[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.