Questions about preempting debugging

Hi there,

I find the description in the “CUDA DEBUGGING TOOLS IN CUDA8”( that there is no any limitation when you are trying to debug multiple instances of CUDA applications on the GPU which is linked to monitors. And I also noted that this feature is only available for Pascal and newer architecture.

So, I was wondering that if I use a Maxwell-based Titan X and CUDA 8.0 tools, can I debug CUDA applications with dynamic parallelism on the display GPU. If not, what are the differences between 5.2- and 6.1-based devices.

Btw, I was also wondering, is there any difference between Linux- and MaxOS-based nsight eclipse tools, especially for those “new” features.

Thank you.

Hi, xiaodai

Pascal support “compute preemption” and this is enabled by default. So you can debug on the display GPU directly.
Maxwell do not support this, if you need to debug on Maxwell(display) GPU. You need make sure the software preemption feature is enabled. And this can be set before your launching debugging from UI.

This is the difference between 5.2- and 6.1-based devices.

As to Linux- and MaxOS-based, I basically think there is no much difference about this “new” features.

Best Regards

Thanks for your reply.