WikiBeck

GPU passthrough on Proxmox VE

4 min read

GPU passthrough on Proxmox VE

GPU passthrough gives a VM direct, exclusive access to a physical GPU, using VFIO to hand the PCI device to the guest instead of the host’s own driver. This is the standard way to get near-native GPU performance in a VM — gaming, AI/ML workloads, video editing, or hardware-accelerated transcoding — on Proxmox VE, for AMD, Intel or NVIDIA GPUs.

Requirements and hardware compatibility

  • Proxmox VE 8.x or higher.
  • CPU virtualization extensions with IOMMU support: Intel VT-d, or AMD-Vi (AMD-V).
  • Motherboard/BIOS: must support IOMMU; ideally also “Above 4G Decoding” and “Resizable BAR” (optional, performance-only).
  • GPU:
    • AMD GPUs generally pass through the most cleanly.
    • NVIDIA consumer GPUs need a workaround for “Error 43” (see below); NVIDIA enterprise/datacenter GPUs may support vGPU/mediated-device passthrough instead of exclusive passthrough.
    • Intel integrated GPUs can be passed through too, with more limited use cases.

BIOS: enable IOMMU (VT-d/AMD-Vi), disable CSM for UEFI boot, optionally enable Above 4G Decoding/Resizable BAR, and update to the latest BIOS version before starting.

Steps

1. Enable IOMMU on the host

Edit /etc/default/grub and set, depending on CPU vendor:

# Intel
GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt"
# AMD
GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on iommu=pt"
update-grub
reboot

Verify:

dmesg | grep IOMMU

2. Isolate the GPU (and its audio function) from the host

Identify the GPU’s PCI IDs:

lspci -nnv | grep -i vga -A 1

Note both the GPU (e.g. 0000:01:00.0) and its associated audio function (e.g. 0000:01:00.1) — GPUs almost always have a paired audio device that must be passed through together.

Blacklist the host’s own GPU driver in /etc/modprobe.d/blacklist.conf:

blacklist nouveau
blacklist nvidiafb
blacklist radeon

Bind both device functions to vfio-pci in /etc/modprobe.d/vfio.conf:

options vfio-pci ids=0000:<GPU_ID>,0000:<AUDIO_ID>
update-initramfs -u -k all
reboot

Verify the device is now bound to vfio-pci instead of its normal driver:

lspci -nnk | grep -A 3 'VGA'

3. Install host drivers (NVIDIA only)

Only needed if the host itself also needs to use the GPU (e.g. for its own transcoding) — for pure passthrough this step is normally skipped, since the host shouldn’t touch the device at all once it’s bound to vfio-pci.

apt update && apt install nvidia-driver nvidia-dkms nvidia-headless-no-dkms
reboot

4. Assign the GPU to a VM

In the Proxmox web UI: select the VM → Hardware → Add → PCI Device → pick the GPU → enable All Functions (binds the GPU and its audio device together) and, if it should be the guest’s primary display, Primary GPU. Save and start the VM.

Advanced

  • Above 4G Decoding / Resizable BAR (BIOS): can improve performance for modern GPUs once passthrough is otherwise working.
  • vGPU / mediated-device passthrough: NVIDIA enterprise GPUs (e.g. Tesla-class) with vGPU licensing, or Intel iGPUs with SR-IOV, can split one physical GPU across multiple VMs instead of handing it to one VM exclusively.
  • Error 43 (NVIDIA consumer GPUs in Windows guests): the driver detects it’s running in a VM and refuses to initialize. Workaround — add to the VM’s config (/etc/pve/qemu-server/<VMID>.conf): 
    args: -cpu host,vendor_id=FakeVendorID
    Patched drivers or unlock scripts exist as alternatives but sit in a legal gray area — prefer the vendor-ID spoofing approach above where it’s sufficient.

Troubleshooting

  • Check IOMMU groups — a device sharing a group with something else the host needs can block passthrough of just that one device: 
    find /sys/kernel/iommu_groups/ -type l
  • dmesg is the primary source for driver bind/unbind and VFIO errors.
  • AMD GPUs: watch for “reset bugs” (the GPU failing to reset cleanly when a VM stops/restarts) — usually fixed by a firmware update, occasionally needs a vendor-specific reset workaround.
  • PCI address drift breaking hypervisor passthrough — a separate failure mode where a passthrough config that pins a device by raw PCI bus address breaks after that address gets reassigned to a different device across a reboot; a real risk once passthrough is set up this way.

Sources

Graph View

Backlinks