CUDA with linux nvidia-drivers 302.17
today I've upgraded to nvidia-drivers 302.17 on my dedicated low-power crunching machine, running linux on amd64 platform. Host is configured to only run GPU tasks.
It seems that the CUDA app (BRP4cuda32nv270) is not compatible with this driver release. Tasks run for about 10 seconds with full CPU utilization (but this seems to be normal behavior, does the same on older driver revisions), but after that they just "waiting to run".
I've tried to cancel existing tasks, but newly downloaded WUs have the same problem.
What do you think, is this a known compat issue with new nvidia-drivers release or a local problem? Graphics card is quite new as well...
I'm going to switch back to 295 series driver for now, if you need any support (testing, ...) just tell me, then I switch driver again.
Here are some additional informations:
sb@cruncher ~ $ nvidia-smi -L
GPU 0: GeForce GT 640 (UUID: N/A)
sb@cruncher ~ $ uname -a
Linux cruncher 3.4.2-gentoo-r1 #1 SMP Mon Jun 18 09:25:50 CEST 2012 x86_64 AMD E-350 Processor AuthenticAMD GNU/Linux
sb@cruncher ~ $ eix -c boinc
[I] sci-misc/boinc (7.0.28@06/16/12): The Berkeley Open Infrastructure for Network Computing