CUDA Configuration on Ubuntu 24.04 LTS (Kernel 6.17 HWE) with RTX 2000 Ada

Ubuntu Support Template

Ubuntu Version:
Example: 24.04.4

Desktop Environment (if applicable):
Example: GNOME

Problem Description:
Hello everyone,
I am considering installing an RTX 2000 Ada Generation with 16GB of VRAM on a workstation running Ubuntu 24.04.4 LTS. Running uname -r reports kernel version 6.17.0-14-generic.

From my checks, the 550 driver branch does not compile correctly on kernel 6.16+. I would like to validate with you the following command sequence to enable the compute stack (CUDA) while maintaining system stability.

1. Cleanup and Repository Setup

I want to ensure there are no conflicts with pre-existing modules:

sudo apt purge nvidia* libnvidia* -y
sudo apt autoremove -y
sudo add-apt-repository ppa:graphics-drivers/ppa -y
sudo apt update

2. Driver Installation (Target Kernel 6.17)

Given the kernel version, I intend to install the 590 branch (or the latest available in the PPA) with DKMS support:

# Install driver and headers for kernel 6.17
sudo apt install nvidia-driver-590 dkms linux-headers-$(uname -r) -y

3. CUDA Stack and Memory Tuning

For compute/inference workloads, I will proceed with toolkit installation and swappiness tuning (essential to prevent desktop freezes if models saturate the 16GB of VRAM):

sudo apt install nvidia-cuda-toolkit -y

# Swappiness tuning to better manage RAM offloading
echo 'vm.swappiness=10' | sudo tee -a /etc/sysctl.conf
sudo sysctl -p

4. Verification

nvidia-smi
nvcc --version

Questions for the community:

  1. Do these steps look correct to you?
  2. Driver 590 Stability on Kernel 6.17: Has anyone experienced “GPU falling off the bus” issues with Ada Lovelace cards on this specific HWE kernel? Were you able to resolve them?
  3. Wayland vs X11: For purely CUDA/compute workflows, are there any known drawbacks to using Noble’s default Wayland session with kernel 6.17?
  4. DKMS: Can anyone confirm that the DKMS module properly handles 6.17 kernel micro-updates without requiring manual driver reinstallation?

Thanks for any technical feedback.

Althouth Ubuntu has announced support for the NVIDIA CUDA toolkit and the distribution of CUDA within Ubuntu’s repositories: Canonical announces it will support and distribute NVIDIA CUDA in Ubuntu | Ubuntu

But I checked and found that this is so far more then the announced target:

The others is on the shame of NVIDIA, missing one offical web page to help users querying and selecting the range of supported versions of OS/Kernel/CUDA/Driver step by step for their cards.

Anyway, I make torch.cuda.is_available() return True on my K1200 card in Ubuntu 24.04.4 now.

The details are described in my reply in NVIDIA forum CUDA Configuration on Ubuntu 24.04 LTS (Kernel 6.17 HWE) with RTX 2000 Ada