And when that day comes, I will buy it. But for now I will go with nVidia. Also I badly need same CUDA codebase/documentation on AMD part as well to make such a move. ML is such a breeze when you throw CUDA at it these days.
That said, AMD had a CUDA clone/conversion system a while back. I'm not sure where that's gone.
The ideal would be to leapfrog CUDA and present something more usable and straightforward, especially something where the memory and threading model is really clear. CUDA already seem pretty good but there's always room for improvement.