Von Borg auf Discord:

I just made a pr for the onednn backend (formerly dnnl, formerly mkl-dnn): <https://github.com/LeelaChessZero/lc0/pull/1558>
This gets improved performance on avx512 cpus and recent intel igpus (with fp16).  Can also use bf16 computation if you have one of the few cpus that support it (untested).

Lothar