(not op) Somewhat, although it's fairly minimal for avx2. The serious downclocking problem comes from avx512 instructions, which the author seems interested in:

(not op) Somewhat, although it's fairly minimal for avx2. The serious downclocking problem comes from avx512 instructions, which the author seems interested in: > This algorithm should map fairly nicely to AVX-512, and should in fact be a bit faster than 2x the speed of AVX2 since a few instructions can be saved In general, you aren't going to make a faster wheel without some trickery (vectorization usually). All the obvious optimizations have already been made in the stdlibs.