How to Install Ubuntu 24.04 Real-Time Kernel (PREEMPT_RT) on a Dual-GPU Laptop

Installing the official Ubuntu real-time kernel (ubuntu-realtime) on Ubuntu 24.04 LTS for low-latency work—such as $1\text{ kHz}$ EtherCAT motor control or robotics—is straightforward, but it often hits a severe snag on laptops with NVIDIA graphics.

During installation, dpkg invokes DKMS to compile proprietary NVIDIA kernel modules against the new PREEMPT_RT kernel. Proprietary drivers (for example nvidia-driver-580) frequently fail to build against hard real-time kernel headers, so apt exits with status 11 and leaves dpkg broken.

This post covers the full install path, why the progress bar sits at 2%, how to bypass the DKMS failure, and how to bring NVIDIA back on the generic kernel without installing it on the realtime kernel.

• Prerequisites
• Step 1: Install ubuntu-realtime from Main Repositories
• Step 2: Why Installation Appears Stuck at 2%
• Step 3: Handling the NVIDIA DKMS Build Failure
  • Why This Happens
  • The Fix: Unbind DKMS and Finish dpkg
• Step 4: Restore NVIDIA on the Generic Kernel Only
• Step 5: Boot into the Real-Time Kernel
• Step 6: Verify PREEMPT_RT and the Dual-Boot Split
• Key Takeaways

Prerequisites

• OS: Ubuntu 24.04 LTS (Noble Numbat)
• GPU setup: Integrated GPU (e.g. Intel UHD Graphics) plus a discrete NVIDIA GPU

Step 1: Install ubuntu-realtime from Main Repositories

On Ubuntu 24.04 LTS, the real-time kernel packages are in the standard archives. No Ubuntu Pro subscription is required:

# 1. Update package lists
sudo apt update

# 2. Install the real-time kernel metapackage
sudo apt install ubuntu-realtime

Step 2: Why Installation Appears Stuck at 2%

During apt install, the progress bar may sit at Progress: [ 2%] for several minutes while extracting linux-realtime-headers.

• Why it slows down: Header packages contain roughly 40,000–50,000 tiny .h and .c files. Updating inode metadata and issuing synchronous disk writes (fsync) for each file creates a heavy I/O bottleneck—even on fast NVMe SSDs.
• What to do: Do not interrupt apt or kill the process. Let extraction finish.

Step 3: Handling the NVIDIA DKMS Build Failure

If proprietary NVIDIA drivers are installed, the install typically stops with an error like:

Error! Bad return status for module build on kernel: 6.8.1-1015-realtime (x86_64)
Consult /var/lib/dkms/nvidia/580.159.03/build/make.log for more information.
dkms autoinstall on 6.8.1-1015-realtime/x86_64 failed for nvidia(10)
run-parts: /etc/kernel/header_postinst.d/dkms exited with return code 11
dpkg: error processing package linux-headers-6.8.1-1015-realtime (--configure)

Why This Happens

Real-time motion control usually runs on integrated graphics (e.g. Intel UHD) rather than the discrete GPU, to avoid System Management Interrupts (SMIs) that inject latency jitter. Ubuntu’s default post-install hooks still try to build DKMS modules for every installed kernel, so a failed NVIDIA build against PREEMPT_RT leaves dpkg locked.

The Fix: Unbind DKMS and Finish dpkg

1. Unregister the NVIDIA DKMS module build job:

sudo dkms remove -m nvidia -v 580.159.03 --all

Replace 580.159.03 with the version shown in your error log.

--all removes built modules for every kernel—not only realtime. After this, nvidia-smi fails on both the generic and realtime kernels until you restore the generic build (Step 4).

1. Resume and finalize the broken package setup:

sudo dpkg --configure -a

dpkg can then configure linux-image-realtime, build the initramfs, and add the PREEMPT_RT GRUB entry without compiling the incompatible GPU modules.

Step 4: Restore NVIDIA on the Generic Kernel Only

Boot (or stay on) the generic kernel—for example 6.8.0-134-generic—and rebuild NVIDIA for that kernel alone. The driver source under /usr/src/nvidia-* is usually still present after dkms remove, so a plain DKMS install is enough; no apt reinstall of nvidia-driver-* is required.

uname -r   # confirm you are on *-generic, not *-realtime
sudo dkms install nvidia/580.159.03 -k $(uname -r)

Use the same version string as in the earlier error log (or dkms status / /usr/src/nvidia-*).

A successful install places modules under /lib/modules/$(uname -r)/updates/dkms/ and may sign them with your MOK key. That does not load them into the currently running kernel, so nvidia-smi can still report:

NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver.

until you reboot into the same generic kernel.

Do not run the same dkms install against a *-realtime uname -r if you want the realtime boot free of proprietary NVIDIA modules.

After reboot on generic:

nvidia-smi
dkms status

Verified on this machine: nvidia-smi works again on *-generic, and dkms status lists nvidia as installed for that generic kernel only.

Step 5: Boot into the Real-Time Kernel

Reboot:

sudo reboot

1. At startup, hold Esc or Shift to open the GRUB menu.
2. Open Advanced options for Ubuntu.
3. Select Ubuntu, with Linux 6.8.x-xxxx-realtime.

On this boot, integrated graphics handle the display. Verified: nvidia-smi is unavailable on the realtime kernel when NVIDIA was never installed for that uname -r—the desired outcome for low-latency work.

Step 6: Verify PREEMPT_RT and the Dual-Boot Split

After login on realtime:

uname -a
nvidia-smi

Expected uname output:

Linux hostname 6.8.1-1015-realtime #16-Ubuntu SMP PREEMPT_RT Fri Jun 26 ... x86_64 GNU/Linux

When PREEMPT_RT appears in the kernel string, hard real-time preemption is active.

Final check after rebooting into each kernel once:

That split is intentional: use the generic kernel when you need the discrete GPU, and the realtime kernel for EtherCAT / motion control on the iGPU.

Key Takeaways

1. Patience on headers: The 2% “freeze” during header extraction is inode-heavy I/O, not a lockup.
2. Integrated graphics for RT: EtherCAT, ROS 2 motion loops, and similar work should stay on the iGPU to avoid GPU-induced latency.
3. Clean up DKMS hooks: If DKMS blocks dpkg, sudo dkms remove -m nvidia -v <version> --all followed by sudo dpkg --configure -a finishes the real-time kernel install cleanly.
4. Restore NVIDIA on generic only: --all also drops the generic-kernel modules. Rebuild with sudo dkms install nvidia/<version> -k $(uname -r) while on *-generic, then reboot. Confirmed result: nvidia-smi OK on generic, unavailable on realtime.