I want to run cuda on a machine running half a s1070 system (it only has 1 pci-e slot unfortunately). Anyway, the machine has no graphics card and is headless in any case so there is no use for X to run.
The problem is that regular users can’t see the cards until root runs something cuda to get the nvidia devices to initialize.
That is, if a user runs device query, only the emulation device is seen. Next run device query as root, the two tesla devices are seen, running device query as a user again, finally the user sees the device.
Is there a way to avoid the super user interaction here?
I did setup on of the cards for the X server (useful for running X on it) which solves this part, but then there is a limit on the kernel runtime for that card and gdb won’t run on that card either, so I don’t like that solution as well.
Any other approaches?