NVIDIA vGPU for Compute

Type: Infrastructure Software Tags: NVIDIA, vGPU for Compute, AI Enterprise, virtualization, MIG, NLS, VMware, KVM, GPU partitioning Related: NVIDIA-vGPU, NVIDIA-AI-Enterprise, NVIDIA-AI-Enterprise-Software, NVIDIA-AI-Enterprise-Infrastructure-Support-Matrix, NVIDIA-AI-Enterprise-VMware-Deployment, NVIDIA-Enterprise-Licensing-Guide, NVIDIA-Data-Center-GPU-Drivers, NVIDIA-MIG, GPUDirect-RDMA, GPU-Direct-Storage, NVLink, NVIDIA-DCGM, NVIDIA-Container-Toolkit Sources: https://docs.nvidia.com/ai-enterprise/release-8/8.0/infra-software/vgpu.html, https://docs.nvidia.com/ai-enterprise/release-8/8.0/infra-software/vgpu/overview.html, https://docs.nvidia.com/ai-enterprise/release-8/8.0/infra-software/vgpu/features.html, https://docs.nvidia.com/ai-enterprise/release-8/latest/infra-software/vgpu/licensing.html, https://docs.nvidia.com/ai-enterprise/release-8/latest/support/support-matrix-8/8.0.html Last Updated: 2026-04-29

Summary

NVIDIA vGPU for Compute is the AI Enterprise-licensed virtualization stack for sharing NVIDIA GPUs across virtual machines for AI training, fine-tuning, inference, ML, and HPC workloads. It uses a hypervisor-host NVIDIA Virtual GPU Manager plus guest drivers, supports time-sliced vGPU, MIG-backed vGPU, and time-sliced MIG-backed vGPU modes, and relies on the NVIDIA License System for vGPU VM licensing.

Detail

Purpose

Use this page for AI Enterprise vGPU for Compute behavior, modes, features, licensing, and virtualization limitations. Use NVIDIA-vGPU for the broader vGPU product concept and NVIDIA-AI-Enterprise-VMware-Deployment for the VMware deployment guide.

Virtualization modes

• Time-sliced vGPU: multiple compute VMs time-share a physical GPU with memory/fault isolation and scheduled compute access.
• MIG-backed vGPU: MIG-capable GPUs expose GPU instances to VMs for spatial isolation and more predictable performance.
• Time-sliced MIG-backed vGPU: multiple vGPUs time-share within a MIG instance; current docs call out Blackwell RTX PRO server GPUs for this dense multi-tenant mode.

Feature surface

• MIG-backed vGPU for hardware-level partitioning and isolation.
• Device Groups for topology-aware provisioning of connected devices.
• GPUDirect RDMA and GPUDirect Storage data paths.
• Heterogeneous vGPU profiles on a single GPU.
• Live Migration, Suspend-Resume, and warm operational workflows where supported.
• Multi-vGPU and peer-to-peer communication.
• NVSwitch, NVLink Multicast, and scheduling policies for virtualized GPU clusters.
• Unified Virtual Memory for CUDA programming in guest environments.

Licensing and limitations

• vGPU for Compute is licensed through NVIDIA AI Enterprise.
• The NVIDIA License System enforces licensing for vGPU for Compute VMs.
• An unlicensed vGPU for Compute VM starts at full capability and then enters a degraded performance state after a short grace period until a valid license is acquired.
• Current licensing guidance says each NVIDIA C-series vGPU on a VM needs its own license from the NVIDIA Licensing System.
• DCGM is not supported on vGPU hypervisor hosts or inside guest VMs for MIG-backed or time-sliced vGPU environments.
• Some large-memory VMs require increased MMIO space for GPUs such as B300 HGX, B200 HGX, H200/H100/H800/H20, L40/L20/L4/L2, RTX Ada, A40/A30/A10/A100, RTX A-series, and V100 families.
• Windows guest VMs support native applications only; containerization-dependent AI Enterprise features are not supported on Windows guest OSes.

NVIDIA context

The Infra 8.0 support matrix lists NVIDIA vGPU for Compute Virtual GPU Manager and Guest Drivers 20.0 as supported. Check NVIDIA-AI-Enterprise-Infrastructure-Support-Matrix before planning a deployment because platform, hypervisor, guest OS, GPU architecture, and feature support vary.

Connections

• NVIDIA-vGPU - broader virtual GPU software context.
• NVIDIA-AI-Enterprise - vGPU for Compute is licensed through AI Enterprise.
• NVIDIA-AI-Enterprise-Software - infrastructure-layer software catalog that includes vGPU for Compute.
• NVIDIA-AI-Enterprise-Infrastructure-Support-Matrix - release-specific support and compatibility checks.
• NVIDIA-AI-Enterprise-VMware-Deployment - deployment path that uses vGPU/vSphere guidance.
• NVIDIA-Enterprise-Licensing-Guide - NLS and AI Enterprise license behavior for vGPU for Compute.
• NVIDIA-Data-Center-GPU-Drivers - driver branch alignment matters for host and guest drivers.
• NVIDIA-MIG - hardware partitioning foundation for MIG-backed vGPU modes.
• GPUDirect-RDMA, GPU-Direct-Storage, and NVLink - data movement and interconnect features exposed where supported, including NVSwitch-backed platforms.
• NVIDIA-DCGM - relevant because current docs call out DCGM limitations in vGPU environments.
• NVIDIA-Container-Toolkit - Linux guest/container workflows depend on container runtime support where supported.

Source Excerpts

• The AI Enterprise vGPU page states vGPU for Compute is licensed only through NVIDIA AI Enterprise.
• The overview describes time-sliced, MIG-backed, and time-sliced MIG-backed vGPU modes.
• The licensing page explains NLS enforcement and degraded behavior when a vGPU for Compute VM is unlicensed.

Resources

• NVIDIA AI Enterprise and NVIDIA vGPU for Compute
• NVIDIA vGPU for Compute Overview
• NVIDIA vGPU for Compute Features
• NVIDIA vGPU for Compute Licensing
• AI Enterprise Infra 8.0 Support Matrix