I am testing the compute exclusive mode with driver version 256.44 on linux 64 bit with Tesla S1070 (or rather 2 GPUs of it):
First I start the Cuda nbody example (from 3.0 Toolkit): nbody -benchmark -n=300000 (which is running on default on device 0)
Then I start a simple program which intention is just to get an automatically chosen gpu device (with compute exclusive mode) by allocating memory on a not specified device; in my case it should then be device 1.
On this simple-program-side everything works as assumed. However my nbody program always crashes with: nbody.cpp(440) : cudaSafeCall() Runtime API error : unspecified launch failure.
Does anyone know why?
If this happens with users codes, that would be really bad. I mean, then one user can abort another user’s program just be starting his own program :-(