Why didn’t AMD talk about their upcoming GPUs in their CES 2025 keynote? Or their new Z-series processors for handheld PCs? What is the Ryzen AI Max? And when can I actually buy a Ryzen 9000X3D chip?
AMD executives only had 45 minutes for their CES 2025 keynote. Fortunately, they also tacked on an additional half-hour or so to field burning questions from a small handful of chip journalists, who crowded around David McAfee, AMD’s corporate vice president and general manager of its Client Channel Business, and Frank Azor, the chief architect of gaming solutions and gaming marketing at AMD.
Fire away, we were told — and we did.
Below, we’ve included a transcript of the conversation, edited for clarity where necessary. Though I don’t identify each reporter by name, they included Paul Alcorn of Tom’s Hardware, Ryan Shrout of Signal65, Marco Chiappetta of Hot Hardware, and myself.
Editor’s Note: The opening of the conversation included “prepared remarks” by the AMD executives, basically explaining that they only had 45 minutes for their keynote and that the company passed over remarks on their RX 9070 graphics cards and RDNA 4 architecture to give a fuller, more comprehensive explanation at a later date.
David McAfee: From a timeline standpoint, it’ll be a little bit later this quarter that we actually begin to roll out RDNA 4 graphics cards. But you know, our focus as we get into this generation is to deliver a really, really compelling value to the end user, with great price-performance.
Lean into all aspects of design efficiency, which is about making it simpler, more cost-effective, more power-effective, to really optimize from silicon all the way through board design so that we can hit the key features that those gamers who are playing enthusiast-class games care about at a price point that they’re going to be really excited to see.
And so I think that we believed, as we built this press conference with the strict time limits, spending five minutes on RDNA 4 was not going to be enough to do it justice. We’ll move it to a separate set of content that comes a little bit later this quarter. Frank, anything to add?
AMD
Frank Azor: We covered everything… There were certain things that we adopted in the press release and we didn’t put into the press conference, like the Z2 [handheld gaming] processor, for example. We tried to include RDNA 4, we really did.
It was going to feed the narrative that we didn’t care about graphics, because we had 45 minutes and we had to rush through. You have to introduce the architecture, all the deltas, all the ray tracing performance, or machine learning performance. Do the positioning of the cards, do FSR [FidelityFX Super Resolution]. Give you a whole overview around FSR, the ISP partners, what’s different about FSR. Show you how FSR works normally. We’d spend 45 minutes to an hour doing that… We started with all this content, and then you’re like, getting it down to the five-minute budget that we had for this.
Journalist: Are you sure it wasn’t because Intel showed half their hand, the Nvidia stuff leaked. You didn’t see some of the competitive offerings and think, maybe it’s not time to talk about this yet?
Azor: All of the above. It isn’t any one thing… It’s not like, oh, we’re just gonna delay because of this. David’s not kidding when he says all the content was in the deck, and part of it is we weren’t doing it justice.
And then you’re like, okay, we’re not gonna do it justice, so the audience is gonna be disappointed. Okay, then you have your competitors making their announcements. Then you have other factors weighing in on this. Okay, so you start looking at all these things, and you put it together, and you say, is it smart or not to include this? It wasn’t smart.
When I say, when we say, later in [the first quarter] we’re going to give you more details, just keep in mind what we just said. We had all the content ready to go. It’s not going to be like March 31.
AMD
McAfee: I’ll also say that I think we also wanted to make sure that our partners at least had enough air cover from us to talk about their products. I think for our board partners, you know this, this is an incredibly important show for them to be able to talk about what they’re doing next gen. And I think they’re all super excited about what’s coming as well.
Journalist: Are you guys cool with partner leaks? Everyone’s going to know the specs just by seeing the cards.
McAfee: I think you’ll see static demos of cards. Everybody loves a good wall of boards. And I think you’ll see that from all of our partners in their spaces. I don’t think you’ll see any live demos, or you better not see any demos from partners — I’ll put it that way. But, you know, power connectors, things like that, TDPs, I think you’ll see a lot of that stuff out there.
Azor: By the way, if you do see demos out there, just know that they don’t have the production [software] driver.
McAfee: All these performance leaks, well, it is accurate for the way the driver performs on the card right now. It is nowhere near where the card will actually perform once we release the full performance driver.
Journalist: Did that also factor into your decision?
Azor: It’s not a readiness issue.
McAfee: We have in house the full performance driver. We intentionally chose to enable partners with a driver which exercises all of the, let’s call it, thermal-mechanical aspects of the card, without really running that risk of leaking performance on critical aspects of the of the product. That’s pretty standard practice.
Journalist: In the past, you’ve emphasized dedicated graphics units on those cards. Now it seems like you’re adopting AI, which is the approach that your competitors are taking. Can you explain that? Why now and what’s shifted in your thinking?
McAfee: I’ll start with that. I think both ray tracing and AI are great examples of that shift. Look a couple of years ago, when the first RTX cards came into the market, there were one or two ray tracing titles and the performance was pretty crummy.
It was more of a technology showcase than it was a real gaming experience. And I think you’ve seen ray tracing change dramatically over the past couple of years. It really has become a much more integral part of so many games today.
Nvidia
I think that ML super resolution is ramping up that same curve pretty quickly as well, where, you know, it’s not for everybody. The purists want every pixel, you know, just brute-force rendered and are not going to be excited about that technology. But it’s also what a lot of gamers are adopting, and I think that especially for those higher resolution gamers that are looking for the combination of high-res gameplay and high frame rates, there’s almost no other way to get there. And I think a lot of gamers are accepting that.
Journalist: And that’s the same approach you’re taking with FSR4?
McAfee: Maybe to talk a little bit about FSR specifically — FSR4 is ML super resolution, and it is built for… as we bring it to market, it will be built for our RDNA 4 architecture. RDNA 4 will bring a pretty massive increase in terms of ML [operations] and compute capability in the shader unit itself. So it is kind of fine-tuned for RDNA 4.
Bringing that to other product families is certainly a possibility for the future, but not something we’re talking about right now, nor committing to a timeline of when that will be available. But as we launch it, it’ll be RDNA 4-focused.
Journalist: My question is on the continuing shortages of 9800X3D parts. This is becoming crazy, and nobody can find these processors pretty much anywhere.
McAfee: What I can say is that we have been ramping our manufacturing capacity — the monthly, quarterly output of X3D parts. That’s 7000X3D as well as 9000X3D. It’s crazy how much we have increased over what we were planning. I will say that the demand that we have seen from 9800X3D and 7800X3D has been unprecedented.
Adam Patrick Murray / Foundry
Azor: Put it this way, we knew we built a great part. We didn’t know the competitor had built such a horrible one. So the demand has been a little bit higher than we had originally forecasted.
McAfee: I think the thing you have to keep in mind is, unlike, you know, building a traditional semiconductor product, it’s basically, you know, 12 to 13 weeks from when you start a wafer to when you get a product out the other end of the machine, and the stacking process adds time to that.
And so it’s longer than a quarter to really ramp, you know, the output of those products, and so we’re working very, very hard to catch up with demand. I think as we go through the first half of this year, you’ll see us continue to increase output of X3D. You know there’s no secret, X3D has become a far more important part of our CPU portfolio than I think we, any of us, would have predicted a year ago.
And I think that trend will continue into the future, and we are ramping capacity to ensure we catch up with that demand as long as consumers want those X3D parts.
Journalist: Is there a gating issue? The silicon or the cache?
McAfee: It is really none of the above. It’s the problem of, you’ve got that lead time to build individually, the CCD wafer, the cache wafer, and then the stacking process that follows adds considerable amount of time as well. It’s a non-trivial timeline when you’re talking about building X3D products.
Journalist: Does the expansion into a broader number of desktop X3D products help or the mobile Fire Range X3D products help? Does it exacerbate the problem or spread the demand around?
McAfee: If I look historically at our 7000X3D products, the 7800X3D was dramatically the highest volume part in that product stack. I think that those 12- and 16-core parts, there are certain types of customers that buy those.
But, you know, I think the reality is, if you are truly a creator, you probably still buy the 9950X. The X3D doesn’t add much at all for creator workloads like, I think, in the data we’ve seen, and you guys will see it as well when you get the chance to review it — it’s like a percent of incremental performance by having the X3D there.
AMD
Because the truth is, X3D is great for games, because it’s all about improving effective memory latency. When it comes to creator apps, it’s about memory capacity or memory bandwidth. The X3D really doesn’t affect either of those parameters in a significant way. So my belief is, in the 9000 series, those higher core count products, there’ll be some demand there, but it’ll still be ten-to-one or more on the eight-core X3D parts because they’re just such a great gaming part. For a pure gamer, there’s nothing else like it.
Journalist: With your APUs, you traditionally have brought them over to a desktop socket eventually. What about the Ryzen AI Max?
McAfee: It would be a different socket, first and foremost. I guess we haven’t released all the details about Ryzen AI Max, so I better be careful what I say. It will not fit in an AM5 socket. I don’t think mechanically it fits in an AM5 socket.
Azor: But you will see one desktop, as we announced.
McAfee: Yeah, there’ll be small-form-factor desktops that are, you know, BGA, soldered-down versions of that product, socketed desktop. I think that’s a much harder problem to solve because there simply isn’t an infrastructure that it drops into. And I think as we disclose more about that product to you guys, you’ll kind of see why, but it’s a little bit tricky from that perspective.
Journalist: I’m getting a lot of questions about what type of memory it takes to support 256GB/s. Is that DDR5X?
McAfee: I’m going to stretch here a little bit. It is certainly an LP5, LPDDR5X product. What I can’t recall is whether that also supports other memory types. And I’m just kind of drawing [a blank], I think it’s the only one. I think it’s only an LP memory interface.
[PR person]: When you think of Strix [Point], it has the bank of 128 right on one side. Strix Halo has two of those banks. That’s one on each side, and that gives you the 256. So I think it’s 4-by-8 or 4-by-16 or however they do the DDR5 part, but it’s two banks now instead of one.
AMD
Journalist: I’d like to hear a bit more about the origins of the Strix Halo (AI Max) part. It seems a bit like Threadripper, where you just went for it with a massive core count, massive threads. Strix Halo seems to pull together the best parts of the CPU, GPU, and NPU in a part that’s optimized for AI.
McAfee: I think with Strix Halo, one of the things that we saw is the way that notebook products with discrete GPUs are built today is really suboptimized. And when you look at that from a performance-power curve across the entire spectrum versus a fully integrated APU design, the discrete design leaves a lot on the table. You’ve got two separate memory subsystems to maintain, just a lot of complexities, power management, all of those things.
Azor: Where’s the cost? Because you’re paying for two memories that don’t get along with one another.
McAfee: So now, by unifying that, it provides the opportunity to deliver a dramatically different notebook experience that scales into what a gaming notebook is capable of doing.
Like that proof point… That was Llama (LLM) performance versus a [GeForce] 4090. It’s a testament to the massive memory footprint that you can put close to the product. To effectively use a 96 MB frame buffer for a GPU is kind of groundbreaking. It opens up a lot of opportunity for our notebook partners and desktop partners, small-form-factor desktop partners, to innovate in ways that they haven’t been able to do in the past with a traditional APU plus dGPU.
In our minds, when it comes to building a great gaming notebook or mobile workstation or small-form-factor desktop — all that power management tech, all that power scalability — I think it points to what a great gaming notebook of the future could be.
I think that this is, this is AMD, kind of creating a category of product here that hasn’t existed in this way in the past. And we think it’s really a unique way to solve that problem that has a lot of benefits.
Journalist: So this is not a one off [one-time product]?
Azor: We are not ready to announce any product, so we are not ready to talk about the future.
Journalist: It would seem like having different GPUs and not needing a discrete GPU would free you from needing to worry about aggressive business practices from others in that space.
McAfee: What I would say about Strix Halo is, I would not expect Strix Halo to be a dominant volume product for us in 2025.
I think this will be a year where we see, you know, some of the initial designs from HP. We see other partners bring additional platforms on board later in in the year. Honestly, the world has been used to APUs plus dGPUs for a really long time. And changing that mindset, educating the consumer, getting the market adoption is going to take time. I don’t know that it changes anything in a massive way in the short term.
You’re totally right though, that by integrating the two together, it goes beyond, I think, even some of the original goals.
Journalist: I have a question out of left field. Nvidia’s GeForce 4090 is so powerful it’s banned from export to China. You guys haven’t put up performance specifications, but do you think it’s powerful enough to be banned from export to China, too?
Azor: I’m not going to address that. But keep in mind with the 4090 — it has 24 GB of memory, right? And you need the entire LLM loaded into memory. The advantage is the memory architecture.
Thiago Trevisan / IDG
McAfee: It has nothing to do with compute.
What I will say, though, is, I am not the expert on, you know, the regulations with the 4090, and its export control laws. I think that the truth, though, is Ryan’s point is right on that, you know. It is about total compute that drives a lot of those regulations, and this product does not match up, from a total compute standpoint, to what a 4090 can do.
Journalist: Do you have a sense in mind of, like, up to what class of GPU you think this displaces?
Azor: So there’s a pretty broad range, if you notice, it was 55 to 120 watts. So there’s a range. Do we? Are we ready to communicate that? … Sure? I don’t know if we’re ready.
[Edited for space.]
Azor: We’re going to scale better with the shared memory than others. It’s really hard to give you a blanket [statement like] “This is going to perform like this from Nvidia” because you’re going to see a lot of differences that you haven’t seen in the past. It’s not a discrete GPU and it’s not a dGPU one-to-one replacement. It’s a different beast, a different animal. So, yeah, it will perform like discrete graphics, but you’re going to see a broad range.
Journalist: So should an enthusiast say that “this crazy amount of memory is going to give me great performance in games?”
Azor: For some games, yeah, some games have large textures, and they load them into memory, and they’ll get a lot of advantage of that size and that speed. You know, it’s not one size fits all. Sports games are probably not going to scale that much.
Journalist: So you’ve put X3D in the gaming class category. You’ve put the AI Max in the gaming class category. Any characterization about which one would be better suited for gaming?
McAfee: I would say that if you are a pure gamer, the… Fire Range 9955HX3D is like that, without a doubt, going to give you the best gaming performance across the board, because of two reasons. Number one, it is a desktop product in a mobile platform, and you’re likely pairing that with a very, very high-end discrete GPU. The AI Max, Max Plus…
Journalist: What’s the difference between the AI Max and AI Max Plus?
McAfee: The Plus makes it better.
Journalist: I walked into that one.
McAfee: What I would say is that the [AI Max] product does not cover the full scale of gaming notebooks, right there. Gaming notebooks that today are built with desktop products in a mobile package with like a 4090, mobile dGPU, that is not the target of the AI Max family. The AI Max family is more mainstream gaming notebooks, mobile workstations. That class of product is the sweet spot that it’s really built for.
Azor: Today, you have to make a sacrifice: If you want a gaming laptop and you want to experiment with Copilot+ and get all the latest features you really don’t have, you get a Strix Halo plus a discrete GPU. What’s unique about Strix Halo is you get everything.
Mark Hachman / IDG
You get an NPU if you want to mess around with that, either as a developer or as a consumer. You get discrete graphics-level performance in an integrated package. And you get an amazing CPU. So it’s kind of like the best of everything, if you want to be at the cutting edge of everything. You don’t know how big that market is. We’ll see in time.
Journalist: Can you say anything about your price targets there?
Azor: That’s really for our OEMs to share. What I’ll say is, you’re going to have ultra small-form-factor desktops like the one that we announced that will probably be at more compelling price points relative to what their dGPU counterparts are going to be, and they’ll offer some unique value propositions like efficiency and size and weight that they can’t compare to on the laptop side.
This is not a value part. This is not a low-end part. This is an ultra-premium part, cutting edge, the most advanced x86 processor arguably out there. They’re not putting it in cheap notebooks. They’re putting them in their best, most premium laptops. But that doesn’t necessarily mean that it’s driven by the part itself. It’s the whole decision that they’re making.
McAfee: I think you will see some application of this in a gaming notebook. I think you will see an equal amount in products that are positioned more as mobile workstation products, because it is really a killer part for that.
Journalist: The HX3D last time was very hard to obtain. If I’m right, there was an exclusivity agreement. Are you guys opening it up to a broader range of partners this time around?
McAfee: What I will say is, I don’t know anything about commercial agreements with that product and customers. It is similar to the prior generation in that it is drop in, pin-compatible, platform-compatible. So any system manufacturer today that’s building with the prior generation, the Zen 4 generation, has a really easy path to upgrade to a Zen 5-based notebook solution.
You know, it’s a pretty small segment of the market, but I don’t know of any reason that there’s any constraints around that particular product.
Journalist: How do you see GPU development moving forward at AMD? Do you see an equal amount of effort put into traditional raster versus AI? Do you see the AI and generated pixels being more important than rasterization for the future? What’s going to be your philosophy in terms of evolving AMD graphics?
McAfee: I think that there’s a pretty fundamental shift that’s happening in the graphics industry right now.
AMD
Moore’s Law is probably a bad analogy to use, but the amount of brute-force rasterization performance improvements that I think we see, and the competition as well, is a fairly muted curve, right? You’re reaching some of these boundaries around rasterization performance that require massive increases in silicon to provide meaningful uplift there.
And I think that you’re also seeing areas where using AI-generated pixels, or AI-enhanced or whatever you want to call it, provides much, much more rapid gains in terms of improving quality and frame rates and generational leaps in performance. I think the way we think about it is in the future, you know, yes, I think it’s absolutely true that you’re going to see more generated pixels, both from us as well as everybody that’s that’s in the graphic space, because it is the path forward in driving more performance and more immersion and better experiences for end users.
We haven’t talked about this a lot, but you know, the journey that we’re on at this moment with RDNA 4 is kind of a new starting point for us, right? Building a reputation for Radeon, just like we had for Ryzen in the early years, where it’s focused on delivering better value for the end user, giving consumers more for their money than what the other guys are delivering, focusing on the features and capabilities that gamers care about most, and being sort of responsive and focused on delivering for the community of enthusiasts that’s out there, is really what we believe we have to build with Radeon. And that’s kind of the focus of the roadmap going forward.
I think as you look beyond this generation, we want to continue to advance that. I think advancing that on all fronts for graphics is going to require — and be driven by — more inclusion of generated pixels to drive a better experience.
Azor: One of the hardest things that our product management teams have to do when it comes to defining our graphics cards is timing the hardware to the ISV ecosystem, and to what consumers are going to be demanding in that time frame in which the cards are going be coming out.
Had we been the first to lean into ray tracing even more than our competition, we would have been going out and asking a bunch of customers to spend a significant amount of money on a technology that they probably weren’t going to be able to capitalize on until two-to-three generations later. And just keep in mind, we’re not the dominant market share provider in the industry. The dominant market share provider in the industry did that, and you really didn’t start to see meaningful gain of ray-tracing games until this last generation, three generations after they first introduced it, and they’re the market share leaders.
So we have to be very careful on why. You see, sometimes people say, “Why are you always trailing?” Well, we’re trailing because we’re following the [Total Available Market] of where the market is, and we’re letting them create some of this market because they are the only ones that really can when you have the kind of position that they have in the industry. We have to time it.
We either have to give you less, somewhere else — so, compromises — or we’d have to raise the price points, which is something they are already doing. So why have two people do exactly the same thing, trying to build these leadership products out there? Which is part of what is different about our graphic strategy moving forward than maybe what we’ve tried to do in RDNA 2 and RDNA 3.
And what you’ll see with RDNA 4 is, it’s much more of a gamer-first design, all about efficiency, all about giving them the feature set for what’s going to matter in this next generation of games.
Can I tell you that in the future, every pixel is going to be ML [machine learning]-generated? Yes, absolutely. It will be in the future. But when that future is? And should we charge a gamer for that technology today so that they can be the seed for that in the future? Those are all the tough decisions we have to face.
Right now, we believe, is the right time for more prevalent ray tracing and to make that investment. And because FSR4 is looking as good as it does, and it is a very efficient way to generate a pixel that is arguably as good as a native pixel, we made the hardware investment to be able to enable that in RDNA 4.
Journalist: When you say you are targeting the sweet spot, not going after the ultra high end, what are we talking? Sub-$1,000? Sub-$700?
Azor: Sub-$1,000 for sure.
McAfee: Significantly below $2,000.
Azor: That’s the focus on RDNA 4. Focus on what gamers actually care about.
Journalist: So do you have one message that you want people coming out of this room to take away about RDNA 4 and Radeon and why it wasn’t in the keynote?
McAfee: I think the biggest thing that I would say is, graphics products and graphics launches are complicated, and there’s a lot that has to be explained to really do it justice. And we wanted to make sure as we launch our RDNA 4, we do it justice, and we cover the hardware improvements, the technology, the software, the FSR, the driver enhancements, like all of that needs to be covered to really satisfy what gaming enthusiasts care about.
Azor: There was a no-win scenario that, at the end of the day, we debated all these different options that our customers, the gamers, the market were going to walk away from today. Had we included it in there for four or five or eight minutes, would they be like, “Wow, that was amazing. I was blown away by it”? No. So why do that?
Let’s give it its proper time. Let’s give it time and let’s win. Let’s put together a recipe it’s going to win — positioning, performance, pricing, the time that it deserves — and show people that we actually do care about gaming and not feed the narrative that “Here’s five minutes in an AI PC keynote. They don’t care about gaming.”