@nomachine

For my local 9950X and 16 threads I am getting ~100 Mkeys/s.

For vast.ai server with AMD EPYC 7642 48-Core 186 threads (up to 192) I am getting 222 Mkeys/s.

Seems like for greater number of threads it is not working properly.

I would expect much more keys/s.

I am currently using and testing AVX-256, and it is slower than AVX-512 by half. The AVX-256 workload here can cause the CPU to downclock significantly, which might negate the performance gains from vectorization. The CPU is throttling as a result.