GPU Passthrough

A hardware specification that allows a single PCIe device, such as a GPU, to present itself as multiple separate virtual functions to the hypervisor. Unlike full GPU passthrough, which dedicates an entire physical GPU to one VM, SR-IOV enables multiple VMs to share a single physical GPU with near-native performance. Each virtual function gets isolated access to hardware resources like memory and command channels. This is the foundation for GPU partitioning in modern data centers, though NVIDIA's consumer cards typically restrict this feature to enterprise SKUs like the A100 and H100.

A hardware unit that translates device-visible virtual addresses to physical addresses, providing memory isolation for direct device assignment. For GPU passthrough to function correctly, the IOMMU (called VT-d on Intel platforms and AMD-Vi on AMD platforms) must be enabled in the system BIOS. It prevents a passed-through GPU from reading or writing memory belonging to the host or other VMs, enforcing security boundaries. Without proper IOMMU grouping, devices sharing the same group cannot be passed through independently.

A Linux kernel framework that exposes direct device access to userspace applications and virtual machines. VFIO is the modern replacement for the legacy KVM PCI passthrough mechanism. It works by binding a physical GPU to the vfio-pci driver, detaching it from the host's graphics stack, and presenting it as a fully isolated device to a QEMU/KVM virtual machine. VFIO relies on the IOMMU for secure memory isolation and supports features like PCIe ACS (Access Control Services) for proper device grouping.

A critical architectural distinction in GPU virtualization:

• Full Passthrough: An entire physical GPU is assigned exclusively to one VM. The guest OS loads the native NVIDIA or AMD driver. This provides the highest compatibility and performance but the lowest density.
• GPU Partitioning: Technologies like NVIDIA MIG (Multi-Instance GPU) or AMD MxGPU slice a single physical GPU into multiple isolated instances at the hardware level. Each partition gets dedicated memory, cache, and compute cores, enabling true multi-tenant GPU sharing without hypervisor software mediation.
• vGPU (Virtual GPU): A time-sliced software approach where a hypervisor driver mediates access, sharing GPU cycles across VMs. This offers flexibility but introduces latency and limits CUDA support.

A kernel patch that forcibly disables PCIe Access Control Services to break IOMMU groups into smaller, isolated units. Many consumer motherboards improperly group multiple PCIe devices—such as a GPU, USB controller, and SATA controller—into a single IOMMU group. Without the ACS override patch, passing through a GPU would require passing through all grouped devices, which is often impossible. This patch is essential for successful GPU passthrough on non-server hardware but carries a security trade-off by bypassing hardware isolation guarantees.

A technology stack that copies the framebuffer from a guest VM's GPU to the host for local display without external capture hardware. Looking Glass uses shared memory via IVSHMEM to transfer rendered frames from the guest GPU to a host application with sub-frame latency. This eliminates the need for a second monitor or a KVM switch when running a Windows VM with GPU passthrough on a Linux host. It supports high refresh rates and minimal CPU overhead, making it viable for interactive workloads like gaming or CAD.