MIG and vGPU: how many VMs one card holds

MIG vs vGPU: what gets sliced

MIG cuts the card in hardware: each instance gets its own compute blocks, its own memory and its own memory bandwidth, so a neighbour on the same card cannot affect you. vGPU works differently: all VMs see the same card and take turns.

MIG-backed vGPU combines the two and assigns a vGPU to a hardware slice: one vGPU can own the whole slice or, since vGPU 19.0, several time-sliced vGPUs can share it. On the RTX PRO 6000 Server Edition, NVIDIA’s RTX vWS sizing guide (updated 19 August 2026) reports that “at the 48 GB profile size, MIG-backed vGPU delivered 20% higher performance than time-sliced vGPU”; its graphics benchmark setup lists DC-48Q and DC-2-48Q.

MIG instances per card

The instance count is fixed by architecture; no setting changes it. Only data-centre cards on HBM split seven ways.

H100 and A100 split into 7, A30 into 4; the RTX PRO 5000 72 GB into 2 × 36 GB, per NVIDIA’s datasheet; for the RTX PRO 4500 the MIG guide names no edition; NVIDIA’s blog of 22 April 2026 and Lenovo’s product guide give the Server Edition two instances of 16 GB, while the workstation card’s datasheet and product page do not mention MIG, so have MIG support on the workstation card confirmed in writing

On the RTX PRO 6000 the profiles are 1g.24gb (four), 2g.48gb (two) and the full 4g.96gb, each with a +gfx variant that supports graphics inside the slice, which MIG did not offer before Blackwell. In a four-way split each 1g.24gb slice gets a quarter of the memory and a quarter of the SMs, according to NVIDIA’s profile table. Which models fit in each slice is in our RTX PRO 6000 MIG profiles guide.

VMs and seats per card, by workload

The 48 and the 12 are two different ceilings. Forty-eight is the graphics ceiling, reached with time-sliced vGPUs inside the four MIG instances, which on vSphere needs VMware Cloud Foundation 9.1 or later (vGPU 19.4 or 20.0 onwards); twelve is the compute ceiling. The sizing guide adds that the best-effort scheduler often supports two to three times as many users when they are not all active at once; the per-session memory maths decides how far that stretches. In operation, if VM video memory keeps hitting the profile ceiling, move to a larger profile rather than adding seats.

We supply all four cards in this table, the RTX PRO 6000 Server Edition, the RTX PRO 5000, the RTX PRO 4500 Server Edition and the L40S, and size virtual-workstation nodes by seats per card. Tell us the workload and the number of seats, and you get a written quote.

vGPU licences, and what happens without one

There are three graphics editions, which NVIDIA’s packaging and licensing guide counts per concurrent user rather than per card, and only one RTX PRO 6000 edition supports vGPU at all. vApps streams applications to one display at 1280×1024 and uses A-series profiles. vPC provides virtual desktops with up to four displays at 2560×1600 or one at 5120×2880, uses B-series profiles and has no CUDA. vWS is for CAD and development, with resolutions up to 7680×4320, Q-series profiles, CUDA and multiple vGPUs per VM. The three are separate licences with different entitlements.

Compute C-profiles are not covered by these three; they are available only through NVIDIA AI Enterprise, licensed per GPU in the server rather than per user. NVIDIA’s vGPU licensing guide maps vApps, vPC and vWS to A, B and Q-series profiles only, and its AI Enterprise documentation requires an AI Enterprise licence for every C-series vGPU. A fleet with both graphics and compute VMs therefore needs two separate licence purchases.

We supply both licence types on the same quote as the cards, so the licence count is clear before you order. Write to us with the number of concurrent graphics users and how many GPUs run compute VMs.

When the licence drops, a graphics vGPU runs at full capability for twenty minutes; after that, NVIDIA’s current client licensing guide (updated 29 September 2026) describes its performance as degraded, and after 24 hours as further degraded, without figures. The guide for vGPU release 16, a release that does not support the RTX PRO Blackwell cards, named a 15 fps cap with degraded CUDA and, after a day, 3 fps with CUDA stopped. Compute (C-series) vGPUs drop to an idle level of compute after twenty minutes. A licensed compute VM that loses the licence server keeps working for seven days by default on NVIDIA’s cloud licence service, a period that is configurable on a DLS; a graphics VM may run without it for up to one day. After reconnection the limits lift, but applications may need restarting.

Hypervisors

NVIDIA’s validated-platforms pages per hypervisor (updated 29 September 2026) mark MIG-backed vGPU on the RTX PRO 6000 and 4500 Blackwell Server Edition as not supported (N/A) on Windows Server and XenServer, and list it on the Linux with KVM page, which names Nutanix AHV and leaves the supported releases to the hypervisor vendor. On Proxmox, MIG instances are created by hand from the CLI, the vGPU layout across slices cannot be changed, and AI Enterprise is not officially supported there. Live Migration requires the same GPU model on both hosts and a target driver from the same release branch that is the same or newer (never older), and does not work between different MIG profiles.

Known problems: vBIOS, compute mode and edition limits

The longest-running issue concerns the RTX PRO 6000 Blackwell vBIOS. Cards shipped with vBIOS 98.02.52.00.02, while MIG on the Workstation Edition requires 98.02.55.00.00, which did not exist at launch. Owners got “Unable to enable MIG Mode: Not Supported” and raised it on the forum. The fix is to update the vBIOS through your supplier, switch the card to compute mode with DisplayModeSelector, then assign profiles, as our MIG runbook describes step by step. Ask for the firmware version before you pay.

Compute mode changes two things. On a workstation where the card also drives the monitor, the display outputs are switched off. According to an answer on NVIDIA’s developer forum (January 2026), BAR1 grows from 256 MB to 64 GB of contiguous address space, and the motherboard must be able to allocate it.

For virtual machines, the main difference between the editions is vGPU: all three RTX PRO 6000 editions do MIG, but only the Server Edition supports vGPU. A workstation card can split into four slices for local containers, but it cannot hand them to virtual machines. Some points apply to all cards. There is no peer-to-peer between MIG instances on different GPUs (since driver R570, P2P between instances on the same GPU works), GPUDirect RDMA works from a GPU instance, and NCCL does not work with MIG. NVIDIA describes MIG as a fit for workloads that do not fully saturate the GPU’s compute capacity, and MIG runs only on Linux with CUDA.

MIG and throughput

Partitioning raises total throughput only when one task leaves the card partly idle.

A single small task runs slower on a slice than on the whole card, because a 1g slice has a seventh of the SMs and an eighth of the memory and its bandwidth. The gain appears only when many slices are busy. The second reason to slice in hardware is predictability: a MIG instance keeps its own SMs, memory and memory bandwidth, so a busy neighbour on the same card does not stretch its response times.

What we supply

We supply five of the cards this article sizes, across the EU with manufacturer warranty: the RTX PRO 6000 Server Edition, the only edition with both MIG and vGPU (passive cooling, so it needs server airflow), the L40S for time-sliced VDI, the RTX PRO 4500 Server Edition for light and medium vGPU users or two MIG instances of 16 GB, the RTX PRO 5000 Blackwell in 48 GB and 72 GB where two isolated instances of 24 GB or about 36 GB each are enough, and the H200 NVL for up to seven isolated compute instances. Quantities and lead time are confirmed in writing with the quote.

FAQ