Interview: AMD's Plans for Ryzen AI, CoPilot+, and Boosting its Brand

AI was of course the main theme at Computex this year, with nearly every major brand trying to make the most of the buzz surrounding this emerging technology. It’s still early days and while consumers are hearing a lot about AI, they might not really be sure they need it in their lives. AMD CEO Dr Lisa Su kicked off Computex 2024 with a keynote highlighting the company’s Ryzen AI chips with integrated NPUs for upcoming laptops, along with several other announcements. Shortly after, Acko Drive was able to chat with AMD's Senior Processor Technical Marketing Manager, Donny Woligroski, to get a bit more context and learn more about the company’s strategy.

Acko Drive: The entire industry – AMD, Intel, Qualcomm, Microsoft – seems to have converged around pushing AI tech first on slim and light laptops, which is different from having flagship desktop CPUs coming out first, and then features trickling down from there. Do you have any thoughts on that? 

Donny Woligroski: From a 10,000-foot perspective, this is a really unique time in that you don't often see all the tech providers, the operating system providers, the ecosystem, all working towards the same goal. I have not seen that, ever. And everybody has, I think, a similar vision of where they want AI to be in 10 years and we're all figuring out the steps to take to get there. We're x86 providers, so Microsoft’s vision is important to us. I personally believe that the OS has to be integrated with AI to make the difference we need to make. So I think Copilot+ is a great first step. It's not the end goal, but it's a really good first step. 

In that context, you know, Microsoft, and AMD have a pretty close partnership, and I'm sure they do with their other vendors as well. I think it was a fairly common idea that everyone could use a new compute resource. That's not to say you can't do AI on a GPU or CPU, but it was identified that if you want to do AI while you're doing other things, we should probably have a new resource. AMD was one of the first ones to do that. The neural processing unit makes the most sense on laptops, because it's very low-power and highly efficiently optimised. You can do a lot of [AI processing] on your GPU, but that would use tens of Watts, where we can do it on an NPU in less than one Watt. 

So the NPU truly enables a mobile option, which I don't think you'd have. And so if Microsoft wants to make an OS that has a certain base level of functionality, then you have to have an efficient engine, because the market is moving towards laptops, right? They're a massive market, we want to support them as best we can. It's just the start of the journey, but it's a really important first step.

AMD CEO Dr. Lisa Su shows off a number of Ryzen AI laptop design wins with several major manufacturers during her Computex 2024 keynote.

Acko Drive:Prior to the CoPilot+ announcement, every company seemed to have agreed to call these things AI PCs. So does this new category fragment the market? Does that give you an advantage or disadvantage?

Donny Woligroski: We have our own definition for Ryzen AI, which is NPU plus CPU plus GPU, all of them with AI acceleration. We think they're all important. People are focusing on the NPU just because it's the new thing, I think. When we go forward, they're probably all going to be equally important. Today, GPUs are the most-used AI accelerators in the world, whether it's for training inference, local LLMs, local Stable Diffusion, or what have you. 

Microsoft has a definition of a CoPilot+ PC; some markets won't have CoPilot+, like China, from what I understand. So it's definitely a nuanced situation. But to my mind, the CoPilot+ step is, like I've said, so important. The only way this works is if the OS is integral to it. If I want to live in a world where I say “Laptop, please project wirelessly to the TV to my right” and it does that automatically without any technical interfacing, I think the OS has to be involved. You're dealing with things that could affect every aspect of your user experience. 

Acko Drive: High-end desktops could use their discrete GPUs to do those things, but because of CoPilot+ marketing, you can’t get these experiences on a desktop right now. So do you think NPUs will scale up across the entire product stack?

Donny Woligroski: From what we know today, I think that’s fairly inevitable. I wouldn't be surprised if in five years, you couldn’t find a non-AI PC. But that's not the same as me saying I think you need an NPU today on a desktop. This first iteration of Microsoft’s OS has some interesting new functionality but I think it remains to be answered if that's the functionality that's going to be key to everybody. And if it turns out that, for example, AI-assisted search is a killer app, and everybody loves it, and everybody needs to have it on a desktop, I think you're gonna see us all go “Okay, we'll give it to you faster than we had planned”. It becomes a higher priority. 

Only two Ryzen AI 300-series processor SKUs have been announced so far.

Acko Drive: What would you say to the serious content creators and gamers, who have powerful desktop PCs? What is their upgrade path if AI features are exclusive to laptops with NPUs?

Donny Woligroski: Who knows what Microsoft's going to do! In two years, they might say, “Yeah, you know what, if you have this powerful of a graphics card on a desktop, that's an AI PC too!” I can't speak to that, I don't have that inside information. I do think that the OS is smart enough to apply any workload to whatever resources make the most sense. If you're on mobile, it would probably do it on the NPU. If you're on a high-powered desktop with multiple graphics cards, oh, yes, obviously, let's do some GPU. We're not worried about battery life. Maybe some low-latency workloads make more sense on the CPU. All of these things work together, but the idea is the OS would control it. 

Acko Drive: On that note, just like high on desktops have discrete PCIe graphics cards. Are we looking at a future with PCIe slot-in NPUs, considering that AMD is marketing its lead in shipping NPUs as a key strength?

Donny Woligroski: I think that's absolutely possible. I don't know if that's going to happen, because if it becomes that standard [requirement], I think it will be included on the processor. But maybe there are use cases, like there's integrated graphics versus [scaling to] discrete. I don't think that's part of anybody's roadmap at this point, though.

From my perspective, we're waiting to see how people adopt CoPilot+ PCs and how the software development community starts to support these resources. I feel very confident that every PC will have an NPU in, say, five years. I don't know how that's [going to play out]. We have engineers who are talking about what we need next; is it a bigger NPU, or maybe better AI acceleration on the GPU? Or is there something else? The next year is probably going to guide us a little better for our roadmaps. It's not a clear path at this point; we're all kind of feeling it out.

Acko Drive: So now you've changed the Ryzen mobile CPU naming scheme, and two Ryzen AI model numbers have been announced. Was there feedback from the market that the previous four-digit scheme was confusing? How do you plan to differentiate models with different features or based on older Zen architectures going forward? 

Donny Woligroski: We revamped [the naming scheme] because we felt that it was important to be able to differentiate something that supports what we're calling the next-gen AI PC or CoPilot+. A consumer can just look at the badge and say, “Oh, this is a Ryzen AI badge”. That's the indicator for human beings that this is the right thing. As far as the model numbering scheme goes, it makes sense to simplify, and that was our goal here. We were going to hit the 10,000s; even Ryzen 9000 is getting close! 

As for filling out the rest of the stack, I can't speak to things we haven't announced. Obviously, though, our goal is to address every part of the market. We have chosen to launch at the very peak of the market, as early adopters but of course, we have plans to support other parts of the market as well.

Our intent is to be clear about what supports next-gen AI and what doesn't. That will be a guiding light. We're not looking to yank older architectures into Ryzen AI.

Ryzen AI 300-Series processors, codenamed "Strix Point", feature new CPU, GPU, and NPU microarchitectures.

Acko Drive: What are your thoughts on benchmarking and reporting AI performance? Is anything you specifically want to highlight to help potential buyers get the right picture about AMD and its competitors? Right now, all the conversation is around TOPS [trillion operations per second], does that necessarily resonate with anybody?

Donny Woligroski: People tend to concentrate on tests that are simpler. I think that happens every generation, or every time there's new tech. It's the same as with multi-core CPUs and how Cinebench and other multi-threaded workloads became more important. We're trying to find workloads that will work across different architectures, such as Arm. I bet that within a few months, the industry will conglomerate around a number of benchmarks, or maybe just one or two. 

I think TOPS is not a great metric, I agree. The only reason [the industry is] using that is there's nothing else. And you have the highest number, right? But once people have their hands on [new PCs with AI capabilities], and you're benchmarking, TOPS becomes far less important than quantifying performance in any other way. 

Acko Drive: AMD has had NPU-enabled products in the market for three years now. Are you seeing strong demand for that, and is that a motivating factor to upgrade? And what has AMD done with its first-mover advantage?

Donny Woligroski: I think it's more basic than that. There's a lot of buzz around AI, and people are starting to feel like “What's that going to do for me?”. That supports some of the feedback we've heard from partners who sell directly. Consumers are really interested, and the early adopters will grab them. If there are really compelling use cases, they become publicised pretty quickly. Then I think you're going to start to see demand from regular human beings, not just enthusiasts. Some people need more education even to be comfortable with the idea of AI. 

There's a few things we have done [with the first mover advantage]. It allowed us to get hardware into the hands of software developers to play with it. It allowed us to test for our own use cases to start to see what is useful, and how to best leverage it, how to show it to software partners like Adobe. But it takes a while for them to learn about the hardware and how to use it. The biggest thing it's bought us was expertise in the consumer space. It led us to today, when we have the best solution, which we might not have had. It was a long-term play, not a short term advantage. But I think it'll pay off in the end. 

Acko Drive: Particularly for the Indian market, but also on a global scale, AMD-based laptop market share and brand awareness are pretty low. What are you doing to push your message and get more models into the market? 

Donny Woligroski: That is definitely [something] we're aware of. We've been working on it. We’ve made huge strides from seven years ago [when the Ryzen CPU series was first launched], but there’s a lot of work to be done. We're very cognisant, very proactive, and we are certainly bolstering our partnerships with OEMs. 

I can say that our technical capability has increased OEM interest in supporting the Ryzen AI 300 series. When they saw what it could do, they were like, “Oh, wow, we kind of need a product with this”, which is really says more than anything I can say because these partners have access to everybody's tech. So that's been very encouraging. 

It's harder to get them to the market than you would expect. We're working on it. I don't have any silver bullets to fix that problem. It's just going to take time to do it organically. And we have to consistently deliver really good products for a period of years before we can grow that share. It worked out really well for the desktop market. It's more nuanced for OEM markets; we have more work to do.

Some responses have been slightly condensed and edited for clarity.