GPU scheduling only affects the result due to buggy optimizations. It's the exact same mechanism as fp rounding error on the CPU or updates to globally shared PRNG state. We use lots of buggy optimizations because they don't matter in practice in most situations (see ex -ffast-math).