Contents

• 1. Overview
• 2. Introduction
• 3. Supported GPUs
• 4. Supported Configurations
• 5. Virtualization
• 6. Concepts
  • 6.1. Terminology
  • 6.2. Partitioning
  • 6.3. CUDA Concurrency Mechanisms
• 7. Deployment Considerations
  • 7.1. System Considerations
  • 7.2. Application Considerations
• 8. MIG Device Names
  • 8.1. Device Enumeration
  • 8.2. CUDA Device Enumeration
• 9. Supported MIG Profiles
  • 9.1. A30 MIG Profiles
  • 9.2. A100 MIG Profiles
  • 9.3. H100 MIG Profiles
  • 9.4. H200 MIG Profiles
  • 9.5. B200 MIG Profiles
  • 9.6. RTX PRO 6000 Blackwell MIG Profiles
• 10. Getting Started with MIG
  • 10.1. Prerequisites
    • 10.1.1. Additional Prerequisites for RTX PRO 6000 Blackwell
  • 10.2. Enable MIG Mode
    • 10.2.1. GPU Reset on Hopper+ GPUs
    • 10.2.2. GPU Reset on NVIDIA Ampere Architecture GPUs
    • 10.2.3. Driver Clients
  • 10.3. List GPU Instance Profiles
  • 10.4. Creating GPU Instances
    • 10.4.1. Instance Geometry
  • 10.5. Running CUDA Applications on Bare-Metal
    • 10.5.1. GPU Instances
    • 10.5.2. GPU Utilization Metrics
    • 10.5.3. Compute Instances
  • 10.6. Destroying GPU Instances
  • 10.7. Monitoring MIG Devices
  • 10.8. MIG with CUDA MPS
    • 10.8.1. Workflow
    • 10.8.2. Configure GPU Instances
    • 10.8.3. Set Up the MPS Control Daemons
    • 10.8.4. Launch the Application
    • 10.8.5. A Complete Example
  • 10.9. Running CUDA Applications as Containers
    • 10.9.1. Install Docker
    • 10.9.2. Install NVIDIA Container Toolkit
    • 10.9.3. Running Containers
  • 10.10. MIG with Kubernetes
  • 10.11. MIG with Slurm
• 11. Device Nodes and Capabilities
  • 11.1. System Level Interface
    • 11.1.1. /dev Based nvidia-capabilities
    • 11.1.2. /proc based nvidia-capabilities (Deprecated)
• 12. Notices
  • 12.1. Notice
  • 12.2. OpenCL
  • 12.3. Trademarks