I’m trying to run a demo application on 9500GT with CUDA 3.2.
The application has 1000200,000 runs, each run does 160 sums on global memory.
The application runs correctly for 100020,000 runs, but with the desired number (1000*200,000) it crashes.
Is it possible that if I run it on a stronger device (GTX 460) it will succeed ?
Possibly, probably not though. It depends on the runtime. How long does it take to finish the 1000*20,000 runs?
If the kernel goes beyond the 2 second runtime or so it will be killed by the driver (a limit when using a graphics card that us connected to a screen). Not sure
Thanks !
The 1000*20,000 runs take about 18-19 seconds.
I did notice that the device has a property “stopOnTimeout” (or something like that) that is True, but I didn’t find a way to change it to false or to increase the timeout value.
Is there a way you know ?
Hi,
yes it is possible to turn of the watchdog timer but this is not recommended.
A better solution is to split the work up and call the kernel multiple times on different data-set.