Buy a workstation with 4× NVIDIA RTX PRO 6000 Blackwell Max-Q in Europe

Who the four-card Max-Q workstation is for

You can buy a workstation with four NVIDIA RTX PRO 6000 Blackwell Max-Q cards from us, built to order and delivered anywhere in the EU. It puts 384 GB of GPU memory into one tower, enough to run a model such as Qwen3-235B-A22B-2507 in FP8 split over all four cards, Llama 3.3 70B for over a hundred conversations at once, or four independent models side by side. Our team sizes the configuration to your workload and budget and replies within one business day.

Companies buy it when one team needs more memory than two cards give and a server room is not available, or not wanted, for the job. Typical buyers are ML teams that evaluate and fine-tune large open models, research labs that keep their data on site, and engineering departments that run a large assistant next to simulation and rendering work. With MIG the four cards also split into up to 16 isolated 24 GB GPUs for a development team.

NVIDIA built the Max-Q for this: the 300 W edition of the RTX PRO 6000 Blackwell, with the full 96 GB and 1,792 GB/s, whose product page says it "enables up to four GPUs in a single system" with "up to 384 GB of combined memory across a four-GPU configuration". The 600 W Workstation Edition does not reach four in the makers' tower specifications, as our article on four RTX PRO 6000 Max-Q in one workstation shows.

Towers whose makers list four Max-Q

Three makers list four Max-Q in their own documents for these towers. We supply the cards for each of them and check power supply, cables and slots before we quote.

HP Z8 Fury G6i specifications page; Lenovo PSREF ThinkStation PX specification (9 October 2026) and ThinkStation PX specifications (version 10.0, 16 March 2026) for the slots; Dell Precision 7960 configurator; all read on 24 October 2026. HP slot count from StorageReview's Z8 Fury G6i review of 23 June 2026.

The towers differ in ways that matter for four cards. HP's Z8 Fury G6i has one processor and wires every GPU slot at PCIe 5.0 x16. Lenovo's ThinkStation PX has two processors, with two of the four PCIe 5.0 x16 slots on each, so a model split over all four cards crosses the link between the sockets, while two pairs of cards each stay on one processor. Dell's Precision 7960 runs two of the four cards at PCIe 4.0 x16, which halves their link bandwidth to the processor. HP and Lenovo also offer their towers for a 5U rack position.

We also build four-card workstations to order on Threadripper PRO or Xeon W platforms, as our AI servers page describes. Either way, send us the tower your company runs or the platform you prefer, and our team replies with a configuration and a written quote.

What the configuration contains

The table shows how we configure a four-card Max-Q workstation and what we change per project.

Our AI servers page (platforms, components, installation); memory and cores by NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026), written for servers and applied here by the same arithmetic.

NVIDIA's configuration guide covers servers only, with at least six physical cores per GPU and system memory of at least twice the total GPU memory. We apply the same rule to workstations: 24 cores and 768 GB for four cards. On an eight-channel platform such as Threadripper PRO, 768 GB fills every channel with one module each.

Processor and PCIe lanes

Four cards at x16 need 64 PCIe lanes before the NVMe drives and the network card take theirs. AMD lists the Threadripper PRO 9975WX, 32 cores at a default TDP of 350 W, with 128 PCIe 5.0 lanes and eight memory channels. Intel lists the Xeon 698X workstation processor, 86 cores at a base power of 350 W and 420 W turbo, with 128 PCIe 5.0 lanes and eight channels. Both carry four cards at full width with lanes to spare.

The slot wiring matters as much as the lane count. In StorageReview's review of 23 June 2026, two Max-Q cards served Llama 3.1 8B in FP8 with vLLM at 23,004 tokens per second at batch 256 in the HP Z8 Fury G6i, with every slot at PCIe 5.0 x16, and at 16,833 in a Dell Precision 7875 whose second card ran at PCIe 4.0 x16. With four cards a tower has more lanes to share, so read its slot table before its processor list.

vLLM handles four cards differently from two. By default its source disables vLLM's custom all-reduce "on more than two PCIe-only GPUs", so a model split over all four cards uses NVIDIA's NCCL, while two TP=2 pairs keep the custom kernel. vLLM 0.30.0 (release notes of 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576), which is switched on only when configured. Our comparison of TP=2, TP=4 and replicas without NVLink explains the layouts, and our guide to RTX PRO 6000 P2P on Linux covers the IOMMU and ACS settings that decide whether peer-to-peer transfers work.

What it runs

By our users-per-card rule, 0.9 × 95.6 GiB − 3 GiB per card, four cards hold 332.2 GiB of weights and KV cache. The rule and the memory per card are the same as for four Server Edition cards, so the full tables in our guide to what four RTX PRO 6000 Server Edition cards run apply here too.

Our estimates by the users-per-card rule: concurrent conversations at full context length, FP8 KV cache for Llama 3.3 70B, 16-bit for gpt-oss-120b; Qwen3-235B-A22B-2507-FP8 checkpoint 236 GB (219.8 GiB) as in our four-card Max-Q article.

A forum build log on Level1Techs, updated in September 2026, gives published figures for four Max-Q at 300 W with SGLang and TP=4: DeepSeek-V4-Flash at 697 output tokens per second and MiniMax M3 at 1,076, both at 80 concurrent requests with 2K-token prompts. These come from one system. Models above about 330 GiB of weights, such as GLM-4.5 in FP8 at 361 GB, need eight cards.

Power on 230 V

Four cards at 300 W make 1,200 W. A 350 W processor brings the total to 1,550 W. With 15 per cent for memory, drives, fans and power-supply losses, the workstation draws about 1.8 kW at the wall under full load by our estimate, and up to about 2 kW with more memory and drives. The Level1Techs build log above gives a reading for one such tower: about 1,650 to 1,760 W on its UPS under full GPU load, with the processor capped at roughly 200 W.

Our estimate from NVIDIA's 300 W board power and the 350 W rating of the Threadripper PRO 9975WX and Xeon 698X, plus 15 per cent; 1 W = 3.412 BTU/h.

• Circuit. One 230 V 16 A circuit carries one four-card tower, planned at 80 per cent, but not two of them, nor one together with a printer room or a kitchen. Lenovo's power configurator for the ThinkStation PX recommends a dedicated circuit for each power supply.
• Power supplies. These supplies reach their full rating only at 200 V or more: HP's 1,700 W unit gives 1,350 W at 115 V, and StorageReview notes that HP's dual 2,700 W option covers a 350 W processor and four Max-Q. Lenovo's configurator gives 1,850 W at 12 A on 200 to 240 V for one PX supply and 2,350 W at 15 A in team mode. On a European socket each of them runs at full rating.
• Cords. A C13 appliance cord is rated at 10 A, which is 2.3 kW at 230 V and close to this load. StorageReview's review of the Dell Precision 7960 lists a C19 cord for its 2,200 W unit, rated at 181 to 240 V. Check the cord and the outlet with the configuration.

Our article on Max-Q power on 230 V works through circuits, cords and UPS sizing for two, four and eight cards.

Cooling, noise and the room

Four blower cards side by side are what the Max-Q was designed for, and the three towers above list them. Published measurements show the limits. A Level1Techs build log of four Max-Q measured 91 °C on the two inner cards at 300 W, with throttling between 88 and 92 °C. Puget Systems limits its own desktops to three Max-Q with a free slot between them and states that "four 300W cards are hard to cool". The makers list four cards for their own chassis and fan design, and a tower with four cards needs free space at the back, where the blowers exhaust, and a room that removes about 2 kW of heat.

No maker of these towers publishes noise figures for a four-Max-Q configuration under GPU load, and HP's page for the Z8 Fury G6i speaks of smart fan control without decibel values. A four-card tower under continuous load belongs in a technical room or an office with air conditioning; our guide to a GPU server in an office covers noise and heat in shared rooms.

Software delivered with the workstation

On request the workstation arrives with the operating system, NVIDIA driver, CUDA and a container runtime installed, with MIG instances set up if you plan them.

Our AI servers page; MIG prerequisites from NVIDIA's MIG user guide as cited in our MIG profiles article.

MIG on the Max-Q also needs a current vBIOS and compute display mode, set with NVIDIA's DisplayModeSelector, which turns the card's display outputs off; our RTX PRO 6000 MIG profiles list the details. As of October 2026 the R580 branch has long-term support to June 2028, with 580.178.04 as the first R580 driver fixed by NVIDIA's security bulletin of 30 September 2026. The Max-Q is not on NVIDIA's vGPU list; virtual machines with vGPU need the Server Edition.

When a rack server with Server Edition cards is the better buy

Four RTX PRO 6000 Server Edition cards in a rack server give the same 384 GB and the same model tables, with passive cards at up to 600 W, redundant power supplies, out-of-band management and vGPU from release 19.0. Their 1,597 GB/s are about 11 per cent below the Max-Q's 1,792 GB/s for decoding. Our inference node has four Server Edition cards, two AMD EPYC processors, 768 GB of DDR5 ECC and four NVMe drives. By our estimate it draws about 3.4 kW at 600 W per card and 2.7 kW at the 450 W cap, which needs a server room and, at 600 W, a 32 A or three-phase feed.

Choose the rack server when the service runs continuously for several departments or for customers, when the machine should not stand near people, or when virtual machines need vGPU. Choose the tower for one team without a server room. The configuration is on our page to buy a 4× RTX PRO 6000 Server Edition server. Putting Max-Q cards into a rack server works only in a chassis whose maker lists them; for continuous operation in a rack we recommend the Server Edition, as our page on the Max-Q in a rack server explains.

How ordering works with us

1. You describe the workload: models, users, data, the room and the timeline. Sizing is free of charge.
2. We propose the configuration and check that cards, tower, power supply and cooling fit the room and the circuit.
3. You receive a written quote within one business day, with price, lead time and warranty terms per component. Quantities and lead time are confirmed in writing with the quote.
4. You order on one EU contract with an EU invoice from Eurokommerz in Vienna.
5. We assemble the workstation and test it before delivery (power-on and basic load test).
6. We deliver anywhere in the EU, with commissioning on site on request; other destinations on request through our logistics partner.
7. Every component carries the manufacturer warranty, handled through us; DOA units are replaced, and the assembly is covered by us.

For your own acceptance test after delivery, use the checks in our GPU server acceptance testing guide, which apply to a four-card tower as well.

What to send us for a quote

• Models you plan to run, with quantisation and context length
• How many people will use the machine, and whether through one service or MIG instances
• Other work on the same machine: fine-tuning, embedding, data preparation, rendering
• The tower model your company standardises on, if any, and the operating system
• The room: circuit rating, cooling and where the tower will stand
• Delivery site, invoicing entity and timeline

Describe these points in the form below, and our team replies with a configuration and a written quote within one business day.

What we supply

We supply the RTX PRO 6000 Blackwell Max-Q as cards for towers whose makers list four, and build four-card Max-Q workstations to order, together with two-card workstations, described on our page to buy a 2× RTX PRO 6000 Max-Q workstation, and Server Edition rack servers for continuous operation. Everything comes on one EU contract and invoice with manufacturer warranty, handled through us.

FAQ