NVIDIA RTX PRO 6000 Blackwell Max-Q or Server Edition in a 4-GPU server

Max-Q or Server Edition: the decision for four cards

For a four-GPU rack server, choose the NVIDIA RTX PRO 6000 Blackwell Server Edition unless the server maker lists the Max-Q for that exact model and the work runs on bare-metal Linux. Both editions carry 96 GB of GDDR7 with ECC, so four cards hold the same models and the same number of conversations; the differences are cooling, power range, memory bandwidth, virtualisation support and which servers take the card. The table sets them side by side.

NVIDIA Max-Q datasheet 5349650 (June 2026); Server Edition product brief SP-12355-001_v02; NVIDIA GPUs supported by vGPU (2 October 2026), read on 24 October 2026; NVIDIA AI Enterprise 8.2 support matrix (2 September 2026), as cited in our Max-Q rack article.

Four cards of either edition give 384 GB, of which our users-per-card rule counts 332.2 GiB as usable: 0.9 × the 95.6 GiB the driver reports, less 3 GiB per card. Llama 3.3 70B in FP8 with an FP8 KV cache holds 156 concurrent conversations of 8K as two tensor-parallel pairs, on either edition. The full model list is in our guide to what a 4× RTX PRO 6000 Server Edition server runs, and every figure there applies to four Max-Q as well.

Cooling and airflow with four cards side by side

A four-card server is built around one kind of airflow. Its fans push air from the front through the heat sinks and out at the back, and a passive card depends on that air. NVIDIA's product brief for the Server Edition gives the minimum airflow through each card's heat sink as 43 CFM at a 25 °C inlet, 65 CFM at 35 °C and 125 CFM at 45 °C, in either direction, so four cards at a 35 °C inlet need at least 260 CFM through their heat sinks alone. The server maker sizes fans and ducts for exactly that.

The Max-Q has its own blower, which NVIDIA's datasheet describes as "Active" cooling and Puget Systems as a fan that draws air from inside the case and pushes it out through the bracket. Four of them side by side compete for intake air. In a Level1Techs forum thread of May and June 2026, one participant reports that the two inner cards of four adjacent Max-Q throttled in his tests and that one free slot between cards avoided it, and another that four cards side by side overheated whatever he tried; a four-card build log of July 2026 on the same forum gives 91 °C for the two inner cards at 300 W. Puget limits its desktops to three Max-Q with a slot gap. A server maker that lists four Max-Q has worked out spacing and fan control for them; in a chassis that lists only the passive card, nobody has.

The bracket direction matters too. The blower exhausts through the card's bracket, so the bracket has to face the hot aisle, as it does in a rear riser of a front-to-back chassis. Our guide to the Max-Q in a rack server has the full checklist of slots, cables and BMC fan control.

What makers list for four-card servers

The lists below were read on 24 October 2026. They show the pattern a buyer meets: the common four-card servers take the Server Edition, and the systems that take the Max-Q are a separate, short list.

Lenovo Press LP2364 (8 May 2026) and LP2263 (28 July 2026); HPE ProLiant Compute DL380a Gen12 QuickSpecs; Supermicro SYS-422GA-NRT page; Lenovo PSREF ThinkStation PX (9 October 2026).

Lenovo's two guides are mirror images: the SR650a V4 takes the Server Edition, four of them only in the version "Slot Capped at 450W", and shows N for the Max-Q; the SR655 V3 and SR665 V3 show the value 3 for the Max-Q, without a legend in the guide, and N for the Server Edition; we confirm the Max-Q count in Lenovo's configuration before we quote. HPE's DL380a Gen12 QuickSpecs list the Server Edition, the RTX PRO 6000D, the H200 NVL, the L40S, the L20 and the L4, and no Max-Q. Supermicro's SYS-422GA-NRT lists both: its datasheet gives up to eight double-width 600 W cards and names eight Server Edition, but states no count for the Max-Q, which we have Supermicro confirm for the exact configuration. More Server Edition servers are in our list of Server Edition compatible servers, and the Max-Q systems are covered on our page about buying a Max-Q rack server.

A typical case: a server picked for the Max-Q that takes only the Server Edition

A company plans a four-card inference server for its own assistant and works out the configuration around four Max-Q: 384 GB, 1,200 W for the cards, a 2U chassis because the rack has little free height. The team picked the 2U server because its maker advertises four double-width GPUs. When the maker's GPU list is checked, that model takes four RTX PRO 6000 Server Edition cards in a 450 W capped version and marks the Max-Q as not supported, which is exactly the position of Lenovo's SR650a V4 in the two guides above.

There are three ways out, and the first is usually the right one.

1. Keep the server, change the edition. Four Server Edition cards at the 450 W cap hold the same models and conversations. The cards draw 1,800 W instead of 1,200 W, about 2.7 kW at the wall by our estimate, which still fits a 230 V 16 A feed per side, and vGPU becomes possible later.
2. Keep the Max-Q, change the chassis. In 2U that means three cards on Lenovo's SR655 V3 or SR665 V3, 288 GB instead of 384 GB; four cards mean a server that lists it, such as Supermicro's 4U SYS-422GA-NRT once Supermicro confirms the Max-Q count, or a rack-mounted workstation such as the Lenovo ThinkStation PX.
3. Run the Server Edition at a lower limit. Its product brief gives a 300 W minimum, and nvidia-smi can set a limit between that and the strapped maximum, as our guide to the RTX PRO 6000 power limit shows; the brief adds that a limit set this way has to be applied again after each driver load. We found no published measurement of the card's speed at 300 W.

We check the maker's list for the exact model before we quote, so a mismatch shows up in the quote, before anything is delivered. Tell us which four-card server you have chosen, and we confirm which edition it takes and at what power.

Bandwidth and speed: 1,792 GB/s against 1,597 GB/s

Token generation for a dense model reads every weight once per token, so its upper bound per conversation is memory bandwidth divided by the bytes read. For Llama 3.3 70B in FP8, about 70.6 GB per token, that gives by our arithmetic about 25 tokens per second on one Max-Q and about 23 on one Server Edition, and about 51 against 45 for a tensor-parallel pair. The Max-Q's bandwidth is 12 per cent higher, and it keeps that lead whatever the Server Edition's power limit.

Prompt processing is bound by the Tensor Cores instead, and there the power limit counts. In a test on the Level1Techs forum, a 600 W RTX PRO 6000 Workstation Edition limited to 300 W processed a 512-token Llama 3.3 70B prompt at 906.64 tokens per second against 1,734.59 at 600 W in llama-bench, while generation fell only from 20.54 to 19.29 tokens per second. StorageReview's roundup of October 2026, run on a Workstation Edition card, measured the 450 W cap at 2 to 5 per cent lower throughput at short context and 4 to 9 per cent at long context. NVIDIA gives the Server Edition 120 TFLOPS of FP32 against the Max-Q's 110.

For chat and agents with short prompts, the Max-Q's bandwidth gives slightly faster replies per user. For long documents, RAG with large contexts or batch jobs, a Server Edition at 450 or 600 W processes prompts faster. We found no published test of both editions in the same server, so these are directions from the specifications and from single-card tests.

vGPU, NVIDIA AI Enterprise and MIG

Virtualisation separates the two editions most clearly. NVIDIA's list of GPUs supported by vGPU, updated on 2 October 2026, names the RTX PRO 6000 Server Edition from release 19.0 and its liquid-cooled version from 20.0, and no Max-Q. On vSphere, the Server Edition needs ESXi 8.0 Update 3g or later for vGPU. Lenovo's guide for the Max-Q states "vGPU software support: No support". A Max-Q server can still give whole cards to virtual machines through PCIe passthrough, but a thread on the Proxmox forum with hosts of two to eight cards reports reset failures after passthrough, with the Max-Q among the cards most often affected; our page on the Max-Q in virtual machines collects what is known.

NVIDIA AI Enterprise follows the same line. Its 8.2 support matrix lists the RTX PRO 6000 Server Edition as supported with servers on the NVIDIA-Certified Systems list, and the Max-Q is not in it. Open-source serving engines on bare metal need no NVIDIA licence, so this matters for NIM, for compute vGPU profiles and for enterprise support from NVIDIA. Profiles per card are in our guide to Server Edition vGPU profiles.

MIG works on both. Each card splits into up to four 24 GB instances, so four cards give up to sixteen isolated GPUs for containers on one Linux host. The Max-Q first has to be switched to compute mode with NVIDIA's DisplayModeSelector, which turns its display outputs off; the Server Edition ships with the outputs off.

Power and feeds for four cards

The cards set most of the load. The table uses our method from the four-card guides: cards plus two 280 W server processors, plus about 15 per cent for memory, drives, fans and power-supply losses.

Our estimate from NVIDIA's board power and AMD's 280 W default TDP for the EPYC 9355; feeds planned at 80 per cent, 2.94 kW for 16 A and 5.89 kW for 32 A at 230 V.

All of it becomes heat, 3.412 BTU/h per watt: about 6,800 BTU/h for 2 kW and 11,600 for 3.4 kW. The Max-Q saves power against a Server Edition at 450 or 600 W, and nothing against a Server Edition held at 300 W. Rack servers with 2,400 or 3,200 W supplies, such as HPE's DL380a Gen12, take C19 cords only. Feed types, cords and UPS sizing are in our page on Max-Q power on 230 V and in our guide to GPU server power on 16 A, 32 A and three-phase.

Which edition for which four-card server

• Shared inference service in a server room or data centre: Server Edition, at the 450 W cap in 2U or at 600 W in 4U. This is our inference node: four Server Edition cards, two AMD EPYC processors, 768 GB of DDR5 ECC and four NVMe drives.
• Virtual machines, virtual desktops or NIM: Server Edition, on a server from the NVIDIA-Certified Systems list.
• A tight power budget per rack: Server Edition at a 300 or 450 W limit in a server that lists it, or the Max-Q where the chassis lists it.
• A development team on bare metal that standardises on the Max-Q: Max-Q, three cards in Lenovo's 2U pair, four in a rack-mounted workstation, or Supermicro's SYS-422GA-NRT with the Max-Q count confirmed by Supermicro.
• Four cards under a desk: Max-Q in a tower, as in our guide to buying a 4× Max-Q workstation.

We supply both editions and build four-card servers with them to order, with manufacturer warranty on one EU contract and invoice. Send us the models, the peak users and your rack position, and our team replies within one business day with the edition, chassis and power mode for it. The ready-made four-card configuration is described on our page about buying a 4× Server Edition server.

FAQ