GPU partitioning

GPU partitioning, also called GPU virtualization or GPU-P, divides one physical GPU into fractions and gives each virtual machine a dedicated fraction. In Windows Server 2025 Hyper-V it runs on SR-IOV, which gives each VM a hardware-backed security boundary and a dedicated share of the GPU. Discrete Device Assignment (DDA) is the contrast case. DDA passes an entire PCIe device to one VM and works on Windows Server 2016 and later. A GPU is either assigned by DDA or partitioned, never both.

Microsoft lists these partitionable GPUs: NVIDIA A2, A10, A16, A40, L2, L4, L40, L40S, RTX Pro 6000 Blackwell Server Edition and AMD Radeon PRO V710. The partition count comes from a set the OEM predefines, and a Microsoft sample shows 16, 8, 4 and 2. Live migration needs an IOMMU with DMA bit tracking, such as 4th Gen Xeon Scalable or EPYC 7003 and later. EPYC 7002 can partition but cannot live migrate. Live migration on NVIDIA GPUs also needs the driver from vGPU software 18.x or later.

Partition for many small VMs, such as VDI and edge inference. Use DDA for one VM that needs the whole card. Keep every GPU in a cluster the same make, model, size and partition count.

Sources

Last verified October 30, 2026.