AVX2's Haswell-or-newer requirement is exactly why real software like FFmpeg and x264 perform runtime CPUID dispatch, shipping multiple compiled code paths and picking the fastest one the actual CPU supports. The frequency-throttling behavior you note here is a real, documented Intel phenomenon that led some server operators to explicitly disable wide vector instructions in latency-sensitive workloads.