AMD CEO Lisa Su: We're seeing 30 times more performance with our Helios systems

요약 보기

AMD CEO 리사 수는 AI 인프라 시장이 향후 4~5년간 2조 달러 규모로 성장한다고 전망했다. Helios 랙 시스템과 Venice 프로세서가 본격 양산에 들어갔고, 최신 Helios는 이전 세대 대비 30배 이상 성능이 향상됐다고 밝혔다. OpenAI·Meta가 MI450을 멀티 기가와트 규모로 도입하며, Anthropic도 2027년 상반기부터 Helios 기반 MI450 2기가와트를 배치할 계획이다. 가격 질문에는 토큰당 달러 비용을 강조하며, 경쟁 제품 대비 생산성 10~15% 이상, 토큰당 비용 우위로 최대 30% 이점이 있다고 주장했다.

다듬은 글

※ 자동 자막 원문(다듬기 전)

01. Tom, thank you. I'm currently in San Francisco at AMD's "Advancing AI" event with AMD CEO Lisa Su. Lisa, thank you for having us. John, I'm so glad to be here. Thank you for being here. Well, I have to outline the situation a little. Intel is due to release its financial results in a few minutes. This week, we received strong signals about demand for AI infrastructure from Google just last night.

02. And then the massive panic attacks in the market over Kimmy K3 and how that could affect demand for the premium segment in the AI ​​space, whether it's software or hardware. So, with that in mind, tell me about the significance of this moment for Helios and AMD. Yes, first of all, John, this is all true.

03. But what we actually see, you know, is much broader. I mean, this is an incredible moment for AI. You know, we constantly monitor our markets, spend time with our customers. And, you know, from what we're seeing today, the high-performance computing and AI market is going to reach $2 trillion in total market opportunity over the next four to five years.

04. So this is really huge in scale. And I'm very pleased to announce this morning that our Helios rack system, as well as our new Venice processors, are now in full production. And this is a huge step forward in terms of AI capabilities. We are seeing incredible numbers, such as over 30x performance improvements in our latest Helios systems compared to the previous generation.

05. And we're talking, you know, from a customer perspective, about tremendous interest and excitement. You know, we had Anthropic, OpenAI, Meta, AT&T, Cisco, just a lot of our customers who came together. You were literally walking towards this, Pensando, I remember our conversations in Austin , San Jose, etc.

06. You immediately announce these big clients for Helios. How much of this is the target volume? And how much for pilot tests ? Well, I think it's safe to say that with a system as complex as Helios, people don't do this kind of work for the sake of, so to speak, "pilot projects," because it's a huge amount of effort when large companies that develop fundamental models need to direct their resources to optimizing the latest models for our equipment.

07. Of course, we've already talked about OpenAI and Meta deploying MI450 systems at multi- gigawatt scales. And this week we announced a collaboration with Anthropic, who plan to deploy up to 2 gigawatts of MI450 capacity based on Helios, starting in the first half of 2027 . So these are large-scale deployments, and we work side by side with our customers' engineering teams: in the hardware, software, and data center areas.

08. I can say that it is a huge amount of work, but everything is going extremely well. So this is a truly joyful day for us . I don't know if you've announced prices, but I've seen estimates and speculation that the Helios rack will cost over 5 million, and that's a premium price segment. How much of this price is due to TSMC raising the cost of chip supply?

09. And what part is due to general shortages and limited supplies ? Well, the most important thing, John, is to focus on the token per dollar metric. Actually, this is what the whole industry is talking about. This is about dollar tokens, and from our point of view, we win in both aspects. Our Helios systems are more productive than competitive counterparts, say, by 10–15% or more.

10. Additionally, we provide a lower cost per token. Thus, our clients receive an advantage of up to 30% in calculating the number of tokens per dollar. This is a huge indicator. When we talk about inference, everyone is eager to use the capabilities of artificial intelligence in an increasing number of applications.

11. We are now at a stage where everyone's budgets for token usage are growing rapidly. Including our own, by the way . Our token spending is increasing every month. But we also see a tremendous increase in productivity because of this. So, this is an opportunity for optimization. New technologies are available to enable you to harness more AI computing power, delivering better performance at a lower cost. And, frankly, it has a multiplier effect.