Hi, I am sitting next to (and have access to more) multigpu boxes running Ubuntu+CUDA3.1. My office machine has a GTX295 and a TeslaC2050 card, both functioning. I use the machine to help develop my python bindings, PyCULA a GPGPU LAPACK wrapper. The GPUs have a lot of downtime, and considering the Tesla was donated to our team from Nvidia for some combined computing projects, I would like to do my part and contribute its downtown to some BOINC projects.

I downloaded BOINC, and am running AQUA, Milkyway, and GPUGRID. I have CUDA3.1 and the mating nvidia 256devdrivers. It is obvious to me that BOINC is not using my cards. I get the following in the messages:

Mon 27 Sep 2010 02:35:06 PM EDT||NVIDIA GPU 0: Tesla C2050 (driver version unknown, CUDA version 3010, compute capability 2.0, 2687MB, 1028 GFLOPS peak)
Mon 27 Sep 2010 02:35:06 PM EDT||NVIDIA GPU 1 (not used): GeForce GTX 295 (driver version unknown, CUDA version 3010, compute capability 1.3, 896MB, 596 GFLOPS peak)
Mon 27 Sep 2010 02:35:06 PM EDT||NVIDIA GPU 2 (not used): GeForce GTX 295 (driver version unknown, CUDA version 3010, compute capability 1.3, 895MB, 596 GFLOPS peak)

I have read various posting regarding multigpu hiccups in windows, but what needs to be done in Linux64? If you guys help me make this fly, I will glady compile a wiki/blog post for others...

Thanks,
Garrett Wright
Temple University Mathematics Research Assistant

By default BOINC uses your best coprocessor only. To have it use all, make a cc_config.xml file in your BOINC Data directory (where client_state.xml file resides) and add into it:

<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
</options>
</cc_config>

Save changes, then exit & restart BOINC to use it. Just re-reading the config file will not work, since GPU decisions are only acted upon at the start of BOINC.

---

Ahhh, I could not determine if that was for windows only. Let me restart! Thanks.

Success, I hear the fans kicking on! Terrific! Thanks again.