I’m setting up a Linux Centos 6.4 box using a Quadro 4000 to drive a display and a K20 to do some CUDA crunching. To that end, I grabbed the latest driver from Nvidia, compiled it, and everything works fine. However, I run into problems once I try to run this setup with the real time kernel. To that end, I apply the well-known nvidia-rt patch to the Nvidia driver code and then compile the module. If I disable the K20 (in the BIOS), the system boots up and runs fine. Enabling the K20 results in a black screen after runlevel 3.
There was one hack I made yesterday that worked just one time: I added a BusID field to the Device block in xorg.conf to hopefully earmark the Quadro to drive the display. Again, this worked only once. Also, I had to use the PCI BusID listed in nvidia-settings rather than the value printed in lspci (why the discrepancy?).
So, is this a familiar problem for anyone? I’d be grateful for any pointers or suggestions.