Comment on The Nvidia DGX-1 Deep Learning Supercomputer in a BoxparentComments−dharma110yIt 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
Comments
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