Is concurrent cuda-gdb sessions possible in Kepler 3.5 ?

I understand that multiple CUDA processes can run concurrently, using CUDA MPS(nvidia-cuda-mps-control, formerly known as CUDA proxy) in Kepler 3.5 GPUs.

Can anyone confirm if the multiple “cuda-gdb” sessions can happen concurrently in Kepler 3.5 using CUDA MPS ?

If the answer to the above question is NO, then is it possible to concurrently run multiple processes while “ONE” cuda-gdb session is in progress?

Any quick information will be greatly appreciated, as we have decide on buying Kepler 3.0 card vs Kepler 3.5 card based on the above support.


CUDA-GDB is not supported with MPS.

However, without MPS, multiple CUDA-GDB instances can debug separate processes, as long as the set of installed GPUs is exclusively partitioned between the processes using the CUDA_VISIBLE_DEVICES environment variable.

Thanks for the clarification

One additional detail: As of CUDA 5.5, single GPU debugging can be enabled as a beta feature on SM3.5 devices in CUDA-GDB. This means that multiple CUDA-GDB instances can debug separate processes on the same SM3.5 device when this feature is enabled. The processes do not run concurrently, but debugging one process no longer fully blocks another process when this feature is enabled. Again, this is without MPS.