GPUs on Proxmox: PCIe passthrough and NVIDIA vGPU explained

GPUs in Proxmox have gone from a tinkering topic to an enterprise feature: since NVIDIA officially supports Proxmox VE as a vGPU platform, AI and VDI workloads can be virtualized cleanly. There are two fundamentally different paths - this guide puts both in context (as of July 2026).

Path 1: PCIe passthrough - one card, one VM

With passthrough, Proxmox hands the entire GPU to one VM. Prerequisites: IOMMU (Intel VT-d or AMD-Vi) must be present on the CPU and enabled in the BIOS/UEFI. On kernels before 6.8, intel_iommu=on belongs in the kernel parameters; since 6.8 it is the default on Intel systems. Add the VFIO modules, and the GPU needs its own IOMMU group. For the VM configuration, machine type q35 and OVMF (UEFI) are the established best practice.

The limits are clear: one card belongs to exactly one VM, the host loses access to it, and a VM with a passed-through device can not be live migrated. In return, passthrough costs nothing extra and works with consumer cards too.

Path 2: NVIDIA vGPU - one card, many VMs

Since vGPU software 18, Proxmox VE is an officially supported NVIDIA vGPU platform (Proxmox announcement of March 19, 2025). As of July 2026: vGPU 18 is already EOL (March 2026), current are vGPU 20 (production, supported until March 2027) and vGPU 19 (LTS, until July 2028). The Proxmox wiki lists tested combinations, including PVE 9.2 with vGPU 19.5 and 20.1.

Supported are the GPUs on the official NVIDIA list - datacenter and pro cards such as the RTX A5000 or the RTX PRO 6000 Blackwell. From the Ampere generation on, SR-IOV is required; Proxmox ships the pve-nvidia-vgpu-helper and the pve-nvidia-sriov service for this. The vGPU instances are attached to VMs as mediated devices via PCI resource mappings - which lets you define them cleanly per node across the cluster.

Licensing and costs

NVIDIA licenses per concurrent user (CCU) as an annual subscription or perpetual license, or alternatively per GPU per year. The editions: vApps, vPC and RTX vWS with list prices of 10, 50 and 250 US dollars per CCU per year (subscription). Licenses are served by a license server - CLS in the NVIDIA cloud or DLS in your own datacenter. Without a valid license, the vGPU throttles after 20 minutes. Important for emergencies: NVIDIA support tickets require a vGPU entitlement and a Proxmox subscription.

Consumer GPUs: the honest assessment

GeForce cards officially get no vGPU. Community unlocks like vgpu_unlock only work with older architectures, Intel iGPU SR-IOV needs unofficial DKMS modules, and Intel's GVT-g supports only old iGPUs. Legitimate for the homelab - for production we advise against it: fragile updates, no vendor backing, no support path.

Step by step to a GPU VM

1. Enable and verify IOMMU: turn on VT-d/AMD-Vi in the BIOS, set kernel parameters depending on the version, check IOMMU groups.
2. Choose the path: exclusive card via passthrough or shared datacenter GPU via vGPU.
3. Set up drivers and mapping: bind VFIO, or install the NVIDIA host driver plus SR-IOV and PCI resource mapping.
4. Assign to the VM and test: q35 + OVMF, install guest drivers, run real load.

The key commands

Verify IOMMU (PVE wiki: PCI(e) Passthrough):

dmesg | grep -e DMAR -e IOMMU -e AMD-Vi
# Show IOMMU groups of all PCI devices:
pvesh get /nodes/{node}/hardware/pci --pci-class-blacklist ""

For passthrough, load the VFIO modules (since kernel 6.2 without the obsolete vfio_virqfd):

printf 'vfio\nvfio_iommu_type1\nvfio_pci\n' >> /etc/modules
update-initramfs -u -k all && reboot

For vGPU on SR-IOV cards (Ampere and newer) per the PVE wiki: NVIDIA vGPU:

apt install pve-nvidia-vgpu-helper
pve-nvidia-vgpu-helper setup
# Install the vGPU KVM driver from the NVIDIA licensing portal with DKMS:
./NVIDIA-Linux-x86_64-<version>-vgpu-kvm.run --dkms
# Enable virtual functions (adjust the PCI ID, @ALL for all GPUs):
systemctl enable --now pve-nvidia-sriov@0000:01:00.0.service

Then map the virtual functions in the GUI under Datacenter → Resource Mappings with "Use with mediated devices" enabled, and assign them to the VM as a PCI device with an mdev type - via CLI for example qm set <VMID> -hostpci0 01:00.4,mdev=nvidia-660.

In practice: 1,200 users on vGPU

That this holds up is shown by St. Pölten UAS: its "CampusCloud" on Proxmox serves around 1,200 users, with peaks of 200 concurrent, on time-sliced vGPUs running on H100, H200 and RTX PRO 6000 hardware. Students get a GPU VM in under two minutes, GPU utilization is significantly higher than with dedicated workstations - and the platform choice was made deliberately against vendor lock-in.

How WZ-IT does it

GPU virtualization is part of a bigger picture for us: we plan and operate GPU clusters for AI workloads end to end - from hardware selection through GPU servers to complete AI infrastructure with an inference stack. The Proxmox layer underneath is covered by Managed Proxmox from €179.90 per node per month. Whether passthrough, vGPU or a dedicated AI server is the right path is something we clarify in a free initial consultation.