Current data - threads-nps efficiency up to 32 threads

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Modern Times
Posts: 2492
Joined: Thu Jun 07, 2012 9:02 pm

Re: Current data - threads-nps efficiency up to 32 threads

Post by Modern Times » Sat Oct 25, 2014 6:22 am

Crafty does not perform well on the particular AMD architecture. I've run tests on my AMD FX-8350. Komodo performed easily the best, Crafty was the worst.

The 32 cores on his machine are 16 modules, with 2 cores per module. In order to maximise performance, Windows is probably allocating 1 core per module until it runs out of modules. So on 16 threads, 1 thread per module, maximum efficiency. Once you make the jump to 32 threads, then it is 2 threads per module of 2 cores. Crafty performs badly in that situation, but other engines do OK. That is my theory. I think that is what it does on a single socket desktop machine with this CPU architecture, but dual socket I can't be sure.

fastgm
Posts: 492
Joined: Mon Aug 19, 2013 4:57 pm
Contact:

Re: Current data - threads-nps efficiency up to 32 threads

Post by fastgm » Sat Oct 25, 2014 3:52 pm

I have put the bench logs here:
http://www.fastgm.de/schach/crafty-bench.zip

As described in the Excel sheet the test time was 90 seconds - 5 repetitions.

Post Reply