RTX PRO 6000 MIG profiles: the 24, 48 and 96 GB instances on every Blackwell edition

RTX PRO 6000 MIG profiles at a glance

The RTX PRO 6000 Blackwell supports MIG with three instance sizes: up to four 1g.24gb instances, up to two 2g.48gb instances or one 4g.96gb instance per card. NVIDIA’s MIG user guide, last updated on 13 October 2026, lists 11 profiles, the three sizes and their variants, in one table for the Workstation Edition, the Max-Q Workstation Edition and the Server Edition.

NVIDIA MIG user guide, “Supported MIG Profiles”, table “GPU Instance Profiles on RTX PRO 6000”, updated 13 October 2026. Reported memory and SMs: nvidia-smi mig -lgip output on an RTX PRO 6000 Server Edition with driver 580.95.05, published by NETWAYS Web Services on 10 December 2025.

The gigabytes in a profile name are nominal. On the card in that published output, a 1g.24gb instance reported 23.62 GiB, a 2g.48gb 47.38 GiB and a 4g.96gb 95.00 GiB, so plan with those figures and confirm them with nvidia-smi mig -lgip on your own card.

MIG support on the Server Edition, Workstation and Max-Q

The guide’s list of supported GPUs names all three editions with 96 GB, compute capability 12.0 and a maximum of four instances, and also lists the RTX PRO 5000 Blackwell (48 GB) and the RTX PRO 4500 Blackwell (32 GB) with two each. The editions differ in what has to happen before the first instance exists.

NVIDIA MIG user guide, “Supported GPUs” and “Getting Started with MIG”, updated 13 October 2026; Max-Q datasheet 5349650 (June 2026); NVIDIA AI Enterprise 8.2 Blackwell vGPU reference (2 September 2026); Server Edition product brief SP-12355-001_v02 (27 June 2025) for the default Display Off mode; vGPU support matrices as in our vGPU profiles guide.

The guide states that “Display mode by default will be set to graphics, this must be set to compute before MIG can be enabled for Workstation Edition and Max-Q Workstation Edition GPUs”, with DisplayModeSelector 1.72.0 or later, after which the card drives no monitor. The Max-Q datasheet gives the same split as the guide, “Up to 4x 24 GB”, “Up to 2x 48 GB” or “Up to 1x 96 GB”. Driver checks, the display-mode switch and the nvidia-smi commands are in our MIG setup guide for RTX PRO Blackwell. MIG mode is “only persistent as long as the driver is resident”, so the layout is recreated at every start-up.

How the slices add up: memory, SMs and engines

The RTX PRO 6000 is divided in quarters. Each 1g.24gb instance holds a quarter of the memory, the L2 cache and the SMs, a 2g.48gb holds half and a 4g.96gb the whole card. In the published -lgip output, a quarter has 46 SMs and a half 94, so four quarters add up to 184 SMs while two halves give the full 188. Copy engines follow the size, with 1, 2 and 4 per instance.

The media engines decide the variant. The full card has four video decoders (NVDEC), four encoders (NVENC), four JPEG decoders and one optical flow accelerator (OFA), and a plain 1g.24gb gets one of each except the OFA. The 1g.24gb+me profile carries the card’s single OFA, so it exists once. The +me.all variant “Allocates all available media engines to this instance”, so only one fits a card. The -me variant “Excludes all media engines for pure compute workloads”, which suits LLM serving and embedding models.

Which RTX PRO 6000 MIG profiles can share one card

With four quarters, four layouts fill a card: four 1g.24gb instances, two 2g.48gb, one 2g.48gb with two 1g.24gb, or one 4g.96gb. The static MIG Manager configuration in NVIDIA’s GPU Operator repository, read on 16 October 2026, defines that third layout as “all-balanced” for the RTX PRO 6000 with device ID 0x2BB5, with two 1g.24gb and one 2g.48gb. There is no size between 24 and 48 GB, so a model of 30 GB takes a half-card instance.

MIG mode is set per GPU, so a server can run some cards split and others whole. The guide shows the placements only as a diagram, and nvidia-smi mig -lgipp lists them on the card itself; create the half-card instance first, so the quarters fill the places it leaves. In Kubernetes, a node with several profiles or with whole cards next to split ones needs the device plugin’s mixed strategy, which our Kubernetes GPU sharing guide explains.

For every layout, NVIDIA’s guide states that “NCCL is currently not supported with MIG” and that “CUDA IPC across GPU instances is not supported”, and with driver R570 only peer-to-peer between instances on the same GPU works. A model split across several cards therefore runs on whole cards with MIG off.

Graphics and compute MIG on Blackwell

NVIDIA’s deployment considerations say that “No graphics APIs are supported” in MIG instances, with one exception, “RTX Pro 6000 Blackwell GPUs where certain MIG profiles support graphics.” These are the +gfx profiles, described as “Adds support for graphics APIs (new in GB20X)”. NVIDIA calls the feature Universal MIG, which “enables both compute and graphics workloads to run on the same GPU with hardware isolation.”

A +gfx instance can run rendering or visualisation jobs that use graphics APIs such as OpenGL or Vulkan, and on the Server Edition the graphics vGPU types for virtual desktops sit on these instances. The H200 table in the same guide has no +gfx profiles, as our H200 NVL MIG profiles reference shows. The guide does not list which +gfx and plain profiles can share one card, so check such a layout with nvidia-smi mig -lgipp first.

What fits in a 24, 48 or 96 GB MIG instance

An instance has to hold the model weights and the KV cache of the conversations it serves. Our sizing guides give the cache 0.9 × the memory the device reports, minus 3 GiB for CUDA graphs and activations, minus the weights. Applied to the reported figures, that leaves 18.3 GiB in a 1g.24gb instance, 39.6 GiB in a 2g.48gb and 82.5 GiB in a 4g.96gb for weights and cache. In an FP8 cache, one conversation at a full 8K context takes 0.625 GiB on Qwen3-14B, 1 GiB on Qwen3-32B and 1.25 GiB on Llama 3.3 70B: 2 × layers × KV heads × head dimension × 8,192 tokens at one byte.

Our estimate with the sizing rule above, FP8 KV cache, full 8K context. Weights: summed safetensors files of Qwen/Qwen3-14B-FP8, Qwen/Qwen3-32B-FP8 and nvidia/Llama-3.3-70B-Instruct-FP8 on Hugging Face; layers and KV heads from each config.json, read on 16 October 2026. Whole card: 95.6 GiB reported, as in our sizing guides.

A 24 GB instance suits embedding and reranking models, speech-to-text and small task models; gpt-oss-20b, at 13.8 GB in its checkpoint files, leaves about 5.4 GiB of cache room. A 14B model fits, but with four conversations it serves one developer or a test. Splitting also costs capacity, because every instance loads its own copy of the weights. Four 1g.24gb instances with Qwen3-14B hold 16 conversations in total, two 2g.48gb hold 78, and the same model on the whole card holds 108. A 32B model in FP8 needs a half-card instance and leaves room for only seven conversations there, so a shared 32B assistant belongs on a whole card. Llama 3.3 70B in FP8 needs 96 GB. In a 4g.96gb instance it loses one conversation against the card with MIG off, because the instance reports 95.00 GiB instead of 95.6.

We supply the RTX PRO 6000 in all three editions as cards or in AI servers built to order. Tell us the models and users per instance, and we reply within one business day with a configuration, MIG layout per card included, and a quote.

MIG-backed vGPU on the RTX PRO 6000 Server Edition

Only the Server Edition hosts MIG-backed vGPU, and NVIDIA’s MIG-backed vGPU page says that on this card “up to four MIG slices can be created on a single GPU”. The AI Enterprise 8.2 reference lists 11 MIG-backed compute types for it, from DC-4-96C to DC-1-8C, under “Required license edition: NVIDIA AI Enterprise”. DC-1-24C, DC-2-48C and DC-4-96C each fill one instance of 24, 48 or 96 GB, and the smaller types are time-sliced inside an instance. The full tables, graphics types, licences and hypervisor versions are in our RTX PRO 6000 Server Edition vGPU profiles guide. NVIDIA’s page notes that live migration between different MIG profiles is not supported and that “The target host must have a matching MIG configuration”, so hosts in one cluster should share a layout. NVIDIA’s Server Edition product brief adds that the default Display Off mode “is required to run NVIDIA Virtual GPU software”.

Worked example: one 8-card Server Edition server for three departments

The example is our estimate, not an NVIDIA reference design. One server with eight RTX PRO 6000 Server Edition cards serves three departments of a company with about 1,000 staff. We assume up to eight developers running jobs at the same moment, a customer service assistant with RAG and a peak of about 60 conversations, and a document queue for finance and legal.

Cards 1 and 2 run four 1g.24gb-me instances each, eight isolated GPUs for the developers, each for a model up to 14B in FP8 or a notebook. Cards 3 and 4 stay whole with Qwen3-32B in FP8 for customer service, 51 conversations at 8K each and 102 together. Card 5 runs the balanced layout: two 1g.24gb instances for a 0.6B embedding model and a 0.6B reranker, about 1.2 GB each in BF16, and a 2g.48gb with Qwen3-14B for ticket summaries, 39 conversations. Card 6 runs two 2g.48gb instances with Qwen3-32B, one for finance and one for legal, so the two queues share no memory or SMs. Card 7 runs Llama 3.3 70B in FP8 on the whole card for harder documents, and card 8 stays whole as a reserve and test bed.

That gives 13 MIG instances on four cards and four whole cards, which needs the mixed strategy in Kubernetes. If one customer service card fails, card 8 takes a third copy of the assistant; if the developers grow to twelve, it takes four more quarters instead, and the reserve moves to a second server. Our eight-card RTX PRO 6000 allocation covers power, PCIe topology and a second server at 2,000 staff.

We build servers with up to eight RTX PRO 6000 Server Edition cards and check the rack, power and airflow before we quote. Describe your departments, their models and peak use in the form below, and we propose the MIG layout card by card.

What we supply

We supply the NVIDIA RTX PRO 6000 Blackwell in all three editions, Server Edition, Workstation Edition and Max-Q, with manufacturer warranty, on one EU contract and invoice, as cards from our RTX PRO 6000 page or in AI servers built to order. Operating system, drivers, CUDA and a container runtime are installed on request. MIG on bare metal needs no licence; where instances back virtual machines, the NVIDIA AI Enterprise or vGPU licences come on the same invoice. Models, RAG and the platform on top are our Private AI/ML service, with engineering by our partner Vixen.UNO.

FAQ