kernel time out after ~8s on GTX285

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello All,

I have an early 2008 Mac Pro with a GTX285 and am attempting to develop CUDA codes
for scientific computing. It is the only graphics card installed. I have run into a show stopper.
I get the error - once my problem size is large enough, but still modest w.r.t. what the
card can handle –

 the launch timed out and was terminated

I note that my display is attached to the 285. Even if I detach the display and access the machine
remotely I still get the error. In case it matters, I am running 10.6.6 and boot into the 64-bit
OS kernel.

SO here are 2 questions:

  1. Is there anyway to disable the kernel timeout? I see have you can query the card and see that
    the timeout is set, but I don’t see any documentation on changing this ‘bit’ with a cuda call.

  2. If I install, in addition to the 285, a ATI 2600XT (the default card for this machine) and attach
    the monitor to it, will the OS/drivers recognize this setup and disable the kernel timeout on
    the 285? It seems to me that this should work, but I’ve seen posts online that folks have
    not been successful, although they are few and it could just be that they are not doing
    something correctly.

I’ll mention that I do have an 8800GT (which shipped with my machine) but I can’t install it
along with the 285 since there are only 2 6-pin power connectors on the mobo and the 285
uses them both. SO I would have to look for a used 2600XT.

I’ve scoured the web and this forum about this issue and have not found any definitive answers.
Unfortunately, I believe the answer to my first question is no, but believe that the second scenario
is something that should be supported by Apple/Nvidia (if not, I suggest Apple and Nvidia call each
other and work this out!)

I had a similar problem on a linux machine, but there, but exiting X one can run as long as you want.

If Nvidia is serious about CUDA on Mac (and I think they are) this is an issue that deserves to be
address at the vendor level.

It would be great, in addition to anyone else, that an Nvidia engineer could respond to this thread!!
I have Intel compilers, and when posting on their forums it is often the case that an actual Intel
engineer responds with great info and solutions.

Any help would be GREATLY appreciated!

Thanks, Hans

2

Hello Hans,

since nobody else answered, I’ll post what I know. I vaguely remember that on Macs the timeout indeed cannot be disabled even if no display is attached. However, that usually should not pose a problem since you can split the work between multiple kernel invocations, each working on a smaller grid.