From your links, gold/plat has a ~15% clock speed hit when using ymm muls, then another ~20% hit when going to zmm muls. So for 512-bit wide to not slow things

From your links, gold/plat has a ~15% clock speed hit when using ymm muls, then another ~20% hit when going to zmm muls. So for 512-bit wide to not slow things down on Intel's chips, you need ~30% runtime of all cores to already be in AVX.