3 ms·

Further to this, if you are running a program where compute is mainly a sparse linear solve, you are memory bandwidth limited and on modern systems like AMD Epy

Further to this, if you are running a program where compute is mainly a sparse linear solve, you are memory bandwidth limited and on modern systems like AMD Epyc you will very very likely see improved performance if you undersubscribe cores by 2x - e.g. using max 64 threads per 128 cores.

bee_rider 3y ago
That’s interesting. I think it must be the case but just to triple check my intuition, I guess you want the threads spread out, so you can get every chiplet in on the action?
- csdvrx 3y ago
  No, I think it's more about how cores may share infrastructure (like M cache etc)