It's both poor software support and poor performance, e.g. Vega 20 has 15TFlops in theory, in practice one could get around 7.5TFlops. And only use models suppo

It's both poor software support and poor performance, e.g. Vega 20 has 15TFlops in theory, in practice one could get around 7.5TFlops. And only use models supported by ROCm, custom CUDA kernels used by many models nowadays might not work after cross-compilation to AMD instructions.