When my CUDA code runs w/o NSight, I get the “Driver has stopped responding and was recovered” message and the program hangs. Obviously a memory boundary issue, which I am trying to track down.
When I run it with NSight and memory checking enabled, it finishes normally with no error messages.
This is really frustrating. So I wonder, what does NSight change that would allow memory violations w/o affecting the program exec?
And what kind of memory violations skate by w/o being detected by “CUDA Memory Checker”?
I had this problem with the previous version of NSight and hoped that 4.0 would fix it, but no joy.