33.3a MIG-backed vGPU: using MIG instances as the underlying resource for vGPU¶

🌐 Context Introduction¶

In modern AI infrastructure, engineers often need to share a single physical GPU across multiple virtual machines (VMs) while maintaining performance isolation and security. Traditionally, vGPU (virtual GPU) allowed splitting a GPU into multiple virtual devices, but each partition had fixed resource boundaries.

With NVIDIA Hopper architecture (H100 and later), a powerful new capability emerged: MIG-backed vGPU. This combines two technologies: - MIG (Multi-Instance GPU) — hardware-level partitioning of a GPU into isolated instances - vGPU (Virtual GPU) — software-level GPU virtualization for VMs

The result is that each MIG instance becomes the underlying resource for a vGPU device, giving VMs dedicated, isolated GPU slices with predictable performance.

⚙️ How MIG-backed vGPU Works¶

• MIG creates hardware-isolated GPU instances (e.g., 1g.10gb, 2g.20gb, 3g.40gb profiles)
• Each MIG instance is assigned to a vGPU device via the NVIDIA virtual GPU Manager
• The hypervisor (e.g., VMware vSphere, Red Hat Virtualization) presents these vGPU devices to VMs
• Each VM sees a dedicated virtual GPU backed by a real MIG partition
• No other VM can access that MIG instance — full isolation is guaranteed at the hardware level

🛠️ Key Benefits for AI Workloads¶

• Deterministic performance — MIG instances provide guaranteed compute and memory slices
• Enhanced security — hardware-level isolation between tenants (no "noisy neighbor" effects)
• Flexible sizing — mix different MIG profiles (1g, 2g, 4g) across VMs on the same physical GPU
• GPU oversubscription prevention — each vGPU maps 1:1 to a MIG instance, preventing resource contention
• Supports live migration — vGPU with MIG backing can be live-migrated between hosts (with compatible GPUs)

📊 Visual Representation: MIG-Backed vGPU hardware slicing¶

This diagram displays how combining MIG hardware blocks with vGPU virtualization provides clean, isolated slices to VMs.

📊 Comparison: MIG-backed vGPU vs. Traditional vGPU¶

🕵️ Configuration Overview (Conceptual)¶

To set up MIG-backed vGPU, engineers follow this general workflow:

1. Enable MIG mode on the physical GPU (via NVIDIA tools)
2. Create MIG instances using desired profiles (e.g., 1g.10gb for small AI inference VMs)
3. Install NVIDIA vGPU Manager on the hypervisor host
4. Configure vGPU devices in the hypervisor, selecting MIG-backed vGPU type
5. Assign vGPU devices to VMs (one MIG instance per vGPU per VM)
6. Install NVIDIA guest drivers inside each VM

Each VM then operates as if it has its own dedicated GPU, but the underlying resource is a hardware-partitioned MIG slice.

✅ Practical Considerations for New Engineers¶

• Check GPU compatibility — MIG-backed vGPU requires Hopper (H100, H200) or newer architecture
• Plan MIG profiles carefully — match MIG size to workload requirements (e.g., training vs. inference)
• Understand licensing — vGPU features may require NVIDIA AI Enterprise or vGPU software licenses
• Monitor MIG utilization — use NVIDIA tools to verify each MIG instance is properly assigned to its VM
• Test isolation — run concurrent workloads in different VMs to confirm no performance interference

📌 Summary¶

MIG-backed vGPU represents the evolution of GPU virtualization for AI infrastructure. By combining hardware-level MIG partitioning with software-level vGPU abstraction, engineers can deliver secure, predictable, and flexible GPU resources to virtual machines — essential for multi-tenant AI environments, cloud deployments, and enterprise data centers.

This approach simplifies resource management while maximizing GPU utilization, making it a cornerstone of modern AI infrastructure operations.