Nvidia just has been the dominant player in workstation GPUs for so long that it’s almost impossible to see AMD catching up. The hardware is one thing, but I don’t see AMD closing the gap even if they had a GPU that could perform as well as Nvidia’s. Just look at some of the benchmarks on stuff like blender and unreal engine. And that’s not to mention all of the work that’s gone on for optimizing ML and deep learning models for CUDA.
I won’t pretend like I’m any sort of expert on GPUs, but I feel that Intel and AMD need to team up to create some standard software implementation for AI/ML, and then support it instead of just relying on open source (like with Pytorch and ROCm)