I recently had to install CUDA 10.0 on Ubuntu 18.04.5. Since it was a bit of a pain, I decided to write a guide. Have fun and save your data before you do this.

I worked with a clean installation directly on my machine (i.e., no virtual machine) with Ubuntu 18.04.5 as the only OS. Depending on your installation mode (virtual machine, dual boot, etc.) you may have to slightly tweak this guide.
Please drop me a note, if this guide worked for another mode of installation, too, so I can mention it.

A brief note on notation: The necessary steps will be contained in an enumerated list, the paragraphs mostly give the reasons for the way I chose to do this.

Make sure to avoid data loss

In figuring this out, I had to reinstall Ubuntu several times, among else because my mouse and keyboard were not responding after rebooting after installing CUDA at one point. Therefore, I strongly recommend you do a backup of all important data before starting.

I wrote this tutorial with to the best of my knowledge, but I do not take any responsibility for any data loss or other damages that occured due to this tutorial.

  1. Take necessary precautions w.r.t. data loss.

Install Ubuntu 18.04.5

  1. Obtain a bootable USB-stick or a similar medium with Ubuntu 18.04.5

  2. During installation choose not to install third-party software, since this includes Nvidia-drivers (V. 450) which that are not the version that is required by CUDA 10.0 (V. 410).

  3. Install all available updates during and after finishing installation.

Fix missing firmware

I had to install two missing firmware files for the CUDA installation to work. Please check whether the files /lib/firmware/rtl_nic/rtl8125a-3.fw and /lib/firmware/rtl_nic/rtl8168fp-3.fw exist. (They may, if you upgraded from an older version, but should not, if you did a fresh install.) If they are missing, I suggest following comment #3 here, which solved the issue for me:

  1. Download the missing firmware from git.

  2. sudo update-initramfs -u

  3. Reboot to reload the kernel.

Install xserver-xorg-core

According to comment #24 by kopkop here:

“the package xserver-xorg-video-nvidia-418_418.39-0ubuntu1_amd64.deb requires “xserver-xorg-core”, but the actual package installed natively on Ubuntu 18.04.2 is named “xserver-xorg-core-hwe-18.04”. If you choose to install “xserver-xorg-core” manually, it will uninstall gnome-desktop if you are using desktop edition which is not acceptable.”

In our case, the problem package will be xserver-xorg-video-nvidia-410, but the issue is otherwise the same. I did not follow kopkop’s suggested solution, though, but simply installed xserver-xorg-core and then manually reinstalled ubuntu-desktop. Do not reboot until after ubuntu-desktop is installed ;)

This problem only seems to occur on freshly installed Ubuntu 18.04.X, but not for versions that were upgraded, say from 18.04.1 to 18.04.2. (compare comment #23 by kriegalex in the linked thread above).

  1. Install xserver-xorg-core: sudo apt-get install xserver-xorg-core

    You will see something like this:
    $ sudo apt-get install xserver-xorg-core
    [...]
    The following packages will be REMOVED:
    ubuntu-desktop [...]

    Do not reboot at this point!

  2. Reinstall ubuntu-desktop – simply sudo apt-get install ubuntu-desktop
    This does not remove xserver-xorg-core.

  3. Now reboot.

Revert the kernel to version 4.15.0

CUDA 10.0 requires kernel version 4.15.0. I tried to install it with 5.4.62 (the newest at the time of writing) and 5.4.60 initially, but the system always failed to detect the NVIDIA drivers after installation. I eventually reverted to kernel version 4.15.0 therefore, which worked.
I suspect that a newer 4.X kernel could also work, but I did not try it. Drop me a note, if you did.

We will install the old kernel, boot into it, remove the newer kernels (to be sure) and then set 4.15.0 as default for booting.

  1. Install kernel version 4.15.0 by running

    1. sudo apt-get install linux-generic

  2. We want to access grub to boot into the older kernel. If you can already access grub during booting, you can skip ahead to 3. I am loosely following the top-comment here:

    1. sudo gedit /etc/default/grub

    2. comment out GRUB_TIMEOUT_STYLE=hidden and set GRUB_TIMEOUT to a convenient number of seconds (e.g., 5) or to -1 to avoid automatic booting.

    3. save and exit

    4. sudo update-grub

  3. Reboot into kernel 4.15.0 and remove the newer kernels:

    1. Reboot and take note which kernels aside from 4.15.0. are present in the grub menu. We will remove them.

    2. Boot into 4.15.0 (in “Advanced options for Ubuntu”).

    3. I am not sure, whether it is strictly necessary to remove the newer kernels, I just did it to be sure. (If installing CUDA with the newer kernels present worked for you, drop me a note.)
      sudo apt remove linux-image-5.4.0.62-generic* linux-headers-5.4.0.62-generic* linux-image-5.4.0.42-generic* linux-headers-5.4.0.42-generic*

  4. Set 4.15.0 as default in grub. You can do this either by indicating the desired number in GRUB_DEFAULT (attention: the counter starts from 0) or by indicating the exact name, which is more robust. I will do the latter.
    I am loosely following the anser of flaysomerages here and this comment.
    (This step may reintroduce entries in the grub menu for the other kernels. I found it not to be an issue as long as the default is set correctly.)

    1. Grub 2.02 is installed on Ubuntu 18.04.5 by default. Please verify by running grub-install --version. (If your version is below 2.00 you have to enter something like “Advanced options for Ubuntu>Ubuntu, with Linux 4.15.0-132-generic” below.)

    2. get the corresponding identifier(s) from grub.cfg:
      sudo view /boot/grub/grub.cfg You want the identifiers corresponding to the entries “Advanced options for Ubuntu” and “Ubuntu, with Linux 4.15.0-132-generic”

    3. Set them accordingly in /etc/default/grub: sudo gedit /etc/default/grub
      For me it was
      GRUB_DEFAULT="gnulinux-advanced-f88b7f87-95fa-4373-ba9f-a363ef3c147a>gnulinux-4.15.0-132-generic-advanced-f88b7f87-95fa-4373-ba9f-a363ef3c147a"

    4. (optional) revert the changes to /etc/default/grub made in 2.2., i.e. comment in GRUB_TIMEOUT_STYLE=hidden and set GRUB_TIMEOUT=0

    5. save and exit

    6. sudo update-grub

  5. (optional) You can check that this has worked by running uname -r

Install CUDA 10.0

Now, we can finally install CUDA 10.0.
I will walk you through the package manager installation, closely following the steps in the official CUDA 10.0 guide. I use their titles and numbering for reference and copy the relevant commands, so you can stay on one webpage, if you so wish.
I marked the steps you can skip, if you followed this guide with (skip).

However, I strongly advice that you also look into the official guide since I will focus on what is relevant for this setup only, i.e., I will be skipping explanaitions and some post-installation steps.

Also, if you run into any other issues, have a look into the release notes.

Preinstallation actions

  • 2.1. Verify You Have a CUDA-Capable GPU

    lspci | grep -i nvidia
    This should give you something like this:
    06:00.0 VGA compatible controller: NVIDIA Corporation GP102 [GeForce GTX 1080 Ti] (rev a1) 06:00.1 Audio device: NVIDIA Corporation GP102 HDMI Audio Controller (rev a1)

  • (skip) 2.2. Verify You Have a Supported Version of Linux

    uname -m && cat /etc/*release
    Since you should have the same setup as I have, you should see:
    x86_64
    DISTRIB_ID=Ubuntu
    DISTRIB_RELEASE=18.04
    DISTRIB_CODENAME=bionic
    DISTRIB_DESCRIPTION="Ubuntu 18.04.5 LTS"
    NAME="Ubuntu"
    VERSION="18.04.5 LTS (Bionic Beaver)"
    ID=ubuntu
    ID_LIKE=debian
    PRETTY_NAME="Ubuntu 18.04.5 LTS"
    VERSION_ID="18.04"
    HOME_URL="https://www.ubuntu.com/"
    SUPPORT_URL="https://help.ubuntu.com/"
    BUG_REPORT_URL="https://bugs.launchpad.net/ubuntu/"
    PRIVACY_POLICY_URL="https://www.ubuntu.com/legal/terms-and-policies/privacy-policy"
    VERSION_CODENAME=bionic
    UBUNTU_CODENAME=bionic

  • 2.3. Verify the System Has gcc Installed
    gcc is not installed by default, so we have to install it.
    You can simply install the package build-essential (it will be installed later, anyway.) or you can install gcc directly:
    sudo apt install gcc

    With gcc --version you should now get:
    gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
    Copyright (C) 2017 Free Software Foundation, Inc.
    This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

    In the official CUDA guide it is indicated that gcc 7.3.0 is required (see Table 1), but 7.5.0 worked fine for me, as well.

  • (skip) 2.4. Verify the System has the Correct Kernel Headers and Development Packages Installed

    uname -r
    Should give you
    4.15.0-132-generic.

    If it did not, but you installed kernel version 4.15.0 in the previous step, make sure you booted the correct kernel.

    sudo apt-get install linux-headers-$(uname -r)

    You should not need to install anything since we just got the headers by installing linux-generic.

  • 2.5. Choose an Installation Method

    I’ll walk you through the package manager installation. There. Choosing done ;P

  • 2.6. Download the NVIDIA CUDA Toolkit

    1. Download the Toolkit and the patch from the official archive. Click through the options, we need the local .deb.

    2. Verify the files. The checksums can be found here

    md5sum cuda-repo-ubuntu1804-10-0-local-10.0.130-410.48_1.0-1_amd64.deb
    md5sum cuda-repo-ubuntu1804-10-0-local-nvjpeg-update-1_1.0-1_amd64.deb

  • (skip) 2.7. Handle Conflicting Installation Methods

    Skipped since there is no CUDA on a clean install.

Installation

  • 3.6.2. Install repository meta-data. (The public key is installed in the next step.)

    sudo dpkg -i cuda-repo-ubuntu1804-10-0-local-10.0.130-410.48_1.0-1_amd64.deb
    sudo dpkg -i cuda-repo-ubuntu1804-10-0-local-nvjpeg-update-1_1.0-1_amd64.deb

  • 3.6.3. Installing the CUDA public GPG key

    sudo apt-key add /var/cuda-repo-10-0-local-10.0.130-410.48/7fa2af80.pub

  • 3.6.4. Update the Apt repository cache

    sudo apt-get update

  • 3.6.5. Install CUDA

    sudo apt-get install cuda
    The patch will be installed automatically.

Postinstallation actions

  • 7.1.1. Environment Setup

    Add the path to CUDA to the PATH variable.
    Citing Saurabh’s answer in this thread:

    For permanently setting the path:
    $ gedit ~/.bashrc
    the file will load. In that file go to bottom and paste this::

    (We have to use 10.0 instead of 8.0 in the next step.)

    export PATH=/usr/local/cuda-8.0/bin:$PATH
    save and close.

  • 7.2.1. Install Persistence Daemon

    Reboot, if you have not done so, before.

    The command suggested in the CUDA guide did not work for me. I instead followed the suggestion by andorjkiss here. (This needs to be run as root user, sudo alone does not work.)

    $ sudo -i
    # nvidia-smi -pm 1
    Enabled persistence mode for GPU 00000000:06:00.0.
    All done.
    # exit

    You should not experience any issues with nvidia-smi, if you followed my guide. Check if the persistencing worked with

    systemctl status nvidia-persistenced.service

    you should see something like this:

    ● nvidia-persistenced.service - NVIDIA Persistence Daemon
    Loaded: loaded (/lib/systemd/system/nvidia-persistenced.service; static; vendor
    Active: active (running) since Sun 2021-01-17 11:34:52 CET; 12min ago
    Process: 1092 ExecStart=/usr/bin/nvidia-persistenced --user root --no-persistenc
    Main PID: 1093 (nvidia-persiste)
    Tasks: 1 (limit: 4915)
    CGroup: /system.slice/nvidia-persistenced.service
    └─1093 /usr/bin/nvidia-persistenced --user root --no-persistence-mode -

  • 7.2.2. Install Writable Samples

    /usr/local/cuda-10.0/bin/cuda-install-samples-10.0.sh <dir>

    According to the CUDA guide (see Section 7.2.2.) you can also install the package cuda-samples-10-0, which will create a read-only copy of the samples in /usr/local/cuda-10.0/samples. I did not try it, but wanted to mention it for the sake of completeness.

  • 7.2.3. Verify the Installation

    (I strongly recommend doing this to be on the safe side.)

    • 7.2.3.1. Verify the Driver Version

      cat /proc/driver/nvidia/version

      You should get

      NVRM version: NVIDIA UNIX x86_64 Kernel Module 410.48 Thu Sep 6 06:36:33 CDT 2018
      GCC version: gcc version 7.5.0 (Ubuntu 7.5.0-3ubuntu1~18.04)

    • 7.2.3.2. Compiling the Examples

      1. The CUDA toolkit should not be installed, yet. You can check with

        nvcc -V

        and install it with

        sudo apt install nvidia-cuda-toolkit

        Now, nvcc -V should give you

        nvcc: NVIDIA (R) Cuda compiler driver
        Copyright (c) 2005-2017 NVIDIA Corporation
        Built on Fri_Nov__3_21:07:56_CDT_2017
        Cuda compilation tools, release 9.1, V9.1.85

      2. Move to <dir>/NVIDIA_CUDA-10.0_Samples and

        make

        This step took roughly 15min on my setup.

      3. Afterwards run
        <dir>/NVIDIA_CUDA-10.0_Samples/bin/x86_64/linux/release/deviceQuery

        Check that the device was found (and that it actually is the one on your system – top of the block) and that the tests passed (last line).

      4. and run

        <dir>/NVIDIA_CUDA-10.0_Samples/bin/x86_64/linux/release/bandwidthTest

        Again, the main point is that the tests passed (bottom of the entry).

And there you go, CUDA 10.0 is now installed on your Ubuntu 18.04.5.