NVIDIA MIG on the RTX PRO 6000: one GPU split into four

How Multi-Instance GPU works on the RTX PRO 6000 Blackwell: the 24, 48 and 96 GB profiles, the commands, what is isolated and when MIG is the wrong tool.

Bare circuit board of an NVIDIA GeForce RTX 5090 Founders Edition with the GB202 chip in the centre, the same chip used in the RTX PRO 6000 Blackwell, next to the board's rear side
Photo: Nvidia RTX 5080 5090 FE PCB by 极客湾Geekerwan, Wikimedia Commons, CC BY 3.0

Multi-Instance GPU, or MIG, lets one NVIDIA data centre GPU act as several smaller GPUs. Each slice gets its own share of compute, its own memory and its own memory bandwidth, isolated in hardware. On the RTX PRO 6000 Blackwell, NVIDIA's documentation lists up to four slices of 24 GB, two of 48 GB, or one full 96 GB GPU. Below is what each layout is good for, the commands to set it up, and when you should not use MIG at all.

Which layout for whom

  • Four teams or four small models that must not disturb each other → 4 × 1g.24gb. Each slice has a quarter of the compute and 24 GB of memory: enough for a 7B to 14B model in 8-bit, a vision model, or a development sandbox.
  • Two services of medium size → 2 × 2g.48gb. Room for a 30B-class model in FP8 or a 70B model in 4-bit per slice, with the other half kept for a second service.
  • One large model, long prompts or many users → no MIG, the full 96 GB. Splitting the card only takes memory away from the KV cache.
  • Many models on many cards → a whole server. Our server in Sofia has eight RTX PRO 6000 Blackwell Server Edition cards, 768 GB of GPU memory in total. With MIG on every card that is up to 32 isolated 24 GB GPUs. It rents from $1.80 per GPU-hour, or about $10,510 a month for the whole machine. It is rented into early 2027 and we take reservations for the period after. See the server.

What MIG isolates

A MIG instance is not a software quota. The GPU's streaming multiprocessors, its memory and the paths to that memory are divided in hardware. A job in one slice cannot read another slice's memory, and a job that fills its slice does not slow down the others through memory bandwidth contention. That is the difference from the two other ways to share a GPU:

Method Memory isolation Performance isolation Typical use
Time-slicing No No, jobs take turns Light, bursty development work
MPS (Multi-Process Service) Limited Partial, compute share can be capped Several processes of one trusted user
MIG Yes, in hardware Yes, fixed share of compute and bandwidth Separate users, teams or production services

The price of that isolation is rigidity. A slice cannot borrow idle compute from its neighbour, and the layout only changes when the slices are empty.

The RTX PRO 6000 profiles

NVIDIA's MIG user guide lists these profiles for the RTX PRO 6000 Blackwell (Workstation, Max-Q Workstation and Server Edition, all with 96 GB):

Profile Memory Compute Max instances
1g.24gb 1/4 (24 GB) 1/4 4
2g.48gb 1/2 (48 GB) 1/2 2
4g.96gb full (96 GB) full 1

Each profile also comes in variants. +gfx adds graphics API support, which is new on this generation: NVIDIA calls it Universal MIG, so one card can serve rendering and AI work side by side. -me leaves out the video engines for pure compute, and +me.all gives all of the card's video engines to one slice.

For comparison, the H100, H200 and B200 split into up to seven instances (for example 1g.23gb on the B200). The RTX PRO 6000 stops at four, but each of its smallest slices still has 24 GB.

The chip under the RTX PRO 6000 is GB202, the same die as in the GeForce RTX 5090. MIG is enabled only on the professional card.

How to set it up

The commands come from NVIDIA's MIG user guide and run as root:

# 1. Turn on MIG mode on GPU 0
nvidia-smi -i 0 -mig 1

# 2. See the available profiles
nvidia-smi mig -lgip

# 3. Create four 24 GB instances with their compute instances
nvidia-smi mig -i 0 -cgi 1g.24gb,1g.24gb,1g.24gb,1g.24gb -C

# 4. Check the result; every slice gets its own MIG UUID
nvidia-smi -L

A process then sees only its slice when you set CUDA_VISIBLE_DEVICES to that MIG UUID, or when Kubernetes hands the slice to a pod through the NVIDIA device plugin.

Four details catch people out:

  • Driver and firmware. NVIDIA gives R575 (575.51.03 or later) as the minimum driver for MIG on the RTX PRO 6000, and minimum vBIOS versions for the workstation editions.
  • Display mode on workstation cards. The Workstation and Max-Q editions ship in graphics display mode. They have to be switched to compute mode with NVIDIA's DisplayModeSelector tool first, which turns off their display outputs. The Server Edition has no display outputs to worry about.
  • Nothing survives a reboot. On Hopper and newer GPUs, MIG mode lasts only while the driver is loaded, and the instances are never persistent. NVIDIA recommends its mig-parted tool, run as a systemd service, to recreate the layout at start-up.
  • Changing the layout means emptying the GPU. Instances are destroyed with nvidia-smi mig -dci and then -dgi, which only works when nothing is running on them.

When MIG is the wrong tool

  • One model that needs the whole card. Llama 3.3 70B in FP8 already takes about 70 GB of the 96 GB. A 48 GB slice cannot hold it, and a smaller split leaves no room for cache. Our VRAM calculation shows the numbers.
  • Training across several GPUs. MIG slices do not join into one large GPU, and a slice cannot span two cards. For multi-GPU training, use full cards.
  • Uneven load. If one service is busy and the other idle, MIG keeps the idle quarter idle. Time-slicing or a single shared inference server such as vLLM (see vLLM vs Ollama vs SGLang) uses the card better.
  • You need a slice smaller than 24 GB. The RTX PRO 6000 has no 1/7 profiles like the H100.

Where MIG fits well is a server shared by people who should not see each other's work: several customers, several departments, or a production model next to a test environment. It also gives each user a predictable speed, which is easier to plan and to bill. For running your own models in the EU, see private LLM hosting.

Questions

Does MIG make the GPU faster? No. It divides the card. Four slices together do about the same work as the full card, sometimes a little less. What you gain is isolation and better use of a card that one small model would leave mostly idle.

Can I mix sizes on one card? Yes, as long as the slices fit: one 2g.48gb and two 1g.24gb instances, for example. nvidia-smi mig -lgipp lists the valid placements.

Does MIG work in virtual machines? NVIDIA's vGPU software can put MIG slices behind virtual machines on supported GPUs. That needs the vGPU software and its licence; check NVIDIA's support list for your card and version before planning on it.

Is MIG available on the RTX 5090? No. The RTX 5090 uses the same GB202 chip, but NVIDIA lists MIG profiles only for the RTX PRO 6000 Blackwell and its data centre GPUs.

Sources

  • NVIDIA, MIG User Guide: supported MIG profiles: RTX PRO 6000 Blackwell profiles 1g.24gb (4 instances), 2g.48gb (2), 4g.96gb (1), the +gfx, +me, +me.all and -me variants, Universal MIG; up to seven instances on H100, H200 (1g.18gb) and B200 (1g.23gb), checked 17 November 2026.
  • NVIDIA, MIG User Guide: getting started: nvidia-smi commands, minimum driver R575 (575.51.03) for the RTX PRO 6000, DisplayModeSelector for the workstation editions, no persistence of MIG mode and instances on Hopper and newer, mig-parted, checked 17 November 2026.
  • NVIDIA, RTX Blackwell PRO GPU architecture whitepaper: the GB202 GPU, checked 17 November 2026.
  • NVIDIA, RTX PRO 6000 Blackwell Server Edition: 96 GB GDDR7, MIG with up to four fully isolated instances, each with its own memory, cache and compute, checked 17 November 2026.
SHARE