It doesn't look like AMD's HIP effort compiles to OpenCL, but to C++, that works on specific AMD GPUs (Fiji R9 Fury X, R9 Fury, R9 Nano). Also, much of the spe

It doesn't look like AMD's HIP effort compiles to OpenCL, but to C++, that works on specific AMD GPUs (Fiji R9 Fury X, R9 Fury, R9 Nano). Also, much of the speed gains for ML on NVidia hardware come from CuDNN - there is no equivalent for OpenCL or AMD hardware