“On Zen 4 it’s very fast, and the intrinsic functions are reasonably easy to use.”
AVX-512-BW masked loads and stores let Finch lowercase 64 bytes per operation and mask off the tail of a string with no scalar fallback loop. Across chunks from 1 byte to 1 KB it stays near the fastest, without the troughs of autovectorized code. He ran it on an AMD Ryzen 9 7950X. Intel’s own chips ship AVX-512 inconsistently.