NVIDIA vGPU for AI Workloads: Profiles, Licensing, MIG and VM Placement

NVIDIA vGPU allows multiple virtual machines (VMs) to share a physical GPU. For AI workloads, the main deployment decisions are GPU memory allocation, licensing, scheduling policies, hardware partitioning and VM placement.

Time-sliced vGPU improves resource utilization, while MIG-backed vGPU provides stronger hardware isolation and more predictable performance. The appropriate configuration depends on workload requirements and the virtualization platform.

What Is NVIDIA vGPU?

NVIDIA vGPU is virtualization software that allows multiple VMs to share a physical GPU. A Virtual GPU Manager runs on the hypervisor host, while a compatible NVIDIA driver runs inside each VM.

Each VM receives a virtual GPU with an assigned profile. Supported configurations can also assign multiple vGPUs to one VM.

NVIDIA vGPU supports virtualization platforms including VMware, Citrix and Nutanix, subject to GPU, hypervisor and software compatibility.

For enterprises running AI development, inference or workstation applications in VMs, vGPU allows physical GPU resources to be shared without dedicating an entire GPU to every VM.

How Does NVIDIA vGPU Profiling Work?

A vGPU profile defines the GPU framebuffer memory allocated to a VM and the capabilities available to it.

Profile sizes range from 512 MB on older Maxwell GPUs to 96 GB on supported Blackwell GPUs, depending on the hardware and software product. Supported configurations are listed in the NVIDIA vGPU packaging guide.

For example, an 8 GB profile allocates 8 GB of GPU framebuffer memory to a VM. However, 8 GB of memory does not guarantee a proportional share of GPU compute performance.

GPU processing cycles, cache and memory bandwidth may be shared or partitioned depending on whether the deployment uses time slicing or hardware-based MIG partitioning.

What Does NVIDIA vGPU Licensing Cost?

NVIDIA offers three primary vGPU software products for virtual desktops and applications:

These are NVIDIA's suggested prices, not guaranteed purchase prices. Perpetual licenses require an initial five-year Support, Updates and Maintenance Subscription (SUMS) purchase.

Licensing follows a concurrent-user (CCU) model. For virtual desktops, NVIDIA generally requires one license per running VM. Session-based deployments count concurrent users.

For example, licensing 300 concurrently running vPC desktops at the annual list price would cost:

300 concurrent desktops × $50 = $15,000 per year

AI training, inference and other CUDA compute workloads using C-series vGPU profiles are licensed separately through NVIDIA AI Enterprise vGPU for Compute.

How Does vGPU License Management Work?

NVIDIA manages applicable licenses through its NVIDIA License System, using either a Cloud License Service (CLS) or Delegated License Service (DLS).

For products subject to software license enforcement, failure to obtain a license causes progressive performance restrictions:

• First 20 minutes: Full GPU performance remains available.
• After 20 minutes: Frame rates are limited to 15 fps, and CUDA performance degrades.
• After 24 hours: Frame rates fall to 3 fps, CUDA applications stop functioning, and GPU resource allocations are restricted.

Performance returns after a valid license is obtained, although some applications may require restarting.

These restrictions are documented in the NVIDIA vGPU licensing FAQ. EULA-only licensing does not apply software-enforced performance restrictions, while vGPU for Compute has separate enforcement requirements.

How Do Time-Sliced vGPUs Work?

Time-sliced vGPU allows multiple VMs to share GPU processing cycles. NVIDIA's scheduler distributes processing time according to three policies:

Scheduling frequency and time-slice duration affect latency and throughput. Longer execution intervals can reduce context-switching overhead for compute-heavy CUDA workloads, while shorter intervals can improve responsiveness.

NVIDIA documents the policies and supported configuration methods in its vGPU scheduling guide.

Available settings and their configuration requirements vary by GPU, hypervisor and software release. Scheduling changes may require removing existing vGPU instances, reloading the driver or restarting the host.

How Does NVIDIA MIG Relate to vGPU?

NVIDIA Multi-Instance GPU (MIG) partitions supported physical GPUs into separate hardware instances with dedicated portions of compute capacity, L2 cache and memory bandwidth.

Unlike conventional time slicing, MIG provides hardware-level resource separation that improves performance predictability.

Three configurations are relevant:

Time-sliced MIG-backed vGPU is available only on supported GPUs, hypervisors and software releases.

For example, the NVIDIA RTX PRO 6000 Blackwell Server Edition supports up to four MIG instances. Compatible configurations can support multiple time-sliced vGPUs within each instance.

The maximum number of VMs per physical GPU therefore depends on both MIG partitioning and supported vGPU density.

MIG can also be used in bare-metal and container environments without vGPU. GPU-cloud operators must distinguish hardware partitioning from VM virtualization when selecting a deployment model.

What Limits vGPU VM Placement?

VM placement depends on the GPU model, vGPU profile, partitioning mode and virtualization platform.

For example, Mirantis OpenStack for Kubernetes (MOSK) 25.2 has several restrictions:

1. Profile compatibility: A physical GPU configured for time slicing supports only one vGPU profile type at a time.
2. Repartitioning: Changing GPU partitions requires affected vGPU instances to be removed first.
3. Live migration: MOSK does not support live migration of VMs with attached vGPUs.
4. GPU attachment: vGPUs cannot be attached to or removed from running VMs.
5. Multiple vGPUs: MOSK does not support assigning multiple vGPUs to one VM.

These restrictions are specific to MOSK 25.2, not universal NVIDIA vGPU limitations. Other supported hypervisors can provide capabilities such as live migration and multiple vGPUs per VM.

Operators must maintain compatible spare GPU capacity for maintenance and VM restarts, particularly when live migration is unavailable.

Which AI Workloads Should Use vGPU, MIG or Passthrough?

The appropriate configuration depends on GPU utilization, latency requirements, resource isolation and workload size.

Time slicing improves utilization but can introduce contention. MIG-backed configurations provide stronger hardware-level isolation between tenants, although workloads remain limited to their allocated resources.

Full-GPU passthrough assigns a physical GPU directly to one VM, preventing other VMs from sharing it through vGPU. Bare-metal deployment removes the virtualization layer entirely. Both suit workloads that require full accelerator capacity.

For enterprises and GPU-cloud providers, the decision is not simply vGPU versus MIG. It is choosing the appropriate combination of time slicing, MIG partitioning and full-GPU allocation to meet performance, isolation and utilization requirements.

Frequently Asked Questions

Can NVIDIA vGPU Be Tested Before Purchasing Licenses?

NVIDIA offers evaluation licenses for up to 128 users for 90 days across its vGPU software products, subject to registration and eligibility requirements.