AMD is held back by their interconnect and firmware disadvantage compared to nvidia. They’ve been trying really hard to create their own cuda, but rocM and HIP still aren’t very popular especially for research.
And their repeated refusal to either implement CUDA or reimplement everyone's CUDA libraries on their own platform. They say that AMD never misses a chance to miss a chance.
Yes. We have a quad MI300A server and run several inference models on it. For $107k it has saved us so much money on tokens already and it's a heck of a lot faster than cloud services.
> Have you ever actually had anyone work with these chips? Developer ux on amd is terrible.
Just how much of dev ux do you need? A foundational library, of course, but as the AI companies keep saying, their models can vibe-code what's needed for those chips anyway.
I am about to spend $20M, if I buy anything other than Nvidia, and things go wrong, I am going to get blamed, and if things go right I will get no credit. This is why AMD is making no progress outside of very narrow cases and supercomputing.
I thought that Nvidia's moat was more in CUDA? Hardware is hard but we've already seen other companies like Google design neural processors with compute efficiency close to Nvidia.
Google would have to start selling them (the real ones, complete with interconnect) to third parties. If google does that, though, Nvidia is done.
Unlike AMD, Google can actually ship software. AMD has never shipped good software other than drivers (maybe) in the entire history of the company, including both ATi's history and true AMD. They have always relied on Intel to provide the software.
Only thing holding them back is fab capacity which nVidia keeps buying in bulk to keep them small.