MIG on a Workstation GPU: Splitting an RTX PRO 6000's 96GB Into Isolated Instances

Bottom Line

• The RTX PRO 6000 Blackwell is the first workstation card with MIG. Profiles: 4 x 1g.24gb, 2 x 2g.48gb, or 1 x 4g.96gb. All three variants, per NVIDIA’s MIG user guide.
• Enabling it is real work. A vBIOS your reseller has to send you, a compute-only firmware mode that disables the card’s display outputs, Linux, an R580-class driver, ReBAR on.
• Each 24GB slice gets a quarter of the SMs and roughly a quarter of the bandwidth, about 450 GB/s. A used RTX 3090 has the same 24GB at 936 GB/s.
• MIG buys isolation, not capacity. One person running two models should not split the card.
• Instances cannot be resized live. Reconfiguring means destroying and recreating them. MIG mode does not persist across reboots on this architecture.

What MIG Actually Is

Multi-Instance GPU partitions one physical GPU into several hardware-isolated GPUs. Each instance gets its own slice of streaming multiprocessors, its own memory, its own share of the memory controllers and L2 cache, and its own fault domain. A process on one instance cannot see, address or starve another. To the driver each instance is a separate CUDA device.

That is different from what most people do today, which is run two models on one card and let them share. Sharing works. It gives no guarantees. A runaway context on one model can evict the other, and a crash in one process can take the device down for both. MIG exists for the cases where that is unacceptable.

Until this generation MIG was a datacentre feature: A100, A30, H100, H200, B200. The RTX PRO 6000 Blackwell brings it to a card you can put in a tower under a desk, and NVIDIA’s own Max-Q datasheet leads with it.

The Profiles

From NVIDIA’s MIG user guide, table for the RTX PRO 6000 Blackwell (96GB):

Two things the table implies that the marketing does not state.

Bandwidth is sliced with memory. MIG partitions the memory controllers along with the memory, so a 1g.24gb instance gets a quarter of the memory slices and, by the design of the partition, roughly a quarter of the 1,792 GB/s: about 450 GB/s. We have not seen NVIDIA publish a per-instance bandwidth figure for this card, so treat 450 GB/s as arithmetic on the partition, not a measurement. Token generation is bandwidth-bound, so a 24GB slice generates tokens at roughly a quarter of the full card’s rate.

No instance can be bigger than the card, and no model can span instances. MIG isolates in both directions. A 75GB model does not run on two 48GB instances. If you might need the whole card for one model, you leave MIG off for that job, which means destroying the instances.

What It Costs to Turn On

This is the section that decides whether MIG is for you. Every item below is documented in NVIDIA’s own developer forums or user guide, and several of them cost owners weeks in 2025.

1. A vBIOS your card probably did not ship with. The minimum is 98.02.55.00.00 on the Workstation Edition and 98.02.6A.00.08 on the Max-Q. Cards from the first retail wave carried 98.02.52.00.02, and nvidia-smi -mig 1 on those answers “Not Supported”. NVIDIA forum staff direct owners to contact the reseller with the serial number, for a PNY card that means PNY, to obtain the field upgrade tool. NVIDIA does not distribute it directly, and one staff reply floated an RMA as the faster path. Ask the seller which vBIOS the card carries before you buy it for MIG.

2. Compute-only display mode, which kills the display outputs. MIG needs the large BAR1 address space that the card’s graphics firmware mode does not expose. NVIDIA’s displaymodeselector --mode=compute switches the firmware, and NVIDIA’s own guide states that on a single-card workstation this disables physical display output. The four DisplayPort 2.1 connectors go dark. You need an integrated GPU or a second card for a monitor, or you run the box headless.

3. Linux. The MIG user guide lists CUDA-supported Linux distributions as the operating environment. A forum owner who tried the mode switch on Windows got the switch to work and MIG itself to answer “Not Supported”. Treat MIG on this card as Linux-only until NVIDIA documents otherwise.

4. A current driver stack and ReBAR. R580-class driver with CUDA 13, and Resizable BAR enabled in system firmware. Both are quick. Both are also things a workstation vendor may ship disabled.

5. Re-creation on every boot. On Hopper and later, enabling MIG mode no longer needs a GPU reset, which is good. It is also no longer persistent across reboots, so you need a systemd unit that enables MIG and recreates your instances at startup, and every service that depends on them has to wait for it.

Add it up and MIG on this card is a day of setup and a dependency on your reseller, on top of a card that lists at $13,250 as of August 2026.

Who Should Split the Card

The honest summary: MIG is a multi-tenant feature. If the word “tenant” does not describe your situation, the unsplit card is faster and simpler.

The 24GB Slice Against a 24GB Card

The comparison that puts the feature in proportion. As of August 2026:

Four 3090s are cheaper and faster per instance. They are also four cards, a server board with the lanes to feed them, roughly 1,400W of GPU, and a dedicated 20A circuit. One RTX PRO 6000 in MIG mode is four isolated 24GB devices in one slot on one power connector, with ECC and a warranty. That is what the price buys. It does not buy throughput.

What Runs in a Slice

A 1g.24gb instance is, for model-fitting purposes, a 24GB GPU with a quarter of the bandwidth. Our best local LLM for 24GB picks apply for weights. Adjust expectations for speed: a dense 27B at Q4 that runs at 30-plus tok/s on a full card lands near a quarter of that on a slice. A 3B-active MoE such as Qwen3.6-35B-A3B is the model class that stays comfortable, because it reads few bytes per token. The MoE vs dense at 24GB page is written for exactly this memory shape.

A 2g.48gb instance is a 48GB GPU at ~900 GB/s, which is close to a 3090’s bandwidth with double the memory. It runs a dense 70B at Q4 with the same context ceiling as any 48GB card.

The Card

The RTX PRO 6000 Blackwell is the one workstation card that does any of this. 96GB GDDR7 with ECC, 1,792 GB/s, PCIe 5.0 x16, MIG up to four instances on all three variants. NVIDIA lists it at $13,250 as of August 2026; street runs higher.

NVIDIA RTX PRO 6000 Blackwell 96GB — buy it for MIG only if the multi-tenant table above describes you, and ask the seller for the vBIOS version before you pay. For a 24/7 shared box the 300W Max-Q variant is the better fit; it has the same MIG profiles and a lower vBIOS floor to clear, and our comparison explains the rest. We do not have a link for the Max-Q; 300W total board power on the listing is how you identify it.

If the goal is several isolated 24GB devices and the budget is not five figures, the cheaper isolation is several cards:

EVGA GeForce RTX 3090 24GB — $1,000–1,300 used, 936 GB/s, and a whole card per tenant. Two of them is the 48GB tier; four is a multi-GPU rig with its own problems.

Enabling It, In Outline

On a card with a qualifying vBIOS, on Linux, with ReBAR on and no display attached to it:

# 1. Switch the card to compute-only mode (disables its display outputs), then reboot
sudo ./displaymodeselector --mode=compute --gpu=0

# 2. Enable MIG mode
sudo nvidia-smi -i 0 -mig 1

# 3. List the profiles the driver offers
nvidia-smi mig -lgip

# 4. Create four 24GB GPU instances, each with its compute instance
sudo nvidia-smi mig -i 0 -cgi 1g.24gb,1g.24gb,1g.24gb,1g.24gb -C

# 5. Confirm; each instance appears as its own CUDA device
nvidia-smi -L

Put steps 2 and 4 in a systemd unit that runs before your inference services. To go back to one 96GB device: destroy the compute instances, destroy the GPU instances, nvidia-smi -i 0 -mig 0. Every workload on the card stops while you do it.

The Decision

See Also

• RTX PRO 6000 Max-Q vs Workstation Edition — the datasheet where the MIG rows come from
• Is 96GB of VRAM Enough in 2026? — what the unsplit card runs, and why most owners should keep it unsplit
• Best Local LLM for 96GB VRAM — the model picks for the whole card
• Can You Put Four RTX PRO 6000 Max-Q Cards in One Workstation? — sixteen instances, and the circuit they need
• MoE vs Dense on a 24GB Card — the right model shape for a quarter-bandwidth 24GB slice
• Motherboard and CPU for a Multi-GPU Rig — the other way to get four isolated devices
• The Electrical Requirements Nobody Puts in an AI Build Guide — why four cards is a wiring job and one card is not

Sources

• NVIDIA Multi-Instance GPU User Guide, “Supported GPUs”: RTX PRO 6000 Blackwell Server Edition, Workstation Edition and Max-Q Workstation Edition, 96GB, maximum 4 instances. “Supported MIG Profiles”: MIG 1g.24gb (1/4 memory, 1/4 SM, 4 instances), 2g.48gb (1/2, 1/2, 2), 4g.96gb (full, 1), plus +gfx and media-engine variants
• NVIDIA MIG User Guide, “Getting Started”: from Hopper onward enabling MIG no longer requires a GPU reset and MIG mode is no longer persistent across reboots; supported environment is CUDA-supported Linux distributions; on RTX PRO Blackwell GPUs setting display mode to compute disables physical display output
• NVIDIA Display Mode Selector tool page: compatible with all three RTX PRO 6000 Blackwell variants; supported GPUs must operate in compute-only mode with physical display outputs disabled for MIG
• NVIDIA developer forums, “RTX PRO 6000 Blackwell Workstation Edition — vBIOS too old for MIG (98.02.52.00.02)”: minimum vBIOS 98.02.55.00.00, driver 580.159.03 with CUDA 13.0, nvidia-smi “Not Supported” error, NVIDIA staff directing owners to the reseller with the serial number for the field upgrade tool
• Level1Techs forum thread on MIG support for the RTX PRO 6000, citing NVIDIA documentation for the Max-Q minimum vBIOS of 98.02.6A.00.08 and the two-step enablement (vBIOS update, then displaymodeselector compute mode)
• Exxact support guide on enabling MIG on RTX PRO 5000/6000 Blackwell: driver 580.95 or newer, CUDA 13.0, Resizable BAR enabled (read via search summary; the page itself refused our fetch)
• NVIDIA RTX PRO 6000 Blackwell product page: 96GB GDDR7 ECC, 1,792 GB/s, 600W, dual slot, four DisplayPort 2.1, MIG up to four instances
• Per-instance bandwidth (~450 GB/s) is our arithmetic on a quarter partition of 1,792 GB/s and is labelled as derived; NVIDIA publishes no per-instance figure we could find
• Prices as of August 2026 from our hardware price reference: RTX PRO 6000 $13,250 NVIDIA list (VideoCardz, wccftech); used RTX 3090 $1,000–1,300 (ResalePrices, gpudojo)