View Single Post
Old 2016-09-15, 20:15   #115
bsquared
 
bsquared's Avatar
 
"Ben"
Feb 2007

340410 Posts
Default

Quote:
Originally Posted by xathor View Post

While it looks like a 64 physical core machine with 4 threads per core for a total of 256 threads, it is not. You'll find that things will scale to 192 threads and anything over that will hit a bottleneck. Try doing an omp_get_num_procs().
This is clearly application dependent. I've seen good scaling all the way to 240 cores on several applications on Xeon Phi cards (4-thread x 60 cores) and would expect the same here on KNL.

Quote:
Originally Posted by xathor View Post
The issue is that a core will flip-flop between a thread constantly. Even if that thread doesn't have any work assigned to it, it will just wait. If you purchased your system from Colfax they have some online resources that will explain this in detail.

https://youtu.be/aItWbUtoi_c is the link to the Colfax video explaining this. Pay attention to 3:40 onwards.

Essentially, you can consider the KNL processor to be a 128 core machine at half the clock speed, so 600Mhz.
That is true of the gen1 KNC cards, and that video you linked to is specific to KNC. But it's my understand that isn't true now with KNL. The Atom cores are much more forgiving, providing out-of-order processing (previously in KNC it was strictly in-order) and back-to-back instructions per thread.
bsquared is online now   Reply With Quote