Bloomberg Audio Studios, podcasts, radio, news.
Welcome back to Bloomberg Tech. Lisa Su, Helios with MI455X, AMD's first rack scale system solution, but inside it, AMD's first and the world's first 2nm chip of that type. A lot was made of it when you actually just stood on stage and held it in your hand for the first time. Why is it significant?
Well, first of all, Ed, it's great to be here with you at CES. I think CES is always a great way to kick off the year because you get so much perspective. So, it was fun giving the keynote last night. Look, Helios is a massive system. You can see it in the background here. And MI455 is an incredibly powerful chip. The context I would give, Ed, is that one of the things we're so clear about is that the demand for AI compute is continuing to increase. We have seen that over the last 5 years when you think about how much new capabilities have come on board. We've now seen a real inflection in the number of people who are using AI. Today, there are probably more than a billion active users using AI, and we expect that to scale to over 5 billion users over the next 5 years. For all of that, you need compute, and lots of it. From that standpoint, MI455 is a significant leap forward in technology capability, made up of 2 and 3nm chips, 320 billion transistors, just a lot of performance.
What's the timeline for it to be deployed in the real world, and who will be the principal first user of it?
You'll see it in the second half of '26, and it will ramp from there. We have very strong partnerships. Greg Brockman was on stage with us last night talking about all the use cases they see. We've announced a partnership with Oracle, many others as well.
So given that it is 2nm, it's in full production now, getting ready to...
We are absolutely getting ready to ship it.
That's at one end of the scale and spectrum. At the other, you have MI440X, which is for small data centers. I'm trying to simplify it, but it's basically an enterprise product. What were you trying to solve for with that?
What we're trying to solve for is that the world is very heterogeneous. You have all kinds of use cases for AI, from the very biggest cloud data centers doing large scale training and inference, to enterprise applications, as well as supercomputers. We have a family of chips. At the highest end is our MI455 for the cloud environment. But we announced last night an MI440, which uses the same basic building blocks but is now really focused on enterprise applications, so you can go into current data centers with the new technology. We're excited about that as well. Enterprises are starting to increase their adoption of AI. In some cases, they want their own control of their data centers in on-prem environments.
What are they doing with it though? We've been so fixated on frontier models with hundreds of billions of parameters and the scale of infrastructure needed for that. With MI440, we're talking about something slightly different. Could you explain what the demand is from those enterprises, what they want with the technology?
You see many enterprises now using AI throughout their business processes, whether in their workflow. Even at AMD, we're using AI through every part of our development process. We're seeing a lot of applications in financial services and healthcare. These are areas, especially in financial services, where people don't want everything necessarily in the cloud. They'd like to have their own on-prem deployments or private cloud deployments. In this case, you don't want to have to build a brand new data center for every new generation of chip. MI440 allows us to use some of those existing data centers and upgrade with the new capabilities.
Welcome. If you're watching us on Bloomberg Television or listening on Bloomberg Radio, we're live in Las Vegas with AMD's CEO Lisa Su, talking about the latest generation of accelerators. What makes this generation of AMD accelerators the better option, particularly for on-prem and at the edge, over what Nvidia is offering right now?
The best way to think about it, Ed, is that we're at an inflection point for AI. We're seeing AI now in every part of compute, from the largest models like ChatGPT, Gemini, and Grok, to many use cases in new capabilities like video production, entertainment, healthcare, and drug discovery. Our claim to fame is outstanding performance at a very advantageous total cost of ownership. The other thing we believe in strongly is an open ecosystem and deep partnerships with our overall ecosystem. When you put those things in perspective, we have a great set of applications that will take advantage of these newest generation chips.
You mentioned that Greg Brockman, OpenAI's president, was on stage with you last night. One of the basic points he made was that there are tools and functions they would love to release, but they're compute constrained. I often ask you to quantify demand, but is there a way to quantify the severity of the lack of compute, the deficit that's out there right now?
Let me give you some numbers to ground what we think the demand environment looks like. Today, we have about a billion active users, and we're ramping to five billion over the next 5 years. We have about 100 zettaflops of compute around the world. We think we have to increase compute by another 100 times over the next four or five years. I introduced a term last night: the yottaflop. A yottaflop is 10 to the 24th flops, a one followed by 24 zeros. That's another 100 times more compute than we have today. What are you going to use all that compute for? The models we have today are great, but they can get so much better. AI is going to be everywhere and for everyone, making our businesses and each of us more productive. We're still in the very early innings of unlocking the power of AI.
So, where we stand is that there's a compute deficit, and software has hit the limits of what current generation compute can offer. Help us understand the bottlenecks and barriers to deploying that compute. A lot is about memory chips. What else? Energy, electricity? What's crossing your desk, Lisa, that gives you pause and makes you say this is a problem right now?
Our job as a technology industry is to push the bleeding edge. When we think about MI455 deploying 2nm and 3nm chips with the latest generation high bandwidth memory, and really deploying these big systems, the important thing is that the entire ecosystem comes together and plans for this next big inflection in compute. That's exactly what we're doing. We're working very closely with the entire supply chain to ensure we have the resources to expand this compute environment. Some of the things you mentioned are constrained.
Which is most severe, Lisa?
I don't think it's any one thing. We're all looking at how to build faster. Power is one area where we see it being brought online as fast as possible. From a silicon standpoint, we're ramping production with our partners. From a memory standpoint, our partners are ramping as well. It's not any one thing; all of these have to go in tandem, which is why partnership is so important in this business.
We started this conversation talking about Helios, AMD's first rack scale architecture and infrastructure. Could you talk about the future and how much of the content you want to own in a server? We started this story with the GPU. If you look at what Nvidia is doing, they want to increasingly own all of what's inside the box. Is that something AMD is focused on too?
What's most important for us is to ensure we have turnkey solutions that are very easy for our customers to deploy. When you think about how to use all this AI compute most effectively, you want it to go into the data center and be up and running on day one. For that, you have to optimize a full system. But from that standpoint, we are very focused on an open ecosystem. We design the CPUs, GPUs, and some networking elements, but we also work with a broad ecosystem of partners and industry standards. It's about ensuring we get the best of all worlds when we put our solutions together.
Looking ahead to MI500 in 2027, that has 1,000 times the performance of the MI300 generation. Something's coming that's a thousand times better. How did you make it a thousand times better?
It is just incredible engineering at every level. MI455 is 10 times better than the chip we launched 6 months ago, the MI355, and MI500 is another 10x. On top of that, we are using the most advanced technology. We have a very clear focus on hardware, software, and system co-design. It is clearly pushing the bleeding edge of capabilities.
What is the status of AMD's ability to sell products into China right now?
China is an important market for us. We sell a broad range of chips into China, including our PCs and other embedded chips.
In the data center context, of course.
