01:40:14] these things. [01:40:15] The second point [01:40:16] is we are big believers [01:40:18] in data sovereignty, [01:40:20] and data sovereignty [01:40:21] for us means [01:40:22] not being tied [01:40:23] to a specific chipset, [01:40:25] not being tied [01:40:26] to a specific model [01:40:27] or a specific set [01:40:28] of development tools. [01:40:30] You need, [01:40:30] as an enterprise, [01:40:31] to manage your data. [01:40:32] Your data is your fuel [01:40:34] in terms of how [01:40:35] you implement AI [01:40:36] across that broader enterprise. [01:40:38] What we've done [01:40:39] with AMD, [01:40:40] and AMD's been [01:40:41] such a great partner [01:40:42] in this, [01:40:43] is leveraging [01:40:44] open source models, [01:40:45] leveraging different sets [01:40:46] of chipsets, [01:40:47] building our own models, [01:40:49] training other models. [01:40:50] We're one of the few [01:40:51] companies in the world [01:40:52] that's actually post-trained [01:40:53] a model on AMD [01:40:55] and then been able [01:40:56] to drive equivalency [01:40:57] in terms of performance [01:40:58] and accuracy [01:40:59] with other more expensive [01:41:00] chipsets [01:41:01] in the marketplace. [01:41:03] So as we have done that, [01:41:04] we've been able [01:41:05] to drive down token costs [01:41:07] essentially through moneyballing [01:41:08] across all of these [01:41:10] different models [01:41:10] and leveraging [01:41:11] that human capital [01:41:12] in a way [01:41:13] that's made a big difference [01:41:14] to our company. [01:41:16] And so increasingly, [01:41:17] even though token consumption [01:41:18] is going up, [01:41:19] we're able to manage [01:41:20] that underlying token cost [01:41:22] at enterprise scale, [01:41:23] which is not a small thing. [01:41:26] Yeah, it's just, [01:41:26] it's a bunch of great work. [01:41:28] Another key area [01:41:29] that's sort of groundbreaking, [01:41:31] which is exciting also, [01:41:32] is you're the first [01:41:33] telecom company [01:41:34] to actually train [01:41:35] a model on AMD [01:41:36] and also you're driving [01:41:38] fully open source models [01:41:40] in telecoms. [01:41:40] So tell us about [01:41:41] that experience [01:41:41] and why you felt [01:41:42] like that was needed. [01:41:44] We felt like the industry [01:41:45] was moving down a path [01:41:47] where folks were only [01:41:49] going to use [01:41:49] closed source models [01:41:50] to solve solutions. [01:41:52] And while closed source models [01:41:53] are an important ingredient [01:41:54] in that, [01:41:55] we didn't want to lose sight [01:41:56] of the fact that [01:41:56] open source models [01:41:57] were going to be [01:41:58] a very significant part [01:41:59] of the solution. [01:42:00] So we announced [01:42:01] what we call [01:42:02] Otel 1.0, [01:42:04] that's the open telco AI model, [01:42:06] a bit over a year ago [01:42:07] or so now, [01:42:08] and we had over 18 million [01:42:10] downloads of that model [01:42:12] since then. [01:42:13] We're excited [01:42:14] to announce today [01:42:15] the launch of Otel 2.0, [01:42:17] which is a more advanced [01:42:18] set of models [01:42:19] that we have trained [01:42:20] on AMD, [01:42:21] and we're now making [01:42:23] that available [01:42:23] via open source [01:42:24] across the broader industry. [01:42:27] And it, I think, [01:42:27] is going to give folks [01:42:28] an opportunity [01:42:29] to see [01:42:29] that not only [01:42:31] do you need [01:42:31] to think about [01:42:32] leveraging models, [01:42:33] but you also need [01:42:34] to think about [01:42:34] training those models [01:42:35] and how they work [01:42:36] across your enterprise. [01:42:38] So that's just great. [01:42:39] We certainly appreciate [01:42:40] the partnership, [01:42:41] and, you know, [01:42:41] we really feel like [01:42:43] this is germane [01:42:44] to the conversation. [01:42:44] It's exactly what [01:42:46] we've been talking [01:42:46] about today, [01:42:47] and I just want [01:42:48] to thank you again [01:42:48] for joining us [01:42:49] and the partnership. [01:42:50] Thank you. [01:42:51] You're a big part of it. [01:42:52] Thank you. [01:42:57] Okay, so let me close [01:42:59] with one last point. [01:43:01] We've established [01:43:02] ourselves as a leader [01:43:03] in enterprise computing [01:43:05] by listening to our customers [01:43:06] and delivering [01:43:07] the right solutions [01:43:08] for their workloads. [01:43:10] Today, we're extending [01:43:11] that leadership [01:43:12] to help customers [01:43:14] accelerate their AI journey. [01:43:17] This is our vision [01:43:18] for powering AI everywhere. [01:43:21] And to expand a little bit [01:43:23] beyond the data center [01:43:24] into physical [01:43:26] and client AI, [01:43:29] it's my pleasure [01:43:30] to introduce [01:43:30] Mr. Jack Yoon, [01:43:32] who leads our computing [01:43:33] and graphics group. [01:43:35] Thank you. [01:43:40] Thank you, Dan. [01:43:42] It's great to see [01:43:44] so many friends, [01:43:45] partners, [01:43:45] and developers [01:43:46] here in San Francisco. [01:43:47] Today, we have seen [01:43:49] how Andy is building [01:43:50] the compute foundation [01:43:51] for AI in a data center. [01:43:54] But the full AI transformation [01:43:56] requires two more ingredients. [01:43:58] Intelligence that becomes personal [01:43:59] and intelligence that can [01:44:01] understand and act [01:44:02] in the physical world. [01:44:04] Those are the two frontiers [01:44:05] I want to explore [01:44:06] with you here today. [01:44:07] Let's begin [01:44:08] with personal AI. [01:44:10] Enterprise AI [01:44:11] is already reshaping [01:44:12] the data center, [01:44:14] but agentic AI [01:44:15] will extend far beyond it. [01:44:17] Agents are the next [01:44:18] great productivity multiplier. [01:44:20] The jump from software [01:44:22] that responds [01:44:23] to software [01:44:24] that completes work. [01:44:25] A personal agent [01:44:26] can double [01:44:27] what one person can do. [01:44:29] A team of agents [01:44:29] can multiply that by 10. [01:44:31] And as agents [01:44:32] gain more capabilities, [01:44:34] often can grow exponentially. [01:44:36] But that promise [01:44:37] has a compute consequence. [01:44:39] Unlike a single AI query, [01:44:41] as agent reasons, [01:44:43] calls tools, [01:44:45] in accordance to other agents [01:44:46] and run nonstop, [01:44:50] agents become more capable [01:44:51] and pervasive. [01:44:52] So will token consumption. [01:44:55] Meaning that demand [01:44:56] will require every layer [01:44:57] of the computing ecosystem [01:44:58] to scale together. [01:45:00] Data centers remain [01:45:01] the foundation [01:45:02] for frontier AI [01:45:03] in the most demanding workloads. [01:45:05] But edge-inclined systems [01:45:07] can extend that capacity, [01:45:09] putting more intelligence [01:45:10] closer to where data [01:45:12] is created [01:45:12] and where the work gets done. [01:45:16] The PC is already [01:45:17] one of the most powerful [01:45:18] compute engines [01:45:19] ever placed [01:45:20] in the hands [01:45:21] of an individual. [01:45:22] Across billions of devices, [01:45:24] enormous computing capability [01:45:26] is already deployed [01:45:27] every single day. [01:45:28] Yet its potential [01:45:29] remains largely untapped. [01:45:32] Putting that capacity [01:45:33] to work [01:45:33] changed the economics [01:45:34] of computing. [01:45:36] More workloads [01:45:36] can run on systems [01:45:38] that are already in place, [01:45:39] reducing overall infrastructure costs [01:45:42] and allowing data centers [01:45:43] to focus on the largest, [01:45:45] most demanding task. [01:45:46] And local computing [01:45:47] brings inherent advantages. [01:45:50] Your data stays with you. [01:45:52] Your applications [01:45:52] remain responsive [01:45:53] and your system [01:45:55] keeps working [01:45:56] even when your network [01:45:58] does not. [01:45:59] And thanks to [01:45:59] a new generation [01:46:01] of smaller, [01:46:02] more capable AI models [01:46:04] will begin to unlock [01:46:05] that potential. [01:46:07] And the unlock [01:46:08] is happening faster [01:46:10] than anyone expected. [01:46:12] Last August, [01:46:13] GPT OSS [01:46:15] needed 120 billion parameters [01:46:17] to score 80 [01:46:18] on GPQA. [01:46:20] Just seven months later, [01:46:22] QN 3.5 scored even higher [01:46:25] with only 9 billion. [01:46:27] Better performance [01:46:28] with 13 times fewer parameters. [01:46:32] This is not [01:46:32] an incremental improvement. [01:46:34] It's a dramatic shift [01:46:35] in the compute required [01:46:37] to deliver [01:46:38] advanced intelligence. [01:46:40] And this is a trend [01:46:41] now an outlier. [01:46:43] Smaller open models [01:46:45] are closing [01:46:45] the frontier gap [01:46:46] at extraordinary speed. [01:46:49] QN 3.6 [01:46:50] 27 billion parameter miles [01:46:52] now outperforms [01:46:54] leading frontier miles [01:46:56] from the previous generation [01:46:57] on graduate level reasoning [01:46:59] at a fraction [01:47:00] of the size and cost. [01:47:02] That changes [01:47:03] is what is possible. [01:47:06] Workloads once reserved [01:47:08] for the cloud [01:47:09] can now run [01:47:10] on path compute engines [01:47:11] designed for personal AI compute. [01:47:15] Let's map [01:47:16] these model performance [01:47:18] breakthroughs [01:47:18] to any platforms. [01:47:20] Ryzen AI 400 [01:47:22] supports miles [01:47:24] up to 24 billion parameters. [01:47:27] More than enough [01:47:28] for QN 3.5 [01:47:29] 9 billion [01:47:30] with substantial headroom. [01:47:33] Ryzen AI Max [01:47:34] extends that capacity [01:47:36] to most as large [01:47:37] as 200 billion parameters [01:47:40] running natively [01:47:41] on a personal system. [01:47:43] This is not reduced AI [01:47:44] compressed to fit [01:47:46] on a PC. [01:47:47] It's a new class [01:47:48] of personal computing [01:47:49] built to run [01:47:51] powerful models [01:47:52] where the work [01:47:52] actually happens. [01:47:54] and this is exactly [01:47:56] what led us [01:47:57] to build [01:47:57] Ryzen AI Halo. [01:47:59] We wanted to give developers [01:48:01] a platform [01:48:02] powerful enough [01:48:03] to run these miles locally [01:48:04] yet simple enough [01:48:06] to use every day. [01:48:08] With 120 gigabytes [01:48:09] of unified memory [01:48:10] and support [01:48:12] for miles [01:48:12] up to 200 billion parameters [01:48:14] developers can build, [01:48:17] test, [01:48:18] and iterate locally [01:48:19] directly on their desk. [01:48:22] Ryzen AI Halo [01:48:24] embodies the future [01:48:25] of personal AI [01:48:26] for everyone. [01:48:27] A future [01:48:28] that shortens [01:48:28] the distance [01:48:29] between an idea [01:48:30] and a working model. [01:48:33] For developers, [01:48:34] that means [01:48:34] a dramatically [01:48:35] more efficient [01:48:36] and optimized path [01:48:37] from experimentation [01:48:38] to deployment. [01:48:41] And as Vanti [01:48:41] showed earlier, [01:48:43] rockam.ai [01:48:44] is making [01:48:44] hardware programming [01:48:45] radically more accessible. [01:48:48] Developers can understand, [01:48:51] modify, [01:48:52] and optimize code [01:48:53] for rockam [01:48:54] even without [01:48:55] deep rockam expertise. [01:48:58] And better yet, [01:48:59] the same open software [01:49:00] foundation extends [01:49:02] across all of AMD's hardware. [01:49:05] Developers can use [01:49:06] the frameworks [01:49:06] and tools [01:49:07] they already know, [01:49:08] start locally [01:49:09] on Ryzen AI Halo [01:49:10] and scale across [01:49:12] the full AMD compute platform [01:49:14] without rebuilding their work. [01:49:16] And we're not stopping there. [01:49:18] Today, [01:49:19] we're expanding our partnership [01:49:20] with Hug and Face [01:49:21] to bring the open AI ecosystem [01:49:24] directly on Ryzen AI Halo. [01:49:28] Power from compute [01:49:29] only creates value [01:49:30] when developers [01:49:31] can quickly access [01:49:33] the right models, [01:49:35] optimize them for the platform, [01:49:36] and turn them [01:49:37] into real applications. [01:49:39] Together, [01:49:41] AMD and Hug and Face [01:49:42] will deliver [01:49:43] Halo-optimized performance [01:49:45] for open models [01:49:46] and genetic workflows. [01:49:51] How cool is that? [01:49:54] And along with [01:49:55] co-engineered libraries [01:49:56] and toolkits [01:49:57] designed to help developers [01:49:58] spend less time [01:49:59] configuring infrastructure [01:50:00] and more time building. [01:50:03] And later this year, [01:50:05] every Ryzen AI Halo box [01:50:07] will include a full year [01:50:09] of Hug and Face Pro, [01:50:11] putting the power [01:50:12] of partnership [01:50:13] in developers' hands [01:50:14] from day one. [01:50:16] Freedom [01:50:16] and more power [01:50:18] to the user. [01:50:20] This is how [01:50:21] we accelerate innovation [01:50:22] in the AI era. [01:50:25] And we're already [01:50:25] pushing that vision [01:50:27] further [01:50:28] with Gorgon Halo. [01:50:30] And seeing it's believing, [01:50:32] look at how small [01:50:34] and beautiful [01:50:35] disk boxes. [01:50:41] We've increased [01:50:42] unified memory [01:50:43] from 128 [01:50:45] to 192 gigabytes, [01:50:48] the largest [01:50:49] unified memory pool [01:50:50] in this class, [01:50:51] and increased [01:50:52] mouse support [01:50:53] from 200 billion [01:50:54] to 300 billion [01:50:55] parameters. [01:50:57] Thank you, Laura. [01:51:01] And this is [01:51:02] no longer limited [01:51:03] to a single platform [01:51:04] form factor. [01:51:06] Together [01:51:06] with the world's [01:51:07] leading OEMs, [01:51:09] AMD now powers [01:51:10] the broadest portfolio [01:51:11] of personal AI systems [01:51:12] from laptops [01:51:13] and workstations [01:51:14] to many PCs [01:51:16] and developer platforms. [01:51:17] That means developers, [01:51:19] enterprises, [01:51:21] and creators [01:51:21] can choose [01:51:22] the right system [01:51:23] for the way [01:51:23] they want to build [01:51:24] and work. [01:51:26] Personal AI [01:51:27] is not a concept, [01:51:28] it is a category, [01:51:30] and is available [01:51:30] right now on AMD. [01:51:33] We now have [01:51:34] all the pieces. [01:51:36] Models are efficient enough. [01:51:38] Hardware is powerful enough. [01:51:40] Software stack [01:51:41] is open and ready. [01:51:43] It opened harnesses [01:51:44] like Lemonade Server [01:51:45] can route workloads [01:51:46] to the right model [01:51:47] and right compute. [01:51:48] The technology is here. [01:51:50] The next challenge [01:51:51] is taking it [01:51:52] from one developer's desk [01:51:53] to thousands of users [01:51:55] across an enterprise. [01:51:56] Cisco sees and believes [01:51:58] in that future [01:51:58] the same way we do. [01:52:00] To take us deeper [01:52:01] into how we turn [01:52:02] that vision [01:52:02] into enterprise-scale reality, [01:52:05] please join me [01:52:05] in welcoming to the stage [01:52:07] Jitu Patel, [01:52:08] President and Chief Product [01:52:09] Officer at Cisco. [01:52:14] Jitu, how are you doing? [01:52:15] How are you? [01:52:16] You look great. [01:52:17] You too. [01:52:17] Thank you for joining me here today. [01:52:18] Congratulations. [01:52:19] Thank you. [01:52:20] Thank you. [01:52:21] Now, Jitu, [01:52:21] you and I talked [01:52:22] how we're at inflection point. [01:52:24] And for the past few years, [01:52:26] Energy has spent time [01:52:27] just building intelligence. [01:52:29] The next phase [01:52:30] is how we deploy [01:52:31] this intelligence [01:52:32] at scale enterprises. [01:52:33] You spent so much [01:52:34] of your time [01:52:35] talking to CIOs, [01:52:36] technology leaders. [01:52:37] What are they telling you? [01:52:38] What's changing? [01:52:39] Well, firstly, [01:52:40] it's a really exciting time [01:52:41] to be alive in tech. [01:52:42] And if you just take [01:52:43] a step back right now [01:52:44] and see what's happening, [01:52:45] there's actually [01:52:46] a fundamental shift [01:52:47] that's happening [01:52:48] in the patterns of inferencing. [01:52:49] So if you think [01:52:50] about the patterns [01:52:51] of inferencing [01:52:51] that existed [01:52:52] during the chatbot era [01:52:53] that was human-led, [01:52:55] it was very, very spiky. [01:52:56] You ask a question, [01:52:57] you get an answer. [01:52:58] What you're starting [01:52:58] to see with agents [01:52:59] is they're working [01:53:00] seven by 24. [01:53:02] They're very consumptive [01:53:04] on bandwidth [01:53:04] and on infrastructure. [01:53:05] So you're starting [01:53:06] to see a much more [01:53:07] persistent pattern [01:53:08] of kind of demand [01:53:09] signal for infrastructure. [01:53:10] You know, [01:53:11] we usually joke [01:53:13] around internally, [01:53:14] humans click, [01:53:14] but agents swarm. [01:53:15] Yes. [01:53:16] You know? [01:53:17] And so as you start [01:53:18] to see this, [01:53:19] what you're going to have [01:53:20] is a tremendous amount [01:53:21] of, [01:53:23] and there's two big concerns [01:53:24] that are also there, [01:53:25] which is one is [01:53:26] every customer [01:53:26] is concerned about security [01:53:28] and every customer [01:53:30] is concerned about cost. [01:53:31] So what you're starting [01:53:32] to see happen [01:53:33] is inferencing [01:53:35] is not just going to be [01:53:36] limited to the data center. [01:53:38] Inferencing is going [01:53:38] to be distributed everywhere [01:53:39] and there's a whole new [01:53:40] class of computing [01:53:41] that's emerging [01:53:42] with this desk-side computing [01:53:44] that you just announced [01:53:46] where there's going [01:53:47] to be inferencing [01:53:47] that you're going to have [01:53:48] a laptop or a desktop [01:53:49] where humans work [01:53:51] and they might have [01:53:52] a desk-side computer [01:53:53] where agents are kind [01:53:54] of going out [01:53:54] and running their jobs [01:53:56] on behalf of humans. [01:53:58] So that's what [01:53:59] we're seeing essentially. [01:54:00] And we're fully aligned G2 [01:54:01] and that's exactly [01:54:02] the shift we're seeing. [01:54:03] You know, [01:54:03] if AI is moving closer [01:54:05] to employees, [01:54:06] compute has to move [01:54:06] closer as well, right? [01:54:08] The PC is not only [01:54:09] just a productivity device [01:54:10] because the intelligence node [01:54:12] in the enterprise, [01:54:14] but deploying intelligence [01:54:15] everywhere creates [01:54:16] a new channel for CIOs [01:54:17] that you and I [01:54:18] talk about so much. [01:54:19] We do, yeah. [01:54:19] How do you operate [01:54:20] thousands of these systems [01:54:22] with security, governance, [01:54:24] and control [01:54:24] enterprises require? [01:54:26] G2, how is Amy and Cisco [01:54:28] solving this [01:54:29] and what are we building together? [01:54:31] Yeah, so this is [01:54:32] an exciting time [01:54:33] because what you're [01:54:34] starting to see is [01:54:35] there's a few things [01:54:36] that need to really be done. [01:54:38] You can't just go out [01:54:39] and deploy your [01:54:40] desk-side computers [01:54:41] and hope that everything [01:54:43] works out in the enterprise. [01:54:44] You need to make sure [01:54:45] that it's governed, [01:54:46] it's managed. [01:54:47] And so there's a few things [01:54:48] that are really important. [01:54:49] Number one, [01:54:50] you have to have [01:54:50] an appropriate amount [01:54:51] of network bandwidth [01:54:52] to satiate the needs [01:54:53] of the agents [01:54:54] that you're going to run locally. [01:54:56] So this notion [01:54:57] of network infrastructure [01:54:58] that can get modernized [01:55:00] to keep up with the needs [01:55:02] that you're going to go out [01:55:03] and generate [01:55:03] is going to be very important. [01:55:04] That's number one. [01:55:05] Number two, [01:55:06] like I said, [01:55:07] every customer is really [01:55:08] worried about token costs. [01:55:10] And so we need to make sure [01:55:11] that we can actually contain [01:55:12] the costs of an agent's [01:55:14] consumption of tokens [01:55:15] when it actually starts [01:55:16] to go awry. [01:55:16] Up and down when needed. [01:55:17] Exactly. [01:55:18] Number three then is [01:55:20] how do we monitor agent behavior [01:55:22] for safety and security [01:55:24] so that if it does start [01:55:25] to do things [01:55:26] that we don't want it to do, [01:55:27] we can in runtime [01:55:28] provide enforcement guardrails. [01:55:30] And then fourthly, [01:55:32] it's just the overall safety [01:55:33] and security [01:55:34] that you need to have [01:55:35] from an apparatus perspective [01:55:36] to get this entire thing humming [01:55:38] in the way that you want it to hum. [01:55:39] Yes, G2, [01:55:40] let's hear how you and I [01:55:40] think very alike. [01:55:41] So, I mean, [01:55:42] the opportunity is not just [01:55:44] being at every device [01:55:45] as you and I talk. [01:55:46] How do we enable [01:55:46] enterprise to deploy, [01:55:48] manage and scale out [01:55:49] with confidence? [01:55:50] Why don't we take a look [01:55:50] at what we've been building [01:55:51] in the past few weeks and months? [01:55:53] Yeah, it's exciting. [01:55:55] So basically what we've done [01:55:56] is there's a full stack [01:55:58] that you can start to think about. [01:55:59] So what you folks have done [01:56:00] with Halo is you've got [01:56:02] an isolated secure agent sandbox, [01:56:06] you've got intelligent routing, [01:56:07] you've got an MCP, [01:56:09] core set of integrations. [01:56:13] And then what we've done [01:56:14] above that is we've made sure [01:56:16] that you have the right level [01:56:17] of security policy enforcement. [01:56:20] We've also got the right level [01:56:21] of observability [01:56:22] so that it's an end-to-end [01:56:24] resilient infrastructure stack. [01:56:26] So it's observability [01:56:27] of how is your apparatus [01:56:30] and the infrastructure working? [01:56:32] Is the agent behavior [01:56:33] being done in the right way? [01:56:35] Can you observe the agent behavior? [01:56:36] And are the tokenomics [01:56:38] within the guidelines [01:56:39] that you want them to be? [01:56:41] And then what we have [01:56:42] is a single unified management plane, [01:56:46] a control plane [01:56:47] that can look at every single [01:56:49] dimension that you have [01:56:51] and be managed [01:56:52] within one environment. [01:56:54] And so that's kind of [01:56:55] the full stack. [01:56:56] And we are really, really excited [01:56:57] about the partnership. [01:56:58] You know, us too. [01:56:59] Our teams love working [01:56:59] with your team, G2. [01:57:00] Yeah. [01:57:00] How is this all managed? [01:57:02] How does the CAO [01:57:03] deploy this at scale? [01:57:04] Yeah. [01:57:04] So let's take a look [01:57:05] at what this looks like, right? [01:57:06] Because what we would have [01:57:08] is this is Cisco Cloud Control, [01:57:10] which is the overall management plane. [01:57:12] So if you happen to have Halo devices [01:57:15] on every desktop, [01:57:17] which it might actually get to, [01:57:19] what you want to make sure [01:57:21] that you do is you are able [01:57:22] to go out and manage [01:57:23] your entire estate [01:57:26] of not just your Halo fleet, [01:57:28] but also what you have running [01:57:29] on the data centers [01:57:30] as well as what you have running [01:57:32] in the cloud. [01:57:32] So you can provide full visibility [01:57:35] for that entirety of the estate. [01:57:38] And then once you've got that, [01:57:39] what we also have is this notion [01:57:41] of how do we go out [01:57:42] and monitor tokenomics? [01:57:45] And so is the agent behaving [01:57:47] the way that it needs to behave? [01:57:49] Or is the agent actually going out [01:57:51] and being overly consumptive? [01:57:53] At which point you should be able [01:57:55] to isolate any individual Halo device [01:57:58] so that you can make sure [01:57:59] that you can quarantine it. [01:58:00] No, I love this. [01:58:00] Then we actually see [01:58:01] the return of invested capital. [01:58:02] It could pay for itself [01:58:03] in three or six months [01:58:04] even quicker. [01:58:05] Exactly. [01:58:05] So what you see over here basically [01:58:07] is you can tell [01:58:08] how every single agent [01:58:10] is performing. [01:58:11] and whether or not [01:58:12] they're consuming [01:58:13] way too much [01:58:13] kind of token costs. [01:58:15] And if they are, [01:58:16] how do you actually contain [01:58:17] the cost as you're going through it? [01:58:18] No, I love this G2. [01:58:19] Now, everyone else is wondering [01:58:21] when can they buy this? [01:58:23] Yeah. [01:58:23] When is this open for everyone? [01:58:25] Well, the good news [01:58:26] is you can buy the Halo [01:58:27] already today. [01:58:29] It does. [01:58:29] And then what we're doing [01:58:31] is we're making sure [01:58:31] that we provide [01:58:32] this entire management apparatus [01:58:34] that's currently [01:58:36] in early availability [01:58:37] for a select set of customers. [01:58:39] but we'll have it [01:58:40] in general availability [01:58:41] in the U.S. [01:58:43] in early fall. [01:58:45] So we're excited [01:58:46] to make sure [01:58:46] that we can provide [01:58:47] not just the compute [01:58:48] but all of the apparatus [01:58:49] that's needed [01:58:50] to go out and manage this [01:58:51] so that you can have inference [01:58:52] running in the cloud, [01:58:54] you can have inference [01:58:55] running in your data center [01:58:56] privately [01:58:56] or you can have inference [01:58:58] running on your desk side. [01:58:59] G2, we're so grateful [01:59:00] for a partnership [01:59:01] with you and Cisco [01:59:02] and we're so excited. [01:59:03] Thank you so much. [01:59:03] What you just saw [01:59:11] brings the vision [01:59:12] and scaling [01:59:12] AI seriously and securely [01:59:14] from personal devices [01:59:15] all the way [01:59:16] to the cloud [01:59:17] but in the age of AI [01:59:19] intelligence [01:59:19] will be pervasive [01:59:20] and it will be everywhere. [01:59:22] The next frontier [01:59:23] is where AI [01:59:24] leaves the screen [01:59:25] enters the world [01:59:26] and acts. [01:59:28] That frontier [01:59:29] here is physical AI [01:59:30] the ultimate application [01:59:32] of agentic frameworks [01:59:33] in the digital world [01:59:35] an agent plans [01:59:37] and completes a task [01:59:38] in the physical world [01:59:40] it must also sense motion [01:59:41] navigate uncertainty [01:59:43] and protect people [01:59:45] here intelligence [01:59:47] does not just produce [01:59:48] an answer [01:59:49] it produces an action [01:59:50] when answers [01:59:52] become actions [01:59:53] latency becomes safety [01:59:54] and reliability [01:59:55] becomes trust [01:59:56] that is why [01:59:58] the promise [01:59:59] of physical AI [01:59:59] is so incredibly powerful [02:00:01] it is not about [02:00:02] replacing human potential [02:00:04] it is about amplifying it [02:00:06] giving surgeons [02:00:07] greater precision [02:00:08] keeping people [02:00:10] out of harm's way [02:00:11] making farms [02:00:12] more productive [02:00:13] supply chains [02:00:15] more resilient [02:00:15] and factories [02:00:17] more efficient [02:00:18] the best technology [02:00:20] does not make people [02:00:20] smaller [02:00:21] it gives them [02:00:22] much greater reach [02:00:23] for more than 20 years [02:00:25] AMD has helped [02:00:27] power the core capabilities [02:00:28] that robotics depend on [02:00:30] from sensing and functional safety [02:00:32] to deterministic control [02:00:34] and precision motion [02:00:36] that foundation [02:00:37] is already trusted [02:00:38] by many of the companies [02:00:39] to find their future robotics [02:00:41] and now [02:00:42] that foundation [02:00:44] is coming alive [02:00:45] in extraordinary ways [02:00:46] these are not machines [02:00:48] following a script [02:00:49] they are beginning [02:00:50] to perceive the world [02:00:51] reason through uncertainty [02:00:53] and act with purpose [02:00:55] the promise of physical AI [02:00:58] already in motion [02:00:59] the next era of robotics [02:01:02] is not programmed [02:01:03] it is autonomous [02:01:05] while traditional robots [02:01:07] execute instructions [02:01:08] autonomous robots [02:01:09] are given an objective [02:01:11] and from there [02:01:12] they determine how to achieve it [02:01:14] and they do this [02:01:15] all at once [02:01:16] and all in real time [02:01:18] the robot must orchestrate [02:01:20] intelligence [02:01:21] motion [02:01:22] and safety [02:01:23] continuously [02:01:24] and every part [02:01:26] must respond [02:01:27] at the speed [02:01:28] of the world around it [02:01:29] that demands [02:01:31] an entirely new [02:01:32] kind of robotics brain [02:01:33] and to solve that [02:01:35] we rethought [02:01:36] its architecture [02:01:37] from the ground up [02:01:38] today [02:01:39] I'm so excited [02:01:40] to introduce [02:01:41] the AMD Korea [02:01:42] AI system [02:01:43] on module [02:01:44] powered [02:01:45] by Ryzen AI [02:01:46] embedded [02:01:47] X100 [02:01:48] it brings CPU [02:01:55] GPU [02:01:56] NPU [02:01:58] and unified memory [02:01:59] together [02:01:59] in one compact [02:02:01] open standard [02:02:02] architecture [02:02:03] this allows [02:02:04] robots [02:02:05] to process [02:02:05] perception [02:02:06] AI reasoning [02:02:08] and do it [02:02:09] all simultaneously [02:02:10] and in real time [02:02:12] the architecture [02:02:13] is different [02:02:14] but the results [02:02:15] are decisive [02:02:16] in third party [02:02:18] testing [02:02:18] against NVIDIA [02:02:19] Jensen Thor [02:02:20] Korea delivers [02:02:22] 3.4 times [02:02:23] better real time [02:02:24] results [02:02:25] 2.3 times [02:02:27] more concurrent [02:02:27] agents [02:02:28] and 1.6 times [02:02:30] more CPU [02:02:31] capacity [02:02:31] that means [02:02:33] faster reactions [02:02:34] and greater [02:02:36] headroom [02:02:36] for the robot [02:02:37] to take on [02:02:37] more complex work [02:02:38] and all this [02:02:40] without compromising [02:02:41] control [02:02:41] or safety [02:02:42] but [02:02:44] to find [02:02:45] the next era [02:02:46] of robotics [02:02:46] takes more [02:02:47] than a powerful [02:02:48] brain [02:02:48] developers [02:02:50] need a complete [02:02:51] path [02:02:51] from idea [02:02:52] to a machine [02:02:53] operating [02:02:54] in the real [02:02:55] world [02:02:56] today [02:02:57] we are introducing [02:02:59] AMD's [02:03:00] Korea AI [02:03:01] robotics [02:03:02] developer platform [02:03:03] the industry's [02:03:10] first turnkey [02:03:11] open [02:03:12] fully [02:03:13] integrative [02:03:14] robotic system [02:03:15] thank you [02:03:17] Laura so much [02:03:18] at the center [02:03:24] is the [02:03:24] Korea AI [02:03:25] system [02:03:25] on module [02:03:26] paired with [02:03:27] a robotics [02:03:28] carrier card [02:03:29] designed [02:03:29] for sensing [02:03:31] and connectivity [02:03:31] built [02:03:32] on Rockham [02:03:33] and Ross [02:03:34] 2 [02:03:34] this platform [02:03:35] gives developers [02:03:36] everything they need [02:03:37] to move [02:03:37] from concept [02:03:38] to prototype [02:03:39] in days [02:03:40] and the same [02:03:41] module [02:03:42] carries forward [02:03:43] into production [02:03:43] with no migration [02:03:45] or redesign [02:03:46] needed [02:03:47] at the finish line [02:03:47] you can build [02:03:49] faster [02:03:49] and move [02:03:50] from imagination [02:03:51] to deployment [02:03:52] with absolute [02:03:53] confidence [02:03:53] this is the [02:03:55] breath [02:03:56] only AMD [02:03:56] can bring [02:03:57] to Tom's [02:03:57] robotics [02:03:58] from the [02:03:59] first signal [02:04:00] to the [02:04:00] final motion [02:04:01] AMD [02:04:02] Korea [02:04:03] for the [02:04:03] brain [02:04:04] Verso [02:04:05] for the [02:04:05] spine [02:04:06] Zinc [02:04:07] for the [02:04:07] joints [02:04:08] and Spartan [02:04:09] for the [02:04:10] sensors [02:04:10] each layer [02:04:11] purpose [02:04:12] built [02:04:12] for its [02:04:13] role [02:04:13] working [02:04:14] together [02:04:15] in one [02:04:15] intelligent [02:04:16] system [02:04:16] connecting [02:04:17] perception [02:04:18] and high [02:04:19] level [02:04:19] reasoning [02:04:20] to real [02:04:21] time [02:04:21] control [02:04:22] the future [02:04:24] will not [02:04:24] be defined [02:04:25] by one [02:04:25] model [02:04:25] one [02:04:27] machine [02:04:27] or one [02:04:28] company [02:04:29] it will [02:04:29] be built [02:04:30] in the [02:04:30] open [02:04:30] across [02:04:32] silicon [02:04:32] software [02:04:34] systems [02:04:35] and an [02:04:36] ecosystem [02:04:36] moving forward [02:04:37] together [02:04:38] from the [02:04:39] cloud [02:04:39] to the [02:04:39] PC [02:04:40] and now [02:04:41] into the [02:04:41] physical [02:04:42] world [02:04:42] AMD [02:04:43] continues [02:04:43] to build [02:04:44] the [02:04:44] computer [02:04:44] foundation [02:04:45] for [02:04:45] intelligence [02:04:46] everywhere [02:04:47] with [02:04:48] laser [02:04:48] focus [02:04:49] on [02:04:49] the [02:04:49] future [02:04:49] so [02:04:50] our [02:04:50] solution [02:04:50] can [02:04:50] augment [02:04:51] and [02:04:51] expand [02:04:52] what [02:04:52] people [02:04:53] are [02:04:53] capable [02:04:53] of [02:04:53] achieving [02:04:53] the [02:04:55] next [02:04:55] era [02:04:55] of [02:04:56] AI [02:04:56] will [02:04:56] not [02:04:56] just [02:04:57] unfold [02:04:57] behind [02:04:57] a [02:04:57] screen [02:04:58] it [02:04:58] will [02:04:59] unfold [02:04:59] in [02:04:59] the [02:04:59] world [02:04:59] around [02:05:00] us [02:05:00] and [02:05:01] now [02:05:01] to bring [02:05:02] it [02:05:02] all [02:05:02] together [02:05:02] and [02:05:03] take [02:05:03] us [02:05:03] to [02:05:03] what [02:05:03] comes [02:05:04] next [02:05:04] please [02:05:05] welcome [02:05:05] back [02:05:05] Lisa [02:05:06] to [02:05:06] the [02:05:06] stage [02:05:07] hi [02:05:13] Lisa [02:05:13] all right [02:05:16] thank you [02:05:16] Jack [02:05:17] and [02:05:17] it's [02:05:18] been [02:05:18] a [02:05:18] big [02:05:19] day [02:05:19] we've [02:05:19] covered [02:05:20] a lot [02:05:20] from [02:05:21] our [02:05:21] data [02:05:21] center [02:05:22] products [02:05:22] to [02:05:23] enterprise [02:05:23] AI [02:05:23] to [02:05:24] personal [02:05:24] and [02:05:24] physical [02:05:25] AI [02:05:25] and [02:05:25] the [02:05:25] software [02:05:26] that [02:05:26] ties [02:05:26] it [02:05:26] all [02:05:26] together [02:05:27] this [02:05:28] is [02:05:28] the [02:05:28] strongest [02:05:29] product [02:05:29] portfolio [02:05:30] in our [02:05:30] history [02:05:31] and [02:05:31] we [02:05:32] believe [02:05:32] it's [02:05:32] the [02:05:32] strongest [02:05:33] portfolio [02:05:33] in the [02:05:34] industry [02:05:34] but [02:05:35] it's [02:05:35] really [02:05:35] just [02:05:36] the [02:05:36] beginning [02:05:36] because [02:05:36] what [02:05:37] our [02:05:37] customers [02:05:38] count [02:05:38] on [02:05:38] is [02:05:39] not [02:05:39] just [02:05:39] today [02:05:40] they [02:05:40] really [02:05:41] count [02:05:41] on [02:05:41] us [02:05:41] looking [02:05:42] ahead [02:05:42] and [02:05:43] pushing [02:05:43] the [02:05:43] bleeding [02:05:44] edge [02:05:44] of [02:05:44] technology [02:05:44] so [02:05:45] I [02:05:45] want [02:05:45] to [02:05:50] use [02:05:51] in [02:05:51] 2028 [02:05:52] we're [02:05:53] going [02:05:53] to [02:05:53] introduce [02:05:54] Florence [02:05:54] Florence [02:05:55] brings [02:05:56] the [02:05:56] next [02:05:56] gen [02:05:57] 7 [02:05:57] 7 [02:05:57] cores [02:05:58] it's [02:05:59] leading [02:05:59] edge [02:05:59] process [02:05:59] technology [02:06:00] it's [02:06:01] a new [02:06:01] set [02:06:02] of [02:06:02] AI [02:06:02] compute [02:06:02] extensions [02:06:03] to [02:06:03] really [02:06:04] ensure [02:06:05] that [02:06:05] we [02:06:05] have [02:06:05] all [02:06:05] of [02:06:05] the [02:06:06] AI [02:06:06] capability [02:06:06] and [02:06:07] it [02:06:07] supports [02:06:07] the [02:06:07] latest [02:06:08] memory [02:06:08] technologies [02:06:09] and [02:06:09] we're [02:06:10] not [02:06:10] stopping [02:06:10] there [02:06:10] we're [02:06:11] already [02:06:11] deep [02:06:12] in [02:06:12] development [02:06:12] of [02:06:13] Ravenna [02:06:13] our [02:06:14] 8th [02:06:14] generation [02:06:15] Epic [02:06:15] family [02:06:16] built [02:06:16] on [02:06:16] Zen 8 [02:06:17] and [02:06:17] that [02:06:17] family [02:06:18] is [02:06:18] already [02:06:18] well [02:06:19] under [02:06:19] development [02:06:19] for [02:06:20] 2030 [02:06:20] now [02:06:22] to [02:06:22] give [02:06:22] you [02:06:22] a [02:06:22] little [02:06:23] bit [02:06:23] of [02:06:23] a [02:06:23] flavor [02:06:24] Florence [02:06:24] is [02:06:25] again [02:06:25] going [02:06:26] to [02:06:26] extend [02:06:26] our [02:06:27] biggest [02:06:27] advantage [02:06:27] with [02:06:28] Epic [02:06:28] it's [02:06:28] all [02:06:29] about [02:06:29] product [02:06:30] breadth [02:06:30] and [02:06:31] having [02:06:31] the [02:06:31] right [02:06:32] foundation [02:06:33] and [02:06:33] also [02:06:34] the [02:06:34] right [02:06:34] family [02:06:35] for [02:06:35] each [02:06:35] workload [02:06:36] so [02:06:36] you [02:06:36] have [02:06:36] Florence [02:06:37] you [02:06:37] have [02:06:38] Ferrara [02:06:38] you [02:06:38] have [02:06:39] Fidenza [02:06:39] and [02:06:40] it's [02:06:40] built [02:06:41] on [02:06:43] that [02:06:43] is [02:06:43] optimized [02:06:44] for [02:06:44] these [02:06:44] different [02:06:45] workloads [02:06:45] that's [02:06:46] giving [02:06:46] the [02:06:46] customers [02:06:47] the [02:06:47] right [02:06:48] CPU [02:06:48] for [02:06:49] every [02:06:49] workload [02:06:50] and [02:06:50] that [02:06:50] is [02:06:51] our [02:06:51] commitment [02:06:51] with [02:06:52] Epic [02:06:52] now [02:06:53] moving [02:06:54] to [02:06:54] instinct [02:06:54] we [02:06:55] are [02:06:55] continuing [02:06:56] our [02:06:56] cadence [02:06:56] of [02:06:57] delivering [02:06:57] a [02:06:58] new [02:06:58] generation [02:06:59] every [02:06:59] single [02:07:00] year [02:07:00] MI [02:07:01] 500 [02:07:02] brings [02:07:02] next [02:07:03] generation [02:07:03] HBM [02:07:04] it [02:07:05] brings [02:07:05] a [02:07:05] larger [02:07:06] scale [02:07:06] up [02:07:06] domain [02:07:07] because [02:07:07] scaling [02:07:08] is [02:07:08] everything [02:07:08] and [02:07:09] it [02:07:09] introduced [02:07:10] new [02:07:10] copper [02:07:11] and [02:07:11] optical [02:07:11] interconnects [02:07:12] and [02:07:13] with [02:07:13] MI [02:07:14] 600 [02:07:14] it's [02:07:15] powered [02:07:15] by [02:07:15] our [02:07:15] CDNA [02:07:16] next [02:07:16] architecture [02:07:17] and [02:07:17] it's [02:07:17] already [02:07:18] deep [02:07:18] in [02:07:18] development [02:07:18] for [02:07:19] 2028 [02:07:20] now [02:07:21] to give [02:07:21] you [02:07:22] an [02:07:22] idea [02:07:22] of [02:07:23] what [02:07:23] we [02:07:23] see [02:07:23] from [02:07:24] a [02:07:24] performance [02:07:24] standpoint [02:07:25] every [02:07:26] generation [02:07:27] of [02:07:27] instinct [02:07:27] has [02:07:28] delivered [02:07:28] a [02:07:28] major [02:07:29] big [02:07:29] step [02:07:29] up [02:07:30] in [02:07:30] performance [02:07:30] and [02:07:31] with [02:07:31] MI [02:07:31] 455 [02:07:32] we've [02:07:32] talked [02:07:33] a lot [02:07:33] about [02:07:33] it [02:07:33] delivering [02:07:34] 35 [02:07:34] times [02:07:35] more [02:07:35] inference [02:07:36] throughput [02:07:36] compared [02:07:37] to [02:07:37] MI [02:07:38] 355 [02:07:38] but [02:07:39] we're [02:07:39] really [02:07:40] really [02:07:40] excited [02:07:41] about [02:07:42] what we [02:07:42] see [02:07:43] is [02:07:43] another [02:07:44] opportunity [02:07:44] to take [02:07:45] a major [02:07:46] step [02:07:46] up [02:07:47] and [02:07:47] actually [02:07:47] bend [02:07:47] the [02:07:48] curve [02:07:48] again [02:07:48] so [02:07:49] MI [02:07:49] 500 [02:07:50] will [02:07:50] deliver [02:07:51] the [02:07:51] largest [02:07:51] generational [02:07:52] leap [02:07:52] in the [02:07:53] history [02:07:53] of [02:07:53] instinct [02:07:53] putting [02:07:54] us [02:07:54] on track [02:07:55] to [02:07:55] deliver [02:07:55] more [02:07:56] than [02:07:56] 2,000 [02:07:57] times [02:07:57] higher [02:07:57] inference [02:07:58] throughput [02:07:58] in [02:07:59] just [02:07:59] four [02:08:00] years [02:08:00] now [02:08:07] what [02:08:08] I [02:08:08] can [02:08:08] tell [02:08:08] you [02:08:08] is [02:08:09] we [02:08:09] are [02:08:09] working [02:08:09] with [02:08:10] a [02:08:10] number [02:08:10] of [02:08:10] our [02:08:10] customers [02:08:11] already [02:08:11] on [02:08:11] this [02:08:12] design [02:08:12] point [02:08:12] MI [02:08:13] 500 [02:08:13] is [02:08:14] super [02:08:14] exciting [02:08:15] and [02:08:15] the [02:08:15] feedback [02:08:16] we're [02:08:16] getting [02:08:16] from [02:08:17] customers [02:08:17] is [02:08:17] just [02:08:18] fantastic [02:08:18] now [02:08:19] put [02:08:20] the [02:08:20] CPU [02:08:21] GPU [02:08:21] and [02:08:22] the [02:08:22] networking [02:08:22] roadmaps [02:08:23] together [02:08:23] and [02:08:24] you [02:08:24] can [02:08:24] now [02:08:24] see [02:08:25] a [02:08:25] complete [02:08:25] cadence [02:08:26] of [02:08:26] our [02:08:26] rack [02:08:27] scale [02:08:27] systems [02:08:27] so [02:08:28] you [02:08:28] should [02:08:28] expect [02:08:29] from [02:08:29] AMD [02:08:29] a [02:08:30] new [02:08:30] Helios [02:08:31] system [02:08:31] every [02:08:32] year [02:08:32] delivering [02:08:33] more [02:08:33] performance [02:08:34] more [02:08:35] capacity [02:08:35] so [02:08:36] that [02:08:36] we [02:08:36] can [02:08:36] run [02:08:36] larger [02:08:37] models [02:08:37] more [02:08:38] agents [02:08:39] at [02:08:39] better [02:08:39] total [02:08:40] cost [02:08:40] of [02:08:40] ownership [02:08:40] and [02:08:41] that [02:08:41] gives [02:08:42] our [02:08:42] customers [02:08:42] the [02:08:43] complete [02:08:43] predictable [02:08:44] roadmap [02:08:44] to [02:08:45] plan [02:08:46] and [02:08:46] scale [02:08:46] their [02:08:46] infrastructure [02:08:47] for [02:08:47] years [02:08:48] to [02:08:48] come [02:08:48] so [02:08:49] lots [02:08:50] of [02:08:50] exciting [02:08:50] things [02:08:51] in [02:08:51] store [02:08:51] over [02:08:51] the [02:08:53] with [02:08:53] that [02:08:54] let [02:08:54] me [02:08:54] wrap [02:08:54] this [02:08:55] up [02:08:55] for [02:08:55] this [02:08:55] morning [02:08:56] today [02:08:57] I [02:08:57] hope [02:08:57] you [02:08:57] saw [02:08:58] a [02:08:58] little [02:08:58] bit [02:08:59] about [02:08:59] how [02:08:59] we're [02:09:00] really [02:09:00] looking [02:09:01] at [02:09:01] that [02:09:02] compute [02:09:02] from [02:09:03] every [02:09:03] angle [02:09:03] so [02:09:04] every [02:09:04] aspect [02:09:04] of [02:09:05] AI [02:09:05] compute [02:09:05] whether [02:09:06] you're [02:09:06] talking [02:09:06] about [02:09:07] Helios [02:09:08] or [02:09:08] MI455 [02:09:09] or [02:09:09] Venice [02:09:09] for [02:09:10] the [02:09:10] world's [02:09:11] largest [02:09:11] AI [02:09:11] systems [02:09:12] or [02:09:12] you're [02:09:12] talking [02:09:13] about [02:09:13] MI430 [02:09:14] and [02:09:14] MI350 [02:09:15] for [02:09:15] sovereign [02:09:16] and [02:09:16] enterprise [02:09:17] AI [02:09:17] you [02:09:18] saw [02:09:18] the [02:09:18] passion [02:09:19] in [02:09:19] BOMC [02:09:19] with [02:09:19] Rockham [02:09:20] AI [02:09:20] it's [02:09:21] so [02:09:21] good [02:09:21] to [02:09:22] see [02:09:22] how [02:09:22] much [02:09:22] has [02:09:23] come [02:09:23] through [02:09:23] that [02:09:23] platform [02:09:24] in [02:09:24] terms [02:09:25] of [02:09:25] making [02:09:25] developers [02:09:26] make [02:09:26] it [02:09:26] much [02:09:27] easier [02:09:27] for [02:09:27] developers [02:09:28] to [02:09:28] access [02:09:29] the [02:09:29] AMD [02:09:30] platform [02:09:30] and [02:09:31] then [02:09:31] Jack [02:09:32] talked [02:09:32] about [02:09:32] Gorgon [02:09:33] Halo [02:09:33] and [02:09:33] our [02:09:34] new [02:09:34] KREA [02:09:34] AI [02:09:34] platforms [02:09:35] really [02:09:36] bringing [02:09:36] together [02:09:36] the [02:09:37] leadership [02:09:37] of [02:09:38] the [02:09:38] end [02:09:38] to [02:09:38] end [02:09:38] AI [02:09:39] story [02:09:39] including [02:09:40] PCs [02:09:40] and [02:09:41] physical [02:09:41] AI [02:09:41] so [02:09:42] lots [02:09:43] of [02:09:43] new [02:09:43] information [02:09:43] but [02:09:44] I [02:09:44] want [02:09:44] to [02:09:44] say [02:09:44] a [02:09:45] very [02:09:45] very [02:09:46] special [02:09:46] thank [02:09:46] you [02:09:47] to [02:09:47] all [02:09:47] of [02:09:47] our [02:09:47] partners [02:09:48] who [02:09:48] joined [02:09:48] us [02:09:48] today [02:09:49] because [02:09:49] it [02:09:49] is [02:09:50] really [02:09:50] through [02:09:50] those [02:09:51] partnerships [02:09:51] that [02:09:52] we're [02:09:52] able [02:09:52] to [02:09:52] do [02:09:53] the [02:09:53] most [02:09:53] amazing [02:09:54] things [02:09:54] together [02:09:55] so [02:09:56] if [02:09:56] I [02:09:56] just [02:09:56] leave [02:09:57] you [02:09:57] with [02:09:57] a [02:09:57] final [02:09:57] thought [02:09:58] when [02:09:59] I [02:09:59] think [02:09:59] about [02:10:00] where [02:10:00] AI [02:10:00] is [02:10:01] today [02:10:01] the [02:10:02] biggest [02:10:02] change [02:10:03] that [02:10:03] we [02:10:15] are [02:10:16] personal [02:10:16] lives [02:10:17] and [02:10:18] I [02:10:18] have [02:10:18] to [02:10:18] say [02:10:19] that [02:10:19] I [02:10:19] spent [02:10:19] my [02:10:20] entire [02:10:20] career [02:10:20] in [02:10:21] tech [02:10:21] believing [02:10:22] that [02:10:22] high [02:10:23] performance [02:10:23] computing [02:10:23] can [02:10:24] make [02:10:24] the [02:10:24] world [02:10:25] an [02:10:25] incredibly [02:10:26] better [02:10:26] place [02:10:26] and [02:10:27] I [02:10:27] have [02:10:27] never [02:10:28] ever [02:10:28] believed [02:10:29] that [02:10:29] more [02:10:29] than [02:10:29] I [02:10:29] do [02:10:29] today [02:10:30] at [02:10:31] AMD [02:10:31] what [02:10:32] we're [02:10:32] focused [02:10:32] on [02:10:33] is [02:10:33] building [02:10:33] the [02:10:33] technology [02:10:34] the [02:10:35] roadmaps [02:10:35] and [02:10:36] the [02:10:36] partnerships [02:10:36] and [02:10:37] there's [02:10:38] never [02:10:38] been [02:10:38] a [02:10:38] more [02:10:39] exciting [02:10:39] moment [02:10:40] than [02:10:40] today [02:10:40] for [02:10:41] our [02:10:41] 30,000 [02:10:42] plus [02:10:42] engineers [02:10:43] this [02:10:44] is [02:10:44] really [02:10:45] the [02:10:45] next [02:10:45] phase [02:10:45] of [02:10:45] AI [02:10:46] this [02:10:46] is [02:10:46] where [02:10:47] we [02:10:47] make [02:10:48] AI [02:10:48] give [02:10:49] the [02:10:49] AI [02:10:49] the [02:10:50] opportunity [02:10:50] with [02:10:51] all [02:10:51] the [02:10:51] tools [02:10:51] and [02:10:52] all [02:10:52] the [02:10:52] capabilities [02:10:52] to [02:10:53] really [02:10:53] bring [02:10:54] meaningful [02:10:55] real [02:10:55] world [02:10:55] impact [02:10:56] and [02:10:56] I [02:10:56] can [02:10:56] tell [02:10:57] you [02:10:57] I [02:10:57] could [02:10:57] not [02:10:58] be [02:10:58] more [02:10:58] excited [02:10:58] to [02:10:59] build [02:10:59] all [02:10:59] of [02:11:00] that [02:11:00] together [02:11:00] with [02:11:01] you [02:11:01] our [02:11:02] ecosystem [02:11:02] thank [02:11:03] you [02:11:03] so [02:11:03] much [02:11:04] for [02:11:04] joining [02:11:04] us [02:11:04] today