RTX PRO 6000 Server Edition vs Workstation vs Max-Q: specs, power, MIG and vGPU compared

What NVIDIA’s family table leaves out

The RTX PRO 6000 Server Edition, Workstation Edition and Max-Q have the same GPU and 96 GB of GDDR7. Against the Workstation Edition, the Server Edition for rack servers is passively cooled, has 1,597 instead of 1,792 GB/s and is the only one of the three with vGPU. The Max-Q differs from the Workstation Edition in power and cooling: it runs at 300 W instead of 600 W, with a blower and about 12% less FP32. NVIDIA’s official family page compares the three versions in six rows: memory, DisplayPort count, power, bus, form factor and cooling. It does not list memory bandwidth, CUDA core counts, MIG or vGPU, although these are what separate the versions in production. The figures are spread across three product pages, three PDFs and a whitepaper that does not mention the Server Edition, so the table below puts them side by side; the full data sheet of all three editions is in our article on RTX PRO 6000 Blackwell specs. We supply all three editions. Describe your task and our team will help you choose between them.

Workstation, Max-Q and Server Edition specifications side by side

Memory bandwidth, 1,792 vs 1,597 GB/s: effect on LLM speed

All three use a 512-bit bus, so the difference comes from the memory clock. The desktop versions run GDDR7 at 28 Gbit/s; for the server card, the rated 1,597 GB/s on a 512-bit bus works out at roughly 25 Gbit/s. Token generation reads every weight for each token and scales almost linearly with bandwidth, so expect about 11% fewer tokens per second on the Server Edition. Prompt processing is Tensor-Core-bound; NVIDIA lists both cards at about 4 PFLOPS of FP4, the Workstation Edition with sparsity and the Server Edition without a sparsity note, but the Server Edition’s lower boost clock (2,430 MHz in its product brief against 2,617 MHz; NVIDIA’s 120 TFLOPS FP32 rating implies about 2,490 MHz) puts it a few per cent behind.

The gap does not change whether a model fits. All three carry 96 GB, and the largest performance differences between cards appear where a model fits on one card but not on the other. GamersNexus logged a 928% gain over an RTX 5090 on Llama 3.3 70B because the model fits in 96 GB but not in the 32 GB of the RTX 5090.

Cooling and how many cards fit in one system

The Workstation Edition has flow-through cooling with two axial fans and depends on the chassis airflow. This works well for one card: GamersNexus measured 82 °C on the core, 88 °C on the memory and about 32.5 dBA (September 2025). With two cards, the first card’s exhaust air feeds the second, and the airflow arithmetic sets the limit, so two cards is the realistic maximum. The Max-Q has the same GPU and memory on a smaller board, with a blower and half the power limit. NVIDIA states up to four per workstation, and Dell rates its Pro Precision 9 T6 for five 300 W RTX PRO cards; the cost is about 12% less FP32. The Server Edition has no fan and relies on the airflow of a server qualified for it; without that airflow it throttles. NVIDIA also lists a liquid-cooled, single-slot Server Edition for its 6U MGX design, supported from vGPU 20.0; the figures in this article are for the air-cooled card. Which rack servers list the Server Edition, with their cable, power and fan rules, is in our list of compatible servers; HPE’s ten-card benchmark system is described in our article on RTX PRO 6000 benchmarks, while HPE’s QuickSpecs give eight as the orderable maximum. Some makers also list the Max-Q in rack servers, as our Max-Q rack server guide shows.

The 450 W limit set by the 12V-2×6 cable

A card held at 450 W can look like a defective one. All three versions use the 16-pin 12V-2×6 connector, whose SENSE pins tell the card how much power the supply can deliver, and the card lowers its own power limit to match. It then reports Max Power Limit: 450.00 W, and nvidia-smi -pl 600 changes nothing, because the cable’s sense pins declare 450 W. The 600 W cable is a separate part that the server vendor has to list for that chassis and slot, so put it in the BOM. Sometimes 450 W is the vendor’s deliberate choice to double card density. At 450 W, generation, which is bound by memory bandwidth, loses little; prompt processing, which is bound by the Tensor Cores, loses more.

If the cards are going into a server you already run, we check power, cooling and the PSU cable before the quote, so a 450 W limit does not come as a surprise after delivery. Tell us the server model and the slots.

MIG on all three editions, vGPU only on the Server Edition

MIG (up to four isolated instances) is listed for all three editions and runs on Linux only. On the desktop versions it is not enabled by default. It needs driver 575.51.03 or later, a current vBIOS (at least 98.02.55.00.00 on the Workstation card, 98.02.6A.00.00 on the Max-Q) and compute display mode, set with DisplayModeSelector 1.72 or later. On a card that drives the monitor, switching to compute mode turns the display outputs off. Users on NVIDIA’s developer forum reported launch units with an older vBIOS, so “Unable to enable MIG Mode: Not Supported” was a common error at first setup. vGPU is available on the Server Edition only: up to 48 VMs per card for virtual desktops (4 MIG instances × 12 time-sliced graphics vGPUs); for NVIDIA AI Enterprise compute vGPUs the limit is 3 per MIG slice, 12 per card. On VMware, NVIDIA’s vGPU 20 release notes list time-sliced, MIG-backed vGPUs for VMware Cloud Foundation 9.1. Lenovo’s product guide for the Max-Q states “vGPU software support: No support.”

Which RTX PRO 6000 edition for which use

We also build AI servers to order with two or four Server Edition cards. Send us the chassis and the number of cards in the form below, and we say which edition it takes.

Published benchmarks between editions, and NVLink

We found no published test of the Server Edition against a desktop edition on one bench. For the two desktop editions there is one. Puget Systems ran the Max-Q and the Workstation Edition on the same Threadripper PRO workstation in July 2025 and measured the Max-Q 5 to 14 per cent slower in content-creation applications. We found no mention of NVLink on NVIDIA’s product pages, but its product brief for the Server Edition lists NVLink as “Not supported”, and Lenovo’s product guide states “NVLink support: No”, so multi-card traffic is PCIe Gen5 on every edition.

With more than one card, the published figures come from different systems, models and software. They show what each edition does in the chassis built for it, not a ranking between the editions.

StorageReview, HP Z8 Fury G6i review (23 June 2026) and HPE ProLiant DL380a Gen12 review (6 November 2025, modified 14 October 2026); Level1Techs forum build log of a four-Max-Q workstation (July 2026, updated September 2026), UPS reading with the processor capped at about 200 W. MLPerf® Inference: Datacenter v6.1, closed division, Llama 2 70B (llama2-70b-99.9), offline scenario, entries 6.1-0050 (HPE ProLiant Compute DL380a Gen12), 6.1-0036 (Dell PowerEdge XE7740), 6.1-0013 (Cisco UCS C845A M8) and 6.1-0070 (MiTAC G4520G6), each with eight Server Edition cards, retrieved from mlcommons.org on 24 October 2026; results verified by MLCommons Association. Per card is our division of the system result by eight, not an MLPerf metric. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Read next to the specification table, these figures say three things. The Max-Q runs four to a workstation, but the inner cards sit close to the 88 to 92 °C at which that build’s cards throttled, so slot spacing and room cooling decide whether they hold their clocks; with the processor capped, the wall reading lies just below our estimate of 1.8 to 2 kW for such a tower. The Server Edition scales to eight cards per server with almost the same result across makers: the four MLPerf® results lie within 2.6 per cent of each other. And the slot wiring counts for every edition: the same two Max-Q delivered 27 per cent less in the tower whose second slot ran at PCIe 4.0, which StorageReview puts down to the slots and the processor together. Our comparison of the Max-Q and the Server Edition in a four-GPU server, the guide to what four Server Edition cards run and the one on eight Server Edition cards give the detail.

What we supply

We supply all three editions of the RTX PRO 6000 (Workstation Edition, Max-Q and Server Edition) across the EU, on one EU contract and invoice, with manufacturer warranty, as cards or in GPU servers built to order; the Workstation Edition and the Max-Q are also available as OEM versions. Where the Server Edition will run virtual machines, the vGPU licence goes in the same quote. Quantities and lead time are confirmed in writing with the quote.

FAQ