Home lab serverPart 4 of 13

Build note

Ubuntu 24.04 on server hardware

Ubuntu can install a new kernel and still leave a machine that has no working NVIDIA module for that kernel. The package can be present, the reboot can finish, and the desktop can still start without the GPU stack you expected. On this P920, the kernel and the driver have to be treated as one update.

The host that has to keep booting

The system is Ubuntu 24.04.4 LTS on the 6.8 kernel line. The hardware is a dual Xeon workstation with two RTX cards, and the proprietary NVIDIA driver is built through DKMS. DKMS is the system that rebuilds an external kernel module against each installed kernel’s headers.

The inventory captured three kernels with the NVIDIA module built:

Kernel NVIDIA module
6.5.0-35-generic 595.99.02
6.8.0-142-generic 595.99.02
6.8.0-146-generic 595.99.02

Keeping more than one kernel is useful on a workstation. If the newest kernel does not boot cleanly, GRUB still has an older one. It is not a full rollback: the driver package, the initramfs and the running kernel can still disagree.

Check the pieces before the reboot

Before a kernel update, I want three answers, not one:

uname -r
dkms status
nvidia-smi

uname -r names the kernel that is running now. dkms status shows which kernel versions have a module built. nvidia-smi checks that the userspace tools can talk to the loaded NVIDIA module and enumerate the cards.

The check after reboot is just as important. A package being installed does not prove that its module is loaded. On this host, an attempted move to driver 590 left driver 595 installed and running after the reboot. nvidia-smi still answered, so the machine looked healthy, but it was not testing the branch I thought I had selected. The loaded module version, not the package name in a terminal scrollback, is the evidence.

The safety margin is not a snapshot

Secure Boot is disabled on this machine, so unsigned DKMS modules are not being held up by a key-enrollment step. The root filesystem is ext4 and does not have system snapshots. The retained kernels are the safety net for a kernel that fails to start; they do not restore a driver package or undo a broken config.

That is why I do not treat a driver branch change as a routine update. First I check which package and module are active. Then I keep the previous kernel available until the new one has booted and nvidia-smi works on both cards.

The rule that would have prevented the next week

Check the loaded module after every driver change. A package name says what was installed; /proc/driver/nvidia/version, dkms status and nvidia-smi tell me what actually booted. Keep an older kernel, but do not mistake it for a system snapshot.

How this was made

  • Written by: Rupesh, from his own kernel and DKMS inventory, with AI drafting and editing.
  • Versions: from uname -r, dkms status, /proc/driver/nvidia/version and nvidia-smi records on rtr-p920.
  • AI-assisted sections: structure and copy editing.