GPU Virtualization: Passthrough, vGPU, SR-IOV & MIG in VMs
Giving GPUs to VMs on your own hosts: VFIO passthrough, NVIDIA vGPU licensing and schedulers, SR-IOV, MIG-backed vGPU, KubeVirt and live migration limits.
Overview
Why this topic
Containers on a node all run on the host kernel, and some GPU work needs a kernel of its own: a tenant that brings its own OS image, a regulated workload that must sit in a VM, a virtual workstation, or a team that wants root on the box. On your own hardware that means a hypervisor, usually KVM, and a decision about how each VM gets its GPU.
The choices differ in isolation, density, whether the VM can move between hosts, and what they cost in licenses. An interviewer wants to hear that you know the plumbing (IOMMU, VFIO, virtual functions), the NVIDIA layer on top (vGPU Manager, types, licensing, schedulers), and the operational edges: what survives a reboot, what blocks live migration, and how to prove a VM is getting the GPU time it paid for.
Interview questions
- Pass an H100 through to a KVM guest end to end. What breaks along the way?
- A team wants VMs with GPUs. How do you choose between passthrough, time-sliced vGPU and MIG-backed vGPU?
- VMs on time-sliced vGPUs see their throughput swing when neighbors are busy. Explain the scheduler and how you would tune it.
- Run GPU virtual machines on Kubernetes with KubeVirt. What does the NVIDIA GPU Operator do for you?
- You need to patch hypervisor hosts that run vGPU VMs. What can live migrate, what cannot, and how do you plan the work?
- Compute VMs on vGPU run normally for twenty minutes after boot, then slow sharply. Debug it.
Sign up free to read the full topic
More AI Infrastructure topics
- GPU Architecture & AI Hardware
- GPU Cluster Fabrics: InfiniBand & RoCE
- GPU Driver & Software Stack Release Validation
- GPU Fleet Health: XID Errors, ECC & DCGM
- GPU Node Bring-up & Burn-in
- GPU Quotas, Utilization & Chargeback On-prem
- GPUs on Kubernetes: GPU Operator, Device Plugin, MIG & Time-slicing
- LLM Inference: KV Cache, Batching & Scheduling
- Manufacturing Test Data & Yield Pipelines
- Mixture of Experts (MoE) Inference
- Multi-tenant GPU Clusters & Confidential Computing
- Post-silicon Validation & Bring-up Labs
Start free on Cariara · Coding problems · System design studio