Case StudyHow Enkrypt AI ships its AI security platform into air-gapped KubernetesRead the case study
Skip to main content

Glossary

Virtual GPU (vGPU): Hardware Virtualization for Shared GPU Resources


What Is a Virtual GPU (vGPU)?

A virtual GPU, or vGPU, is a share of one physical GPU assigned to a virtual machine. A hypervisor and a vendor driver split the card. Each virtual machine gets a fixed slice of GPU memory and scheduled access to the compute engines. The guest sees a normal GPU and runs its usual drivers and applications.

vGPU began with virtual desktops and graphics workstations. It now also serves AI, inference, and general compute. The most common implementation is NVIDIA vGPU software, which runs on hypervisors such as VMware vSphere, KVM, and Citrix.

Why does vGPU matter now?

Because GPUs are expensive and most workloads do not need a whole card. Giving every virtual machine a dedicated GPU leaves much of the hardware idle. vGPU raises density by letting several machines share one device.

Cloud providers have moved the same way. Fractional GPU virtual machines now let teams rent a slice of an accelerator for small inference, remote workstations, and development. The same economics drive GPU fractioning in Kubernetes. vGPU applies that idea at the hypervisor layer.

How does vGPU work?

A vendor manager driver runs on the hypervisor host. It divides the physical GPU into vGPU instances. Each instance is assigned to one virtual machine. The guest installs a matching guest driver.

Two resources are divided differently:

  • Memory. Each vGPU gets a fixed framebuffer size set by its profile. It is not shared or overcommitted.
  • Compute. The GPU time-slices between vGPUs using a scheduler. Common policies are best effort, equal share, and fixed share.

A profile defines the framebuffer size and the intended workload type. Profiles are grouped by use, such as virtual workstations, virtual PCs, and compute for AI. You pick a profile per virtual machine.

vGPU vs GPU passthrough vs MIG, what is the difference?

The three approaches trade isolation, density, and cost.

ApproachHow it shares the GPUIsolationBest fit
PassthroughOne whole GPU to one virtual machineComplete, nothing is sharedLarge training or one heavy workload
vGPUHypervisor slices memory, time-slices computeVM boundary, shared compute enginesVDI, workstations, many small VMs
MIGHardware partitions into isolated instancesHardware-enforced memory and computeMulti-tenant inference with strict limits

On supported data-center GPUs, vGPU and MIG can be combined. Each vGPU is then backed by a MIG instance, which gives the VM dedicated compute as well as dedicated memory. For a deeper look at hardware partitioning, read optimizing NVIDIA GPUs with partitioning in Kubernetes.

When should you use vGPU?

Use vGPU when workloads already run in virtual machines and each needs only part of a GPU. Typical cases are virtual desktop infrastructure, remote engineering workstations, and small AI inferencing services in a virtualized data center. It also suits teams that need VM-level isolation and live migration support.

Skip it for large training jobs that need every GPU on a node. Passthrough is simpler there. For Kubernetes-native sharing without a hypervisor, software fractioning or MIG is usually the lighter path.

What are the limits of vGPU?

vGPU adds licensing cost, since NVIDIA software is licensed per user or per GPU and needs a license server. Compute is time-sliced, so a noisy neighbour can affect latency. Memory is fixed per profile, so resizing means reassigning the profile. Support also depends on the card, driver, and hypervisor combination.

How CloudRaft helps

CloudRaft designs and operates GPU infrastructure on Kubernetes, helping you choose between vGPU, MIG, and software fractioning for your workloads. Talk to our team.

Conclusion

A vGPU shares one physical GPU across virtual machines with a fixed memory slice each. It raises utilization for virtual desktops and small compute jobs, at the cost of licensing and shared compute. Match it to your isolation needs, and compare it against MIG and passthrough before committing.