GPU passthrough

GPU passthrough hands a physical graphics card to one virtual machine using the CPU's IOMMU, VT-d on Intel and AMD-Vi on AMD. The IOMMU isolates the card as its own PCIe device and removes it from the host's control. The guest then loads the card's normal driver and talks to the hardware directly. Nothing else on the box, the host included, can use that GPU while the assignment stands.

Overhead against bare metal is small, typically under 5 percent on compute and rendering, since the hypervisor leaves the data path once the device is assigned. A common build puts two cards in one workstation under Proxmox VE or VMware ESXi. One stays with the host, the second passes through to a Windows guest running CAD or simulation. The mechanism carries different names by platform. Hyper-V calls it Discrete Device Assignment, and KVM calls it VFIO passthrough.

Speccing for this means confirming the board and CPU put the target GPU in a clean IOMMU group. A card grouped with unrelated devices cannot be assigned on its own. Server and workstation boards separate those groups far more predictably than consumer desktop boards.

Sources

Last verified August 29, 2026.