AMD rolls out new rival to Nvidia’s AI chip

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

Chip designer AMD has launched sales of a new chip it hopes will break Nvidia’s dominance of the artificial intelligence processor market, which it predicted would be worth $400bn by 2027.

AMD’s MI300X chip was “the most advanced AI accelerator in the industry”, chief executive Lisa Su said at an event in San Jose, California, claiming it outperforms Nvidia’s current offering.

Until recently, AMD was predicting that the market for AI chips would reach $150bn by 2027. But “it’s really clear that the demand is just growing much, much faster”, Su said on Wednesday.

Nvidia’s H100 chip — which Su used as a point of comparison for the MI300X throughout her presentation — has so far dominated the market, with demand outstripping supply as companies like Amazon, Meta, Microsoft and Google use them to develop generative AI tools. It has been a boon for Nvidia, whose quarterly revenues tripled year-on-year in the most recent quarter.

AMD expects its MI300 chips to be its fastest product to hit $1bn in sales, a target it believes it will achieve around mid-2024. A variant of the MI300X, the MI300A, will cater specifically to supercomputing.

When AMD initially announced its new chip in June, analysts were downbeat on its chances of catching up to Nvidia in the current generation of AI chips. Nvidia is already planning to launch a new H200 chip next year that it claims will be a “game-changer”.

AMD’s Su invited Microsoft chief technology officer Kevin Scott and Meta AI’s senior director of engineering Ajit Mathews onstage to talk about how they were incorporating the MI300 into their AI workloads.

One group planning to use AMD’s chips is OpenAI — the generative AI start-up whose ChatGPT product sparked a massive interest in the technology — which will incorporate AMD’s new chips in the latest version of its Triton AI software.

“OpenAI is working with and in support of an open ecosystem,” OpenAI engineer Philippe Tillet said in a statement. “We plan to support AMD’s GPUs including MI300” in the latest release of Triton.

In another example of AMD’s efforts to go head-to-head with Nvidia, it also debuted its ROCm 6 software platform, which will compete with Nvidia’s proprietary platform Cuda. “Software actually is what drives adoption,” Su said.

Speaking to reporters following her speech, Su said the anticipated $400bn market for AI processors in 2027 left plenty of space for AMD. “We think we could get a nice piece of that,” she said. Included in that estimate is China, which is facing a US government crackdown on exports of advanced AI chips.

Su said AMD “spends a lot of time with the [Biden] administration” and US commerce department. “We understand for the most advanced chips, [export restrictions] are important for us to have — from a national security standpoint.”

What was important, she said, was “finding a balance”.

“On raw specs, MI300X dominates H100,” Dylan Patel and Daniel Nishball at semiconductor consulting firm SemiAnalysis wrote on Wednesday following the announcement. OpenAI’s announcement in particular was “a big deal”, they said, as both OpenAI and Microsoft would now be using AMD chips for work on AI inference.

Nvidia did not immediately respond to a request for comment.