Microsoft Is Helping Finance AMD’s Expansion Into AI Chips

This sparked AMD’s share price increase today. Am not suprised, it’s very good news for AMD and a stramp of approval from an industry heavyweight.
Will be interesting to see if this is about graphics cards or better x86/x64 chips. Time will tell.

5 Likes

Is this confirmed? I had the impression that this was a false alarm – or at least there was at one point a report that AMD and MS were working more closely, maybe that AMD was actually working on MS’s in house AI chip, which was debunked.

It’s probably at least in part about a push to make AMD’s software stack for AI more robust.

An observation: MS almost missed the Internet/browser world, they did miss mobile devices. Social media was never a big thing for them either. I’m sure they are determined to not miss AI. Good for them.

A google search for “microsoft amd” reveals a few other stories :

and a paywalled one Bloomberg - Are you a robot?

No insight from me here, just initially read the Yahoo finance story.

It doesn’t matter. The government threw a lot of money at AMD as part of the contracts for Frontier, El Capiitan and other supercomputers… Well, not actually at AMD, but to make it possible to run existing CUDA code on AMD Instinct GPUs–and much of that within the national labs to port existing applications from older supercomputers.

ROCm 6 is CUDA compatible on a software level, and tools can easily be retargeted to other than AMD CPUs and AMD Radeon or Instinct GPUs.

Note that Frontier is the current #1 on the Top 500 list, Aurora is #2 and expected to be upgraded to #1 later this year–and the be succeeded by El Capitan in the #1 position. But Aurora is an Intel system with Intel GPUs. Can you port ROCm to Aurora? Sure. CUDA? nVidia would have to do that, and I haven’t been following the Aurora software stack development. I’m retired now, but I would, for any new large app, plan to target ROCm. Well, look at the libraries listed here: https://www.amd.com/en/developer/resources/infinity-hub.html If one of those libraries fits your needs, 90% or more of your software development may have just gone away. Similar libraries exist for CUDA, but which large machine are you going to run on? For most supercomputer apps, they are either NP or doubling all dimensions is a minimum cost of a factor of 16 (3 dimensions of space, one of time, or an FFT at n*log2n per dimension). Tesla is building some high-end supers but not all will run CUDA.

1 Like

Good to hear from you Bob. Hope all is going well for you in retirement…doc