NVIDIA vGPU on Hyper-V with Windows Server 2025: GPU partitioning, DDA, cards and licences

Which Windows Server versions and cards support NVIDIA vGPU on Hyper-V

NVIDIA vGPU on Hyper-V needs Windows Server 2025: NVIDIA's validated-platforms page for vGPU software R595 on Windows Server, updated on 29 September 2026, states that NVIDIA vGPU is supported only on that version, through GPU partitioning (GPU-P). On Windows Server 2022, NVIDIA supports its GPUs only in DDA deployments, where one VM gets a whole card. Of the cards we supply, the NVIDIA RTX PRO 6000 Blackwell Server Edition, the RTX PRO 4500 Blackwell Server Edition, the L40S and the L4 are listed for both modes.

NVIDIA Virtual GPU Software R595 for Windows Server, validated platforms, release notes and product support matrix, updated 29 September 2026; R580 (vGPU 19) validated platforms for Windows Server, updated 29 September 2026; all read 25 October 2026.

The Windows Server documentation on GPU partitioning, last updated in July 2025, names the RTX PRO 6000 Blackwell Server Edition, the L40S, the L4 and older NVIDIA data-centre cards. The RTX PRO 4500 Server Edition arrived later, in NVIDIA's vGPU release 20.0, and is not yet on that older list; NVIDIA's own list is the newer one. We read both on 25 October 2026.

This page is about Hyper-V on Windows Server with the cards we supply. The same questions for other hypervisors are answered in our guides to NVIDIA vGPU on Proxmox VE and to GPUs in VMware vSphere, and the overview of vGPU support by hypervisor and card puts them side by side. Whether the host should run Windows Server at all, or Linux, is the subject of our comparison of Windows Server and Linux for a GPU server.

GPU partitioning and DDA in plain terms

The two modes give a VM a GPU in different ways, and one physical card uses one of them at a time. The Windows Server documentation states it directly: a GPU is assigned "either as DDA or as partitionable GPU, but not both".

GPU partitioning (GPU-P) splits one card into several partitions, and NVIDIA's Virtual GPU Manager on the host turns each partition into a vGPU of a fixed type, such as a 24 GB Q-series vGPU for a CAD workstation or a 2 GB B-series vGPU for an office desktop. The vGPUs share the card's compute by time slicing and each gets its own share of the memory. Hyper-V assigns the partitions to VMs automatically; the documentation notes that a specific partition cannot be chosen for a specific VM.

Discrete Device Assignment (DDA) hands the whole card to one VM, which then runs the NVIDIA driver as if the card were in a physical machine. This is the only mode on Windows Server 2022, and on Windows Server 2025 it remains the simplest route for a VM that needs all 96 GB of an RTX PRO 6000 Server Edition, or several cards.

The guest operating system matters for the choice. NVIDIA's support matrix for a Windows Server 2025 host lists Windows 11, Windows 10, Windows Server 2025 and 2022 guests, Ubuntu 24.04, 22.04 and 20.04, Red Hat Enterprise Linux 9.8, 9.6, 9.4 and 8.10, and SUSE Linux Enterprise Server. Its footnotes add two limits that are easy to miss: for Ubuntu 24.04 and 22.04 and for RHEL 9.x, "GPU pass through is not supported", so these guests take a GPU-P vGPU, not a DDA card; for SUSE Linux Enterprise Server, "only GPU pass through is supported". vApps runs only in Windows guests.

Hyper-V DDA on Windows Server 2022: one whole GPU per VM

On a Windows Server 2022 host, NVIDIA's R580 page supports all GPUs that its vGPU software supports in DDA deployments "only"; the R595 page says the same. A 2022 host can therefore give each VM a whole card, but it cannot share one card between several desktops. The guest list is shorter too: Windows 11 and 10, Windows Server 2022, Ubuntu 20.04, RHEL 8.8 and 8.10 and SUSE Linux Enterprise Server.

For a VDI project this decides the platform. Sixteen office desktops on one RTX PRO 4500 Server Edition need GPU-P, and with it Windows Server 2025 on the host. A department that only needs two or three GPU VMs with a whole card each can stay on 2022 with DDA. Our guide to replacing A40 and A16 vGPU hosts covers the hardware side of moving an older VDI farm to current cards.

Setting up GPU-P on a Windows Server 2025 host

NVIDIA's vGPU user guide (updated 29 September 2026) describes the host setup for Windows Server in one section, and the steps below follow it. They assume Windows Server 2025 with the Hyper-V role, a virtual switch and a VM already in place.

1. Firmware. Enable virtualisation support (VT-d or AMD-V), SR-IOV and Above 4G Decoding. On AMD processors also enable ARI, ACS and AER, which applies to the dual AMD EPYC platforms in our reference configurations.
2. Virtual GPU Manager. Extract NVIDIA's package; the Windows Server files are in the folder GridSW-Windows-Server-GPU-P. In an elevated PowerShell, pnputil /add-driver adds nvgridswhostserver.inf with the options /subdirs /install /reboot. After the reboot, nvidia-smi should show the cards.
3. Partitions. Get-VMHostPartitionableGpu lists the cards that support GPU-P with their names and possible partition counts. Set-VMHostPartitionableGpu with -Name and -PartitionCount sets the count per card, which NVIDIA defines as "the maximum number of vGPUs that can be added to the GPU". The count therefore follows the vGPU type; NVIDIA's example table for the A16 gives one partition for A16-16Q and sixteen for A16-1Q.
4. vGPU series. "By default, only Q-series vGPUs may be created on a GPU", and the Virtual GPU Manager allows only one series per card. For B-series (vPC) or A-series (vApps), set the registry value GridGpupProfileType in the card's driver key to 2 for A-series or 3 for B-series (1 is Q-series) and reboot the host.
5. The VM. With the VM shut down, NVIDIA's steps set -GuestControlledCacheTypes $true, 1 GB of low and 32 GB of high memory-mapped I/O space with Set-VM, then Add-VMGpuPartitionAdapter with minimum, maximum and optimal values for VRAM, encode, decode and compute, within the limits that Get-VMHostPartitionableGpu reports. Get-VMGpuPartitionAdapter confirms the adapter.

Two rules from NVIDIA's known product limitations shape the layout. Windows Server "does not support a mixture of different types of time-sliced vGPUs on the same GPU", so all vGPUs on one card share a series and a memory size. And the guest driver must be compatible with the host: a guest driver newer than the host's Virtual GPU Manager, or two or more major releases older, leaves the vGPU disabled with error code 43. In vGPU 20.2 the Windows Server host package is version 596.84, the Windows guest driver 596.86 and the Linux guest driver 595.91.07. NVIDIA's bulletin of 30 September 2026 makes 19.6 and 20.2 the first fixed releases for CVE-2026-47489, as our article on NVIDIA driver branches and CUDA versions sets out.

Partition counts per card

The partition count is the number of VMs a card serves. Our profile guides give the counts NVIDIA publishes for each card; the figures below are taken from them.

Our guides to RTX PRO 6000 Server Edition and RTX PRO 4500 Server Edition vGPU profiles, from NVIDIA's vGPU user guide and vPC sizing guide; maximum densities, not recommended deployment points.

Because Windows Server allows one series and one memory size per card, a host that serves both CAD users and office desktops puts them on separate cards: for example two RTX PRO 6000 Server Edition cards with DC-24Q for eight CAD seats and two RTX PRO 4500 Server Edition cards with 2 GB vPC desktops for 32 office users. The detailed profile tables are in our guides to RTX PRO 6000 Server Edition vGPU profiles, RTX PRO 4500 Server Edition vGPU profiles, L40S vGPU profiles and L4 vGPU profiles.

We supply the RTX PRO 6000 and RTX PRO 4500 Server Editions, L40S and L4, alone or in servers for Hyper-V hosts; send us the number of desktops and CAD seats per host, and we reply with a card layout and the licence lines.

Licences for vGPU on Hyper-V

The licence follows the vGPU series, as on every hypervisor. NVIDIA's validated-platforms table lists vWS, vPC and vApps for time-sliced vGPUs on our cards and vWS and vApps for DDA; the client licensing guide adds that vPC is not available in pass-through mode. Q-series vGPUs need RTX vWS, B-series vPC or vWS, and A-series vApps. NVIDIA's packaging guide describes these editions per concurrent user (its introduction also mentions a per-GPU annual subscription for vGPU software), and our guide to NVIDIA vGPU licensing with RTX vWS, vPC and vApps explains the counting and the terms.

The guests obtain their licences from NVIDIA License System, either the cloud service or a Delegated License Service appliance on your network; NVIDIA's release notes state that this release family is "incompatible with all releases of the NVIDIA vGPU software license server", the older on-premise product. Our article on the NVIDIA vGPU licence server describes both options.

Compute profiles are the gap on Hyper-V. NVIDIA's known product limitations for Windows Server state that NVIDIA vGPU for Compute is not supported on Windows Server and that "C-series vGPU types are not available". The NVIDIA AI Enterprise 8.2 support matrix, updated 2 September 2026, lists VMware vSphere, Red Hat Enterprise Linux with KVM, Ubuntu and Nutanix AHV as hypervisors, and no Windows hypervisor. As our AI Enterprise licensing guide notes, compute vGPUs are licensed per GPU through AI Enterprise, so on Hyper-V that licence line does not apply to vGPU at all.

AI and compute VMs on Hyper-V

For inference or development VMs, three routes remain on a Windows Server 2025 host.

• A whole card by DDA. The VM gets the full card and the normal driver. Windows guests are supported; among Linux guests, the matrix leaves Ubuntu 20.04, RHEL 8.10 and SUSE Linux Enterprise Server for pass-through on a 2025 host.
• A full-card Q-series vGPU. DC-96Q on the RTX PRO 6000 Server Edition gives one VM the whole card through GPU-P, under an RTX vWS licence, which NVIDIA's licensing guide names for CUDA work and AI development in VMs. This route also covers Ubuntu 24.04 and RHEL 9 guests. NVIDIA supports CUDA Toolkit 13.2 in the guests of this release family.
• Several vGPUs in one VM. NVIDIA supports up to 16 vGPUs per VM on Windows Server 2025, on Blackwell cards only as time-sliced Q-series vGPUs on a card without MIG. NVIDIA's pages make no statement on peer-to-peer traffic between such vGPUs, so we would not plan a tensor-parallel model across them.

Two limits apply to all three. MIG cannot be used: the R595 table marks MIG-backed vGPU as not supported for the RTX PRO Server Editions on Windows Server. And nested virtualisation is not supported, so a guest cannot run the Hyper-V role itself; NVIDIA lists virtualisation-based security and WSL 2 in the guest as supported. For a GPU server that mainly runs language models with vLLM, a Linux host is usually simpler, as our Windows Server or Linux comparison explains.

Live migration, clusters and maintenance

Live migration with GPU-P is new in Windows Server 2025. The Windows Server documentation states: "Beginning with Windows Server 2025, live migration is supported with GPU partitioning." It needs processors with IOMMU DMA bit tracking, for example "AMD EPYC 7003 and later (Milan)" and "4th Gen Intel Xeon SP (Sapphire Rapids)"; EPYC 7002 (Rome) supports GPU-P but not live migration with it. On hosts without that capability, VMs "are automatically restarted where GPU resources are available". The same page describes GPU-P as "designed for standalone servers", with live migration between standalone hosts for planned downtime; for clustering against unplanned downtime it requires Windows Server 2025 Datacenter. All hosts need the same GPU model and the same partition count per card, and the migration traffic falls back to TCP/IP with compression, which costs host CPU time.

NVIDIA adds its own conditions. Migration is disabled for a VM that uses unified memory, a CUDA debugger or a profiler, and unified memory "is enabled by default and cannot be disabled" on the vGPUs that support it: in Linux guests that means the full-card types DC-96Q on the RTX PRO 6000 Server Edition and DC-32Q on the RTX PRO 4500 Server Edition. A Linux VM with a full-card vGPU therefore does not live-migrate and is shut down for host maintenance. NVIDIA also lists a known issue in which a VM migrated from a 20.1 host to a 20.0 host fails, so update the target hosts first.

DDA VMs do not live-migrate at all; the Windows Server documentation on clustered VMs says that live migration of VMs using DDA GPUs "isn't currently supported". A failover cluster with GPU pools restarts them on another node with a free GPU from the pool, which needs Windows Server 2025 Datacenter and "the same make and model of the GPUs across all the servers in your cluster". For maintenance, keep one free card of each type somewhere in the cluster. If you are planning a Hyper-V cluster with GPUs, describe the VM types and how many of each, and we propose cards per host and the licences per type.

Branches: vGPU 19 or vGPU 20

NVIDIA's vGPU release table gives vGPU 19, on the R580 branch, long-term support until July 2028, and vGPU 20, on R595, production support until March 2027. NVIDIA's R580 page for Windows Server lists the RTX PRO 6000 Server Edition since release 19.1 and does not list the RTX PRO 4500 Server Edition, which arrived in 20.0. A Hyper-V farm with RTX PRO 6000 Server Edition, L40S or L4 cards can stay on the long-term branch; a farm with RTX PRO 4500 Server Edition cards runs vGPU 20 until NVIDIA names its successor branch.

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 licences (RTX vWS, vPC, vApps) 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