RTX PRO 6000 vGPU profiles on the Blackwell Server Edition: time-sliced and MIG-backed types

RTX PRO 6000 vGPU profiles: the short answer

The RTX PRO 6000 Blackwell Server Edition offers two families of vGPU profiles. Time-sliced profiles share the whole card in turns, with a fixed frame buffer per VM, and MIG-backed profiles place each vGPU on a GPU instance of one, two or four of the card’s 24 GB slices. On Blackwell, the MIG-backed types cover graphics as well as compute. NVIDIA’s MIG-backed vGPU page says that “Universal MIG technology on Blackwell enables both compute and graphics workloads to be consolidated” on the same physical GPU.

NVIDIA names the types DC, then the slice count for MIG-backed types, then the frame buffer in GB and the series letter. DC-8C is an 8 GB time-sliced compute vGPU, DC-1-24C a compute vGPU on one 24 GB slice, and DC-1-6Q a 6 GB graphics vGPU inside one slice. The series letter sets the licence. The vGPU user guide (release 20.0 to 20.2, 30 September 2026) states that “Q-series vGPU types require a vWS license”, that B-series types need a vPC licence or vWS, and that A-series types need vApps. C-series types require NVIDIA AI Enterprise, according to NVIDIA’s AI Enterprise vGPU reference.

How many VMs other cards hold, and how MIG compares with time-slicing in general, is in our guide to MIG and vGPU VMs per card.

Time-sliced vGPU profiles on the RTX PRO 6000

Time-sliced vGPUs divide the 96 GB of frame buffer into equal or mixed sizes and share the GPU’s compute in turns. NVIDIA’s AI Enterprise vGPU reference for Blackwell (release 8.2, updated 2 September 2026) lists seven compute types for the card, with the same maximum in equal-size and mixed-size mode.

Compute types: NVIDIA AI Enterprise 8.2, Blackwell vGPU types reference, RTX PRO 6000 Server Edition table. Graphics: NVIDIA vWS sizing guide (profile sizes, test setup, DC-24Q and DC-12Q examples) and vPC sizing guide, both updated 19 August 2026; graphics maximums are shown only where these guides state them.

The full Q, B and A tables sit in section 9.1.10 of NVIDIA’s vGPU user guide, which lists MIG-backed and time-sliced types for each series. The vWS sizing guide gives the card’s profile sizes as “2GB, 3GB, 4GB, 6GB, 8GB, 12GB, 16GB, 24GB, 32GB, 48GB, 96GB” and notes that “1 GB profiles are not available on Blackwell GPUs.” Its methodology chapter shows one card hosting four DC-24Q vGPUs, and a mixed example with four DC-12Q and two DC-24Q. For the 2 GB office profile, the vPC sizing guide states: “Without MIG-backed Time-Sliced vGPU, the maximum is 32.” Release 19.0 added B-series types with 3 GB of frame buffer on Ada and Blackwell GPUs.

MIG-backed vGPU profiles on the RTX PRO 6000

A MIG-backed vGPU sits on a GPU instance with its own compute blocks and memory. On this card, NVIDIA’s MIG-backed vGPU page says that “up to four MIG slices can be created on a single GPU”. The smaller types are time-sliced inside one instance, and the same page states that “Within each MIG slice, one to three time-sliced vGPUs for Compute with 8 GB frame buffer each can be created”, for up to 12 VMs per card.

NVIDIA AI Enterprise 8.2, Blackwell vGPU types reference (column “Maximum vGPUs per GPU”); per-card count: our reading, the listed maximum times the instances of that size, which matches NVIDIA’s 12 VMs for DC-1-8C.

NVIDIA’s table heads the count column “Maximum vGPUs per GPU” and names instance profiles such as MIG 1g.8gb and MIG 4g.48gb, which do not appear in NVIDIA’s MIG user guide. The MIG guide, updated 11 September 2026, lists 1g.24gb, 2g.48gb and 4g.96gb for this card, each with variants such as +gfx. We read the other labels as time-sliced vGPUs inside those three instance sizes, which fits the frame buffer sums: three DC-1-8C fill one 24 GB slice. Check the counts on your host with nvidia-smi before you plan a layout. Creating the instances on the card is covered in our MIG setup guide for RTX PRO Blackwell.

The graphics types follow the same scheme. NVIDIA’s vWS sizing methodology builds its example on four MIG 1g.24gb+gfx slices: one slice with four DC-1-6Q, one with two DC-1-12Q and two mixed slices, one DC-1-12Q with one DC-1-8Q, and one DC-1-12Q with three DC-1-4Q. With a 2 GB profile, the vPC sizing guide states that “48 users per board is achieved with MIG-backed Time-Sliced vGPU.”

vGPU-48GB and MIG 2g.48gb: the 48 GB profiles on the RTX PRO 6000

NVIDIA’s tables contain no type called vGPU-48GB. A search for it usually means a VM with half the card’s memory. For compute, NVIDIA lists three 48 GB types: DC-48C, two per card, time-sliced over the whole GPU; DC-2-48C, one per 2-slice MIG 2g.48gb instance and two per card, each on its own half of the hardware; and DC-4-48C, two inside the full-card instance. For graphics, the test setups in NVIDIA’s vWS sizing guide use DC-48Q, DC-2-48Q and DC-4-48Q on this card.

The three types differ in isolation. DC-2-48C gives each VM its own compute blocks and memory bandwidth, so one VM cannot slow the other. DC-48C and DC-4-48C share the compute in turns. NVIDIA’s vWS sizing guide reports that “at the 48 GB profile size, MIG-backed vGPU delivered 20% higher performance than time-sliced vGPU” and does not name the types it compared; its graphics benchmark setup lists DC-48Q and DC-2-48Q. On vSphere, DC-4-48C shares an instance and needs VCF 9.1.

vGPU releases and hypervisors for the RTX PRO 6000

NVIDIA’s release notes for vGPU 19.0 list the card as newly supported “on the Red Hat Enterprise Linux with KVM, Ubuntu, and VMware vSphere hypervisors”. The same release added MIG-backed vGPUs for graphics and time-sliced vGPUs inside a GPU instance. Release 19.1 added the card on XenServer and two further hypervisors, and 19.2 added ESXi 9.0.1.0. The liquid-cooled Server Edition followed in vGPU 20.0.

On vSphere the MIG-backed features arrived later. Release 19.4 lists both MIG-backed graphics vGPUs and time-sliced vGPUs inside a GPU instance as new “on VCF 9.1”, and release 20.0 lists the same pair. Releases 19.6 and 20.2 add several MIG-backed vGPUs in one VM on VCF 9.1 or later. NVIDIA’s vSphere support matrix, updated 29 September 2026, lists the card for VCF 9.1, for VCF 9.0.1 or later and for ESXi 8.0 Update 3g or later updates of 8.0. The release notes list graphics inside slices and shared slices on vSphere only from VCF 9.1. On ESXi 8.0 and VCF 9.0, plan graphics users on time-sliced profiles over the whole card and compute VMs on types that fill a whole GPU instance: DC-1-24C, DC-2-48C and DC-4-96C.

Which profile for office VDI, CAD and AI compute VMs

Office VDI uses B-series profiles under a vPC licence. The vPC guide’s 48 per card assumes 2 GB per user and “reflects maximum supported density rather than a recommended deployment point”. The same guide says that “larger profile sizes (2B or 3B) are often needed” for current operating systems. According to NVIDIA’s MIG-backed vGPU page, MIG-backed vGPU supports the common desktop guest operating system on this card and the RTX PRO 4500 Server Edition, but not on the A30, A100, A800, H100, H200, B200 or B300. At 3 GB, dividing the 96 GB frame buffer gives 32 vGPUs per card with or without MIG, by our arithmetic; check that maximum in the B-series table of NVIDIA’s vGPU user guide.

CAD and 3D users take Q-series profiles under vWS, with the frame buffer chosen from the model sizes they open. By frame buffer alone, 12 GB fits eight per card and 24 GB four, but NVIDIA’s vWS sizing guide puts 2 to 4 heavy users on one card with DC-12Q to DC-24Q, and 3 to 6 medium users with DC-4Q or DC-8Q, because they share its compute. AI compute VMs take C-series profiles under NVIDIA AI Enterprise. Use DC-1-24C for four isolated 24 GB inference VMs per card, DC-2-48C for two 48 GB VMs, or DC-96C for a whole card in one VM. For development VMs that share a slice, DC-1-8C gives twelve per card, on vSphere from VCF 9.1. Estate sizing for hundreds of desktops is in our guide to GPU servers for VDI.

We supply the RTX PRO 6000 Server Edition with vWS, vPC or NVIDIA AI Enterprise licences on the same quote. Tell us how many office, CAD and compute users you plan and which hypervisor version you run.

Worked example: two hosts with four RTX PRO 6000 each

The example is our estimate from the profile limits above, not an NVIDIA reference design. Two hosts each carry four Server Edition cards, eight in total, on VMware Cloud Foundation 9.1 so that the MIG-backed graphics types are available. Each card serves one type of user, which keeps one licence type per card.

Our estimate: 3 CAD users per card within the 2 to 4 heavy users of NVIDIA’s vWS sizing guide, 48 desktops per card from its vPC sizing guide, DC-1-24C from the AI Enterprise 8.2 vGPU reference; licence counting as on our licence page.

With one host down, half of each pool of users loses its vGPU. A VM restarts on the other host only where an unused vGPU of the same type is left, and NVIDIA’s MIG-backed vGPU page notes that migration between different MIG profiles is not supported. Keep spare cards or plan an extra host.

For larger estates the same per-card figures give these counts. 100 heavy CAD users at three per card need 34 cards, nine four-card hosts and a tenth as spare; at four per card, NVIDIA’s upper figure, they need 25 cards and eight hosts with a spare. 1,000 office desktops at 3 GB need 32 cards by the frame buffer arithmetic, nine four-card hosts with one spare, and at 2 GB inside MIG slices 21 cards and seven hosts with a spare. The L40S vGPU profiles give a time-sliced alternative for the office part.

We build GPU servers with four RTX PRO 6000 Server Edition cards to order and check the rack, power and airflow before we quote. Send us your host count, hypervisor and user mix through the form below.

Scheduling, and which RTX PRO 6000 editions support vGPU

Time-sliced vGPUs follow one of three schedulers. The vGPU user guide names best effort, equal share and fixed share, and states that “The best effort scheduler is the default scheduler for all supported GPU architectures.” On a MIG-enabled card the setting is made per GPU instance, except on ESXi, where the RmPVMRL registry key sets it for all instances on the GPU.

Of the three RTX PRO 6000 editions, NVIDIA lists only the Server Edition for vGPU. NVIDIA’s support matrices for vSphere and for Red Hat Enterprise Linux with KVM list the RTX PRO 6000 Blackwell Server Edition and its liquid-cooled version, but no Workstation or Max-Q edition. Both of those cards do MIG on bare metal, without vGPU, as our comparison of the three RTX PRO 6000 editions explains.

What we supply

We supply the RTX PRO 6000 Blackwell Server Edition, the edition NVIDIA lists for vGPU, as cards or in AI servers built to order, with manufacturer warranty and on one EU contract and invoice. The vWS, vPC and NVIDIA AI Enterprise licences for the profiles in this guide come on the same quote, sized with the hardware, as described on our NVIDIA AI Enterprise and vGPU licences page. For the office part of a mixed estate we also supply the L40S. Engineering, if you want it, is by our partner Vixen.UNO.

FAQ