Install NVIDIA GPU drivers on N-series VMs running Linux

Caution

This article references CentOS, a Linux distribution that is End Of Life (EOL) status. Please consider your use and plan accordingly. For more information, see the CentOS End Of Life guidance.

Applies to: ✔️ Linux VMs

To take advantage of the GPU capabilities of Azure N-series VMs backed by NVIDIA GPUs, you must install NVIDIA GPU drivers. The NVIDIA GPU Driver Extension installs appropriate NVIDIA CUDA or GRID drivers on an N-series VM. Install or manage the extension using the Azure portal or tools such as the Azure CLI or Azure Resource Manager templates. See the NVIDIA GPU Driver Extension documentation for supported distributions and deployment steps.

If you choose to install NVIDIA GPU drivers manually, this article provides supported distributions, drivers, and installation and verification steps. Manual driver setup information is also available for Windows VMs.

For N-series VM specs, storage capacities, and disk details, see GPU Linux VM sizes.

Supported distributions and drivers

NVIDIA CUDA drivers

For the latest CUDA drivers and supported operating systems, visit the NVIDIA website. Ensure that you install or upgrade to the latest supported CUDA drivers for your distribution.

Note

The latest supported CUDA drivers for original NC-series SKU VMs is currently 470.82.01. Later driver versions are not supported on the K80 cards in NC.

Note

The Azure NVads A10 v5 VMs only support GRID 14.1(510.73) or higher driver versions. The vGPU driver for the A10 SKU is a unified driver that supports both graphics and compute workloads.

Tip

As an alternative to manual CUDA driver installation on a Linux VM, you can deploy an Azure Data Science Virtual Machine image. The DSVM edition for Ubuntu 16.04 LTS pre-installs NVIDIA CUDA drivers, the CUDA Deep Neural Network Library, and other tools.

NVIDIA GRID drivers

Microsoft redistributes NVIDIA GRID driver installers for NV and NVv3-series VMs used as virtual workstations or for virtual applications. Install only these GRID drivers on Azure NV VMs, only on the operating systems listed in the following table. These drivers include licensing for GRID Virtual GPU Software in Azure. You don't need to set up a NVIDIA vGPU software license server.

The GRID drivers redistributed by Azure don't work on most non-NV series VMs like NC, NCv2, NCv3, ND, and NDv2-series VMs but works on NCasT4v3 series.

For more information on the specific vGPU and driver branch versions, visit the NVIDIA website.

Distribution Driver
Ubuntu 20.04 LTS, 22.04 LTS

Red Hat Enterprise Linux 7.9, 8.6, 8.8, 8.9

SUSE Linux Enterprise Server 15 SP2, 12 SP2,12 SP5

Rocky Linux 8.4
NVIDIA vGPU 16.5, driver branch R535(.exe)

NVIDIA vGPU 16.3, driver branch R535(.exe)

Note

For Azure NVads A10 v5 VMs we recommend customers to always be on the latest driver version. The latest NVIDIA major driver branch(n) is only backward compatbile with the previous major branch(n-1). For eg, vGPU 17.x is backward compatible with vGPU 16.x only. Any VMs still runnig n-2 or lower may see driver failures when the latest drive branch is rolled out to Azure hosts.

NVs_v3 VMs only support vGPU 16 or lower driver version.

Visit GitHub for the complete list of all previous Nvidia GRID driver links.

Warning

Installation of third-party software on Red Hat products can affect the Red Hat support terms. See the Red Hat Knowledgebase article.

Install CUDA drivers on N-series VMs

Here are steps to install CUDA drivers from the NVIDIA CUDA Toolkit on N-series VMs.

C and C++ developers can optionally install the full Toolkit to build GPU-accelerated applications. For more information, see the CUDA Installation Guide.

To install CUDA drivers, make an SSH connection to each VM. To verify that the system has a CUDA-capable GPU, run the following command:

lspci | grep -i NVIDIA

Output is similar to the following example (showing an NVIDIA Tesla K80 card):

lspci command output

lspci lists the PCIe devices on the VM, including the InfiniBand NIC and GPUs, if any. If lspci doesn't return successfully, you may need to install LIS on CentOS/RHEL.

Then run installation commands specific for your distribution.

Ubuntu

Ubuntu packages NVIDIA proprietary drivers. Those drivers come directly from NVIDIA and are simply packaged by Ubuntu so that they can be automatically managed by the system. Downloading and installing drivers from another source can lead to a broken system. Moreover, installing third-party drivers requires extra-steps on VMs with TrustedLaunch and Secure Boot enabled. They require the user to add a new Machine Owner Key for the system to boot. Drivers from Ubuntu are signed by Canonical and will work with Secure Boot.

  1. Install ubuntu-drivers utility:

    sudo apt update && sudo apt install -y ubuntu-drivers-common
    
  2. Install the latest NVIDIA drivers:

    sudo ubuntu-drivers install
    

    Reboot the VM after the GPU driver is installed.

  3. Download and install the CUDA toolkit from NVIDIA:

    Note

    The example shows the CUDA package path for Ubuntu 24.04 LTS. Replace the path specific to the version you plan to use.

    Visit the NVIDIA Download Center or the NVIDIA CUDA Resources page for the full path specific to each version.

    wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
    sudo apt install -y ./cuda-keyring_1.1-1_all.deb
    sudo apt update
    sudo apt -y install cuda-toolkit-12-5
    

    The installation can take several minutes.

  4. Reboot the VM after installation completes:

    sudo reboot
    
  5. Verify that the GPU is correctly recognized (after reboot):

    nvidia-smi
    

NVIDIA driver updates

We recommend that you periodically update NVIDIA drivers after deployment.

sudo apt update
sudo apt full-upgrade

CentOS or Red Hat Enterprise Linux

  1. Update the kernel (recommended). If you choose not to update the kernel, ensure that the versions of kernel-devel, and dkms are appropriate for your kernel.

    sudo yum install kernel kernel-tools kernel-headers kernel-devel
    sudo reboot
    
  2. Install the latest Linux Integration Services for Hyper-V and Azure. Check if LIS is required by verifying the results of lspci. If all GPU devices are listed as expected, installing LIS isn't required.

    LIS is applicable to Red Hat Enterprise Linux, CentOS, and the Oracle Linux Red Hat Compatible Kernel 5.2-5.11, 6.0-6.10, and 7.0-7.7. Refer to the Linux Integration Services documentation for more details. Skip this step if you plan to use CentOS/RHEL 7.8 (or higher versions) as LIS is no longer required for these versions.

    wget https://aka.ms/lis
    tar xvzf lis
    cd LISISO
    
    sudo ./install.sh
    sudo reboot
    
  3. Reconnect to the VM and continue installation with the following commands:

    sudo rpm -Uvh https://dl.fedoraproject.org/pub/epel/epel-release-latest-7.noarch.rpm
    sudo yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
    sudo yum clean all
    sudo yum -y install nvidia-driver-latest-dkms cuda-drivers
    

    The installation can take several minutes.

    Note

    Visit Fedora and Nvidia CUDA repo to pick the correct package for the CentOS or RHEL version you want to use.

For example, CentOS 8 and RHEL 8 need the following steps.

sudo rpm -Uvh https://dl.fedoraproject.org/pub/epel/epel-release-latest-8.noarch.rpm
sudo yum install dkms

sudo wget https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo -O /etc/yum.repos.d/cuda-rhel8.repo

sudo yum install cuda-drivers
  1. To optionally install the complete CUDA toolkit, type:

    sudo yum install cuda
    

    Note

    If you see an error message related to missing packages like vulkan-filesystem then you may need to edit /etc/yum.repos.d/rh-cloud , look for optional-rpms and set enabled to 1

  2. Reboot the VM and proceed to verify the installation.

Verify driver installation

To query the GPU device state, SSH to the VM and run the nvidia-smi command-line utility installed with the driver.

If the driver is installed, Nvidia SMI lists the GPU-Util as 0% until you run a GPU workload on the VM. Your driver version and GPU details may be different from the ones shown.

NVIDIA device status

RDMA network connectivity

RDMA network connectivity can be enabled on RDMA-capable N-series VMs such as NC24r deployed in the same availability set or in a single placement group in a virtual machine (VM) scale set. The RDMA network supports Message Passing Interface (MPI) traffic for applications running with Intel MPI 5.x or a later version:

Distributions

Deploy RDMA-capable N-series VMs from one of the images in the Azure Marketplace that supports RDMA connectivity on N-series VMs:

  • Ubuntu 16.04 LTS - Configure RDMA drivers on the VM and register with Intel to download Intel MPI:

    1. Install dapl, rdmacm, ibverbs, and mlx4

      sudo apt-get update
      
      sudo apt-get install libdapl2 libmlx4-1
      
      
    2. In /etc/waagent.conf, enable RDMA by uncommenting the following configuration lines. You need root access to edit this file.

      OS.EnableRDMA=y
      
      OS.UpdateRdmaDriver=y
      
    3. Add or change the following memory settings in KB in the /etc/security/limits.conf file. You need root access to edit this file. For testing purposes you can set memlock to unlimited. For example: <User or group name> hard memlock unlimited.

      <User or group name> hard    memlock <memory required for your application in KB>
      
      <User or group name> soft    memlock <memory required for your application in KB>
      
    4. Install Intel MPI Library. Either purchase and download the library from Intel or download the free evaluation version.

      wget http://registrationcenter-download.intel.com/akdlm/irc_nas/tec/9278/l_mpi_p_5.1.3.223.tgz
      

      Only Intel MPI 5.x runtimes are supported.

      For installation steps, see the Intel MPI Library Installation Guide.

    5. Enable ptrace for non-root non-debugger processes (needed for the most recent versions of Intel MPI).

      echo 0 | sudo tee /proc/sys/kernel/yama/ptrace_scope
      
  • CentOS-based 7.4 HPC - RDMA drivers and Intel MPI 5.1 are installed on the VM.

  • CentOS-based HPC - CentOS-HPC 7.6 and later (for SKUs where InfiniBand is supported over SR-IOV). These images have Mellanox OFED and MPI libraries pre-installed.

Note

CX3-Pro cards are supported only through LTS versions of Mellanox OFED. Use LTS Mellanox OFED version (4.9-0.1.7.0) on the N-series VMs with ConnectX3-Pro cards. For more information, see Linux Drivers.

Also, some of the latest Azure Marketplace HPC images have Mellanox OFED 5.1 and later, which don't support ConnectX3-Pro cards. Check the Mellanox OFED version in the HPC image before using it on VMs with ConnectX3-Pro cards.

The following images are the latest CentOS-HPC images that support ConnectX3-Pro cards:

  • OpenLogic:CentOS-HPC:7.6:7.6.2020062900
  • OpenLogic:CentOS-HPC:7_6gen2:7.6.2020062901
  • OpenLogic:CentOS-HPC:7.7:7.7.2020062600
  • OpenLogic:CentOS-HPC:7_7-gen2:7.7.2020062601
  • OpenLogic:CentOS-HPC:8_1:8.1.2020062400
  • OpenLogic:CentOS-HPC:8_1-gen2:8.1.2020062401

Install GRID drivers on NV or NVv3-series VMs

To install NVIDIA GRID drivers on NV or NVv3-series VMs, make an SSH connection to each VM and follow the steps for your Linux distribution.

Ubuntu

  1. Run the lspci command. Verify that the NVIDIA M60 card or cards are visible as PCI devices.

  2. Install updates.

    sudo apt-get update
    sudo apt-get upgrade -y
    sudo apt-get dist-upgrade -y
    sudo apt-get install build-essential ubuntu-desktop -y
    sudo apt-get install linux-azure -y
    
  3. Disable the Nouveau kernel driver, which is incompatible with the NVIDIA driver. (Only use the NVIDIA driver on NV or NVv2 VMs.) To disable the driver, create a file in /etc/modprobe.d named nouveau.conf with the following contents:

    blacklist nouveau
    blacklist lbm-nouveau
    
  4. Reboot the VM and reconnect. Exit X server:

    sudo systemctl stop lightdm.service
    
  5. Download and install the GRID driver:

    wget -O NVIDIA-Linux-x86_64-grid.run https://go.microsoft.com/fwlink/?linkid=874272
    chmod +x NVIDIA-Linux-x86_64-grid.run
    sudo ./NVIDIA-Linux-x86_64-grid.run
    
  6. When you're asked whether you want to run the nvidia-xconfig utility to update your X configuration file, select Yes.

  7. After installation completes, copy /etc/nvidia/gridd.conf.template to a new file gridd.conf at location /etc/nvidia/

    sudo cp /etc/nvidia/gridd.conf.template /etc/nvidia/gridd.conf
    
  8. Add the following to /etc/nvidia/gridd.conf:

    IgnoreSP=FALSE
    EnableUI=FALSE
    
  9. Remove the following from /etc/nvidia/gridd.conf if it is present:

    FeatureType=0
    
  10. Reboot the VM and proceed to verify the installation.

Install GRID driver on Ubuntu with Secure Boot enabled

The GRID driver installation process does not offer any options to skip kernel module build and installation and select a different source of signed kernel modules, so secure boot has to be disabled in Linux VMs in order to use them with GRID, after installing signed kernel modules.

CentOS or Red Hat Enterprise Linux

  1. Update the kernel and DKMS (recommended). If you choose not to update the kernel, ensure that the versions of kernel-devel and dkms are appropriate for your kernel.

    sudo yum update
    sudo yum install kernel-devel
    sudo rpm -Uvh https://dl.fedoraproject.org/pub/epel/epel-release-latest-7.noarch.rpm
    sudo yum install dkms
    sudo yum install hyperv-daemons
    
  2. Disable the Nouveau kernel driver, which is incompatible with the NVIDIA driver. (Only use the NVIDIA driver on NV or NV3 VMs.) To do this, create a file in /etc/modprobe.d named nouveau.conf with the following contents:

    blacklist nouveau
    blacklist lbm-nouveau
    
  3. Reboot the VM, reconnect, and install the latest Linux Integration Services for Hyper-V and Azure. Check if LIS is required by verifying the results of lspci. If all GPU devices are listed as expected, installing LIS isn't required.

    Skip this step if you plan to use CentOS/RHEL 7.8 (or higher versions) as LIS is no longer required for these versions.

    wget https://aka.ms/lis
    tar xvzf lis
    cd LISISO
    
    sudo ./install.sh
    sudo reboot
    
    
  4. Reconnect to the VM and run the lspci command. Verify that the NVIDIA M60 card or cards are visible as PCI devices.

  5. Download and install the GRID driver:

    wget -O NVIDIA-Linux-x86_64-grid.run https://go.microsoft.com/fwlink/?linkid=874272
    chmod +x NVIDIA-Linux-x86_64-grid.run
    
    sudo ./NVIDIA-Linux-x86_64-grid.run
    
  6. When you're asked whether you want to run the nvidia-xconfig utility to update your X configuration file, select Yes.

  7. After installation completes, copy /etc/nvidia/gridd.conf.template to a new file gridd.conf at location /etc/nvidia/

    sudo cp /etc/nvidia/gridd.conf.template /etc/nvidia/gridd.conf
    
  8. Add two lines to /etc/nvidia/gridd.conf:

    IgnoreSP=FALSE
    EnableUI=FALSE
    
  9. Remove one line from /etc/nvidia/gridd.conf if it is present:

    FeatureType=0
    
  10. Reboot the VM and proceed to verify the installation.

Verify driver installation

To query the GPU device state, SSH to the VM and run the nvidia-smi command-line utility installed with the driver.

If the driver is installed, Nvidia SMI will list the GPU-Util as 0% until you run a GPU workload on the VM. Your driver version and GPU details may be different from the ones shown.

Screenshot that shows the output when the GPU device state is queried.

X11 server

If you need an X11 server for remote connections to an NV or NVv2 VM, x11vnc is recommended because it allows hardware acceleration of graphics. The BusID of the M60 device must be manually added to the X11 configuration file (usually, etc/X11/xorg.conf). Add a "Device" section similar to the following:

Section "Device"
    Identifier     "Device0"
    Driver         "nvidia"
    VendorName     "NVIDIA Corporation"
    BoardName      "Tesla M60"
    BusID          "PCI:0@your-BusID:0:0"
EndSection

Additionally, update your "Screen" section to use this device.

The decimal BusID can be found by running

nvidia-xconfig --query-gpu-info | awk '/PCI BusID/{print $4}'

The BusID can change when a VM gets reallocated or rebooted. Therefore, you may want to create a script to update the BusID in the X11 configuration when a VM is rebooted. For example, create a script named busidupdate.sh (or another name you choose) with contents similar to the following:

#!/bin/bash
XCONFIG="/etc/X11/xorg.conf"
OLDBUSID=`awk '/BusID/{gsub(/"/, "", $2); print $2}' ${XCONFIG}`
NEWBUSID=`nvidia-xconfig --query-gpu-info | awk '/PCI BusID/{print $4}'`

if [[ "${OLDBUSID}" == "${NEWBUSID}" ]] ; then
        echo "NVIDIA BUSID not changed - nothing to do"
else
        echo "NVIDIA BUSID changed from \"${OLDBUSID}\" to \"${NEWBUSID}\": Updating ${XCONFIG}"
        sed -e 's|BusID.*|BusID          '\"${NEWBUSID}\"'|' -i ${XCONFIG}
fi

Then, create an entry for your update script in /etc/rc.d/rc3.d so the script is invoked as root on boot.

Troubleshooting

  • You can set persistence mode using nvidia-smi so the output of the command is faster when you need to query cards. To set persistence mode, execute nvidia-smi -pm 1. Note that if the VM is restarted, the mode setting goes away. You can always script the mode setting to execute upon startup.
  • If you updated the NVIDIA CUDA drivers to the latest version and find RDMA connectivity is no longer working, reinstall the RDMA drivers to reestablish that connectivity.
  • During installation of LIS, if a certain CentOS/RHEL OS version (or kernel) is not supported for LIS, an error “Unsupported kernel version” is thrown. Please report this error along with the OS and kernel versions.
  • If jobs are interrupted by ECC errors on the GPU (either correctable or uncorrectable), first check to see if the GPU meets any of Nvidia's RMA criteria for ECC errors. If the GPU is eligible for RMA, please contact support about getting it serviced; otherwise, reboot your VM to reattach the GPU as described here. Less invasive methods such as nvidia-smi -r don't work with the virtualization solution deployed in Azure.

Next steps