MIG and vGPU: how many VMs one card really 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.

There is also a hybrid. MIG-backed vGPU lays a virtual card over a hardware slice, and inside the slice time-slicing still applies. On the 48 GB profile this mode measured about 20% faster than pure time-slicing, per NVIDIA’s sizing guide.

How many pieces each card gives

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

H200 and A100 split into 7, A30 into 4

On the RTX PRO 6000 the profiles are 1g.24gb (four), 2g.48gb (two) and the full 4g.96gb — each with a +gfx variant, i.e. graphics inside the slice. That did not exist before Blackwell. In a four-way split, 96 GB and 188 SM divide into 23.6 GiB and 46 SM per slice; GPT-OSS-20B on such a slice produces around 80 tokens per second.

Card, scenario, seats

48 and 12 do not contradict each other: forty-eight is the graphics ceiling, twelve the compute one. The sizing guide adds that the best-effort scheduler “often results in a 2–3× oversubscription” when users are not all active at once — the per-session memory maths decides how far that stretches. The monitoring rule: if VM video memory keeps hitting the profile ceiling, increase the profile — not the seat count.

Licences: without one, the card works for 20 minutes

There are three products, licensed per concurrently active user, not per card — and only one RTX PRO 6000 edition supports vGPU at all. vApps streams applications: one display, 1280×1024, A-series profiles. vPC is virtual desktops: four displays up to 5120×2880, B-series, no CUDA. vWS is CAD and development: up to 7680×4320, Q-series, CUDA, multiple vGPUs per VM. In NVIDIA’s price list vWS costs five times vPC and twenty-five times vApps — a product mistake costs more than a card mistake.

Compute C-profiles are not in that trio: they live only in NVIDIA AI Enterprise, licensed per GPU in the server, not per user. Asked directly whether vWS/vPC/vApps are included in AI Enterprise, NVIDIA answered: no — C-profiles only. A mixed fleet means two procurements.

When the licence drops, degradation is two-stage: twenty minutes at full speed, then 1280×1024, three frames per second and CUDA off. After reconnection the limits lift — but applications need restarting.

Hypervisors

Two details break plans most often. 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 and identical driver and hypervisor versions on both hosts — and does not work between different MIG profiles.

What breaks in practice

The longest saga belongs to the RTX PRO 6000 Blackwell. Cards shipped with vBIOS 98.02.52.00.02 while MIG requires 98.02.55.00.00, which did not exist at launch. Owners got “Unable to enable MIG Mode: Not Supported” and asked the obvious question on the forum. The working procedure: update the vBIOS through your supplier, switch the card to compute mode with DisplayModeSelector, then assign profiles. Ask for the firmware version before you pay.

Compute mode has consequences. On a workstation where the card also drives the monitor, video output goes dark. BAR1 grows from 256 MB to 64 GB of contiguous address space — and the motherboard must be able to allocate it.

The main edition limit: all three RTX PRO 6000 editions do MIG, but only Server Edition supports vGPU. A workstation card can split into four slices for local containers — it cannot hand them to virtual machines. Limits common to all cards: MIG disables NVLink and peer-to-peer, GPUDirect between MIG devices is unsupported, NCCL does not work with MIG. NVIDIA puts it plainly: MIG is for workloads that do not need a whole GPU. Linux with CUDA only.

Does slicing make things faster?

Depends on whether one task saturates the card. If not, partitioning raises aggregate throughput.

The last row is the important one: a single task gets worse on a slice, because a slice is literally 1/7 of the card across all three resources at once. The win appears only when there are many busy slices — and scaling is sublinear (the MIGPerf benchmark showed N instances do not deliver N× throughput). The second reason to slice in hardware is predictability: in a seven-tenant test on A100, a protected service held P95 around 1,319 ms on MIG versus 2,499 ms on a shared card.

What we supply

Eurokommerz supplies the RTX PRO 6000 Server Edition (the only edition with both MIG and vGPU — passive cooling, needs server airflow), the L40S for time-sliced VDI, and the RTX PRO 5000 Blackwell in 48 GB and 72 GB when two isolated 24 GB instances are enough — EU-wide, with manufacturer warranty.

FAQ