NVIDIA vGPU on Proxmox VE: RTX PRO Server Edition, L40S and L4 support, setup, licences

Which cards run NVIDIA vGPU on Proxmox VE

NVIDIA supports vGPU on Proxmox VE for the RTX PRO 6000 Blackwell Server Edition, the RTX PRO 4500 Blackwell Server Edition, the L40S and the L4, among other cards: its validated-platforms page for vGPU R595 on Linux with KVM, updated on 29 September 2026, names Proxmox VE as a supported platform and these GPUs as supported devices. The H200 NVL and the RTX PRO 6000 Workstation and Max-Q editions are not on that list and go into a VM by whole-card passthrough. Graphics profiles under RTX vWS, vPC and vApps are the supported case; NVIDIA AI Enterprise with its compute profiles is, in Proxmox's own words, "currently not officially supported with Proxmox VE".

NVIDIA Virtual GPU Software R595 for Linux with KVM, validated platforms and release notes (20.0 to 20.2), updated 29 September 2026; Proxmox wiki "NVIDIA vGPU on Proxmox VE", last edited 13 August 2026; MIG counts from NVIDIA's MIG guide as cited in our MIG articles; all read 24 October 2026.

The RTX PRO 5000 Blackwell 72 GB is on NVIDIA's vGPU list for Red Hat Enterprise Linux with KVM 9.6 only, not on the Linux with KVM page, so it is not a vGPU card for Proxmox. NVIDIA's validated-platforms page for Red Hat Enterprise Linux with KVM notes that the L40S and the RTX PRO 6000 Server Edition ship in display-off mode and must be used in it, which is the factory state of both passive cards.

This page is about running vGPU on Proxmox with the cards we supply. Profile tables per card are in our guides to RTX PRO 6000 Server Edition vGPU profiles, RTX PRO 4500 Server Edition vGPU profiles and L40S vGPU profiles, and the same decision on vSphere is covered in GPUs in VMware vSphere: passthrough or vGPU.

Who supports what: NVIDIA, Proxmox and the subscription

NVIDIA treats Proxmox as one of several Linux with KVM platforms, next to Nutanix AHV, SUSE Linux Enterprise Server, Red Hat OpenStack Platform and others. Its support matrix for that category states that "NVIDIA vGPU software is supported on Linux with KVM platforms only by specific hypervisor software vendors", gives no version numbers and tells readers to "consult the documentation from your hypervisor vendor". The validated-platforms page adds: "For NVIDIA vGPU software support, contact your Linux with KVM hypervisor vendor." The page it links for Proxmox is the Proxmox wiki article on NVIDIA vGPU.

That article sets the terms on the Proxmox side. "Beginning with NVIDIA vGPU Software 18, Proxmox Virtual Environment is an officially supported platform." A support ticket needs "an active and valid Proxmox VE subscription on your cluster" at Basic, Standard or Premium level, plus an active NVIDIA vGPU entitlement. The enterprise package repository also needs a valid subscription per node; the no-subscription repository is meant for evaluation and non-production use. For a company host that serves desktops or GPU VMs to staff, plan both lines: the Proxmox subscription for every node and the NVIDIA licences for the users or GPUs.

Tested Proxmox VE and vGPU versions

Because NVIDIA defers version support to the hypervisor vendor, the Proxmox wiki's table is the reference for which kernel and host driver belong together. The three current rows, as of 13 August 2026:

Proxmox wiki "NVIDIA vGPU on Proxmox VE", supported combinations, last edited 13 August 2026; the 8.4.1 row also lists kernels 6.11.11-2-pve and 6.14.5-1-bpo12-pve. Read 24 October 2026.

The wiki notes that newer releases within a branch should also work on the same or an older kernel. That matters for security. NVIDIA's bulletin of 30 September 2026 lists vGPU releases up to 19.5 and 20.1 as affected by CVE-2026-47489, so 19.6 and 20.2 are the first fixed releases, as our article on NVIDIA driver branches and CUDA versions sets out. A host on the 19.5 row should move to 19.6 within the same branch.

The branch choice depends on the card. NVIDIA's vGPU release table gives vGPU 19, on the R580 driver branch, long-term support until July 2028, and vGPU 20, on R595, production support until March 2027. The RTX PRO 6000 Server Edition runs on both. The RTX PRO 4500 Server Edition and the liquid-cooled RTX PRO 6000 Server Edition arrived in release 20.0, so a Proxmox host with those cards stays on the shorter vGPU 20 branch until NVIDIA names its successor.

Preparing the Proxmox host

The firmware settings are the same as for any vGPU host. The wiki lists VT-d on Intel or AMD-V on AMD, both called IOMMU; SR-IOV; Above 4G decoding; and Alternative Routing ID Interpretation (ARI), which Ampere and later GPUs need. All four cards in the table above are Ada or Blackwell, so all four settings apply.

On the software side, Proxmox ships a helper. pve-nvidia-vgpu-helper is included from pve-manager 8.3.4; pve-nvidia-vgpu-helper setup blacklists the nouveau driver, installs the kernel headers and sets up DKMS. The host driver is NVIDIA's vGPU Manager for Linux KVM, downloaded as described in NVIDIA's vGPU quick start guide ("Choose Linux KVM as target hypervisor when downloading", in the wiki's words) and installed with the --dkms option so that it is rebuilt with each kernel update. Hosts with Secure Boot need the module signed and the key enrolled at the console, which on a server means the BMC's remote console; our guide to the NVIDIA driver with Secure Boot on Linux servers covers the key handling.

On Ampere and newer GPUs, vGPUs are created on SR-IOV virtual functions, which must be enabled before vGPU can be used. NVIDIA's sriov-manage script does this until the next reboot; the persistent way on Proxmox is the systemd service pve-nvidia-sriov@ALL.service, enabled once, or with a PCI address in place of ALL for a single card. lspci -d 10de: then lists the virtual functions next to the physical GPU.

We supply the RTX PRO 6000 and 4500 Server Editions, the L40S and the L4, as cards or in servers built to order; send us your node count, Proxmox VE version and users per host, and our team replies with cards and licence lines.

Giving a VM its vGPU

In the web interface, the virtual functions become a resource mapping under Datacenter, Resource Mappings, with "Use with mediated devices" enabled; at VM start Proxmox takes the first free function. On the command line, qm set attaches a function as hostpci0 with an mdev= type such as nvidia-660, and pvesh get on the mapping lists the types the driver currently offers, which can change with what is already running on the card. The VM has to be shut down while the vGPU is added.

Three details from the wiki save time in practice. The built-in noVNC and SPICE console cannot show the vGPU's display, so a graphics VM needs remote desktop software inside the guest. The guest driver has to match the host driver per NVIDIA's compatibility tables, and on Linux guests the CUDA Toolkit may need a separate installation. And Windows Fast Startup leads to a blue screen on the next boot with the vGPU disabled, so it is turned off in desktop images.

MIG-backed vGPU on the RTX PRO 6000 Server Edition

The wiki documents MIG-backed vGPU on the RTX PRO 6000 Server Edition, with two limits: "MIG instances must be created and managed manually from the CLI", and the placement of vGPUs is handled by the driver and cannot be changed. MIG is switched on per GPU with nvidia-smi -i 0 -mig 1, the available instance profiles are listed with nvidia-smi mig -lgip, and graphics vGPUs need the profiles with the +gfx suffix; the wiki's example creates four of them with nvidia-smi mig -cgi 47,47,47,47 -C. Some profiles exist only on MIG: the wiki names a 2 GB vWS profile among them.

Four MIG instances of 24 GB give four VMs fully separated compute and memory bandwidth; time-sliced vGPUs inside the instances raise the count, up to 48 office desktops per card with 2 GB profiles per NVIDIA's vPC sizing guide, against 32 on time-sliced profiles over the whole card. Our guide to MIG and vGPU VMs per card has the counts for every card. The RTX PRO 4500 Server Edition splits into two MIG halves of 16 GB, but the wiki's MIG section was tested on the RTX PRO 6000 Server Edition only; the L40S and L4 have no MIG and run time-sliced profiles, from L40S-1Q (32 per card) to L40S-48Q (one per card) in NVIDIA's user guide.

Licences on Proxmox: vWS, vPC, vApps, and no AI Enterprise

On Proxmox the licence follows the vGPU series, as on any hypervisor. Q-series profiles need RTX vWS, B-series vPC or vWS, and A-series vApps, per NVIDIA's vGPU user guide; the editions, how NVIDIA counts them and their terms are in our guide to NVIDIA vGPU licensing with RTX vWS, vPC and vApps. The guest obtains its licence from NVIDIA's License System, either the cloud service or a Delegated License Service appliance on your own network, as our article on the NVIDIA vGPU licence server explains. The wiki adds one Proxmox-specific trap: the guest clock must be synchronised by NTP, or the guest cannot request a licence.

Compute profiles are the gap. The C-series types, such as DC-24C on the RTX PRO 6000 Server Edition, come with NVIDIA AI Enterprise, licensed per GPU. The Proxmox wiki states that AI Enterprise drivers behave much like vGPU drivers and the same steps should apply, but that "NVIDIA AI Enterprise is currently not officially supported with Proxmox VE", and NVIDIA's AI Enterprise 8.2 support matrix lists vSphere, Red Hat Enterprise Linux with KVM, Ubuntu, Nutanix AHV and Red Hat OpenStack, without Proxmox. A Proxmox cluster for data scientists or inference VMs therefore has two supported routes: whole cards by passthrough, with the data-centre driver in the guest, or Q-series vWS profiles, which NVIDIA's vGPU licensing covers for CUDA work in a virtual workstation, as our AI Enterprise licensing guide notes. Which NVIDIA software needs AI Enterprise at all is in our overview of free NVIDIA software and AI Enterprise.

Live migration and maintenance

Proxmox VE 8.4, released on 9 April 2025, added live migration for running VMs "which have mediated devices such as NVIDIA vGPU in use", if "hardware and driver support for live migration is equally available on the target node". NVIDIA's own conditions come on top: the target needs the same GPU model and a driver from the same release branch that is the same or newer, and a VM cannot move between different MIG profiles, as our guide to MIG and vGPU per card summarises. In practice that means upgrading the target host first, keeping one card type per cluster or per HA grouping, and keeping free vGPU capacity of each type on another node for maintenance.

GPU passthrough or vGPU on Proxmox

Passthrough hands the whole card to one VM through VFIO, and the guest runs the normal NVIDIA driver. It is the only route for the H200 NVL and for the RTX PRO 6000 Workstation and Max-Q editions on Proxmox, and the simpler one for a VM that needs a full 96 GB card for a model, or several cards for tensor parallelism. It needs no vGPU licence; AI Enterprise software running in the VM, such as NIM in production, still needs its per-GPU licence. Passthrough on hosts with several cards has its own known problems: a Proxmox forum thread of 2025 reports lockups and failed resets after guest shutdown, with the Max-Q most affected, as our article on the RTX PRO 6000 Max-Q in virtual machines describes.

For many small desktops or development VMs, vGPU is the better fit; for a few large compute VMs, passthrough. A cluster can run both on separate cards or hosts, because a card can be passed through or host vGPUs but not both at once. If you are planning a Proxmox GPU cluster, describe the VM types and how many of each, and we propose cards, hosts and licences per type.

What we supply

We supply the RTX PRO 6000 Server Edition, the RTX PRO 4500 Server Edition, the L40S and the L4 as cards with manufacturer warranty, and build them into AI servers to order, assembled and tested before delivery, with the operating system, drivers, CUDA and a container runtime installed on request. vGPU and NVIDIA AI Enterprise licences go on the same quote and invoice as the hardware, as our licensing page explains; quantities and lead time are confirmed in writing with the quote, on one EU contract and invoice.

FAQ