Skip to main content

GPU & MIG Configuration

Overview​

NVIDIA Multi-Instance GPU (MIG) lets you partition a single physical GPU (for example, NVIDIA A100/H100) into multiple isolated GPU instances. Each instance has dedicated compute, memory, and bandwidth.

In Armada Bridge, you can configure and manage MIG profiles directly from the UI, without manual CLI setup.

MIG configuration helps you:

  • Increase GPU utilization by running multiple workloads on one GPU
  • Isolate workloads for predictable performance
  • Allocate GPU resources based on workload needs

MIG Purpose and Benefits​

Why Use MIG?​

  • Resource Sharing - Share single GPUs across multiple applications
  • Improved Utilization - Maximize GPU efficiency
  • Cost Optimization - Better resource allocation
  • Isolation - Independent workloads don't interfere
  • Flexibility - Configure instances based on workload needs

MIG Profiles​

Available MIG profiles include:

  • 1g.10gb - 1 compute slice + 10 GB memory
  • 1g.10gb+me - 1 compute slice + 10 GB memory (media extension variant)
  • 1g.20gb - 1 compute slice + 20 GB memory
  • 2g.20gb - 2 compute slices + 20 GB memory
  • 3g.40gb - 3 compute slices + 40 GB memory
  • 4g.40gb - 4 compute slices + 40 GB memory
  • 7g.80gb - Full GPU profile (80 GB)
note

Available MIG profiles depend on the GPU model in the selected node (for example, H100 80GB).

Prerequisites​

  • Tenant Admin access
  • A running Kubernetes cluster (for example, JupyterHub with KAI scheduler template) with supported NVIDIA GPUs

Enable MIG Configuration​

Step 1: Open MIG Settings for a Node​

  1. Open the cluster details (click cluster name or ellipsis menu → View).

    Cluster Name

  2. Go to the Nodes tab.

    Cluster Nodes

  3. Click the required node and select Enable MIG to open available MIG profiles.

    Cluster MIG

    Cluster MIG Profile

Step 2: Select MIG Profiles and Apply​

  1. Select required MIG profile(s) using the plus (+) icon.
  2. Choose profiles based on workload requirements (for example, Jupyter server needs).
  3. Click Apply.

Select Cluster MIG Profile

Step 3: Monitor and Verify​

  1. Wait for profile creation to complete (typically 1–2 minutes).
  2. During processing, UI shows: Applying MIG configuration changes. Please wait...
  3. Verify the warning is cleared and profiles appear in the node view.

Cluster MIG Process State

Select Cluster MIG Profile

Edit MIG Profile​

Step 1: Open Edit MIG​

  1. In the node MIG view, click the Edit icon.

Cluster Edit MIG

Step 2: Update Profiles​

  1. Add profiles using plus (+) or remove profiles using minus (–).
  2. Review the updated profile allocation.
  3. Click Update.

Remove Cluster MIG Profile

Step 3: Monitor and Verify​

  1. Wait for update to complete (typically 1–2 minutes).
  2. During processing, UI shows: Applying MIG configuration changes. Please wait...
  3. Verify the warning is cleared and updated profiles are displayed.

Cluster MIG Process State

Cluster MIG Success State

Disable MIG Configuration​

Step 1: Start Disable MIG​

  1. In the node MIG view, click Disable MIG.

Disable MIG

Step 2: Confirm Disable MIG​

  1. In the confirmation pop-up, click Disable MIG.

Disable MIG Pop Up

Step 3: Monitor and Verify​

  1. Wait for disable operation to complete (typically 1–2 minutes).
  2. During processing, UI shows: Applying MIG configuration changes. Please wait...
  3. Verify the warning is cleared and GPU returns to full-capacity mode.

Disable MIG State

Disable MIG Success State

GPU Allocation for Workloads​

Allocate to JupyterHub​

When creating JupyterHub servers:

  1. Select server environment
  2. Choose GPU Type
  3. Select from available MIG profiles or full GPU
  4. Start server with GPU resources

Allocate to Deployments​

When deploying models:

  1. Configure deployment resources
  2. Select GPU Allocation
  3. Choose MIG profile or full GPU
  4. Deploy model

Allocate GPU

Best Practices​

MIG Configuration​

  • Use smaller profiles for inference-only workloads
  • Reserve full GPUs for training jobs
  • Monitor GPU utilization regularly
  • Adjust profiles based on actual usage patterns

Performance Considerations​

  • MIG profiles have slight performance overhead
  • Full GPUs provide maximum performance
  • Benchmark workloads before production deployment
  • Consider scaling horizontally instead of vertically

Monitoring GPU Usage​

View GPU Metrics​

Monitor GPU utilization:

GPU Metrics

  • GPU memory usage
  • Compute utilization
  • Temperature
  • Power consumption

Optimize Allocation​

Based on metrics, you can:

  • Adjust MIG profiles
  • Add/remove GPU instances
  • Rebalance workloads
  • Plan capacity upgrades

Next Steps​

Last updated on