Hacker News Reader

Yes, can be about 1.6x speedup: https://github.com/Voultapher/sort-research-rs/blob/main/wri... (cpp_intel_avx512).

Disclosure: I am the main author of vqsort, which is also a vectorized Quicksort but portable.

The 512 bit registers are the least important part of AVX-512. The much more important parts are the 32 registers, mask registers (these are great for quicksort), compressed displacement (which shrinks the instruction size for unrolled code), and a bunch of other goodies.

You don't need avx512 to implement a bitonic sort. I worked on several implementations with plain avx.

Or SSE2+.

AVX-512's VPCOMPRESS* instructions are very useful for binary-radix sorting though.

It can often be more than you would expect from the width increase, as lots of masking tools and instructions allow you to vectorize things more efficiently than could be done before.