I don’t really pay attention to AMD cards anymore. I’m not a gamer and every time over the past five years I’ve put a toe in the water of the disaster that is ROCm, their drivers, “support” in frameworks, etc I run back to CUDA.
It really seems like AMD has ceded everything other than supercomputer cluster ML to Nvidia and CUDA. They do well in gaming and they can sell and support the extremely high end of ML with mostly bespoke applications that have teams of professional support and institutional backing. When you’re spending $100 million or whatever on hardware flop/$ really counts and you can throw people at making it work.
In the low to mid ranges of ML the 25% or even 50% savings on hardware is quickly lost when compared to the human time sink that is the entire AMD ML “ecosystem”. If you’re buying 10 GPUs you’re talking about saving what, $20K? That “savings” gets wrecked really quickly when you look at salaries and time spent chasing yet another bug/quirk/showstopper with ROCm.
Frankly, Nvidia doesn’t exactly foster goodwill (just look at any HN thread when they come up). I suspect that a decent number of people are willing to throw time and pull requests into the ROCm ecosystem just out of spite for Nvidia. Good for them because Nvidia needs a serious competitor!
I’m rooting for AMD but I’ve always been in the position of “it needs to work now in the shortest and cheapest path of least resistance” which has always been, and will be for the foreseeable future, CUDA.