@Anhuarcin It could be possible that there are code paths generated for runtime dispatch, so that if the CPU detection shows AVX support, then it will branch to use that - I can imagine that IPP does stuff like that to squeeze out performance.