NVIDIA MIG (Multi-Instance GPU) on the RTX PRO 6000 Blackwell splits the 96 GB card into up to four hardware-isolated instances: four of 24 GB (1g.24gb), two of 48 GB (2g.48gb) or one of 96 GB (4g.96gb). The MIG setup is three commands: sudo nvidia-smi -i 0 -mig 1 to enable it, nvidia-smi mig -cgi ... -C to create instances, and a Docker --gpus '"device=0:0"' selector to hand one instance to a container. It needs driver 575.51.03 or later, and on workstation and Max-Q cards a minimum vBIOS and a switch from graphics to compute mode.
This guide covers the profiles, the prerequisites you should confirm with support before planning around MIG, the commands as NVIDIA documents them, two layouts worth using, and an honest comparison with simply running several containers on an unpartitioned card. It applies to the MassiveGRID servers with the RTX PRO 6000 Blackwell; the RTX PRO 4000 does not support MIG.
What is NVIDIA MIG, and what does it isolate?
Multi-Instance GPU (MIG): a feature of selected NVIDIA GPUs that partitions one physical card into several GPU instances at the hardware level. NVIDIA describes the RTX PRO 6000 Blackwell as supporting "up to four fully isolated instances", where "each MIG instance has its own high-bandwidth memory, cache, and compute cores". To software, each instance looks like a separate, smaller GPU with its own UUID.
Without MIG, every process sees all 96 GB and they compete: one can allocate the whole memory, and a heavy job slows the rest. With MIG, a workload in a 24 GB instance cannot take more than 24 GB and cannot starve its neighbours of compute, because the slices are fixed.
MIG is a way to share a card, not to make it faster: it turns one large resource into up to four predictable small ones.
MIG profiles on the RTX PRO 6000 Blackwell
NVIDIA's supported-GPU list gives the RTX PRO 6000 Blackwell (Server, Workstation and Max-Q Workstation Editions, 96 GB) a maximum of four instances. The supported-profiles page lists these profiles for all three editions:
| Profile | Memory per instance | Share of the card | Instances possible | Typical use |
|---|---|---|---|---|
1g.24gb | 24 GB | 1/4 | 4 | Four tenants or services with a quantised model each |
2g.48gb | 48 GB | 1/2 | 2 | Two mid-size models, or production and staging |
4g.96gb | 96 GB | Whole card | 1 | The full card in MIG mode |
1g.24gb+gfx, 2g.48gb+gfx, 4g.96gb+gfx | 24, 48, 96 GB | 1/4, 1/2, 1 | 4, 2, 1 | Instances with graphics capability |
1g.24gb+me, 1g.24gb+me.all, 2g.48gb+me.all | 24 or 48 GB | 1/4 or 1/2 | 1 each | An instance that holds the media (encode and decode) engines |
1g.24gb-me, 2g.48gb-me | 24 or 48 GB | 1/4 or 1/2 | 4 or 2 | Instances without media engines |
The profile name encodes the slice: 1g.24gb is one compute slice with 24 GB. NVIDIA's datasheet for the Max-Q card summarises it as "MIG Instance Types: Up to 4x 24 GB".
The RTX PRO 4000 has no MIG. The RTX PRO 4000 Blackwell SFF Edition in the 24 GB server is not on NVIDIA's MIG supported-GPU list, which names the RTX PRO 6000, 5000 and 4500 Blackwell only. On that card you run workloads side by side without hardware partitioning.
Prerequisites: confirm these three things before you plan around MIG
The store listing for the MassiveGRID RTX PRO 6000 servers names Multi-Instance GPU as a feature of the card. NVIDIA attaches three conditions to using it on RTX PRO Blackwell cards, and two of them concern the card itself, not your software. The servers carry the RTX PRO 6000 Blackwell Max-Q edition, so the Max-Q rows apply.
| Requirement (NVIDIA MIG User Guide) | Value | How to check |
|---|---|---|
| Driver | R575 branch, 575.51.03 or later | nvidia-smi prints the driver version. The current driver, 620.70.06, satisfies it. |
| Minimum vBIOS, Max-Q Workstation Edition | 98.02.6A.00.00 (the Workstation Edition needs 98.02.55.00.00) | nvidia-smi --query-gpu=vbios_version --format=csv |
| Display mode | Workstation and Max-Q cards ship in graphics mode and must be set to compute mode before MIG can be enabled | Changed with NVIDIA's DisplayModeSelector tool, version 1.72.0 or later |
NVIDIA's guide says that if a vBIOS update is needed you should "contact your reseller or system provider". It gives the display-mode command as:
sudo ./DisplayModeSelector --gpumode=compute --gpu=<GPU_ID>and warns that compute mode disables physical display output where the card is the primary display adapter. The reverse is --gpumode=graphics.
What to do: before you design a multi-tenant setup around MIG, ask MassiveGRID support to confirm the vBIOS version and the display mode of the card in your server, and whether the mode change is something you run or they do. Run the vBIOS query yourself as well; it takes one command. Do not flash a vBIOS or change the display mode on a remote server without that confirmation. If MIG is not available on your card, the unpartitioned alternative described below needs none of this.
Enable MIG and create instances with nvidia-smi
These are the commands from NVIDIA's "Getting Started with MIG" page. They assume the driver is installed (see installing NVIDIA drivers and Docker on Ubuntu 24.04), the three prerequisites are met, and nothing is using the GPU: stop Ollama, vLLM and any GPU containers first.
1. Enable MIG mode on GPU 0.
sudo nvidia-smi -i 0 -mig 1
nvidia-smi -i 0 --query-gpu=pci.bus_id,mig.mode.current --format=csvCheck: the query prints Enabled. NVIDIA notes that from the Hopper generation onwards, enabling MIG mode no longer requires a GPU reset. If you see "pending enable state ... In use by another client", a process still holds the GPU.
2. List the profiles and placements your card offers.
nvidia-smi mig -lgip
nvidia-smi mig -lgippThe first command prints each profile with its ID, memory and how many instances are free; the second prints where each profile can be placed. Trust this output over any table, including the one above.
3. Create GPU instances and their compute instances. NVIDIA's example is sudo nvidia-smi mig -cgi 9,3g.20gb -C on an A100, mixing a profile ID and a profile name; -C creates the compute instance in each GPU instance. With the RTX PRO 6000 profile names, four equal instances are:
sudo nvidia-smi mig -cgi 1g.24gb,1g.24gb,1g.24gb,1g.24gb -CTwo halves are -cgi 2g.48gb,2g.48gb -C. NVIDIA is explicit that enabling MIG mode alone is not enough: "without creating GPU instances (and corresponding compute instances), CUDA workloads cannot be run on the GPU".
4. Verify.
sudo nvidia-smi mig -lgi
nvidia-smi -LCheck: -lgi lists four GPU instances with their profile and placement, and nvidia-smi -L prints the physical GPU followed by four MIG devices, each with a UUID of the form MIG-xxxxxxxx-.... Those UUIDs are what you hand to workloads.
5. Change the layout. Stop the workloads, destroy the compute instances and then the GPU instances, and create a new set:
sudo nvidia-smi mig -dci && sudo nvidia-smi mig -dgiPersistence. On this GPU generation nothing survives a reboot. NVIDIA states that MIG mode is "no longer persistent across system reboots" on Hopper and later GPUs, and that created MIG devices are not persistent either. After every reboot or driver reload you must enable MIG and create the instances again before any workload starts. NVIDIA points to its MIG Partition Editor (mig-parted) for this, including a systemd service that recreates the geometry at startup.
Give each container or process its own MIG instance
With the NVIDIA Container Toolkit installed, Docker selects a MIG device in one of two documented formats: <GPUDeviceIndex>:<MIGDeviceIndex> or the MIG-<UUID> string. NVIDIA's examples:
sudo docker run --gpus '"device=0:0"' \
nvidia/cuda:13.4.3-base-ubuntu24.04 nvidia-smi -L
sudo docker run --runtime=nvidia \
-e NVIDIA_VISIBLE_DEVICES=MIG-c7384736-a75d-5afc-978f-d2f1294409fd \
nvidia/cuda:13.4.3-base-ubuntu24.04 nvidia-smiCheck: inside the container, nvidia-smi -L lists exactly one MIG device and nvidia-smi shows a memory total of about 24 GB, not 96 GB. Replace the example UUID with one from your own nvidia-smi -L output; device=0:1, 0:2 and 0:3 address the other instances.
A real service follows the same pattern. Ollama's documented GPU container command is docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama. For a per-tenant instance, change three things: the device, the volume and name, and the port, which you bind to the loopback interface because Docker's published ports bypass ufw:
sudo docker run -d --gpus '"device=0:0"' -v ollama-a:/root/.ollama -p 127.0.0.1:11434:11434 --name ollama-a ollama/ollama
sudo docker run -d --gpus '"device=0:1"' -v ollama-b:/root/.ollama -p 127.0.0.1:11435:11434 --name ollama-b ollama/ollamaProcesses outside containers use the same UUIDs through CUDA_VISIBLE_DEVICES=MIG-<UUID>, as NVIDIA's bare-metal example shows. Each tenant then gets its own port behind a TLS reverse proxy with its own credentials; none of these ports is opened in the firewall directly.
Two layouts that make sense on one 96 GB card
Four 24 GB tenants. Four 1g.24gb instances give four teams, customers or environments a GPU each with a hard memory limit. A 24 GB slice holds what a 24 GB card holds: on the Ollama library (sizes on 16 November 2026) gpt-oss:20b is a 14 GB download, qwen3:14b 9.3 GB and qwen3.8:27b 18 GB, the last one with only 6 GB left for context. Arithmetic on the price: $1,656.15 per month divided by four is $414.04 per 24 GB instance, and by two is $828.08 per 48 GB instance, with the server's 20 CPU cores, 128 GB of RAM and NVMe shared between them.
A model, embeddings and speech on one card. A retrieval or assistant stack usually has three GPU consumers: the chat model, an embedding model and speech-to-text. One layout to try is a 2g.48gb instance for the language model and two 1g.24gb instances for the rest. Confirm that your card accepts the mix with nvidia-smi mig -lgipp; the documented counts are per profile. Sizing from published figures: llama3.3:70b is a 43 GB download, which fits a 48 GB instance only with a short context, while a 27B to 32B model at 18 GB to 20 GB leaves it plenty of room. The embedding model bge-m3 is 1.2 GB. OpenAI's Whisper README lists about 10 GB of VRAM for the large model and about 6 GB for turbo, so either fits a 24 GB instance; self-hosting Whisper on a GPU covers that service. The benefit of MIG here is that a transcription batch cannot push the chat model out of memory.
MIG or several containers on one unpartitioned card?
You can also leave the card whole and start several containers with --gpus all. Both approaches run several workloads on one GPU; they differ in what happens when the workloads disagree.
| Question | MIG instances | Unpartitioned card, several containers |
|---|---|---|
| Memory limit per workload | Hard: 24 GB or 48 GB per instance | None: any process can allocate all 96 GB, and the next one fails to start or crashes |
| Compute | Each instance has its own compute slice | Shared; a heavy job slows the others |
| Unused capacity | Stays unused: an idle 24 GB instance cannot lend its memory or compute | Shared freely: one busy workload can use the whole card |
| Largest model | Limited by the instance: 24 GB or 48 GB | Up to the full 96 GB |
| Setup and reboots | Driver, vBIOS and compute-mode prerequisites; instances recreated after every reboot | Works as soon as the driver and container toolkit are installed |
| Good for | Tenants who must not affect each other, fixed budgets per team, mixed services with predictable needs | One team, one main model, workloads you size and trust yourself |
MIG wins on predictability and loses on efficiency. One vLLM server on the whole card, batching requests from all users, serves a single model to many people better than four isolated copies in 24 GB instances, and it can load models that no slice holds. For one model and one organisation, do not partition; the guide to running vLLM on the full 96 GB is the better path. Partition when isolation is the requirement.
Sometimes neither is right. If each tenant needs a GPU only for a few hours, paying by the hour costs less. On-demand 24 GB slices of the same card were listed on 16 November 2026 at $0.69 per hour on RunPod Secure Cloud, $1.13 at Azure in East US (NC24lds_xl, a quarter of a card with 24 vCPUs) and $1.29376 at Google Cloud in Iowa (g4-standard-12); 48 GB slices were $1.39, $2.44 (NC72lds_xl) and $2.58752 (g4-standard-24). Around the clock, 730 hours of a 24 GB slice is $503.70, $824.90 or $944.44 against $414.04 for a quarter of the dedicated server, and a 48 GB slice is $1,014.70, $1,781.20 or $1,888.89 against $828.08 for half. The break-even for a quarter of the server is about 600, 366 and 320 hours a month respectively. And if every tenant needs a whole machine with its own CPU, memory and disks, four RTX PRO 4000 servers at $449.99 per month each give stronger separation than any partitioning of one card. The GPU dedicated server plans list both options, and the RTX PRO 6000 server page goes through the 96 GB configurations.
What goes wrong with MIG
| Symptom | Cause | Fix |
|---|---|---|
| Enabling MIG fails or reports it is not supported | The card is still in graphics mode, or the vBIOS is below the minimum | Check the vBIOS version with the query above and ask support about the display mode before doing anything else. |
| "MIG mode is in pending enable state ... In use by another client" | A process or container still holds the GPU | Stop Ollama, vLLM and GPU containers, check nvidia-smi for remaining processes, run the enable command again. |
| MIG is enabled but workloads see no GPU | No GPU instances or compute instances exist yet | Create them with -cgi ... -C and confirm with nvidia-smi -L. |
| Everything worked until the reboot | MIG mode and instances are not persistent on this GPU generation | Recreate them at boot with mig-parted or your own systemd unit, before Docker starts. |
| A container sees the whole GPU or all instances | Started with --gpus all | Use --gpus '"device=0:N"' or NVIDIA_VISIBLE_DEVICES=MIG-<UUID>. |
| A model that ran on the whole card fails with out of memory | It needs more than the instance holds | Use a larger profile, a smaller or more heavily quantised model, or a shorter context. |
Security note. MIG isolates GPU memory and compute. It does not isolate the host: tenants who share a server still share its kernel, disks and network unless you separate them with containers, users and firewall rules. Give tenants an HTTPS endpoint and a key, not a shell on the machine, and keep every model port on 127.0.0.1 behind the reverse proxy.
Frequently Asked Questions
How many MIG instances does the RTX PRO 6000 Blackwell support?
Up to four. NVIDIA's MIG documentation lists the 96 GB RTX PRO 6000 Blackwell (Server, Workstation and Max-Q Workstation Editions) with the profiles 1g.24gb (four instances), 2g.48gb (two instances) and 4g.96gb (one instance), plus graphics and media variants. For comparison, the same list gives the H100 seven instances.
Does MIG work on the RTX PRO 6000 Blackwell Max-Q out of the box?
Not without checks. NVIDIA requires driver 575.51.03 or later, a minimum vBIOS of 98.02.6A.00.00 on the Max-Q Workstation Edition, and the card switched from graphics mode to compute mode with DisplayModeSelector 1.72.0 or later. Query the vBIOS with nvidia-smi --query-gpu=vbios_version --format=csv and ask MassiveGRID support to confirm the card's mode and vBIOS before you plan a deployment around MIG.
Does the RTX PRO 4000 Blackwell support MIG?
No. The RTX PRO 4000 Blackwell, including the SFF Edition in the 24 GB server, is not on NVIDIA's list of MIG-capable GPUs; the Blackwell workstation cards on that list are the RTX PRO 6000, 5000 and 4500. Several workloads can still share the 24 GB card, without hardware isolation.
Do MIG instances survive a reboot?
Not on this GPU generation. NVIDIA states that from Hopper onwards MIG mode is no longer persistent across reboots and that created MIG devices are not persistent either. Recreate them at boot with NVIDIA's mig-parted tool or a systemd unit that runs the nvidia-smi mig commands before your workloads start.
Is MIG faster than running several models on the whole GPU?
No. MIG divides the card; it does not add capacity, and an idle instance cannot lend its share to a busy one. Use it when workloads must not affect each other. For one model serving many users, a single server on the unpartitioned 96 GB card is the more efficient layout. Both options run on the RTX PRO 6000 servers on the GPU dedicated servers page.
A 96 GB RTX PRO 6000 you can partition yourself
The MassiveGRID RTX PRO 6000 server gives you the whole card and the whole machine: 96 GB of GDDR7 ECC, a Core Ultra 7 265, 128 GB of DDR5 and 2x 1.92 TB NVMe for $1,656.15 per month, with root access, a 1 Gbps port with unlimited traffic and a dedicated IPv4 address. Run it as one 96 GB GPU or, once the MIG prerequisites are confirmed, as up to four 24 GB instances.
See the RTX PRO 6000 ServerFurther Reading
- NVIDIA MIG User Guide: Getting Started with MIG (prerequisites, commands, containers)
- NVIDIA MIG User Guide: supported GPUs
- NVIDIA MIG User Guide: supported MIG profiles
- NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition product page
- Ollama documentation: Docker with NVIDIA GPUs
- OpenAI Whisper README: model sizes and VRAM
- RunPod pricing (Secure Cloud RTX PRO 6000 MIG instances, checked on 16 November 2026)
- Azure Retail Prices API (NC RTX PRO 6000 BSE v6 fractional sizes, East US)
- Google Cloud accelerator-optimized machine pricing (G4 fractional GPU sizes)