nvidia

Wrapper scripts to install proprietary NVIDIA drivers on GNU/Linux systems.

Michael McMahon

The process for installing proprietary NVIDIA drivers is unnecessarily complicated across different GNU/Linux operating systems. These scripts make complex installs easier.

Format: nvidia (Version) (Video Output)

Variations:

• cudaburnin.sh runs gpu_burn for one hour. This requires one of the other nvidiaXXX scripts to be run first.
• nvidiaXXX.XXgpu.sh installs NVIDIA drivers and cuda for standard video output from the GPU. Use this as the default even if you are using a compute node.
• nvidiaXXX.XXmb.sh installs NVIDIA drivers and cuda for non-standard video output from the motherboard. This is only useful if you have to use VGA output for IPMI remote management. cuda performance takes a small hit. This is not recommended.
• nvidiaXXX.XXAAAlan.sh installs from a local FTP server.
• nvidiafedora26-1.sh DISCONTINUED
• nvidiafedora26-2.sh DISCONTINUED
• nvidianew.sh is an unfinished development version for full automation. (Help wanted)

Compatible operating systems:

• CentOS 6
• CentOS 7
• Debian 8
• Debian 9
• RHEL 6
• RHEL 7
• Scientific Linux 6
• Scientific Linux 7
• Ubuntu 14.04 Desktop
• Ubuntu 14.04 Server
• Ubuntu 16.04 Desktop
• Ubuntu 16.04 Server
• Ubuntu 17.10 Desktop
• Ubuntu 17.04 Server

By using these scripts, you agree to all EULAs presented by NVIDIA.

Prerequisites

Prerequisites for this script:

0. Disable secure boot.

   On Supermicro boards, set the JPG1 pin to the 1-2 setting to enable the BIOS to switch between Video Output modes. In BIOS, set Advanced > PCIe/PCI/PnP Configuration > VGA Priority to either Offboard for GPU video output or Onboard for motherboard video output. Cuda testing shows that Offboard improves performance slightly with GNU/Linux drivers even if the only GUI is a command shell. The only time that Onboard should be used is when the machine will be controlled through IPMI. This exception to the rule can be avoided by remotely controlling the system through SSH.

1. Install the system (with Compatibility Libraries and Development Tools or build-essential if applicable).

2. Update all software.

   • CentOS / Scientific Linux

   su
   yum update -y
   

   • Fedora dnf -y update

   • Debian based systems

   sudo apt update
   sudo apt upgrade -y
   sudo apt-get dist-upgrade -y
   

3. Install Compatibility Libraries and Development Tools or build-essential

   • CentOS / Scientific Linux 6

   yum groupinstall -y Development\ Tools
   yum groupinstall -y Compatibility\ Libraries
   

   • CentOS / Scientific Linux 7

   yum groups install -y Development\ Tools
   yum groups install -y Compatibility\ Libraries
   

   • Debian based systems

   sudo apt-get install -y ledmon build-essential
   

4. Boot into the correct runlevel with nomodeset rdblacklist nouveau.

   • Reboot and edit grub temporarily (press arrow keys up and down repeatedly during boot)
   • Press e on the top entry to edit temporarily. Edit the line that starts with linux. Add these entries around words like 'ro quiet': nomodeset rdblacklist nouveau 3
   • Note: Ubuntu Desktop requires editing /etc/default/grub and running update-grub or backing up and editing /boot/grub/grub.cfg with: nomodeset rdblacklist nouveau 2 text

Running the script

To run these scripts:

• Boot into your GNU/Linux distro with runlevel 2 or 3.
• Login as root.
• Download the script and change to its directory.
• bash nvidia39087gpu.sh
• Read the output of the script and look for errors. nvidia-smi should run at the end and display all of the NVIDIA cards present.
• reboot
• After rebooting, run nvidia-smi again to ensure the install succeeded.

Troubleshooting

• If nvidia-smi fails to load or all of the video cards are not listed, the installer may have ran into a problem. Check the /var/log/nvidia-installer.log file for help and more details. If you cannot solve the problem, contact NVIDIA support.

• Warnings about missing 32 bit libraries are OK on 64 bit systems.

• If secure boot is enabled, you may see "ERROR: Unable to load the kernel module 'nvidia.ko'. This happens most frequently when this kernel module was built against hte wrong or improperly configured kernel sources."

• These scripts will need to be reinstalled with each kernel update if DKMS is not used.

• If gcc is missing, follow the prerequisites section and install Development Tools or build-essential metapackages.

• If kernel headers are missing, install the kernel headers for your kernel.

  Debian based systems: sudo apt install -y linux-headers-$(uname -r)

  Red Hat based systems: yum install -y kernel-devel

• Some cards will show up twice in nvidia-smi.

• If everything seems to be installing correctly and the card does not work, try using a different driver from http://www.nvidia.com/object/unix.html