As the GPGPU scene trajectory seems dismal[1] for the foreseeable future wrt the DX, this seems like the best hope.

[1] Fragmentation, at best C++ dialects, no practical compiler tech to transparently GPU offload, etc