I would like to know if there is any way to indicate cuda-gdb which card to use. My problem is the following. Currently I have 3 cards in my computer, detected by the driver as follows:
GPU 0: Quadro FX 5800
GPU 1: nForce 980a
GPU 2: Tesla C1060
When using cuda-gdb the debugger detects GPU 0 as being used by the X system and sends the program to the very limited integrated card instead of my Tesla. Is there any way to change this behavior? Alternatively, is there a way to change the order the system detects the cards?
Assuming this is Linux, the nvidia-smi tool will let you set the 980a to “compute disabled”.
Doing that just causes the debugger to fail D:
error: The CUDA driver failed initialization.
edit:This person here was having a similar problem on a different context. tmurray mentioned a future new functionality to address this byt I don’t know if it’s yet available.
edit2:oh, I guess he meant this. Getting 3.1 beta kind of defeats the purpose though since cuda-gdb was not yet available for 3.1 last time I checked. Is there a way to change the order the system recognizes the cards? I tried playing around a little with mknod and the /dev/nvidia* devices but that didn’t work.