I recently found some rather weird thing on Ubuntu Feisty which could be a problem in later versions of CUDA. To make it clear, AFAIK CUDA runs fine out of the box, but who knows what will happen as CUDA develops further…
The linux-restricted-modules-$KERN_VERSION packages contain an older version of the Nvidia kernel module, which will be build on boot time and placed in /lib/modules/$KERN_VERSION/volatile. The Nvidia installer will place the kernel module (using the same name) in /lib/modules/$KERN_VERSION/kernel/drivers/video. As a result, modprobe will always load the old driver instead of the newer drivers which is not what one might expect after installing the CUDA drivers.
To circumvent this, I did the following:
(1) Disable build of Nvidia module in /etc/default/linux-restricted-modules-common
=> DISABLED_MODULES=“nv”
(2) Remove already built modules:
# rm -rf /lib/linux-restricted-modules/$KERN_VERSION/nv*
(3) Reboot.
Substitute $KERN_VERSION with your kernel version string (at my box currently: 2.6.20-16-generic).
I won’t say it’s a bug. It’s simply a result from using non-Debian packages (the stuff provided by Nvidia) with Debian. Hopefully Nvidia will release deb packages for Debian derivates some time…
I also hope that NVIDIA will release packages of CUDA for Ubuntu and keep them up to date for each release of the distribution or CUDA. It seems reasonable considering how popular Ubuntu is.
Can anyone from NVIDIA comment on the possibility of future Ubuntu support?