If you look at the ROCm API you'll see that it's pretty much exactly that, a CUDA compatibility layer, but an identical API means little if the functionality be

If you look at the ROCm API you'll see that it's pretty much exactly that, a CUDA compatibility layer, but an identical API means little if the functionality behind it has different quirks and caveats and that's harder to assess. I am rooting for ROCm but I can't justify betting on it myself and I suspect most of the industry is in the same boat. For now.