I read in the CUDA documentation and various forum posts that there is a 5 second time limit for kernels running on cards with an X display attached. I created a test program that launches one kernel with an infinite loop to see how the system reacts to the long running kernel. The following code causes my entire system (running 32-bit Ubuntu 8.04) to freeze, and require a reboot. Is this the expected behavior? I’ve seen several forum posts indicating that a timeout message should be displayed but I am not seeing that.
The watchdog timeout behavior will be highly dependent on your particular HW, OS & code. Assuming that this only reproduces with a single GPU while X is running, then this is expected behavior. If it reproduces without X running, then there might be a bug somewhere, and you’ll need to attach an nvidia-bug-report.log from your system.