It's a host of things, generally I've seen:

  - Poorly defined APIs/types that allow vendors to diverge.
  - Legit driver bugs.
  - Vendors preferring optimizations to following the spec(including skipping memzero requirements).
  - Hardware that doesn't behave how it should(fp32 falling back to fp16 silently for instance).

Overall you don't really win any points in the market for correctness, instead performance and efficiency are highly valued.