The vqsort README says it is a non-issue on that generation CPU (Skylake-X) based on their benchmarks: https://github.com/google/highway/tree/master/hwy/contrib

The vqsort README says it is a non-issue on that generation CPU (Skylake-X) based on their benchmarks: https://github.com/google/highway/tree/master/hwy/contrib/sort#study-of-avx-512-downclocking https://github.com/google/highway/tree/master/hwy/contrib/so... At least on their low clock speed server CPU (<=3GHz), downclocking was hard to measure compared to clock speed variability they got with std::sort. The 10980xe used for windows benchmarks here normally boosts much higher, so bigger differences could be expected. The author of OP mentioned measuring clock speeds with perf and seeing some difference.