What's the story with ML stuff on AMD consumer GPUs nowadays? Gotten better, or still "buy NVidia for that"?
The NAVI line was never officially supported by ROCm (though ROCm 3.7 in August seemed to get some compiles working for the 5700 XT, a year after its release).
-------
Generally speaking, ROCm support only comes to cards that have a "Machine Intelligence" analog. MI50 is similar to Radeon VII, so both cards were supported in ROCm. The Rx 580 was similar to MI8, and therefore both were supported.
The Rx 550 had no similar MI card, and had issues that are yet resolved today. Rx 570 had no MI card, but apparently its similar enough to the 580 that I'm not hearing many issues.
In effect: AMD only focuses their software development / support on Machine Intelligence cards. The cards that happen to line up to the MI-line happen to work... but the cards outside of the MI line are spotty and inconsistent.
Based on the ISA alone, I'd prefer compute-applications to move to RDNA frankly. Its clearly a better designed architecture.
------------
I can't find any public information on CDNA. I've even emailed ORNL for information, but they responded back saying that NDAs prevent them from saying anything.
Whether CDNA is based on NAVI or Vega will be a big question as Frontier launches next year. I hope its Navi based, I really do, because its just better. But I'd "understand" if ROCm can't be ported easily to Wave32 or other issues (RDNA is a huge architectural change).
Direct ML is still in its infancy and is quite slow right now, AMD really needs to step up their game if they want to compete.
FP32/64 performance is going to be capped, as with prior generations.
Do you mean "enterprise ML"? Like, if I ran an ML company I wouldn't be looking at these cards?