MIG and vGPU on the RTX PRO 5000, 4500 and 4000 Blackwell and RTX Ada: what NVIDIA lists

Which of the eleven cards can be shared between users or jobs

Three of the eleven NVIDIA professional cards on this page can be divided between users or jobs by NVIDIA's own documents, as of 10 November 2026. The RTX PRO 5000 Blackwell with 48 GB splits into two MIG instances of 24 GB under Linux, with a minimum driver and the display mode set to compute. The RTX PRO 4500 Blackwell Server Edition splits into two MIG instances of 16 GB and runs NVIDIA vGPU software from release 20.0. The RTX 5000 Ada Generation runs vGPU software from release 16.1 and has no MIG. For the RTX PRO 4500 Blackwell Workstation Edition the MIG guide leaves the answer open. The RTX PRO 4000 and 2000 Blackwell, the RTX 4500, 2000 and 2000E Ada, the RTX A1000 and the RTX A400 are in neither list: for these seven NVIDIA lists neither MIG nor vGPU.

MIG, Multi-Instance GPU, cuts a card into fixed instances with their own memory. NVIDIA vGPU software shares a card between virtual machines under a hypervisor. How the two differ in practice, and what vGPU means for licences, is the subject of our comparison MIG and vGPU: how many VMs one card holds. This page answers the narrower question an administrator has before an order, whether the card is on the list.

Two documents decide it. The first is the section "Supported GPUs" of NVIDIA's MIG User Guide, last updated 4 November 2026. The second is the list "NVIDIA Virtual GPU Software Supported GPUs", last updated 2 October 2026, which says of itself that it names only GPUs that support at least one release of NVIDIA vGPU software.

MIG and vGPU support card by card

NVIDIA MIG User Guide, Supported GPUs (4 November 2026); NVIDIA Virtual GPU Software Supported GPUs (2 October 2026); NVIDIA datasheets 5349550 (June 2026) and 5192801 (April 2026); product brief PB-12790-001_02 (20 April 2026); NVIDIA's RTX 5000 Ada product page. Each card name links to its spec page.

The MIG guide says the technique is supported from the Ampere generation on and then lists the supported products. The generation alone is therefore not enough: a card has to be in the table, and no RTX Ada card and no RTX A card is. A search for "RTX 5000 Ada MIG" or "RTX PRO 4000 Blackwell MIG" ends with a no for that reason, and the same holds for the RTX PRO 4000 Blackwell SFF Edition.

The cards in this table are available to order from us across the EU; send us the card and the number of users or virtual machines and we reply with a written quote within one business day.

RTX PRO 5000 Blackwell MIG with two instances of 24 GB

The datasheet of the NVIDIA RTX PRO 5000 Blackwell (June 2026) gives the two layouts for the 48 GB card: up to 2 × 24 GB or 1 × 48 GB. The MIG guide's profile table names them 1g.24gb, available twice, and 2g.48gb, available once; each exists as a +gfx variant too, and the 24 GB profile has further variants for the media engines.

The guide ties MIG on this card to four conditions:

• Linux. The guide supports MIG on Linux only.
• Driver 575.57.08 or later. NVIDIA serves two copies of the guide. The one updated on 4 November 2026 gives 575.57.08, an older copy of 13 October 2026 gave 575.51.03; plan on the higher figure. The current releases of branches R580, R595 and R620 are all above it.
• vBIOS 98.02.73.00.00 as the minimum.
• Display mode set to compute, with NVIDIA's DisplayModeSelector tool in version 1.72.0 or later. NVIDIA writes its warning that compute mode turns the physical display outputs off for the RTX PRO 6000; we would plan the RTX PRO 5000 the same way and keep remote access or a second display adapter for the console.

The commands, the reboot behaviour and the way back to graphics mode are in our step-by-step guide to MIG on RTX PRO Blackwell cards. What fits into a 24 GB or a 48 GB instance is worked through in RTX PRO 6000 MIG profiles, whose instance sizes are the same.

The 72 GB version of the RTX PRO 5000 is a different order in both documents. Its datasheet gives up to 2 × 36 GB, yet the MIG guide does not contain the card; the vGPU list contains only this version, from release 20.2, and not the 48 GB card. Our comparison of the RTX PRO 5000 with 48 GB and 72 GB covers the rest.

RTX PRO 4500 Blackwell MIG and the row without an edition

Two different cards carry this number, and the MIG guide does not say which one it means. Its table has one row, "RTX PRO 4500 Blackwell", with 32 GB, two instances and a minimum driver of 575.64. The profile table lists 1g.16gb twice and 2g.32gb once. No vBIOS minimum is given for the row.

For the passive NVIDIA RTX PRO 4500 Blackwell Server Edition the card's own documents settle it. The datasheet of April 2026 states up to two MIG instances at 16 GB each, and the product brief of 20 April 2026 lists MIG as supported with up to two instances and asks for driver R595 or later.

For the NVIDIA RTX PRO 4500 Blackwell Workstation Edition they do not. NVIDIA's MIG guide lists an RTX PRO 4500 Blackwell without naming the edition; the Workstation Edition datasheet does not mention MIG. Its product page is silent as well. We advise against building a deployment on MIG with the workstation card. If a project depends on it, get NVIDIA's confirmation in writing before the order.

The two cards differ in everything around the split as well: 165 W and passive cooling in a single slot on the Server Edition, 200 W and active cooling in two slots on the Workstation Edition, and no display output stated for the Server Edition. Our comparison RTX PRO 4500 or RTX PRO 4500 Server Edition sets them side by side.

We supply both RTX PRO 4500 cards and the RTX PRO 5000 across the EU; tell us how many users or jobs should share a card and which hypervisor or operating system runs on the host, and we reply with a written quote within one business day.

vGPU on the RTX 5000 Ada and the RTX PRO 4500 Server Edition

NVIDIA's vGPU list contains two of the eleven cards. The NVIDIA RTX 5000 Ada Generation has been supported since release 16.1 and is marked with full support; its datasheet names NVIDIA vPC and vApps and NVIDIA RTX Virtual Workstation. One setting is easy to miss: the display ports are on by default, and NVIDIA's product page says to turn them off when the card is used with vGPU software. The RTX 4500 Ada, which looks like the smaller sibling, is not on the list. That difference matters when an Ada fleet is extended for virtual desktops, as our comparison of the RTX 4500 Ada and RTX 5000 Ada explains.

The Server Edition is supported from release 20.0, also with full support, for RTX Virtual Workstation and virtual PC, with SR-IOV and 16 virtual functions. NVIDIA's vGPU 21.0 documentation lists MIG-backed vGPU for it, so a virtual machine can sit on one of the two 16 GB instances. The types and the desktops per card are in our guide to RTX PRO 4500 Server Edition vGPU profiles.

The release branch decides the host driver. As of November 2026 NVIDIA's vGPU index lists three: vGPU 21 on driver branch R620, a long-term branch whose release 21.0 dates from November 2026, with an end of life of November 2029; vGPU 20 on R595, newest release 20.3, to March 2027; and vGPU 19 on R580, a long-term branch with 19.7 as its newest release, to July 2028. The types reference of vGPU 21.0, updated 5 November 2026, has sections for both cards. The Server Edition is not in vGPU 19. Host and workstation drivers are the subject of our guide to the NVIDIA RTX Enterprise driver on Windows 10 and 11.

SR-IOV and passthrough on the RTX A1000 and other workstation cards

Queries such as "RTX A1000 SR-IOV" come from administrators who want to hand a small card to a virtual machine. NVIDIA's documents say little here, and we do not fill the gap with assumptions.

SR-IOV appears in NVIDIA's vGPU 21.0 documentation as part of vGPU: virtual functions that give virtual machines with vGPUs IOMMU protection. The page names no GPU. For workstation cards outside the vGPU list, the RTX A1000 and RTX A400 among them, we found no NVIDIA statement on SR-IOV.

Passthrough of a whole card to one virtual machine is documented in two places. The vGPU documentation states that in pass-through deployments all GPUs that support NVIDIA vGPU software are supported, which among the eleven cards of this page means the Server Edition and the RTX 5000 Ada. The table of GPUs supported for device passthrough in the data-centre driver notes of 620.70.06 names none of the eleven; it lists XenServer, VMware vSphere, KVM on Red Hat Enterprise Linux 9 and Hyper-V as hypervisors. For the other nine cards NVIDIA makes no statement. Our conclusion is that passthrough on them is a matter of testing on your own hypervisor, without documented support behind it.

Which card to order for which kind of sharing

Two isolated halves on a Linux workstation. The RTX PRO 5000 Blackwell with 48 GB is the only workstation card of the eleven with documented MIG. Two developers, or one inference service and one test environment, each get 24 GB, under the conditions named above: Linux, the minimum driver and vBIOS, and the display mode set to compute.

Virtual desktops or virtual machines in a rack server. The RTX PRO 4500 Server Edition brings vGPU and MIG-backed vGPU on a passive single-slot card. It needs the airflow of a server chassis.

Virtual workstations on an Ada platform. The RTX 5000 Ada with 32 GB is the one Ada card of the eleven on the vGPU list.

More than two instances per card. None of the eleven offers that. NVIDIA's MIG guide lists the RTX PRO 6000 Blackwell with 96 GB for four instances, in all three editions.

No sharing to plan on. That leaves eight cards: the RTX PRO 4500 Workstation Edition, whose MIG row names no edition, and the seven that are in neither document, the RTX PRO 4000, the RTX PRO 2000, the RTX 4500 Ada, the RTX 2000 Ada, the RTX 2000E Ada, the RTX A1000 and the RTX A400. Our reading is that separate users or jobs then need separate cards. The RTX PRO 4000 is single slot, so four of them occupy four adjacent slots, as our page on four RTX PRO 4000 in one workstation shows; the other one-slot boards are in our comparison of single-slot NVIDIA RTX cards.

If it is not yet clear whether MIG, vGPU or separate cards fit, describe the workloads and the host. The first consultation is free of charge, and the written quote names the card that NVIDIA lists for that use. People who look for a Server Edition of the RTX PRO 5000 or 4000 for this purpose find the cards that exist in our article RTX PRO 5000 Server Edition: which cards exist.

FAQ