I tried both the NVIDIA way, and the source compile way described above, and found both of them troubling. For example, the NVIDIA way didn't recognize my card, immediately making things complicated. On the other hand, the Debian source compile way described above caused one processor core of my dual-core setup to "vanish" when the underlying kernel was recompiled, probably because of some mysterious flag that I should have set and that had (of course) not been mentioned in the HOW-TO. Therefore I cannot advise either of them unless you want to fiddle with (and learn to know) some funny build settings and waste loads of time on stuff you probably never wanted to know.
However, I found a very easy and foolproof way buried in the long NVIDIA HOW-TO of the debian wiki
! It should be perfect for anyone who
a) doesn't need a specific NVIDIA driver version - most can probably live very well with the ones provided in the Debian repositories
b) has an aversion of compiling kernel packages whose compile flags and side effects probably noone fully (!) understands
c) doesn't like to play around with runlevels either
If the above is true for you, try this (without quotes):
- "aptitude install nvidia-kernel-2.6.32-5-686" (version in package name must match the kernel you're running, check using "uname -a" if unsure)
- "aptitude install nvidia-glx" (neccessary only first time)
- "aptitude install nvidia-xconfig" (neccessary only first time)
- "nvidia-xconfig" (neccessary only first time)
- "glxinfo |grep rendering" -> if everything worked, the output should be "direct rendering: Yes"
If you have more than one kernel installed (I grub start the 686 kernel to run the machine, but like to have an additional 486 kernel as an gnome-capable backup if I somehow manage to break the 686 kernel) you need to repeat the steps not marked with "neccessary only first time" after booting the respective kernels.