I've now had two work units in as many days error out with some variation on the above error message, one GPUGRID and one PrimeGrid, both times a significant proportion of the way through the WU.

GPUGRID gave me:
Cuda error: Kernel [pme_fill_charges_accumulate] failed in file 'fillcharges.cu' in line 73 : unspecified launch failure.
PrimeGrid gave:
Cuda error: kernel invocation: unspecified launch failure

The card is a 9600GT, apparently very slightly factory overclocked (to 675MHz, from 650 stock), running under 64 bit Ubuntu Linux with 190.53 drivers and BOINC 6.10.17. Temperature monitor's not gone above 70C under GPUGRID and 65C with PrimeGrid, with the fan at about 50%, according to the nVIDIA settings app.

I figure it can't be a project-specific bug, and those temperatures are a long way under the throttling limits. Could it be the factory overclock, and if so how would I reverse that? Or do I just have a dodgy card and I'll have to put up with it?

190.53 is the latest stable driver. Both crashes were while I wasn't using the computer.
Is there a problem with running multiple GPU projects? I mostly enabled GPU WUs from PrimeGrid as a test after the GPUGRID unit errored out, but I can easily turn them off again.
Edit: worked out how to enable clock speed alteration in the nVIDIA settings program and turned the core frequency down to stock speeds; will see if it keeps crashing.

Just had another errored PrimeGrid unit, having clocked down to stock settings and disabled new GPUGRID units until this gets sorted or otherwise (less lost credit from an error in an hour-long unit than in a day-long unit).

Any other suggestions?

Edit: There are 195.36.03 beta drivers available, might they be worth a try, or are they likely to be less stable?

Down-clock the GPU (and and other overclocked part of the system) to reference speeds. Let us know if the problem persists thereafter.

MJH

The GPU is now running at reference settings (950MHz), yes, and nothing else has been overclocked. I've had one computation failure so far in about 12 hours since I did that, which is the same sort of rate I was experiencing before downclocking.

As I said, I'm not accepting any new GPUGRID units, only PrimeGrid's shorter units. I haven't run a GPUGRID unit since the one that crashed. (I can, of course, turn off PrimeGrid's and turn on GPUGRID's, but since both are giving the same error and PrimeGrid's short units lose me less credit when they crash, I'd prefer to keep it this way around.)

The GPUGRID task was the only GPU WU on the system when it crashed; PrimeGrid's tasks were only fetched after that one errored out, according to the message log. I'm pretty sure we can rule out project-specific bugs. (I'm asking here rather than PrimeGrid because, given the error isn't limited to either project, it makes sense to ask the heavily GPU-centric project, rather than the one with a single GPU app).

I doubt it's a general system hardware problem as I've had no CPU computing errors, but I will run further tests and check things are seated properly etc. Temperatures are at maximum 70C, well below the area I'd expect to see problems, and cooler with PrimeGrid (max 65C).

Assuming I don't find anything, I'll try the BETA drivers and report back.

I've had no errors since switching to the beta drivers, and I've completed both PrimeGrid and GPUGRID units. Given an approximate rate of about 1 error every 12 hours before the driver switch, the chances of having no errors in the 48 hours or so I've been running GPU units since switching drivers are about 2%, so I'm pretty confident this is sorted. Thanks for the help, folks.

Glad to hear your problem seems to be solved!

For future reference, here's another thing one can try when you start getting unexplained errors:

Kill the rabbits! Specifically, the dust bunnies in the computer. Over time, dust accumulates all over, but it's really bad when it starts clogging the cooling fins underneath heatsink fans. The system starts loosing lots of cooling ability when that happens, and unpleasant things happen, including errors.

I prefer to use a vacuum rather than compressed air for this, and to reach under the fan I tape a straw onto the end of the flat vacuum attachment. It works wonders.