Comment by janwas

Gemma.cpp has nested thread pools, one per chiplet, and one across all chiplets. With such core counts it is quite important to minimize any kind of sharing, even RMW atomics.