I’ve got a computer with one half of a tesla S1070 installed (one interface card, two GPUs) and a long running (20 minutes) single GPU matlab code which I’m trying to run twice
System is setup in exclusive mode, cudaThreadExit is called on exit from matlab code, kernel doesn’t ask for a device (the idea of exclusive mode)
A single matlab instance runs through fine
if I run a second instance (or in fact ANY gpu code), the kernel crashes with an unspecified launch error (4).
The crash happens on the call to another GPU code, not on launch (it can run for 15 minutes, I call radixSort from the SDK and the matlab code crashes with unspecified error)
Any ideas on the problem and how to solve it?