I work with the entire CUDA toolkit on a daily basis, and it is also a mess. Nvidia is locked in though, and doesn’t plan any rework anytime soon (you can refer to their own statements on this). Any widespread alternative forces greater competition, and better products as a result.

I’ve never met a single engineer who has worked on any of Intel’s acceleration toolchains, but they are just now getting new devices into the datacenter, so maybe it will gain in popularity.