Maximum Kernel Duration? Is there one?

It there a maximum amount of time that a kernel can run for? I am getting “The launch timed out and was terminated.”

Surely a restriction like this is pretty severe, is there any way around it if it exists.

If you are running CUDA kernels on a device which is also running a display, then there is a display driver imposed time limit (in the 5-10 seconds range depending on operating system). The only away around it is to break running kernels down into a sequence of smaller operations, or run your kernels on a dedicated CUDA device.

So I could circumvent this by plugging my monitor into onboard graphics for example?

Yes, but it depends on what operating system you’re running. If you’re using linux, you can plug the monitor into the onboard graphics (and you may need to only use the console as well, I can’t remember) but then you can run kernels for as long as you’d like. On Windows, you are stuck…it only supports one graphics driver, so I don’t think you could even use both graphics devices at once; I hear this was fixed in Windows 7 (the ability to use multiple drivers at once), but even then it doesn’t allow graphics devices to run the kernels longer than the watchdog timer is set for.