> we have to enstablish a difference between theorical limit and practical limit
Agreed, but more than that, we should also consider what jump first proposed, i.e., OS imposed limits. It is my view that :
OS Imposed Limit = none, infinite
theoretical limit = (total available memory)/((stack usage per thread) + (memory usage per thread))
practical limit <= theoretical limit as your app does not run in isolation or in a perfect environment.
It is in distinct contrast to CUDA and AMD's equivelent where there is a hardware imposed limit but the docs encourage you to run several million threads if possible to give their scheduler the best chance of doing it's job!.
Git
|