Forum > GPU crunching
Latest drivers (NVidia and ATI)
Sutaru Tsureku:
--- Quote from: Sutaru Tsureku on 06 Apr 2010, 02:39:17 pm ---Which nVIDIA driver you tested with CUDA V2.3?
I guess you used only 197.13 with CUDA V2.3 and CUDA V3.0.14 in your test, right?
--- End quote ---
--- Quote from: glennaxl on 06 Apr 2010, 11:23:17 pm ---My test was CUDA v2.3 vs CUDA v3.0.14 on MB_6.08_CUDA_V12_VLARKill_FPLim2048.exe
--- End quote ---
I asked, because from my experiences and test of 190.x , 191.x and 195.x was 190.x the fastest nVIDIA driver (with CUDA V2.3).
196.x I didn't tested.
Now with 197.13 and CUDA V2.3 or V3.0.14 I'm sceptical that it's faster than 190.x + CUDA V2.3 .
It's look like the 197.13 + CUDA V3.0.14 is ~ 1 - 2 % slower than 190.x + CUDA V2.3 .
This sounds not much, but my OCed GTX260-216 have a RAC of ~ 15,000.
And ~ 1 - 2 % are ~ 150 - 300 Cr./day (RAC). This are two or four WUs/day less crunched.
So someone should make a test if 190.x + CUDA V2.3 or 197.13 + CUDA V3.0.14 is faster.
If someone would explain me how to make it with bench WUs, I would make this test in near future.
If someone could do this test faster/earlier as I, it would be also fine.. ;)
glennaxl:
To test grab KWSN Knabench 1.81(Raistmer's modified Benchmark Package
The quick and dirty ;D method:
1) Make sure boinc is running then snooze the gpu leaving the cpu at 100%. (note: LAIM should be off)
2) Extract the KNAbench and prepare the science apps. Read the readme for further instructions
3) Modify Knabench181_nb.cmd -> look for "-verb -st" and append "-device N" where N is the gpu number eg. "-verb -st -device 0" (note: there are 6 instances; its used for text output and parameter)
4) run Knabench181_nb.cmd
if you want to test N gpus at a time, just duplicate the knabench folder and modify the target gpu (-device N)
Long Method:
Make sure nothing is running in background; cpu is idle.
Number of CPUs = Number of KNAbench folder
Number of GPUs = Number of KNAbench folder
so for q6600 and gtx 295 it would have 6 KNAbench folders
follow steps 2 & 3 above; step 3 for the gpu test folders
run Knabench181_nb.cmd for all folders
Sutaru Tsureku:
O.K., here we go..
It was not so easy to understand how to do it.. but I got it.. ;)
The result in short.. I don't know what's faster.. :D
The results on my QX6700 @ 3.14 GHz with manufacturer OCed GTX260-216:
[MB_6.08_CUDA_V12_VLARKill_FPLim2048.exe]
[only GPU calculation, CPU idle]
[fastest in yellow]
197.13 + CUDA 3.0.14
TaskName: PG0395.wu ------ TaskName: PG0444.wu ------ TaskName: PG1327.wu
47.578 secs Elapsed ----- 43.219 secs Elapsed ----- 31.391 secs Elapsed
16.203 secs CPU time --------- 15.609 secs CPU time --------- 14.922 secs CPU time
197.13 + CUDA 2.3
TaskName: PG0395.wu ------ TaskName: PG0444.wu ------ TaskName: PG1327.wu
47.078 secs Elapsed ----- 43.719 secs Elapsed ----- 32.375 secs Elapsed
16.250 secs CPU time --------- 15.813 secs CPU time --------- 15.547 secs CPU time
190.38 + CUDA 2.3 (1st run)
TaskName: PG0395.wu ------ TaskName: PG0444.wu ------ TaskName: PG1327.wu
47.078 secs Elapsed ----- 43.891 secs Elapsed ----- 32.453 secs Elapsed
16.469 secs CPU time --------- 16.172 secs CPU time --------- 16.078 secs CPU time
190.38 + CUDA 2.3 (2nd run)
TaskName: PG0395.wu ------ TaskName: PG0444.wu ------ TaskName: PG1327.wu
47.531 secs Elapsed ----- 43.063 secs Elapsed ----- 31.563 secs Elapsed
16.422 secs CPU time --------- 15.922 secs CPU time --------- 15.141 secs CPU time
The best would be to let run 3 or 5 runs per driver/CUDA version combis for to calculate the average. For to compensate 'runaway' results.
But, no.. not I again.. ;)
Who would like to do it? :)
EDIT: Or maybe also longer WUs for 'benching'?
glennaxl:
Nvidia 257.21 drivers released today.
I havn't done any any benchmarks yet but for multiple gpu users (vista/7), you don't need to extend your desktop (using dummy plugs or real monitors). ALL cuda gpus will be detected.
efmer (fred):
--- Quote from: glennaxl on 15 Jun 2010, 11:49:59 am ---Nvidia 257.21 drivers released today.
I havn't done any any benchmarks yet but for multiple gpu users (vista/7), you don't need to extend your desktop (using dummy plugs or real monitors). ALL cuda gpus will be detected.
--- End quote ---
Nice, already used the beta drivers with these features, but maybe this release driver is even better. Was still a lot slower than to XP 64 version.
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version