Planet 3DNow!

- The german-speaking online magazine for the AMD user since 1999 - 10th biggest Distributed Computing team worldwide

[Jack Huynh @jackhuynh]

✨ Personal AI is the next computing platform.

AI is shifting from something you access to something you build with, locally, at the edge, and across systems.

We’re unlocking new possibilities for developers:

• @AMD Ryzen AI Halo, a local-first developer system, preorder starting in June, develop AI without limits on your desk

• Gorgon Halo with up to 192GB unified memory, supporting 300B+ parameter models locally, run massive models locally

We’re excited to partner with @ClementDelangue🤗, Co-founder and CEO of @huggingface, to advance open-source AI for Ryzen AI.

Our focus is seamless AI, from model to deployment.

Cloud, edge, device. One continuum.

Multi-agent systems, local inference at scale, open models as infrastructure.

This is the next computing era 🚀

[Planet 3DNow! @planet3dnow]

@CatalystMaker Radeon RX 9070 and 9070 XT still have 25.12.1. as latest driver.

[TechEpiphany @TechEpiphanyYT]

$AMD Transcipt from AMD Goldman Sachs Communacopia + Technology Conference 08-Sep-2025

1/2

Okay. Good afternoon, everybody. Welcome to the Goldman Sachs Communacopia Technology Conference. My  name is Jim Schneider. I'm a SIEM Category Analyst here at Goldman Sachs. It's my pleasure to welcome AMD  and the EVP of Datacenter Solutions, Forrest Norrod with us today.

Thanks a lot. Thanks for being here.

QUESTION AND ANSWER SECTION

Q: Thank you. Forrest, maybe I want to start a very big picture and talk about AI as a broad topic. In contrast, the  main speakers here at the conference who are coming at AI from the perspective of applications, you're coming  up from the perspective of infrastructure. What's your high level vision for where AI's going as technology? Why  does the world need it? How useful do you think it's going to be? And do you really think the reality is going to live  up to the level of investment we're seeing today?

A: Well, I think the story is still in the very early innings, but the indications are super positive. AMD has been  honored to be in the discussion with many of the leaders in AI model development for quite a few years. And so  we've had a bit of a good perspective to see the development of the technology and the application in a very early  sense. And so we have -- we've seen it and we've used it ourselves for both business process as well as  engineering development.

I would say our assessment is it's still in its infancy, but super positive. I mean, we're seeing on the software side,  we're seeing some pretty substantial improvements in productivity as well as time to develop code for both  software as well as verification tasks. And increasingly on the hardware design side, we're using it for chip  development as well. So we're seeing, all of the right early indications to say, look, this is going to develop or  deliver sorry, real business value. And I think that's the fundamental question. If it does and we feel confident that  it will. This is going to be a hugely transformative technology.

Q: In terms of the use cases, I think, you as a company have stated that consumer AI applications are well ahead of  enterprise. I think that's borne out by a lot of the data points we see in the market. Now, what consumer use  cases do you see in the industry today that fit you most, in terms of both utility and monetization?

A: I think on the consumer side, I actually am going to pivot that a little bit because I think we're increasingly seeing  more indication of traction on the business side as well. Again, we are an engineering company. And so the thing  that gets us most excited is seeing a productivity enhancements flowing from use of AI as part of the engineering  process. And we're starting to see that really materialize in a major way.

And so I, I think that, the -- look consumer side is still going to be a fascinating area and everybody likes to use  their chat bot, but where this is really going to change the world is, can we change business and development  processes? And that's where I think, we're much more excited at this point.

Q: Yeah. How do you think about some of the potential bottlenecks for AI deployment? Do you think raw computing  power, networking power is still kind of the limiting the pace of AI software deployments? And if so, like, which is  the more of a limiting factor on computer networking?

A: I think, that one of the most interesting challenges from a computer architecture point of view right now is, AI is by  its nature, a very distributed problem, distributed in terms of inside of the GPU and increasingly distributed across  many GPUs and very large systems, particularly we can get to agenting cases. Deploying AI systems means  really deploying a number of different workloads, a number of different models supported by other applications  across a very large network computer. And so that is a -- it's an interesting computer science problem.

Certainly networking and communicating efficiently and effectively across these resources is perhaps increasingly  a large part of this, because slight inefficiencies in distributing the problem and networking and coalescing results  can make major impacts on the effectiveness and the efficiency of the deployment. So we do think the importance  of networking, the importance of distributed systems from a software perspective as well, or going to be dominant  factors in performance of systems going forward.

Q: So I want to pivot to AMD's progress business for a second. I think firstly you've done a great job of ramping your  GPU franchise over the past couple of years, going from very little sales to about $7 billion this year. If the speed  is correct. What are the things that have worked out well for you so far or best so far and what areas have you  seen maybe slower than expected progress?

A: Well, I think look, if you take a look at the way that we've approached the GPU market in many ways similar to  what we did on the CPU side, we took a strategy of building a multi-generational phased approach to gradually  build up the competitiveness and differentiation of our solutions over multiple generations, thinking about how do  we systematically get more and more competitive and take leadership positions in a larger number of workloads.

We started, on MI -- in at least in the MI300 generation, we started with inference. We said, hey look, we're going  to -- we've got to because we're chiplet architecture, we have the ability to have more memory than our  competitor at the time. That translates into more efficient inference, particularly as inference skills out, we're going  to -- go after inference leadership in MI300, MI324 will build out our software ecosystem to make sure that we're  making that capability as accessible, as fast as we can. But then we will systematically build out training capability  in MI355, both on the silicon as well as the software side. And it's all culminating in our MI450 generation, which  we're launching next year, where that is for us our nuestras generation where we believe we are targeting, having  leadership performance across the board, any sort of AI workload via training our inference.

And so everything that we've been doing is been focused on the hardware and the software and increasingly now  at the system level and cluster level as well to build out that capability. So it all intersects and MI450 is perhaps  akin to our Milan moment for people that are familiar with our EPYC roadmap. the third generation of EPYC CPUs  is the one where we targeted having no excuses. It was superior. Rome and Naples were very good chips and  they were highly performing and the best possible solution for some workloads. But Milan is where it was the best  CPU for any x86 workload [indiscernible] (00:28:22) stuff. We're trying to view and plan for MI450 to be the same.  It will be, we believe and we are planning for it to be the best training inference, distributed inference,  reinforcement learning solution available on the market.

Q: Interesting. And so, if you look back and reflect over the past, these lessons you've had over the last 12 to 18  months, has it set up your AMD for success for the next two to three years, let's say. And one of the most  underappreciated points of the AMD's progress, in your opinion?

A: Well, I think, again, we've tried to take a very systematic and thoughtful approach to the problem. Again, gradually  building up our capabilities and making sure that we're delivering value in each step along the way. So, again,  first, getting good at inference and then building up our capability to allow customers to begin training with us. And  then again, all culminating with bringing it all together in the MI450 generation.

I think one of the reasons that we did that is we also acknowledge that I mean, quite candidly, NVIDIA is a  fantastic company. They've done a fantastic job and they were well ahead. And so we had to catch up. And we  also knew that in that time of catching up the season over the last couple of years or last three or four years has  been the most important thing is for the big model companies to get to train the next, each one of them to train  their next frontier model, to get the next level of capability the fastest. And so that's driven most of the industry for  the past few years. And I think that's, again, driven the NVIDIA successes. They had the most mature ecosystem  and they were the -- they are the fastest time to train promise.

We decided -- we with this multi-generational roadmap put the objective in place of okay we are -- when we get to  MI450, we're going to be there the same time as when Vera Rubin was intended to be there. And we're going to be there with that part that's fully performant. The software stack that's fully there, at least for the 80% of the  market that's constituted by the top 20% or so customers.

And so, we've focused on getting there in the MI450 so that for training there's no excuses and then for or there's  no impediment there's no hesitation of, hey, if I'm training, I'll be behind in this generation if I go with AMD. And  that's been the learning for us and that's been the realization. MI300, MI325, MI355 good for inference, a little bit  behind in terms of time to introduction for training. And so that's been the thing that I think has slowed us down on  the training progression . We recognize that fairly early.

Q: Maybe just thinking broadly, your broad kind of remit across data centers. So how do you think about AMD's total  market opportunity in data centers on both the CPU and GPU side? Is there a specific market share? You think  you have the, you know, the right to win, if you will? And what's the threshold of market share you would find kind  of either encouraging or disappointing on the other hand?

A: Yeah, you actually hit a source point with me. There's no such thing as a market share that it's sort of right to win  that. And that's something, if I ever hear that from our internal teams, I say extinguish that from your we have no  fear to share market...

Q: I'll say after in future.

A: Yeah, yeah. because at the end of the day, look customers are going to or are going to buy the best possible  product to meet all of their needs. And if we're not offering that to them, we don't have any right to any portion of  the market. What we've done on the CPU side is, come out, I think, with a compelling roadmap, work very closely  with our customers over time. And we've gotten the most recent quarter reporting the mercury. We're at 41%  share on server CPUs up from essentially zero, when we started this...

Q: Yeah.

A: ..this journey about seven years ago. And our share continues to grow there very rapidly on the CPU. So we've  picked up about 8 points of share in the last 12 months. And if anything, it's picking up speed.

I think on the CPU side, the strength of our roadmap is such the level of our customer engagements, both with the  cloud customers as well as the broad and enterprise customers continues to improve. And I think our share will  continue to grow there. I'm very confident that our share will continue to grow quite strongly on the CPU side, and  we aspire to absolute server CPU leadership in the relatively short period of time.

On the GPU side, look where we again, we've built this roadmap not just at the GPU level, but really at the  solution level with with the right CPU and networking matched to the GPU that we think will deliver not just  performance but a compelling TCO value for the customers. And we aspire with that roadmap to be a meaningful  portion of the market. What that means is, I think if you're not in strongly into the double digit percentage, it's a  20% of the market, you're not meaningful -- you're not a meaningful player. And we certainly aspire to get to be a  meaningful player as an intermediate step and then, of course, continue to grow over time.

Q: Fair enough. Now, you previously shared that's company 2028 AI accelerate TAM of $500 billion. Help us draw a  line from where we are today to that future point in time from a market standpoint. Iis that a sort of a straight line,  is it something that accelerates over time? How do you think about how the market TAM evolves? And is there  anything that really needs to happen in terms of technology, monetization or anything else for that to happen?  And I mean, is that number even now too low?

A: Well, I think, our lead customers continue. And you everyone sees this in the Hyperscalers Capital plans as well.  They continue to be extremely bullish on the long term prospects of AI. And when we first articulated the $500  billion TAM number well over a year ago, I think we got a lot of, raised eyebrows and questions about it. I think it's  much less question today.

Q: Yeah.

A: And again, it's because people are seeing the early results. Now, in the end, if business value does not get  realized by end customers from all of this technologies and this growth rate is going to slow down. But we see  enough evidence that that business value is there, that we're pretty optimistic that this is going to continue to grow  at a pretty rapid pace. And the pace right now, quite candidly, is modulated more by data center and power  availability than anything else.

Q: Yeah. I think it's fair from I think a year or two ago. Now, speaking with investors, I think, probably the most  debated number for AMD is your data center GPU revenue for next year in 2026. Can you maybe help us think  about what the key growth drivers are for that business? And to the extent your visibility on that? What needs to  happen for you to capture your desired goal in terms of revenue scale?

A: Yeah, I think the key for us is, we've obviously just introduced the MI355 a few months ago, but we are  anticipating material revenue from the MI450, which we'll be launching in about a year from now in next year. So  we are expecting to see material contribution there. What's going to drive that? It's really, continuing to work very  closely with our end customers on preparing for their deployments of MI450. We're getting a lot of extraordinarily  positive response from our customers right now.

So, you heard from some of them that are advancing AI day back in June, you heard Sam Altman from OpenAI,  get up and talk about the very close partnership and feedback that they've been providing to us for the last few  years and their excitement over the MI450. You've heard the same thing from Oracle and a few others as well.  We are deeply engaged with quite a number of end customers on ensuring that as we wrap up the design -- the  validation now of the MI450 and the supporting rack level and cluster level infrastructure, that we move that  smoothly through the rest of the validation that we get it ready for production them and then we ramp it efficiently  and effectively into production with them.

One of the things that we've really spent a lot of time and attention to is making sure that the rack level solution  will move to market smoothly with a minimum of hiccups. And so we began -- we have I hope folks will give us  some credit for being very predictable in our execution on the data center side. I think we've got a good track  record of doing what we've said we will do for the last six or seven years. And that's really from where you come  from a very rigorous development process that identifies risks and then takes them down in a very systematic  way.

So a couple of years ago, as we were looking at the MI450, one of the obvious risks was this shifting from  delivering chips to literally having a rack level infrastructure. And so we very quickly decided that to substantially  bolster our capabilities, system level capabilities, we contracted with ZT Systems and then we brought them on  board to begin doing the development of what became our Helio's rack level design, over two years ago. And then  we over the last two years been very systematic in building up the design, proving out subsystem by subsystem,  building out electrical, mechanical signaling, cabling, power, et cetera, subassemblies, prototyping them, proving  them out and getting the whole system ready for production. We've also made some interesting choices. So I  think, specifically to de-risk the design.

If you look at Helios, it's very thoughtfully designed to be as compatible as possible, the data center level with  alternatives that a customer might have. So things like making sure that the ratio of air cooling to liquid cooling  within the rack is equivalent to or similar to NVIDIA. So that customers can build data centers with the right  number of chillers, that's a 18 month lead time items. If we require a substantially different number of chillers for  per 100 megawatts than NVIDIA does, this from customer has to make a decision maybe earlier than they're  willing to make a decision on AMD. So we've worked through that and then we very systematically work through  all of the signal integrity of the cabling. A lot of the issues that we knew from our experience doing the  supercomputers with HP, we designed 0.5 megawatt cabinet systems, years ago with HP, and we learned a lot of  lessons there.

And so if you look at Helios, for example, it's actually larger than an NVL72 rack. It still has that same pod size, 72  GPUs per pod, but it's bigger physically, which is not an issue for our customers because the physical space is  inconsequential, but it's bigger and it's easier -- because of that it's easier to manufacture, it's easier to support,  it's easier to service. And we believe it will be more reliable than a device that is been more focused on density for  density sake.

Q: Interesting. So if you think about that you mentioned a couple of them, the biggest challenges to kind of attaining  the revenue profile that you desire in 2026, what are the risks or the challenges that you, I mean, is it still things  like, software stack? Do you feel like it's customer enablement and do you feel like these things you mentioned,  whether that be full rack solutions or cooling, are now relatively de-minimis risks from a technology standpoint?

A: Well, I think they're all I mean, we're paying very close attention to a long list of items. I think we've got a very  rigorous -- again, a very rigorous de-risking plan in place, development and validation plan in place. I mean,  obviously, yeah, I mean, it is a very complex rack level solution. There is mechanical -- potentially mechanical  issues, potential signal integrity issues, potential thermal issues. And so we're trying to pay attention to all of  those. And I think we've got them all pretty well in hand.

As well on the software side, particularly for the lead customers, the 20 or so customers that really matter that are  going to drive the overwhelming proportion of the capital investment. We've been working very closely with them  to make sure that the software that they require is going to be there in time.

Now, maybe not the full long tail, we're I mean, NVIDIA has done a great job investing in AI for many, many years.  And they've got support for a very long tail of customers. We're not going to be able to quickly match that, but  we're not trying. We're trying to make sure that we are fully there at MI454. The customers that really matter for  the 80%, 85% of the market.

Q: Yeah. Makes sense. Maybe talk about your progress with your sort of new prospects in terms of the the US  Hyperscalers and other customers who are not yet your customers at this point, what key obstacles do you see  from a customer perspective to them adopting a solution today?

A: Well, I think, so fortunately, actually, unlike when we started with the CPU side, all of the major customers, every  one of those ones that we just talked about is already an AMD customer and we've already got a data center  engagement with them. So we've got familiarity with them. They understand us. They're generally all using us on  the CPU side. So we've got a pre-existing relationship in that which is helpful. We're not trying to build that up as  we were at the beginning of the CPU side.

But we've been fortunate enough to have some great relationships on the MI side, on the Instinct side with several  of the major hyperscalers, obviously Microsoft, Oracle, Meta are the ones that are most prominent and public,  OpenAI, of course, as a user. But we've been engaged as well with several others. And I think that you will see in  our next even partially on MI355 and then certainly with MI450, you're going to see the aperture of the end  customers and the hyperscalers open up quite a bit. And so that's based on the work that we're doing with them  right now and the very close collaboration feedback that we're getting from.

Q: And so would you say that's kind of giving you that is that the software piece is kind of giving you there more to  like be to widen aperture of customers not getting to the very longest tail, but kind of certainly widening beyond  the initial set of target customers you had last generation?

A: Yeah, no, absolutely. And so we've been fairly systematic about building out the support for the different  frameworks, the libraries, the various open source projects that are relevant again to these customers. And  something like Jack, for example, Jack support is very important to a number of these customers or Jack's  support was relatively mature a year ago. It's come a long way. And again, we're trying to be systematic about  being fully complete for this targeted set of customers on the software side.

Q: And as you think about your product roadmap going forward, what's kind of driven your confidence to move to this  sort of annual cadence in such a competitive environment? And I guess how are customers helping you prioritize  your order and say your product roadmap today?

A: Well, I think the industry given the excitement and the level of change in innovation, you know, that's that's what's  really driving this this annual cadence, by the way. It's painful for the industry. It's painful for end customers...

Q: Yeah.

A: ...to take products at this level of cadence. And so but it's a competitive imperative. And so one of the things that  we're trying to be thoughtful about doing going forward and we did even on the MI355 to the MI300 generation is  try to make sure that there is commonality in the infrastructure -- commonality in reusing the infrastructure. So it's not a complete rip and redo with every new generation such that we're containing the the change to the things  that that matter, that give performance. But it's it makes it a little easier for the data centers and the customers to  adopt. It's not quite the old tick tock, you know, strategy that we used to have on the server CPU side. But it is  trying to be thoughtful about we've got to maintain this rapid pace to be competitive. How do we make it as easy  as possible on our customers to accommodate consumption of this technology on that pace.

Q: I mean, from a philosophically from a gross margin perspective, how do you think about pricing to the value you're  providing as you continue to sort of up the game on each new generation of technology? In other words, if the raw  performance of MI355 is kind of on par with Blackwell and MI400 series is going to be on par with Rubin. Should  we expect more pricing power and improving gross margins within sort of the data center GPU space for AMD  going forward?

[STH @ServeTheHome]

AMD goes into RDNA 4 at Hot Chips 2025 https://t.co/TuDhYjpQ5E

[AI at AMD @AIatAMD]

ROCm on Windows? Yes, with WSL.

AI developers can now run AMD ROCm on Windows, unlocking easier access to AMD GPUs for AI training, inference, and experimentation — no dual booting required.

Get started with @ComfyUI: https://t.co/szdKyggiS8

[QuakeCon @QuakeCon]

@AMDGaming + @FrameworkPuter = a small, but mighty modular desktop PC 😤

[Framework @FrameworkPuter]

The full video is at https://t.co/KXJGdVX1Fv

[Jeff Geerling @geerlingguy]

I tested four Framework Desktop Mainboards with AMD Ryzen AI Max+ 395 (what a mouthful!) in a private AI cluster.

[Phoronix @phoronix]

The @FrameworkPuter powered by @AMDRyzen Strix Halo is here! Framework Desktop With @AMD Ryzen AI Max Offers Excellent, Linux-Friendly Performance Review

https://t.co/bbmNul2n8C

[Chips and Cheese @ChipsandCheese9]

Hello you fine Internet folks,

Today's article is about running and profiling gaming workloads on AMD's Zen 5 and looking at what we can ascertain about how certain gaming workloads behave on Zen 5.

Hope y'all enjoy!

https://t.co/DswBtxRE8s

https://t.co/J9BvBy5YxH

[AMD @AMD]

Upgrade alert 🚨
Download AMD Software: Adrenalin™ Edition Preview Driver (mainstream support coming soon in 25.8.1 WHQL) today to unlock support for 128B parameters locally in Windows on the AMD Ryzen AI Max+ 395 128GB processor and @lmstudio, the world’s first windows AI PC processor to run @AIatMeta's Llama 4 Scout.

New in this upgrade:

• AMD Variable Graphics Memory - now enables up to 128B parameter support in Vulkan llama.cpp on Windows
• Make full use of the 96GB VGM available on a Ryzen™ AI Max+ 395 128GB machine.
• Available now from @HP, @ASUS, @FrameworkPuter, @CORSAIR.

Learn more: https://t.co/pv1lxu5qLu

[AMD GPUOpen @GPUOpen]

Our #opensource AMD Interactive Streaming SDK has been updated to support 🐧#Linux.

🎮 #FOSS developers can now build custom low-latency streaming solutions for cloud gaming, VDI, and embedded apps on AMD GPUs.

📥Download the update: https://t.co/qUxWUOpr7b

[Chips and Cheese @ChipsandCheese9]

Hello you fine Internet folks,

Before AMD's Infinity Fabric, AMD's interconnect for high core count CPUs was HyperTransport (HT). Today's article is a retrospective look at the HT used in the Magny Cours CPUs.

Hope y'all enjoy!

https://t.co/VWGb8j6aMc

https://t.co/BpVeGkkMEN

[Planet 3DNow! @planet3dnow]

@CatalystMaker https://t.co/AyPZcN6TAF

[Semiconductor News by Dylan Martin @DylanOnChips]

AMD said Tuesday it has entered a strategic alliance with global systems integrator HCLTech to deliver advanced AI and cloud solutions for enterprise customers. https://t.co/7z005actNn

[Chips and Cheese @ChipsandCheese9]

Hello you fine Internet folks,

At AMD's Advancing AI 2025, I had the pleasure of interviewing Alan Smith, AMD Sr. Fellow and Chief Instinct Architect, about CDNA4 and the MI350.

Hope y'all enjoy!

https://t.co/LfknHwaRhj

https://t.co/oY79AimcQ1

https://t.co/gsJVbq0dtF

[AMD GPUOpen @GPUOpen]

We've just released the latest #FSR 3.1.4 in our updated plugin for #UnrealEngine 5, ready to drop into your game! 🎮

🔥 Easily get your game ready for FSR 4
🔥 New upscaler CVars and multiple bug fixes

📥 Download now! https://t.co/fDt1Hv0xiS

[Lisa Su @LisaSu]

So proud to continue our deep partnership with @Xbox @XboxP3 @BondSarah_Bond to bring the next generation gaming experiences to gamers everywhere!!!!! 😀

[Chips and Cheese @ChipsandCheese9]

Hello you fine Internet folks,

Today's article is on AMD's CDNA4 architecture that will be found in the MI350 series.

Hope y'all enjoy!

P.S. We also have an interview with Alan Smith coming up, so look forward to that!

https://t.co/ESKkQmi0jJ

https://t.co/ttNyTKp0pO