How to Properly Get NVIDIA Drivers/CUDA Setup for RTX A6000 Workstation

Properly installing NVIDIA drivers and CUDA on an RTX A6000 workstation running Fedora requires using the official RPM Fusion repositories rather than downloading runfiles manually. This approach ensures seamless integration with the kernel, automatic updates via dnf, and compatibility with Wayland display servers. By leveraging the distribution’s package manager, you avoid the common pitfalls of driver mismatches and broken Xorg sessions that often plague manual installations on enterprise-grade hardware.

Prerequisites and Repository Configuration

Before installing any proprietary software, you must enable third-party repositories because Fedora does not ship NVIDIA drivers in its default free-software-only sets. The standard procedure involves enabling the RPM Fusion free and non-free repositories. These repositories host the necessary packages for graphics acceleration and CUDA toolkit components. First, update your system to ensure all existing packages are current. This reduces the risk of dependency conflicts during the driver installation. Then, install the RPM Fusion release packages. On modern Fedora versions, this is typically a single command that enables both free and non-free sources.

sudo dnf install https://mirrors.rpmfusion.org/fedora/rpmfusion-free-release-$(rpm -E %fedora).noarch.rpm https://mirrors.rpmfusion.org/fedora/rpmfusion-nonfree-release-$(rpm -E %fedora).noarch.rpm

After enabling these repositories, refresh your local package metadata. This step is crucial because it allows dnf to locate the latest NVIDIA driver versions available for your specific Fedora release. Without this update, the package manager might attempt to install outdated or incompatible driver versions.

Installing the NVIDIA Driver

Fedora provides several driver branches: open-source (Nouveau), stable proprietary, and development proprietary. For an RTX A6000 workstation requiring stability and full CUDA support, you should install the nvidia-driver package from the non-free repository. Avoid the nvidia-open package unless you specifically require open-source kernel modules, as they may lack full feature parity with the proprietary branch for professional workloads. Install the driver and the associated firmware packages using the following command:

sudo dnf install nvidia-driver nvidia-firmware

This command pulls in the kernel module, the user-space libraries, and the necessary firmware blobs for the Ada Lovelace architecture found in the A6000. During installation, dnf will handle the generation of the initial ramdisk (initramfs) automatically. You do not need to manually rebuild the boot image as you would with manual runfile installations. Once the installation completes, reboot your system. Upon restarting, verify that the proprietary driver is active by checking the kernel modules loaded at boot. You can confirm this by running:

nvidia-smi

If the command outputs a table showing your GPU model, driver version, and memory usage, the driver installation was successful. If you encounter a black screen after reboot, ensure that Secure Boot is either disabled or that you have signed the NVIDIA kernel modules with a Machine Owner Key (MOK).

Setting Up CUDA Toolkit

With the base driver installed, the next step is to add the CUDA toolkit. NVIDIA provides dedicated repositories for CUDA development on Fedora. These repositories contain the compiler tools, libraries, and samples required for building applications. Add the CUDA repository to your system configuration. You can use the provided script from NVIDIA’s documentation or manually configure the repo file. Ensure you select the correct repository tag corresponding to your Fedora version. After adding the repo, install the CUDA toolkit meta-package:

sudo dnf install cuda-toolkit

This installs the core compiler (nvcc) and runtime libraries. For deeper development environments, you might also want to install the cudnn library if your workflows depend on deep learning frameworks like PyTorch or TensorFlow. Verify the installation by checking the version of the compiler:

nvcc --version

The output should match the CUDA version supported by your installed driver. Note that the driver must be newer than the minimum version required by the CUDA toolkit version you installed. Regularly updating both the driver and the toolkit via dnf maintains this compatibility.

Troubleshooting Common Issues

Users often face issues with Wayland compositors failing to initialize with NVIDIA drivers. Fedora defaults to Wayland, but some older NVIDIA setups prefer Xorg. If you experience flickering or login loops, switch to the X11 session from the GNOME login screen. Alternatively, ensure you are using a recent enough driver version that supports Wayland properly. Another common issue is missing /dev/nvidia* devices. This usually indicates that the kernel module failed to load. Check dmesg for errors related to nvidia. If Secure Boot is enabled, verify that the MOK has been enrolled. In some cases, disabling Secure Boot temporarily resolves loading issues during the first boot after installation. For those managing virtual machines, ensure that the host system’s driver does not conflict with guest passthrough configurations. While this guide focuses on bare-metal workstation setup, similar principles apply when integrating GPUs into KVM/QEMU environments, though additional steps like PCI passthrough configuration are required.

Related Guides

If you are setting up virtualization alongside your workstation, you may find it useful to review how to add the display driver inside virt-manager for VM guests. Additionally, if you encounter input device failures during boot, check our guide on keyboard not working in Windows or Fedora 37 which covers GRUB-level debugging.

Frequently Asked Questions

Why does nvidia-smi fail after installation?

This typically happens if the kernel module did not load due to Secure Boot restrictions or a mismatched kernel version. Rebooting after enrolling the MOK key usually resolves this, or you can check dmesg logs for specific module loading errors.

Can I use Nouveau instead of the proprietary driver?

While Nouveau is available, it lacks full support for the RTX A6000’s professional features and may not provide optimal CUDA performance. The proprietary driver is strongly recommended for workstation tasks requiring stability and hardware encoding capabilities.

How do I update the driver safely?

Simply run sudo dnf update nvidia-driver. Because you installed via RPM Fusion, the package manager handles dependencies and kernel module rebuilding automatically, ensuring no manual intervention is needed after a reboot.

What if my monitor stays black after installing the driver?

This is often a Wayland compatibility issue. Try logging out and selecting "GNOME on X11" from the session menu at the login screen. If that fails, ensure your monitor cable is connected to the GPU ports, not the motherboard.

Do I need to install CUDA separately from the driver?

Yes, the nvidia-driver package contains only the kernel module and runtime libraries. To compile code using nvcc or access full developer toolkits, you must install the cuda-toolkit package from the separate NVIDIA CUDA repository.