When upgrading to CUDA 7.0 I noticed that when I start multiples processes that use a GPU on a machine with multiple GPUs they all get assigned to GPU 0. Because the compute mode is exclusive mode then only the first one runs while the others give “invalid device ordinal”’ errors.
Has anyone noticed something similar?
This shows the problem: https://gist.github.com/FilipeMaia/663353355ed2e1fa763b
I’m running on Linux with driver version 346.46.
I’ve tried both on GTX 680 and GTX Titan Blacks.