What Is NVIDIA MIG?

NVIDIA MIG, short for Multi-Instance GPU, is a hardware feature that divides one GPU into several smaller, fully separate GPUs. NVIDIA introduced it with the A100 in 2020, and it is part of the Hopper and Blackwell data centre GPUs and some Blackwell professional cards.

Without MIG, several jobs sharing a GPU compete for the same cores, cache and memory bandwidth, so one heavy job can slow every other one down. MIG instead carves the GPU into slices along hardware lines. Each instance gets its own streaming multiprocessors, its own share of the L2 cache, its own memory controllers and a fixed block of memory. A fault or a memory-hungry job in one instance cannot affect the others.

To software, each instance looks like a separate GPU with its own device ID. CUDA applications run on it unchanged, and container platforms can hand out instances just as they would whole GPUs.

A GPU is divided into seven compute slices and a matching number of memory slices. An administrator combines them into instances using fixed profiles, named for the compute slices and memory they include.

MIG is limited to certain data centre and professional GPUs. The maximum number of instances, and the memory each gets, depends on the card.

Among the cards we list, the H100 PCIe card and the H200 NVL card divide into up to seven instances. The RTX PRO 6000 Blackwell Server Edition splits into four 24 GB instances, and the RTX PRO 4500 Blackwell Server Edition into two of 16 GB. Some cards from the same family, such as the RTX PRO 4000 and 4500 workstation cards, do not list MIG, so check the datasheet for the exact model.

MIG is one of three common ways to share a GPU, and they solve different problems.

MIG pays off wherever a large GPU would otherwise sit partly idle running one small job. Typical uses include serving several AI inference models from one card, giving developers or students their own isolated GPU, running many video analytics streams on an on-premises server, and consolidating mixed workloads onto fewer data centre GPUs to save power and rack space.

It is less useful when one job needs the whole GPU, or when GPUs must work together over NVLink, because MIG instances do not communicate with each other directly.

MIG is configured from the host with NVIDIA’s standard tools, and the steps are the same on every supported GPU. An administrator first enables MIG mode on the GPU with nvidia-smi, which needs the GPU to be idle and reset. GPU instances are then created from the chosen profiles, followed by compute instances inside them.

MIG mode stays set across reboots, but the instances themselves do not, so most sites define it in a configuration file and apply it at boot with NVIDIA’s MIG partition editor. NVIDIA’s Data Center GPU Manager (DCGM) reports utilisation, memory and errors for each instance, so each workload can be monitored separately.

These are the questions people most often ask before relying on MIG.

We fit MIG-capable GPUs in the servers we build as standard, with the drivers installed and MIG profiles set up to suit your workloads before the system ships.