crash and TDR

Hi, not sure if this is the correct forum…

I’ll try to keep this question a little general.

I have cuda code running very well on development PCs (3 PCs with VS 2013).
Those PCs were configured to allow nsight debugging sessions, so TDR was disabled,
but I’m mostly running in release, or debug without debugging the kernels.

The same code works on production PCs (3 PCs without VS 2013).

I took a little larger dataset.
it works well in the development PCs, but crashes every time at the same point (cudaMemcpy).

So I’ve enabled TDR on the development PC and it also crashes at the same point of code.

any thoughts? should TDR have any effect when not trying to debug the kernels?
does the crash make any sense to you? should I disable TDR using the registry on production PCs?

I can go more into the specifics if necessary.

thanks,
Matan

If your GPU is subject to TDR, TDR is enabled, and your CUDA kernel runs longer than 2 seconds, it will be killed, regardless of what it is doing or whether it’s a debug or release build. If the resulting error is reported on the next call to cudaMemcpy(), you are missing proper error checking for the kernel launch itself.

Thanks!