Generally, the reason to bother with this approach is if you have a project that only needs tensor cores in a tiny part of the code and otherwise benefits from the cross platform nature of OpenCL, so you have a mostly shared codebase with a small vendor-specific optimization in a kernel or two. I've been in that situation and do find that approach valuable, but I'll be the first to admit the modern GPGPU landscape is full of unpleasant compromises whichever way you look.