Now I can’t claim to be an expert by any means but I feel like you are over complicating this. For one, we know in practice the performance is much better when you force enable AVX2 on AMD with MKL; it was benchmarked. For another, that’s already obvious, because even without micro optimizing we’re still talking about codepaths that are able to better fill the processor and support 256 bit integers. Certainly not everything will always be straightforward but MKL seems like one of the single most obvious libraries for this.
Further, Intel has a rich history here. They have been sued and lost over Intel C++ Compiler’s “Cripple AMD” bit. Funny thing is, it's still there. Although now Intel has warnings throughout their pages. Guess removing a strcmp cpuid, “GenuineIntel” was too much work.
Let’s be realistic. Intel wants to win the benchmarks. There’s no need to construct an elaborate reality where this is all done out of concern that they might make performance worse on microarchitectures they did not design.
Further, Intel has a rich history here. They have been sued and lost over Intel C++ Compiler’s “Cripple AMD” bit. Funny thing is, it's still there. Although now Intel has warnings throughout their pages. Guess removing a strcmp cpuid, “GenuineIntel” was too much work.
Let’s be realistic. Intel wants to win the benchmarks. There’s no need to construct an elaborate reality where this is all done out of concern that they might make performance worse on microarchitectures they did not design.