MIG CONFIGURATION

One GPU. Multiple isolated workloads.

gpufleet configures NVIDIA Multi-Instance GPU (MIG) so compatible GPUs can be partitioned into dedicated, schedulable instances for Kubernetes workloads.

Talk to gpufleet →

WHY IT MATTERS IN KUBERNETES

Give each workload a dedicated slice of your GPU.

A single GPU is often too large for a workload, while sharing it without boundaries can create unpredictable performance. MIG provides hardware-level partitions on supported NVIDIA GPUs, allowing Kubernetes to schedule defined GPU slices with dedicated compute and memory resources.

WHAT IT ENABLES

Run multiple right-sized workloads on a supported GPU.
Give teams predictable, isolated GPU resources.
Reduce idle capacity when jobs do not need a full GPU.

HOW WE APPROACH IT

Verify GPU and workload compatibility before partitioning.
Configure the NVIDIA GPU Operator and device plugins for Kubernetes.
Align MIG profiles with the resource requests your workloads make.

WORKLOAD FIT

Is MIG right for your workload?

MIG is a strong fit when several predictable workloads need less than a full GPU and benefit from hardware-isolated compute and memory. It is commonly useful for inference services, notebooks, and smaller batch jobs.

Good fit

Services have stable, measurable GPU memory requirements.
Multiple workloads need isolation on the same physical GPU.
Throughput and utilization matter more than full-GPU access.

Consider a full GPU instead

A job needs all available GPU memory or compute.
Workloads need high-bandwidth communication across GPUs.
Demand changes too quickly for a fixed partition layout.

HARDWARE SUPPORT

MIG requires supported NVIDIA hardware.

Profile sizes vary by GPU model and memory configuration. We verify the exact device and driver combination before defining a plan.

GPU	MIG support	Planning note
A100	Supported	Up to seven GPU instances, depending on profile.
H100	Supported	Hopper-generation profiles and isolation.
H200	Supported	Profiles account for its larger memory capacity.
RTX 4090	Not supported	Use time-slicing or separate full-GPU workloads.
L40S	Not supported	Use time-slicing or separate full-GPU workloads.

MANAGED CONFIGURATION

What MIG management actually includes.

Profile and strategy design

Map workload memory and compute needs to MIG profiles, then choose the GPU Operator's single strategy for one uniform profile type or mixed when nodes must advertise multiple MIG resource types.

MIG Manager enablement

Enable migManager, create the partition configuration, and apply node labels so NVIDIA MIG Manager safely configures the requested geometry on each eligible node.

GPU Operator integration

Integrate the driver, device plugin, GPU Feature Discovery, and DCGM monitoring so Kubernetes advertises the correct resources and schedules pods against profiles such as nvidia.com/mig-1g.5gb.

UTILIZATION EXAMPLE

Turn stranded capacity into schedulable capacity.

BEFORE15%

1 inference service

One service reserves the full A100 while most capacity sits idle.

AFTER90%

7 inference services

Seven right-sized instances share one A100 using isolated MIG slices.

Illustrative utilization scenario; results depend on model, traffic, batching, MIG profiles, and GPU memory requirements.

Sanitized nvidia-smi -L outputREDACTED

GPU 0: NVIDIA A100-SXM4-40GB (UUID: GPU-…redacted)
  MIG 1g.5gb Device 0: (UUID: MIG-…01)
  MIG 1g.5gb Device 1: (UUID: MIG-…02)
  MIG 1g.5gb Device 2: (UUID: MIG-…03)
  MIG 1g.5gb Device 3: (UUID: MIG-…04)
  MIG 1g.5gb Device 4: (UUID: MIG-…05)
  MIG 1g.5gb Device 5: (UUID: MIG-…06)
  MIG 1g.5gb Device 6: (UUID: MIG-…07)

CONTACT

Build a GPU platform that works for your workloads.

We help you select, configure, and operate the Kubernetes GPU capabilities your team needs.

Talk to gpufleet →