Message boards :
Number crunching :
Automatic Selection of Fastest Core doesn't work correct
Message board moderation
Author | Message |
---|---|
Send message Joined: 21 Jun 16 Posts: 3 Credit: 96,784,852 RAC: 1,276 |
I have checked a little bit with the different settings that are possible for Cuda-Cards and ran: dnetc518-win32-x86-cuda31.exe -bench This brings up following: dnetc v2.9109-518-GTR-10092921 for CUDA 3.1 on Win32 (WindowsNT 6.2). Please provide the *entire* version descriptor when submitting bug reports. The distributed.net bug report pages are at http://bugs.distributed.net/ [Jun 21 14:10:25 UTC] nvcuda.dll Version: 31.0.15.5176 [Jun 21 14:10:25 UTC] RC5-72: using core #0 (CUDA 1-pipe 64-thd). [Jun 21 14:10:26 UTC] RC5-72: Benchmark for core #0 (CUDA 1-pipe 64-thd) 0.00:00:00.54 [10,041,063,384 keys/sec] [Jun 21 14:10:26 UTC] RC5-72: using core #1 (CUDA 1-pipe 128-thd). [Jun 21 14:10:27 UTC] RC5-72: Benchmark for core #1 (CUDA 1-pipe 128-thd) 0.00:00:00.46 [10,574,494,120 keys/sec] [Jun 21 14:10:27 UTC] RC5-72: using core #2 (CUDA 1-pipe 256-thd). [Jun 21 14:10:28 UTC] RC5-72: Benchmark for core #2 (CUDA 1-pipe 256-thd) 0.00:00:00.48 [11,201,033,720 keys/sec] [Jun 21 14:10:28 UTC] RC5-72 benchmark summary : Default core : #0 (CUDA 1-pipe 64-thd) Fastest core : #1 (CUDA 1-pipe 128-thd) [Jun 21 14:10:28 UTC] Core #1 is significantly faster than the default core. The GPU core selection has been made as a tradeoff be ... and responsiveness of the graphical desktop. Please file a bug report along with the output of -cp ... only if the the faster core selection does not degrad ... The Automatic-Mode chooses Core 0 to run and the WUs take 345 seconds When I manually change the setting to use Core 1, the WU needs only 312 seconds and I didn't check Core 2 so far. Seems as if the Automatic-Settings needs some tuning. Or can I set this local on each machines ? I have more different Cards then there are different profiles possible |
![]() Send message Joined: 11 Feb 14 Posts: 114 Credit: 7,646,283 RAC: 87 |
There are 12 different CUDA cores, you should run the complete benchmark on each of your cards and than it might be enough to use the four different venues (default/home/work/school) to assign all of them the best cores, your cards are just from 3 different generations. The cores with lower numbers are for old cards, I guess the 2-pipe or 4-pipe cores should perform best (either #5 or #8 in particular, but that's just a guess). ![]() |