What are NVIDIA MPS and MIG?

Two GPU-sharing mechanisms: MPS coordinates CUDA processes through a server, while MIG partitions supported datacenter GPUs into isolated device instances.

Multi-Process Service, or MPS, lets CUDA processes submit through an MPS server so their work can overlap more efficiently than separate contexts that time-slice. NVIDIA's documentation distinguishes the general service from Volta-specific controls such as active-thread percentage. Static MPS SM partitioning is a separate Ampere-and-newer feature, exposed through CUDA_MPS_SM_PARTITION.

Multi-Instance GPU, or MIG, divides supported datacenter GPUs starting with Ampere into hardware-isolated instances. Each instance has dedicated compute and memory resources and appears as a selectable device. MIG therefore provides stronger isolation than MPS or a green context, but it needs supported hardware and administrative reconfiguration.

The three mechanisms must not be conflated. A green context partitions SMs inside one process. MPS coordinates multiple processes, and merely starting its daemon does not prove client throughput or isolation. MIG is unavailable on a Tesla T4, so a T4 experiment cannot supply a MIG result.

Measured

On the Tesla T4 used for day 94, nvidia-smi -q -d compute reported Compute Mode Default. nvidia-cuda-mps-control -d exited 0, and echo quit | nvidia-cuda-mps-control also exited 0. Startup warned that /var/log/nvidia-mps was not writable, so no daemon log was produced.

That capture proves only that the MPS control daemon started and stopped on this Turing node. Day 94 did not launch an MPS client workload, measure MPS scheduling, test its SM controls, or compare MPS throughput. Its timing table measures ordinary streams against two 20-SM green contexts: the short job finished at 21.488 ms shared and 3.044 ms partitioned. MIG is unsupported on this T4 and was not measured.

Diagram: one GPU shown three ways: MPS funnels several process queues through a server; a green context draws fixed SM bands inside one process; MIG exposes separately isolated device instances on supported datacenter hardware.

Related terms

• green context
• CUDA stream
• streaming multiprocessor
• nvidia-smi
• compute capability

Where you meet this

• Day 94, sharing a GPU, which records the daemon lifecycle separately from the green-context benchmark.
• Select which GPU, where MIG instances and visible-device ordinals affect device selection.

Sources

• NVIDIA Multi-Process Service documentation: https://docs.nvidia.com/deploy/mps/index.html (checked 2026-09-01)
• NVIDIA Multi-Instance GPU introduction: https://docs.nvidia.com/datacenter/tesla/mig-user-guide/introduction.html (checked 2026-08-29)

Byline

Written by: pending. Reviewed by: pending. Written on: pending. Last checked on: pending. Verified evidence was captured on 2026-09-02; publication still requires named author and reviewer sign-off.