@neauoire Usual guidance is 2x number of CPUs. Split the data up beforehamf or, depending, have a queue of work and workers peel the work off when they are free (second option makes more sense when workers sometimes block on i/o, mutexes, etc.)