High-capacity GPUs (like the A100, H100, or Blackwell) often sit underutilized in bare-metal environments. Software-based time slicing lacks hardware-level resource guarantees.
NVIDIA Multi-Instance GPU (MIG) solves this by partitioning a single physical card into independent instances at the hardware level, giving each partition dedicated compute, memory, and bandwidth.
Here is a quick step-by-step CLI walkthrough on Ubuntu 24.04:
1. Install Drivers and Verify Support
sudo ubuntu-drivers install
nvidia-smi -q
Look for the MIG Mode section in the output to confirm hardware support. (Install nvidia-container-toolkit if you plan to run Docker).
2. Enable MIG Mode
Ensure no active workloads are attached, then enable MIG mode on GPU 0:
sudo nvidia-smi -i 0 -mig 1
3. List Supported Profiles
Query the supported resource layouts for your GPU model:
nvidia-smi mig -lgip
4. Partition the GPU
Create GPU Instances (GI) and Compute Instances (CI) using your chosen profile string:
sudo nvidia-smi mig -cgi 2g.48gb,2g.48gb -C
Verify active partitions using nvidia-smi mig -lgi.
5. Target MIG Instances in Docker
List generated UUIDs:
nvidia-smi -L
Pass the specific MIG UUID directly to your Docker container:
docker run --rm -it --gpus '"device=MIG-af414487-fcaa-5f42-b210-6f614c9cf780"' nvcr.io/nvidia/pytorch: nvidia-smi
6. Reset Partitions
To revert to single-instance operation:
sudo nvidia-smi mig -dci
sudo nvidia-smi mig -dgi
sudo nvidia-smi -i 0 -mig 0
Managing bare-metal lifecycle, drivers, and hardware provisioning at scale takes operational bandwidth. At MIG servers, we provide pre-configured, dedicated GPU hardware optimized for partitioned AI workloads out of the box.
Top comments (0)