NVIDIA vGPU on Proxmox VE

Introduction

NVIDIA vGPU technology enables multiple virtual machines to use a single supported physical GPU.

This article explains how to use NVIDIA vGPU on Proxmox VE. The instructions were tested using an RTX A5000.

Disclaimer

At the time of writing, Proxmox VE is not an officially supported platform for NVIDIA vGPU. This means that even with valid vGPU licenses, you may not be eligible for NVIDIA enterprise support for this use-case. However, Proxmox VE's kernel is derived from the Ubuntu kernel, which is a supported platform for NVIDIA vGPU as of 2024.

Note that although we are using some consumer hardware in this article, for optimal performance in production workloads, we recommend using appropriate enterprise-grade hardware. Please refer to NVIDIA's support page to verify hardware compatibility .

Hardware Setup

We're using the following hardware configuration for our test:

Test System

CPU	Intel Core i7-12700K
Motherboard	ASUS PRIME Z690-A
Memory	128 GB DDR5 Memory: 4x Crucial CT32G48C40U5
GPU	PNY NVIDIA RTX A5000

Some NVIDIA GPUs do not have vGPU enabled by default, even though they support vGPU, like the RTX A5000 we tested. To enable vGPU there, switch the display using the NVIDIA Display Mode Selector Tool. This will disable the display ports.

For a list of GPUs where this is necessary check their documentation.

The installation was tested on the following versions of Proxmox VE, Linux kernel, and NVIDIA drivers:

pve-manager	kernel	vGPU Software Branch	NVIDIA Host drivers
7.2-7	5.15.39-2-pve	14.1	510.73.06
7.2-7	5.15.39-2-pve	14.2	510.85.03
7.4-3	5.15.107-2-pve	15.2	525.105.14
7.4-17	6.2.16-20-bpo11-pve	16.0	535.54.06
8.1.4	6.5.11-8-pve	16.3	535.154.02

It is recommended to use the latest stable and supported version of Proxmox VE and NVIDIA drivers. However, newer versions in one vGPU Software Branch should also work for the same or older kernel version.

Preparation

Before actually installing the host drivers, there are a few steps to be done on the Proxmox VE host.

Tip: If you need to use a root shell, you can, for example, open one by connecting via SSH or using the node shell on the Proxmox VE web interface.

Enable PCIe Passthrough

Make sure that your system is compatible with PCIe passthrough. See the PCI(e) Passthrough documentation.

Additionally, confirm that the following features are enabled in your firmware settings (BIOS/UEFI):

• VT-d for Intel, or AMD-v for AMD (sometimes named IOMMU)
• SR-IOV (this may not be necessary for older pre-Ampere GPU generations)
• Above 4G decoding
• PCI AER (Advanced Error Reporting)
• PCI ASPM (Active State Power Management)

The firmware of your host might use different naming. If you are unable to locate some of these options, refer to the documentation provided by your firmware or motherboard manufacturer.

Note: It is crucial to ensure that both the IOMMU options are enabled in your firmware and the kernel.

Setup Proxmox VE Repositories

Proxmox VE's comes with the enterprise repository set up by default as this repository provides better tested software and is recommended for production use. The enterprise repository needs a valid subscription per node. For evaluation or non-production use cases you can simply switch to the public no-subscription repository. This provides the same feature-set but with a higher frequency of updates.

You can use the Repositoires management panel in the Proxmox VE web UI for managing package repositories, see the documentation for details.

Update to Latest Package Versions

Proxmox VE uses a rolling release model and should be updated frequently to ensure that your Proxmox VE installation has the latest bug and security fixes, and features available.

You can update your Proxmox VE node using the update panel on the web UI.

Blacklist the Nouveau Driver

Next, you want to blacklist the open source nouveau kernel module to avoid it from interfering with the one from NVIDIA.

To do that, add a line with blacklist nouveau to a file in the /etc/modprobe.d/ directory. For example, open a root shell and execute:

echo "blacklist nouveau" >> /etc/modprobe.d/blacklist.conf

Then, update your initramfs, to ensure that the module is blocked from loading at early boot, and then reboot your host.

Setup DKMS

Because the NVIDIA module is separate from the kernel, it must be rebuilt with Dynamic Kernel Module Support (DKMS) for each new kernel update.

To set up DKMS, you must install the headers package for the kernel and the DKMS helper package. In a root shell, run

apt update
apt install dkms libc6-dev proxmox-default-headers --no-install-recommends

Note: If you do not have the default kernel version installed, but for example an opt-in kernel, you must install the appropriate proxmox-headers-X.Y package instead of proxmox-default-headers.

Host Driver Installation

Note: The driver/file versions depicted in this section are only an example, use the correct file names for the chosen driver you're installing.

To begin you need the appropriate host and guest drivers. See their support page on how to get them . Choose Linux KVM as target hypervisor.

In our case we got the host driver:

NVIDIA-Linux-x86_64-525.105.14-vgpu-kvm.run

To start the installation run the installer, we recommend using the --dkms option, ensuring that the module gets rebuilt after a kernel upgrade:

# chmod +x NVIDIA-Linux-x86_64-525.105.14-vgpu-kvm.run
# ./NVIDIA-Linux-x86_64-525.105.14-vgpu-kvm.run --dkms

Follow the steps of the installer, after that you need to reboot your system, for example with

# reboot

Enabling SR-IOV

On some NVIDIA GPUs (for example those based on the Ampere architecture), you must first enable SR-IOV before you can use vGPUs. You can do that with the sriov-manage script from NVIDIA.

# /usr/lib/nvidia/sriov-manage -e <pciid|ALL>

Since that setting gets lost on reboot, it might be a good idea to write a cronjob or systemd service to enable it on reboot.

Here is an example systemd service for enabling SR-IOV on all found NVIDIA GPUs:

[Unit]
Description=Enable NVIDIA SR-IOV
After=network.target nvidia-vgpud.service nvidia-vgpu-mgr.service
Before=pve-guests.service

[Service]
Type=oneshot
ExecStart=/usr/lib/nvidia/sriov-manage -e ALL

[Install]
WantedBy=multi-user.target

Depending on the actual hardware, it might be necessary to give the nvidia-vgpud.service a bit more time to start, you can do that by adding

ExecStartPre=/bin/sleep 5

just before the ExecStart line in the service file (replace '5' by an appropriate amount of seconds)

This will run after the nvidia-daemons were started, but before the Proxmox VE virtual guest auto start-up.

You can save this in /etc/systemd/system/nvidia-sriov.service. Then enable and start it with:

# systemctl daemon-reload
# systemctl enable --now nvidia-sriov.service

Verify that there are multiple virtual functions for your device with:

# lspci -d 10de:

In our case there are now 24 virtual functions in addition to the physical card (01:00.0):

01:00.0 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:00.4 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:00.5 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:00.6 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:00.7 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.0 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.1 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.2 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.3 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.4 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.5 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.6 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:01.7 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.0 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.1 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.2 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.3 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.4 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.5 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.6 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:02.7 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:03.0 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:03.1 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:03.2 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)
01:03.3 3D controller: NVIDIA Corporation GA102GL [RTX A5000] (rev a1)

Guest Configuration

General Setup

First, set up a VM like you normally would, without adding a vGPU.

After you have configured the VM to your liking, shut down the VM and add a vGPU by selecting one of the virtual functions, and selecting the appropriate mediated device type.

For example:

CLI:

# qm set VMID -hostpci0 01:00.4,mdev=nvidia-660

GUI:

To find the correct mediated device type, you can use sysfs. Here is an example shell script to print the type, then the name (which corresponds to the NVIDIA documentation) and the description, which contains helpful information (such as the maximum available instances). Adapt the PCI path to your needs:

#!/bin/sh
set -e

for i in /sys/bus/pci/devices/0000:01:00.4/mdev_supported_types/*; do
    basename "$i"
    cat "$i/name"
    cat "$i/description"
    echo
done

Since pve-manager version 7.2-8 and libpve-common-perl version 7.2-3, the GUI shows the correct name for the type.

If your qemu-server version is below 7.2-4, you must add an additional parameter to the vm:

# qm set VMID -args '-uuid <UUID-OF-THE-MDEV>'

The UUID of the mediated device is auto-generated from the VMID and the hostpciX index of the config, where the hostpci index gets set as the first part and the VMID as the last. If you use hostpci2 on VM 12345 for example, the generated UUID will be:

00000002-0000-0000-0000-000000012345

After that, you can start the VM and continue with the guest configuration. We installed a Windows 10 and Ubuntu 22.04 VM, but it's similar for other supported operating systems.

Windows 10

First install and configure a desktop sharing software that matches your requirements. Some examples of such software include:

• VNC
  many different options, some free, some commercial
• Remote Desktop
  built into Windows itself
• Parsec
  Costs money for commercial use, allows using hardware accelerated encoding
• RustDesk
  free and open source, but relatively new as of 2022

We used simple Windows built-in remote desktop for testing.

Enabling Remote Desktop in Windows 10

Then you can install the Windows guest driver that is published by NVIDIA. Refer to their documentationto find a compatible guest driver to host driver mapping. In our case this was the file

528.89_grid_win10_win11_server2019_server2022_dch_64bit_international.exe

Start the installer and follow the instructions, then, after it finished restart the guest as prompted.

• 

  Starting NVIDIA driver installation

• 

  Accepting the license agreement

• 

  Finishing the installation

From this point on, the integrated noVNC console of PVE will not be usable anymore, so use your desktop sharing software to connect to the guest. Now you can use the vGPU for starting 3D applications such as Blender, 3D games, etc.

• 

  Unigine Valley

• 

  SuperTuxKart

• 

  Blender

Ubuntu 22.04 Desktop

Before installing the guest driver, install and configure a desktop sharing software, for example:

• VNC
  many options. We use x11vnc here, which is free and open source, but does not currently provide hardware accelerated encoding
• NoMachine
  provides hardware accelerated encoding, but is not open source and costs money for business use
• RustDesk
  free and open source, but relatively new as of 2022

We installed x11vnc in this example. While we're showing how to install and configure it, this is not the only way to achieve the goal of having properly configured desktop sharing.

Since Ubuntu 22.04 ships GDM3 + Gnome + Wayland per default, you first need to switch the login manager to one that uses X.org. We successfully tested LightDM here, but others may work as well.

# apt install lightdm

Select 'LightDM' as default login manager when prompted. After that install x11vnc with

# apt install x11vnc

We then added a systemd service that starts the vnc server on the x.org server provided by lightm in /etc/systemd/system/x11vnc.service

[Unit]
Description=Start x11vnc
After=multi-user.target

[Service]
Type=simple
ExecStart=/usr/bin/x11vnc -display :0 -auth /var/run/lightdm/root/:0 -forever -loop -repeat -rfbauth /etc/x11vnc.passwd -rfbport 5900 -shared -noxdamage

[Install]
WantedBy=multi-user.target

You can set the password by executing:

# x11vnc -storepasswd /etc/x11vnc.passwd
# chmod 0400 /etc/x11vnc.passwd

After setting up LightDM and x11vnc and restarting the VM, you can connect via VNC.

Now, install the .deb package that NVIDIA provides for Ubuntu. Check the NVIDIA documentation for a compatible guest driver to host driver mapping.

In our case this was nvidia-linux-grid-525_525.105.17_amd64.deb, and we directly installed from the local file using apt. For that to work you must prefix the relative path, for example ./ if the .deb file is located in the current directory.

# apt install ./nvidia-linux-grid-525_525.105.17_amd64.deb

Then you must use NVIDIA's tools to configure the x.org confiuration with:

# nvidia-xconfig

Now you can reboot and use a VNC client to connect and use the vGPU for 3D applications.

• 

  Unigine Valley

• 

  SuperTuxKart

• 

  Blender

Note: If you want to use CUDA on a Linux Guest, you must install the CUDA Toolkit manually. Check the NVIDIA documentation which version of CUDA is supported for your vGPU drivers. In our case we needed to install CUDA 11.6 (only the toolkit, not the driver) with the file: cuda_11.6.2_510.47.03_linux.run

Guest vGPU Licensing

To use the vGPU unrestricted, you must adhere to NVIDIA's licensing. Check the NVIDIA documentation for how to do that. Make sure the guest time is properly synced, otherwise the guest will not be able to request a license for the vGPU.