two graphics cards, nvidea just for CUDA, not display

For Ubuntu Hardy Heron…

I’ve got a mobo with two PCI-e slots, one x4 and one x16.

Can I put my existing ATI card in the x4 slot, a 9800GT in the x16 slot, and just use the GPU in the 9800GT for computational purposes and continue to use the ATI card for display?

Any pitfalls? I’m a total newbie at CUDA, but what little I’ve read implies that there might be a performance gain using the 9800GT just for computation.

Thanks,

–arne

In theory, this should work fine.

Well, there was one problem, which is that my mobo didn’t know how to select one out of two PCI-e video cards. Not a hard problem to solve tho’, I had a PCI Quadro 280 (?, pre-CUDA in any case) in my closet o’ stuff. It’s in there and working fine. When I installed the CUDA driver it explicitly said that this card would be ignored by the driver.

I installed the CUDA 2.1 stuff. But the driver does not seem to be working:

arnet@omocha:~/devel/cuda/v_21/sdk/bin/linux/release$ sudo ./deviceQuery

Not loading nvidia module; not used in /etc/X11/xorg.conf

NVIDIA: could not open the device file /dev/nvidiactl (No such device or address).

cudaSafeCall() Runtime API error in file <deviceQuery.cu>, line 59 : initialization error.

The device exists:

arnet@omocha:~/devel/cuda/v_21/sdk/bin/linux/release$ ls -l /dev/n*

crw-rw-rw- 1 root root 1, 3 2008-12-19 18:52 /dev/null

crw-rw---- 1 root video 195, 0 2009-02-20 18:43 /dev/nvidia0

crw-rw-rw- 1 root root 195, 255 2009-02-20 18:43 /dev/nvidiactl

One more thing:

I have specifically NOT enabled restricted drivers on this system because I thought it might interfere with the NVIDIA supplied drivers.

Still dinking around with it… and wondering if there is some problem with trying the setup this way.

–arne

I’ve figured out what’s happening, but not why. The nvidia kernel module is not being installed. If I explicitly install it like so:

sudo insmod /lib/modules/2.6.24-23-generic/kernel/drivers/video/nvidia.ko

Then CUDA runs.

–arne

This is expected behavior. Your NVS-280 is not supported in the non-legacy driver branch, and as such, the driver will not install.

Thanks. This also explains why I can’t run the “nvidia” driver for the 280 card and use CUDA at the same time. I’m guessing it’s one or the other. So I have to use the slow and limited “nv” driver, which isn’t what I had in mind (but it’s working, I’m posting from the system set up this way).

So… While I can get the two cards to work, the fact that one is “legacy” nvidia and one is “new” nvidia means I can’t get it to work the way I want.

Just to help out the economy I may try to pick up an ATI PCI card and see how that flies. The open-source ATI driver is pretty good and this way I can sidestep this whole issue. I still like the idea of having one card for display, and one just for CUDA computational stuff. At this point I think it’s just because I’m stubborn…

–arne