> People replying with "don't bother, just use Nvidia&CUDA" only make the problem worse

No, they don't "only make it worse". It's good advice to a large proportion of people who just need to get work done. AMD/Intel need to hear that, and step up, by providing real, fully-supported alternatives, not leaving their customers to fool around with half-working CUDA imitations. ML is such an important field right now, and they've dropped the ball.