CRC32X works on 8 bytes at a time and has a throughput of one invocation per cycle, whereas the SIMD operates on blocks in parallel (the chromium code does 64 b

CRC32X works on 8 bytes at a time and has a throughput of one invocation per cycle, whereas the SIMD operates on blocks in parallel (the chromium code does 64 bytes an iteration, with a lot of instruction-level parallelism too). Theoretically M1 could have thrown more silicon at it to allow more than one CRC32X invocation per cycle, but that's not very useful if you can achieve the same with SIMD anyway.