| Author |
Message |
|---|
|
|
|
I have 2 GTX570's installed running Win 64 on a i7.
The normal time to complete a PPS cuda wu is about 9min 30 sec / card
When there is a pause / restart of Boinc manager / Boinc doing a benchmark / suspend project etc the cards go in a half speed mode & computing times double. I tried to reduce CPU tasks from 8 cores to 7 but with no effect.
I don't run SLI because it makes no difference in speed & the Nvidia driver installed is the latest
The only way to get them to run @ 100% again is to suspend PrimeGrid, abort all running tasks & do a cold restart of the pc |
|
|
|
|
|
Probably cards switch to 3D-Low level ( 400MHz gpu core frequency). Did you check this? |
|
|
 mikey
Send message
Joined: 17 Mar 09
Posts: 2369
ID: 37043
Credit: 1,064,325,096
RAC: 94,529
                         
|
I have 2 GTX570's installed running Win 64 on a i7.
The normal time to complete a PPS cuda wu is about 9min 30 sec / card
When there is a pause / restart of Boinc manager / Boinc doing a benchmark / suspend project etc the cards go in a half speed mode & computing times double. I tried to reduce CPU tasks from 8 cores to 7 but with no effect.
I don't run SLI because it makes no difference in speed & the Nvidia driver installed is the latest
The only way to get them to run @ 100% again is to suspend PrimeGrid, abort all running tasks & do a cold restart of the pc
Windows has no way to reset/restart a gpu without restarting the whole pc, so when one goes funky you MUST restart the pc to get it back to normal again. You don't need to suspend and abort though, just shutdown Boinc first then restart the pc. Sometimes, as in once in a while, you will lose a workunit but normally it will just come back up and continue on where it left off. One thing to try though is to change gpu drivers, the latest is not always the greatest for us crunchers. Gamers yes, but not always crunchers. |
|
|
|
|
Probably cards switch to 3D-Low level ( 400MHz gpu core frequency). Did you check this?
Thanks. Yes I did. For some reason it drops the speed.
I tried to run Afterburner to lock the clock speed but even that don't help
BTW both cards were in a XP Pro 32 system & I never had this
One thing to try though is to change gpu drivers
Can you suggest a driver version I can try? |
|
|
|
|
|
AFAIK there is a bug in a drivers ( even in the newest ) with that problem. No solution for now with multigpu hosts. :( |
|
|
rroonnaalldd  Volunteer developer
Volunteer tester
Send message
Joined: 3 Jul 09
Posts: 1213
ID: 42893
Credit: 34,634,263
RAC: 0
                 
|
|
The downclock-problems after overheating was known for some driver versions and i thought nVidia had solved this error...
I would try the Cuda4-driver 270.81 for Win32, if your gpu is supported by this driver. I use the same Cuda4-driver 270.41.19 in my linux64-host and had not seen any downclocking in the last months.
____________
Best wishes. Knowledge is power. by jjwhalen
 |
|
|
|
|
|
It's not downclocking by overheating. When you crunching e.g. PG, suspending it and changing to other project locks gpu on 405 MHz. I believe drivers for win 7 are affected.
My hosts are 1 gpu for now, so i can't check new drivers. If anyone find bug-free drivers, let us all know. |
|
|
|
|
|
This is an issue (resume/suspend status) with newer drivers, if you can run 260.99 do it, it does not downclock anything. |
|
|
mikey
Send message
Joined: 17 Mar 09
Posts: 2369
ID: 37043
Credit: 1,064,325,096
RAC: 94,529
|
It's not downclocking by overheating. When you crunching e.g. PG, suspending it and changing to other project locks gpu on 405 MHz. I believe drivers for win 7 are affected.
My hosts are 1 gpu for now, so i can't check new drivers. If anyone find bug-free drivers, let us all know.
Can you run just one gpu project and it will run just fine? |
|
|
|
|
|
Yes, currently I'm using 266.58 drivers, win 7 x64 and no problem with any project ( seti, pg, einstein etc ). |
|
|
rroonnaalldd Volunteer developer
Volunteer tester
Send message
Joined: 3 Jul 09
Posts: 1213
ID: 42893
Credit: 34,634,263
RAC: 0
|
|
I'm using driver 270.41.19 in my linux64-box and have no problems in PG, Collatz and Einstein.
____________
Best wishes. Knowledge is power. by jjwhalen
|
|
|
|
|
|
Feedback
Thanks for the advice!
I switched back to 266.58 (win7 64) & GPU speed remains constant.
Tried it @ DistrRTgen & it also works.
This proves that the latest is not always the greatest
I had a similar one with newer ATI drivers > 5.11 @ Moo where it would use a full CPU core instead of 0.05%
Edit
The downside :(
Performance dropped by 8% |
|
|
|
|
Feedback
Thanks for the advice!
I had a similar one with newer ATI drivers > 5.11 @ Moo where it would use a full CPU core instead of 0.05%
It's bug in ati drivers. ;)
1 gpu in host, ~0% usage on cpu. More than 2 gpu in host, 1 cpu core per gpu core ( 100% usage ). |
|
|
|
|
|
The downclocking problem isn't really due to a Bug in the driver's, it's more due to the driver technology out pacing the Cuda apps,
the real culprit is the the Boinc api embedded in the Cuda apps, the Cuda apps at all the Major Cuda projects predate the lastest api changes,
see Jason G's post at Lunatics for the Technical details: Recent Driver Cuda-safe Project List
From Applications that use coprocessors in the Compute with BOINC Wiki (documentation on creating and operating BOINC projects):
Cleanup on premature exit
The BOINC client may kill your application during execution. This may leave the GPU in a bad state. To prevent this, call
boinc_begin_critical_section();
before using the GPU, and between GPU kernels do
if (boinc_status.quit_request || boinc_status.abort_request) {
// cudaThreadSynchronize(); or whatever is needed
boinc_end_critical_section();
exit(0);
}
To cure the downclocking problems Cuda projects will need to rebuild their Cuda apps
Claggy |
|
|
