Eight NVIDIA RTX PRO 6000 Blackwell at 300 W each: Max-Q or Server Edition capped low

Max-Q or Server Edition at 300 W per card

For an eight-card server with a power budget of 300 W per GPU, choose the NVIDIA RTX PRO 6000 Blackwell Server Edition and cap it at 300 W, unless the chassis maker lists the RTX PRO 6000 Max-Q for that exact model and the work runs on bare-metal Linux. Both reach the same 2.4 kW for the cards and the same 768 GB of GPU memory. The Max-Q keeps more memory bandwidth at that power; the Server Edition keeps the server makers' support, vGPU, NVIDIA AI Enterprise and the option to run at 450 or 600 W later.

NVIDIA Max-Q datasheet 5349650 (June 2026); NVIDIA product brief SP-12355-001_v02 for the Server Edition's power modes and 300 W minimum; NVIDIA GPUs supported by vGPU (2 October 2026); NVIDIA AI Enterprise 8.2 support matrix (2 September 2026), as cited in our four-card comparison; chassis rows from the makers' pages read on 24 October 2026.

The power figure is why the question comes up. An eight-card server at 600 W per card needs about 6.3 kW at the wall by our estimate and a three-phase feed per side, while eight cards at 300 W fit a 32 A single-phase feed. Halving the card power is the way to get 768 GB of GPU memory into a rack position that cannot supply more. The four-card version of the same choice is in our comparison of the Max-Q and the Server Edition in a 4-GPU server.

Which eight-card chassis list the Max-Q

NVIDIA positions the Max-Q as a workstation card: its product page says it "enables up to four GPUs in a single system". Eight of them in one server is a configuration that the server maker has to qualify, and we found few that have. These are the eight-card systems we checked on 24 October 2026.

Supermicro SYS-422GA-NRT datasheet; 2CRSi Godì 1.8SP5-D product page; HPE ProLiant Compute DL380a Gen12 QuickSpecs; Lenovo Press LP2364 (8 May 2026) and LP2263 (28 July 2026); Dell PowerEdge XE7745 technical guide (Rev. A09, September 2026); all read on 24 October 2026.

Supermicro's SYS-422GA-NRT states "Support up to 8 double-width 600W PCIe GPU accelerator cards at 35ºC", connects them through dual-root PCIe 5.0 switches, and names both the Max-Q and the Server Edition in its GPU list, with four 3,200 W supplies in a 3+1 layout. The datasheet names eight cards for the Server Edition and gives no count for the Max-Q, so we have Supermicro confirm the number of Max-Q for the exact configuration before we quote. Supermicro's 3U SYS-322GA-NR also names the Max-Q among its GPUs, without a count for it, and Supermicro's RTX PRO systems page gives that chassis up to four RTX PRO 6000 Blackwell without naming an edition, so on current documents we do not offer it as an eight-card RTX PRO 6000 server.

2CRSi, a French maker, builds the Godì 1.8SP5-D as a 5U air-cooled server with two AMD EPYC 9005 processors and eight double-width slots. Its page names the RTX PRO 5000 Blackwell and the RTX PRO 6000 Blackwell Max-Q at 300 W, rates the slots for GPUs of up to 350 W and notes that its RTX PRO Blackwell Server Edition option excludes the RTX PRO 6000. The page gives no number of Max-Q cards. Power comes from 4+1 supplies of 2,000 W. The operating temperature is given as 10 to 25 °C, against 10 to 35 °C for the Supermicro systems above.

Lenovo's option guide for the Max-Q, LP2364, lists it for two 2U servers with three cards each and marks every other ThinkSystem model as not supported, the eight-card SR675 V3 included. HPE's DL380a Gen12 QuickSpecs list the Server Edition and no Max-Q, and Dell's XE7745 technical guide names eight RTX Pro 6000 at 600 W and does not mention the Max-Q. For Cisco we did not check and claim nothing either way. The rack systems that take the Max-Q in smaller counts are on our page about buying a Max-Q rack server.

Eight blowers in a server: airflow and inlet temperature

A GPU server moves air from front to back through passive heat sinks. NVIDIA's product brief for the Server Edition gives the minimum airflow through each card: 43 CFM at a 25 °C inlet, 51 CFM at 30 °C, 65 CFM at 35 °C and 125 CFM at 45 °C. For eight cards at 35 °C that is at least 520 CFM through the heat sinks, and the server's fans are sized for it.

The Max-Q brings its own blower, which draws air from inside the chassis and pushes it out through the bracket. Eight of them depend on the intake air that reaches each card from its neighbour's side. On the Level1Techs forum, a four-card Max-Q build log of July 2026 reports 91 °C on the two inner cards at 300 W, close to the 88 to 92 °C at which its cards throttled, and Puget Systems limits its desktops to three Max-Q with a free slot between cards. A server maker that lists eight Max-Q has worked out slots, ducts and fan curves for them. In a chassis that lists only passive cards, no one has.

Inlet temperature is where the two listings differ most. Supermicro rates its 4U chassis for eight 600 W cards at 35 °C. 2CRSi's Godì, whose GPU list centres on 300 W cards, is specified for 10 to 25 °C. Dell limits its XE7745 to 30 °C with RTX Pro 6000 cards, and Lenovo's thermal rules for the 2U SR650a V4 allow four RTX PRO 6000 Server Edition at the 450 W limit only up to 25 °C. A server room that holds 27 °C at the rack front, the top of ASHRAE's recommended range, suits the Supermicro chassis and the XE7745, but not the Godì. Measure the inlet at the rack position before choosing; our guide to GPU server inlet temperature explains where and how.

We build eight-card servers only on chassis whose maker lists the card for that model and slot. Tell us the rack position, its feed and the inlet temperature, and our team replies with the edition and chassis that fit, within one business day.

The Server Edition held at 300 W

NVIDIA's product brief SP-12355-001_v02 gives the Server Edition two power modes, set by the strapping of its 16-pin cable: 600 W maximum, the default, or 450 W maximum, each with a 300 W minimum. A limit between the minimum and the strapped maximum can be set with nvidia-smi, and the brief states that such a limit "must be reestablished after each new driver load", while a limit set through the board management interface (SMBPBI) persists across driver loads and reboots. For continuous operation the cap belongs in the server's management controller where the chassis supports it, or in a service that applies it at every boot. Dell's XE7745 guide, for example, states that the server "does not support user power cap", so the method depends on the chassis. The commands are in our guide to Server Edition power at 600, 450 and 300 W.

The cap costs compute. 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 in llama-bench at 906.64 tokens per second against 1,734.59 at 600 W, while token generation fell only from 20.54 to 19.29 tokens per second. Prompt processing at 300 W thus ran at about 52 per cent of the 600 W figure on that card.

The Max-Q is designed for 300 W. NVIDIA rates it at 110 TFLOPS of FP32 and 3,511 AI TOPS within that budget, and the Server Edition's 120 TFLOPS assume its full power. On paper the Max-Q has the edge at 300 W in bandwidth. NVIDIA publishes no compute figure for the Server Edition at 300 W, and we found no published test of a Max-Q against a Server Edition held at 300 W, so how the two compare in compute at that power is not known.

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

Token generation of a dense model reads every weight once per token, so its upper bound per conversation is memory bandwidth divided by the bytes read. The power limit barely changes bandwidth. For Llama 3.3 70B in FP8, about 70.6 GB per token, our arithmetic gives about 25 tokens per second on one Max-Q and about 23 on one Server Edition, or about 51 against 45 for a tensor-parallel pair, before any cost of the exchange over PCIe. Summed over eight cards, the Max-Q server has 14,336 GB/s and the Server Edition server 12,776 GB/s, 12 per cent apart.

For interactive chat and agents, where generation dominates, the Max-Q's bandwidth gives slightly faster replies per user at the same 300 W. For long documents, RAG with large prompts and batch jobs, prompt processing dominates, and there the power limit decides more than the edition. A server that might later be allowed 450 W per card gains most on that side, and only the Server Edition can take that step: its 450 W cap cost 2 to 5 per cent of LLM throughput at short context and 4 to 9 per cent at long context against 600 W in StorageReview's roundup of October 2026, measured on a Workstation Edition card.

vGPU, NVIDIA AI Enterprise and MIG on eight cards

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. Lenovo's guide for its Max-Q option states "vGPU software support: No support". On a Max-Q server, virtual machines get whole cards through PCIe passthrough, and 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. What is known about that is on our page about the Max-Q in virtual machines.

NVIDIA AI Enterprise follows the same line. Its 8.2 support matrix lists the Server Edition as supported with servers on the NVIDIA-Certified Systems list, and the Max-Q is not in it. Serving with open-source engines on bare metal needs no NVIDIA licence, so this matters for NIM, compute vGPU profiles and NVIDIA's enterprise support. Where it is used, it is licensed per GPU, so eight cards need eight licences on either edition. The profiles per card are in our guide to Server Edition vGPU profiles.

MIG works on both editions: each card splits into up to four 24 GB instances, so eight cards give up to 32 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 them off. Details are in our guide to RTX PRO 6000 MIG profiles.

Power, feeds and rack density at 2.4 kW of cards

The table uses the method of our other eight-card guides: cards plus two AMD EPYC 9555 processors at their default TDP of 360 W, plus 15 per cent for memory, drives, network cards, fans and conversion losses. A measurement on the delivered system replaces the 15 per cent.

Our estimate from NVIDIA's board power and AMD's EPYC 9555 page; feeds planned at 80 per cent: 2.94 kW at 230 V 16 A, 5.89 kW at 32 A, 8.87 kW at 400 V 16 A three-phase. Dell estimates about 8 kW at full load for its eight-card nodes with 600 W cards.

At 300 W per card the server is above the 2.94 kW we plan for a 230 V 16 A feed, so it needs a 32 A feed per side, and with A and B feeds each side carries the whole server. The same 32 A feed carries the Server Edition at 450 W, which is the step the Max-Q cannot make. All of that power becomes heat: about 3.6 kW is 12,300 BTU/h and needs roughly 580 CFM of air at an 11 °C rise, against about 1,000 CFM at 6.3 kW, by the 160 CFM per kW in our guide to GPU rack power and cooling.

For the rack, by our estimate a 10 kW air-cooled position takes two eight-card servers at 300 W, 7.2 kW, or one at 600 W. That doubles the GPU memory per rack at the cost of throughput per card, as our guide to the air-cooled AI server in a 10 kW rack shows for other mixes. Feed types, cords and A/B layouts are in our guide to GPU server power on 16 A, 32 A and three-phase, and the Max-Q card counts on 230 V in our page on Max-Q power on 230 V.

What 768 GB holds on either edition

Both editions carry 96 GB per card, of which the driver reports 95.6 GiB. By our users-per-card rule, 0.9 × 95.6 GiB less 3 GiB per card, eight cards hold 664.3 GiB of weights and KV cache on either edition. Llama 3.3 70B in FP8 at 8K holds 312 concurrent conversations as four tensor-parallel pairs with an FP8 KV cache, and gpt-oss-120b holds 152 conversations of 32K as eight copies. The power limit changes how fast those conversations run; the number that fits stays the same. The full model list and the layouts are in our guide to what an 8× RTX PRO 6000 Server Edition server runs, and every count there applies to eight Max-Q as well.

Which edition for which eight-card server

• A shared inference service in continuous operation: Server Edition, capped at 300 W while the feed is 32 A, raised to 450 W when the feed allows it.
• Virtual machines, vGPU desktops or NIM: Server Edition, on a server from the NVIDIA-Certified Systems list.
• A room that cannot hold 25 °C at the rack front: a chassis its maker rates for the inlet temperature you have, such as Supermicro's SYS-422GA-NRT at 35 °C.
• A development team on bare metal that already works on Max-Q workstations: Max-Q, in a chassis that lists it, such as Supermicro's SYS-422GA-NRT once Supermicro confirms the Max-Q count, so that containers and drivers move from the desk to the rack unchanged.
• More than 3.6 kW available at the rack position: Server Edition at 450 or 600 W, as described on our page to buy an 8× RTX PRO 6000 Server Edition server.

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply both editions and build eight-card servers to order, assembled and tested before delivery, with manufacturer warranty on every component on one EU contract and invoice. Send us your models, peak users and the power available per rack position, and we reply with the edition, the power cap and the chassis.

FAQ