GPU Virtualization - Rafay Product Documentation

When Do You Need an NVIDIA AI Enterprise License with GPU Virtualization?

If you're deploying NVIDIA GPUs in a virtualized or cloud-native environment, understanding NVIDIA AI Enterprise (NVAIE) licensing can save you from unexpected compliance issues — or unnecessary spending. The rules vary significantly depending on how your GPUs are deployed. Here's a practical breakdown.

What Is NVIDIA AI Enterprise?

NVIDIA AI Enterprise (NVAIE) is NVIDIA's end-to-end software platform for AI workloads, covering everything from GPU drivers and Kubernetes operators to NIM microservices, NeMo, Triton Inference Server, and RAPIDS. It's licensed on a per-GPU basis and available as an annual subscription or perpetual license with 5-year support.

The key thing to understand: NVAIE licensing is about software access and support, not raw GPU compute. The GPU itself will generally work without a license — but your access to enterprise-grade software, NGC containers, and SLA-backed support depends on it.

Deployment Models and When a License Is Required

1. Bare Metal (No Hypervisor)

License required? Conditionally.

On bare metal, the NVIDIA driver works without an NVAIE license for standard CUDA workloads. However, a license is required to:

• Access NVAIE-gated NGC containers (NIM microservices, enterprise frameworks)
• Use the vGPU for Compute guest driver for licensing enforcement
• Receive enterprise support with SLA

Some GPUs bundle NVAIE automatically. Each NVIDIA H100 PCIe or NVL and H200 NVL GPU includes a 5-year NVAIE subscription that activates with the GPU serial number. Notably, Blackwell (B200/B300) DGX systems do not include NVAIE — licenses must be purchased separately.

2. GPU Passthrough (PCIe Passthrough / VFIO)

License required? Same as bare metal for the guest VM.

In GPU passthrough, the entire physical GPU is assigned to a single VM. The guest VM behaves identically to a bare metal system from a driver perspective. NVAIE licensing requirements inside the VM mirror bare metal:

• Standard CUDA compute works without a license
• Access to NVAIE NGC software and NIM microservices requires a license
• One license per physical GPU passed through

3. NVIDIA vGPU for Compute (Time-Sliced / MIG-backed vGPU)

License required? Yes — software enforced.

This is where licensing is strictly enforced. NVIDIA vGPU for Compute is licensed exclusively through NVAIE. One license is required per vGPU instance assigned to a VM.

How enforcement works:

• When a vGPU VM boots on a supported GPU, it initially operates at full capability
• If it fails to obtain a license from the NVIDIA License System (NLS), performance degrades over time
• The license is checked out from either a Cloud License Server (CLS) or a Delegated License Server (DLS) for air-gapped environments

vGPU for Compute supports up to 16 vGPU instances per physical GPU, with each requiring its own license. It enables advanced capabilities not available in passthrough: live migration, suspend/resume, warm updates, and multi-tenant GPU sharing.

4. Cloud Deployments (AWS, Azure, GCP, OCI)

License required? Depends on the model.

• Pay-as-you-go via CSP marketplace: License is included in the hourly per-GPU price. No separate purchase needed.
• BYOL (Bring Your Own License): Use an on-prem NVAIE license on a certified cloud instance. One license per GPU, checked out from NLS.
• H100 / H200 NVL instances: May include NVAIE bundled depending on the cloud provider's offering.

5. MIG (Multi-Instance GPU)

License required? Depends on the underlying deployment.

MIG allows a single physical GPU (A100, H100, H200, B200) to be partitioned into up to 7 isolated GPU instances, each with dedicated memory and compute. From a licensing perspective, MIG doesn't change the fundamental rules — what matters is how the host is deployed:

• MIG on bare metal — same as bare metal, license required for NVAIE software access
• MIG on vGPU (time-sliced) — each vGPU instance backed by a MIG slice still requires an NVAIE license per vGPU
• MIG on passthrough — the VM receives a MIG-partitioned GPU, same rules as passthrough

MIG on Kubernetes has an additional operational consideration: the NVIDIA GPU Operator's MIG Manager handles reconfiguration automatically by watching the nvidia.com/mig.config node label. When the label changes, MIG Manager stops GPU pods, applies the new MIG geometry, and restarts them. One license is still required per physical GPU regardless of how many MIG slices are created.

6. Kubernetes (Bare Metal or VMs)

License required? Conditionally, same rules as the underlying deployment.

The NVIDIA GPU Operator automates driver and plugin deployment on Kubernetes. If your K8s nodes are:

• Bare metal — NVAIE license needed for NGC-gated software, not for basic CUDA
• vGPU-backed VMs — NVAIE license required per vGPU, software enforced
• Passthrough VMs — same as bare metal inside the VM

The GPU Operator can also manage license token distribution automatically when configured with NLS credentials.

Quick Reference

License Server Options

Cloud License Server (CLS): Hosted by NVIDIA, requires internet connectivity. Easiest to set up.

Delegated License Server (DLS): On-premises, air-gapped environments. You deploy and manage it. Required for secure/sovereign deployments.

Both are managed through the NVIDIA License System portal at nvid.nvidia.com.

Key Takeaways

1. vGPU for Compute always requires a license — it's software enforced and performance degrades without one.
2. GPU passthrough and bare metal only require a license for NVAIE software access, not for basic CUDA functionality.
3. H100 and H200 NVL GPUs bundle 5 years of NVAIE — Blackwell (B200/B300) does not.
4. GH200 and GB200 are bare metal only — vGPU is not supported on these platforms.
5. HGX systems only support full board passthrough to a single VM — no partial passthrough.
6. For production AI workloads, the NVAIE license is almost always worth it — it unlocks NIM microservices, enterprise support SLAs, security patches, and the full NGC catalog.