Most heavy math libraries that use AVX do function multiversioning. When calling the function it does a CPUID instruction check and then picks a code path appro

Most heavy math libraries that use AVX do function multiversioning. When calling the function it does a CPUID instruction check and then picks a code path appropriate for the CPU. Intel's MKL gets flak for this because non-Intel CPU's always fall back to the least optimal path even if they support a more efficient path. I think you'd be hard pressed to find programs using AVX that didn't do some sort of feature check. While AVX debuted with Sandy Bridge in 2011, Intel's Celeron/Pentium branded CPUs don't support those instructions (or didn't for a long time).