185.85 + CUDA 2.2 = no app freezing anymore.

[KarVi]

I would just like to report that since upgrading nVidia drivers to 185.85 and using the 2.2 CUDA dll's, my 9500 doesn't seem to stop processing, or slow down in the middle of WU's, as it has done since starting to run CUDA.

Normally I would have to restart the machine once a day (a good run could go on for 2 days), because the app slowed down, or stopped processing. A reboot was the only way I knew to get back to normal speed.

With the newer drivers and DLL's, it hasn't happened once, and crunching times are also better than before.

Don't know if its the drivers or the DLL's thats doing the trick, but its definetely better.

I recommend upgrading if peoble are having the same problems, its worth a shot.

[Borgholio]

Found the latest graphics drivers, but where are the CUDA 2.2 drivers?

[Pappa]

Found the latest graphics drivers, but where are the CUDA 2.2 drivers?

Raistmer posted them here in this message

http://lunatics.kwsn.net/12-gpu-crunching/v10-of-modified-seti-mb-cuda-opt-ap-package-for-full-multi-gpucpu-use.msg17425.html#msg17425

[Borgholio]

Got'em.  Thanks!

[Borgholio]

I can confirm that, at least for the last 24 hours, no workunits have locked up since I upgraded to 185.85 and Cuda 2.2.  However, my video card literally melted due to a fan failure so any further testing will have to be postponed until I get the RMA in a week or two.  :-/

[chelski]

My CUDA card (9600GSO) also had a process freeze problem on both Cuda 2.0 and 2.1 with BOINC 6.6.20.  Went back to 6.6.7 with Maik's monitoring script for a while.

Going to Cuda 2.2 and the latest built V11 fixed that on the 6.6.20.  Edit:  With driver 185.85, of course

Speed up was >10%, from typical 19 mins to 16-17 mins range.

Incredibly, I can now play a game without turning off Boinc/CUDA thread (Sins of a Solar Empire) and then upon exiting the game, had the satisfaction that CUDA runs normally again (back to the 16-17 mins per WU range)

Thanks to Raistmer and everyone else who have done an incredibly wonderful job of fixing the broken nVidia release and making CUDA a tolerable experience.

[FalconFly]

As usual there's little to no documentation about what files are needed where.

Am I correct that only two Files are needed from the CUDA 2.2 Package (cudart.dll and cufft.dll) ?

Is the cudart.dll supposed to be copied to both ProgramFiles/BOINC as well as BOINC/data Directories, while cufft.dll goes to BOINC/data Directory only ?

Are the .dll's sufficient or has the CUDA 2.2 Developer Package to be installed ?

[Richard Haselgrove]

As usual there's little to no documentation about what files are needed where.

Am I correct that only two Files are needed from the CUDA 2.2 Package (cudart.dll and cufft.dll) ?

Is the cudart.dll supposed to be copied to both ProgramFiles/BOINC as well as BOINC/data Directories, while cufft.dll goes to BOINC/data Directory only ?

Are the .dll's sufficient or has the CUDA 2.2 Developer Package to be installed ?

You need the standard 185.85 driver package, but you don't need the developer package.

Once you have the driver, the only two other files you need are the ones you have identified - cudart.dll and cufft.dll

The important place to put them is the setiathome.berkeley.edu folder, a sub-folder of 'Projects' in the BOINC Data folder. Neither of them needs to go in Program files.

[FalconFly]

Alright, that sounds good.

Searching my own System running the 185.85WHQL Driver though, I did not find these .dlls on my entire System outside of the CUDA 2.2 dev package and the ones obviosuly in the respective BOINC subfolders.
I guess those were in the NVidia temp folder the Driver installation created (and which I deleted) ?

[Raistmer]

@all who had app freeze before
Look for "out of memory"errors and CPU fall back instead of  app hang with 185.85 + CUDA 2.2
I had to restart my dual-GPU Win64 server 2003 host once already.
Symptoms are:

GPU memory query call returns zero both for available and total GPU memory (printed in stderr),
subsequent CUFFT plan creation call fails with out of memory error and CPU fallback occurs.
Cause GPU in question pretty slow one I found this situation much more annoying than silent GPU hang (GPU get idle anyway but CPU become busy with underoptimized CPU stock app).

[FalconFly]

Hm, will do.

On the XP32 Systems CUDA 2.2 now runs and the behaviour (system lag/Video response times) indeed has improved dramatically

On the Vista64 system however, all CUDA workunits immediately errored out after putting the two .dlls into place (now taken from the 64bit CUDA 2.2 Release of course).
After reverting back to the old ones everything is back to normal.

[Raistmer]

On the Vista64 system however, all CUDA workunits immediately errored out after putting the two .dlls into place (now taken from the 64bit CUDA 2.2 Release of course).

"of course" Needed DLLs was posted on this board and there was no x64 version posted cause there was no x64 CUDA MB version released.

[FalconFly]

Oops, my bet was that the .dll's had to interact with the 64bit Driver primarily, thus I used those for assumed compatibility requirements (especially with all the problems in mind I had with the 64bit Driver).

Will switch to the 32bit dlls then

So far, noting no problems with CUDA 2.2 *knock on wood*