Amazon Web Services, the world's largest cloud computing provider, is considering using new artificial intelligence chips from Advanced Micro Devices, though it has not made a final decision, an AWS executive told Reuters.
The remarks came during an AMD event where the chip company
outlined its strategy for the AI market, which is dominated by rival Nvidia.
Despite AMD disclosing some technical specifications for an
AI chip coming later this year that could in some ways beat Nvidia's best
current offerings on some metrics, the news sent shares down after AMD did not
disclose a flagship customer for the chip.
In interviews with Reuters, AMD Chief Executive Lisa Su
outlined an approach to winning over major cloud computing customers by
offering a menu of all the pieces needed to build the kinds of systems to power
services similar to ChatGPT, but letting customers pick and choose which they
want, using industry standard connections.
"We're betting that a lot of people are going to want
choice, and they're going to want the ability to customize what they need in
their data center," Su said.
While AWS has not made any public commitments to use AMD's
new MI300 chips in its cloud services, Dave Brown, vice president of elastic
compute cloud at Amazon, said AWS is considering them.
"We're still working together on where exactly that
will land between AWS and AMD, but it's something that our teams are working
together on," Brown said. "That's where we've benefited from some of
the work that they've done around the design that plugs into existing
systems."
Nvidia does sell its chips piecemeal but is also asking
cloud providers if they are willing to offer an entire system designed by
Nvidia in a product called DGX Cloud. Oracle is Nvidia's first partner for that
system.
Brown said AWS had declined to work with Nvidia on the DGX
Cloud offering.
"They approached us, we looked at the business model,
and it didn't make a lot of sense," Brown said.
Brown said that AWS prefers to design its own servers from
the ground up. AWS started selling Nvidia's H100 chip in March, but as part of
systems of its own design. © Reuters