Hi,

i get an error after few seconds on any task. But only on the 8800 GTS.
GTX 285 work correct.

<core_client_version>6.6.20</core_client_version>
<![CDATA[
<message>
Unzul�ssige Funktion. (0x1) - exit code 1 (0x1)
</message>
<stderr_txt>
# Using CUDA device 1
# Device 0: "GeForce GTX 285"
# Clock rate: 1476000 kilohertz
# Total amount of global memory: 1073741824 bytes
# Number of multiprocessors: 30
# Number of cores: 240
# Device 1: "GeForce 8800 GTS"
# Clock rate: 1188000 kilohertz
# Total amount of global memory: 671088640 bytes
# Number of multiprocessors: 12
# Number of cores: 96
Cuda error in file 'nonbonded.cu' in line 189 : invalid device symbol.

</stderr_txt>
]]>

If you look at the table of cards that can run GPUGRID (here), you'll see that the 8800GTS has an older chipset (G80) that can not perform the necessary calculations.

The error that you see is the correct result with that card.

Mike

____________
Want to find one of the largest known primes? Try PrimeGrid. Or help cure disease at WCG.

how can i limit the treads to the gtx 280?

I posted this:

On GPU Grid a situation arose where a participant has multiple CUDA devices in the system, but only one of them is capable of running GPU Grid tasks. This occurs because GPU Grid requires a later version of CUDA than other projects.

System error information:

i get an error after few seconds on any task. But only on the 8800 GTS.
GTX 285 work correct.

<core_client_version>6.6.20</core_client_version>
<![CDATA[
<message>
Unzul�ssige Funktion. (0x1) - exit code 1 (0x1)
</message>
<stderr_txt>
# Using CUDA device 1
# Device 0: "GeForce GTX 285"
# Clock rate: 1476000 kilohertz
# Total amount of global memory: 1073741824 bytes
# Number of multiprocessors: 30
# Number of cores: 240
# Device 1: "GeForce 8800 GTS"
# Clock rate: 1188000 kilohertz
# Total amount of global memory: 671088640 bytes
# Number of multiprocessors: 12
# Number of cores: 96
Cuda error in file 'nonbonded.cu' in line 189 : invalid device symbol.

It is critical that there be a way to limit or control the scheduling of tasks so that GPU Grid tasks are not scheduled on a GPU that cannot run those tasks even if the targeted GPU can run tasks from other projects.

GPU Grid thread: http://www.gpugrid.net/forum_thread.php?id=962#8668

At risk of getting my head taken off ..... :)

In this kind of situation is it not the Project app responsibility to test suitability before running on a GPU? BOINC could be at this for ever and a day as Projects requirements develop/come on stream, CUDA versions change etc etc. Its almost a case of "hard coding" a special inside BOINC for GPUGRID, which seems a little unlikely.

(My thought is that the app marks it "Device Incompatible" and returns it. The mechanism for that exists already [aka "compute error"], just a case of a new information label so the user [and project] realise whats happened)

Regards
Zy

Boinc has the control to limit the number of CPU's used so why could it not limit the number of GPU's or the ability to specify which GPU's are disabled i.e. CUDA 0, CUDA 1 etc. through the cc_config.xml file.

Rob

Boinc has the control to limit the number of CPU's used so why could it not limit the number of GPU's or the ability to specify which GPU's are disabled i.e. CUDA 0, CUDA 1 etc. through the cc_config.xml file.