Non-linear performance scaling on CPU over 8 physical cores?

[drHuangMHT]

I know it's a dumb idea to run models on CPUs, but I do observe some weird behavior on my dual socket EPYC system.
When I launch one instance of webui I see utilization being around 67%, and not all logical cores are used. As for two instances the time for each iteration is the same but effectively doubled output. However, when I upgrade from 8 cores(phy) to 16 cores(phy) per socket the inference time does not change. Everything is the same as previous 8 cores setup, but performance is also the same despite having doubled amount of CPU cores. Cinebench result scaled as expected.
I heard you guys screaming "buy a used GPU pls", but is this expected?

[drHuangMHT]

Just upgraded from Windows Server 2019 to 2022 and problem solved, sort of. Time took for each iteration is halved, but overall utilization is still around 65% with one instance. Close for now.