Four RTX PRO 6000 Max-Q in one workstation: 384 GB under a desk, and what it takes

Why the Max-Q exists

Four RTX PRO 6000 Max-Q cards fit in one workstation, and makers list that configuration in their specifications, for example HP for the Z8 Fury G6i and Lenovo for the ThinkStation PX. Such a build gives 384 GB of GPU memory and needs 64 PCIe lanes for the cards and, by our estimate, up to about 2 kW at the wall. The RTX PRO 6000 Blackwell comes in three editions with the same 24,064 CUDA cores and 96 GB of GDDR7, as our edition comparison explains. The Workstation Edition runs at 600 W with a double flow-through cooler, the Server Edition is passive for rack chassis, and the Max-Q keeps the full memory and the full 1,792 GB/s at 300 W. NVIDIA’s product page says that “it enables up to four GPUs in a single system”, with “up to 384 GB of combined memory across a four-GPU configuration”.

The lower power limit reduces the clock speed and with it the peak rates: 2,280 MHz boost instead of 2,617, 110 TFLOPS of FP32 instead of 125 and 3,511 AI TOPS instead of 4,000. In a review of July 2025, Puget Systems measured the Max-Q 5 to 13 per cent behind the Workstation Edition in Blender, Octane, Redshift and V-Ray and about 14 per cent behind in DaVinci Resolve’s GPU effects and Unreal Engine. NVIDIA describes the cooler only as “active”; Puget describes it as a blower that takes air from inside the case and pushes it out through the back, which is what lets four cards sit side by side. A double flow-through card releases its heat into the case instead, which, with its 600 W draw, is the likeliest reason the makers in the table below list one or two per machine, or none.

Which towers take four Max-Q cards

Makers’ product pages, configurators and specification documents, as of October 2026; the Pro Precision 9 T6 input-voltage figure from StorageReview’s launch report. HP’s specifications page for the Z8 Fury G6i, read on 24 October 2026, lists up to four Max-Q and names no Workstation Edition. HP’s previous Z8 Fury G5 also took four Max-Q cards or one Workstation Edition.

The towers that take four Max-Q cards do not list four 600 W cards; the most any of these makers lists is two, in Dell’s Pro Precision 9 T6. Every one of these power supplies delivers its full rating on a 230 V European socket; the lower figures in the makers’ tables apply to 100 to 127 V networks.

By the power budget below, the 1,400 W Dell unit cannot carry four cards at full load and a single HP supply would run at its limit, so a four-card order should name the larger or the dual supply.

We build four-card Max-Q workstations to order on Threadripper PRO or Xeon W platforms, with the processor, power supply and cooling sized for this load. Tell us which tower you already run, and we say in advance if four cards will not work in it.

PCIe lanes and processor choice

Four cards at full width need 64 PCIe lanes before the NVMe drives and the network card take theirs. AMD’s Threadripper PRO 9000 WX processors provide 128 PCIe 5.0 lanes, Intel’s Xeon 600 for workstations also 128 on its larger models such as the 658X and 698X, and the Xeon W-3500 series 112. All of them carry four cards at x16.

The slot wiring matters as much as the processor. In a review of 23 June 2026, StorageReview tested two Max-Q cards in HP’s Z8 Fury G6i, which runs every GPU slot at PCIe 5.0 x16, and in a Threadripper-based Dell Precision 7875 where the second card ran at PCIe 4.0 x16. Serving Llama 3.1 8B in FP8 with vLLM at a batch of 256, the HP reached 23,004 tokens per second and the Dell 16,833; the review credits both the slots and the processor’s handling of vLLM’s scheduling. For four cards, the slot table of a tower says more than its processor list.

Power: about 2 kW at the wall

Four cards at 300 W make 1,200 W. A 350 W processor, the rated power of both a Threadripper PRO 9995WX and a Xeon 698X, whose turbo limit is 420 W, brings the total to 1,550 W before memory, drives, fans and the losses in the power supply. At the wall that is roughly 1.8 to 2 kW under full load. This is our estimate from the rated figures. Plan the circuit on it and confirm it by a measurement on the delivered system.

On a 230 V circuit, 2 kW is about 9 A. A 16 A circuit, 3,680 W, carries the machine with room to spare, but not two of them together with the rest of an office. Size an uninterruptible power supply with headroom above 2 kW, and check that the building’s circuit is not already shared with a kitchen or a printer room.

Heat and noise

Practically all of that power ends up as heat in the room: 2 kW is about 6,800 BTU per hour, as much as a 2 kW fan heater running all day. A small office without air conditioning will warm up noticeably, and the machine’s own inlet temperature rises with it. Plan the room’s cooling together with the workstation configuration.

Published tests show where four in a row reach their limit. Puget Systems found in July 2025 that directly adjacent Max-Q cards have little room to draw air; it limits its own desktops to three Max-Q with a slot gap and writes that four 300 W cards are hard to cool. Exxact reports, in an article updated in March 2026, that four Max-Q throttled in a stock configuration and held the full 300 W with its custom cooling, at up to 85 °C. With four cards, choose a tower whose maker lists them, keep the intakes clear and read the inner cards’ temperatures separately during the burn-in.

We have found no independent noise measurement of a four-card Max-Q tower. Ask for the configured system’s acoustic data before it goes next to someone’s desk, and put the rear of the machine, where the blowers exhaust, towards open space.

What 384 GB runs

Weights from the model cards and published checkpoints on Hugging Face, in GB or GiB as published; each card holds 95.6 GiB (about 102.6 GB) as the driver reports it, and conversation counts follow our users-per-card rule (0.9 × 95.6 GiB, less 3 GiB of overhead, less the weights). The Llama 3.1 405B FP8 figure is our sum from the checkpoint’s parameter types. The model card of Qwen3-235B-A22B-Instruct-2507-FP8 serves it on four GPUs; we found no published benchmark of it on four RTX PRO 6000 cards.

The useful configurations are the first five rows: one large mixture-of-experts model across all four cards, two or four smaller models side by side, or a 70B model with room for a team. For published multi-card numbers, in a review of 6 November 2025, StorageReview ran four RTX PRO 6000 Server Edition cards, which have about 11 per cent less bandwidth than the Max-Q, with vLLM splitting the model across all four: gpt-oss-120b in NVFP4 reached 105.8 tokens per second per user with four users, and Llama 2 70B reached 32.9 tokens per second for a single user.

How many people that serves follows from our users-per-card rule: 0.9 × 95.6 GiB, less 3 GiB per card, leaves 332.2 GiB on four cards for weights and KV cache, and the layout decides how much of it is cache. Every copy stores its own weights, so fewer copies leave room for more conversations.

Our estimates by the users-per-card rule, concurrent conversations at full context length, FP8 KV cache for Llama 3.3 70B and 16-bit for gpt-oss-120b: weights 67.7 GiB for Llama 3.3 70B in FP8, 60.8 GiB for gpt-oss-120b and 219.8 GiB (236 GB) for Qwen3-235B-A22B-2507-FP8; cache 1.25 GiB per 8K conversation with an FP8 cache and 1.125 GiB per 32K conversation for gpt-oss-120b.

The full tables for four 96 GB cards are in our guide to what four RTX PRO 6000 Server Edition cards run, and the tower configuration on our page to buy a 4× RTX PRO 6000 Max-Q workstation.

Write to us with the models you plan to keep resident and the number of users, and we check what fits in 384 GB and quote the workstation around it.

Splitting one model over PCIe

The RTX PRO 6000 has no NVLink, so a model split across cards exchanges data over PCIe. vLLM’s documentation recommends tensor parallelism when a model is too large for one GPU but fits one node, and in a note on uneven GPU splits adds: “if the GPUs on the node do not have NVLINK interconnect (e.g. L40S), leverage pipeline parallelism instead of tensor parallelism for higher throughput and lower communication overhead”. Test both on your own model.

Direct transfers between the cards help tensor-parallel serving, in which the cards exchange data at every layer. NVIDIA’s NCCL documentation explains one reason they fail: on bare-metal Linux, “CUDA and the NVIDIA driver stack do not support IOMMU-enabled PCIe peer-to-peer memory transfer”. Check the IOMMU and ACS settings in the BIOS before blaming the cards.

MIG: up to sixteen GPU instances for a team

Each Max-Q splits into up to four isolated 24 GB instances, so the tower can offer sixteen independent GPUs to a team of developers. MIG runs on Linux with driver 575.51.03 or later and a Max-Q vBIOS of 98.02.6A.00.00 or later, and each card has to be switched from graphics to compute mode with NVIDIA’s DisplayModeSelector first. On a card that drives the monitor, that switch turns the display outputs off, so a MIG workstation either runs headless or keeps a small separate card for the screen. Profiles with the “+gfx” suffix keep graphics support inside an instance. vGPU for virtual machines is not available on the Max-Q; it is a Server Edition feature, as our MIG and vGPU guide explains.

Tower or rack server: which to choose

Four Server Edition cards in a rack server offer the same 384 GB with passive cooling, at 1,597 GB/s instead of 1,792 GB/s, a configurable power limit of up to 600 W and vGPU support, and they keep the power, heat and noise out of the office. They need a server room instead, with the airflow that passive cards depend on. Some 2U servers take four of them only at a 450 W limit, as our compatible servers guide shows; our inference node is built with four. Our article on how many GPUs fit in one server covers that side. A tower suits one team without a server room. A rack server is the better choice when the machine is shared across departments or runs continuously for customers.

What we supply

We supply the RTX PRO 6000 Blackwell Max-Q with manufacturer warranty on one EU contract and invoice, as single cards or in a four-card workstation built to order. For a rack, we build the same 384 GB with four Server Edition cards.

FAQ