> There is also the fact that in modern CPU-side development you will find it almost always is a bad idea to try to drop to assembly micro-optimization. Your compiler is going to almost always do a better job optimizing the general case into an optimal set you won't be able to easily implement in raw ASM, especially when you start using vectorization instructions.

I agree. In fact, the fact that it's not even a good idea in most situations to write in assembly on x86 CPUs proves the point even more. GPUs are less willing to spend die space on things like sophisticated branch prediction, complex superscalar architectures, register renaming, and so forth. This effectively means that GPUs punish naive codegen (which is the kind of codegen that humans without in-depth knowledge of the architecture tend to produce with their brains) significantly worse than x86 CPUs do. For an extreme example, look at the vector part of the low-power VideoCore IV architecture [1]: compared to a compiler, I don't trust myself for a second to program this optimally.

[1]: https://github.com/hermanhermitage/videocoreiv/wiki/VideoCor...