2U or 4U for a four-GPU server: power caps, airflow, noise and room to grow

Four cards in 2U or in 4U: what changes

Four NVIDIA RTX PRO 6000 Blackwell Server Edition cards fit a 2U server or a 3U or 4U server, and the cards hold the same models either way. What changes is the power each card may draw, the inlet temperature the server accepts, the number of power supplies and cords, the noise, the rack space and whether the server can grow to eight cards. Choose 2U for a cool room, a 16 A feed and a fixed four-card plan; choose 4U for 600 W per card, warmer air or a later step to eight.

Lenovo Press LP2263 (28 July 2026) and LP2128 (20 October 2026) with Lenovo's SR650a V4 thermal rules; HPE QuickSpecs for the ProLiant DL385 Gen11 and DL380a Gen12; Supermicro SYS-322GA-NR and SYS-422GA-NRT datasheets; all read on 24 October 2026. Wall power by our estimate with two 280 W processors plus 15 per cent.

The models and users are the same in both chassis: four 96 GB cards give 332.2 GiB of usable memory by our users-per-card rule, and our guide to what a 4× RTX PRO 6000 Server Edition server runs lists them. This page covers the chassis.

Power caps: what 2U makers allow per card

The 2U makers handle 2.4 kW of cards in different ways. Lenovo's option guide LP2263 for the Server Edition lists two variants for the 2U ThinkSystem SR650a V4: the 600 W card, of which the server takes two, and a version "Slot Capped at 450W", of which it takes four. Lenovo's server guide, updated on 20 October 2026, gives the server's general limit as four 400 W or two 600 W double-wide GPUs. HPE's QuickSpecs for the 2U ProLiant DL385 Gen11 state that "4x … RTX PRO 6000 BSE … are now supported", list the card among the 600 W GPUs for which its Maximum Performance fan kit is supported, and offer "iLO GPU power capping" from the factory with a 450 W or 550 W enablement kit. Dell's 2U PowerEdge R7725 takes two 450 W double-width cards.

The 3U and 4U servers run four cards at their full 600 W. Supermicro lists four Server Edition cards for its 3U SYS-322GA-NR, HPE's 4U DL380a Gen12 takes two, four or eight, and Supermicro rates its 4U SYS-422GA-NRT for up to eight double-width 600 W cards at 35 °C.

The 450 W cap costs little for interactive inference. StorageReview measured 2 to 5 per cent lower LLM throughput at short context and 4 to 9 per cent at long context in October 2026, on a Workstation Edition card. In MLPerf® Inference v6.0, HPE's DL385 Gen11 ran four Server Edition cards in 2U at 450 W and reached 13,361.9 tokens per second on Llama 2 70B offline, entry 6.0-0054; by our division that is 3,340 per card, 3 per cent below the 3,451 per card of the same server with two cards at 600 W, entry 6.0-0052, with different driver builds.

MLPerf® Inference: Datacenter v6.0, closed division, available category, Llama 2 70B (llama2-70b-99), offline scenario, entries 6.0-0052 and 6.0-0054, retrieved from mlcommons.org on 24 September 2026. Results verified by MLCommons Association. Per card is our division of the system result, 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.

Prompt processing of long documents, batch jobs and fine-tuning lean on compute and lose more under a cap; a compute sweep by Delta Computer kept 86 per cent of its tensor result at 450 W. If those jobs matter, 600 W per card is the reason to choose 3U or 4U. Our guide to Server Edition power at 600, 450 and 300 W explains how the cable and the firmware set the mode.

Airflow and inlet temperature

Four passive cards need air through their heat sinks. NVIDIA's product brief for the Server Edition gives 65 CFM per card at a 35 °C inlet and 43 CFM at 25 °C, so four cards need at least 260 CFM at 35 °C. The whole server at about 3.4 kW needs roughly 540 CFM at an 11 °C rise, by the 160 CFM per kW of our guide to GPU rack power and cooling. A 2U chassis moves that air through half the frontal area of a 4U one, with smaller fans at higher speed.

The fans show it. Lenovo's SR650a V4 has six 60 mm fans, and its thermal rules name ultra fan modules for GPUs of 300 W and more. HPE's DL380a Gen12 uses four fan assemblies, each with one dual-rotor 92 × 56 mm fan and two 40 × 28 mm fans. Supermicro's 3U SYS-322GA-NR has up to five 80 mm fans at the front and six 60 mm fans inside.

The inlet limits follow. Lenovo's thermal rules for the SR650a V4 allow four RTX PRO 6000 Server Edition at the 450 W limit only at an ambient of 25 °C or lower, the same as for two at 600 W. HPE's thermal matrix for the DL385 Gen11 lists 25 °C for the GPU CTO server, and its list of changes states that "25C~30C server ambient temperature is now supported", next to GPU power capping and the Maximum Performance fan kit; ask HPE which GPU configuration that covers. Supermicro rates its 3U SYS-322GA-NR for 10 to 35 °C and its 4U SYS-422GA-NRT for eight 600 W cards at 35 °C. A room that holds 22 °C at the rack front suits both; a room that reaches 27 °C in summer, the top of ASHRAE's recommended range, suits the 3U or 4U server. Our guide to GPU server inlet temperature explains where to measure.

We check the chassis against your room before we quote. Tell us the inlet temperature and the feed at the rack position, and our team replies within one business day with the height units and power mode that fit.

Power supplies, cords and feeds

The number of power supplies grows with the height. Lenovo's SR650a V4 takes up to two supplies, from 800 to 3,200 W, and Dell's R7725 two, up to 3,200 W. HPE's DL385 Gen11 takes one or two, the largest listed being the 1800W-2200W Flex Slot Titanium kit. In 4U, HPE's DL380a Gen12 needs five supplies for two or four double-width GPUs, two for the system board and three for the GPUs, of 2,400 or 3,200 W with C19 cords only. Supermicro's 3U SYS-322GA-NR has three 3,200 W supplies in a 2+1 layout and its 4U SYS-422GA-NRT four in a 3+1 layout.

Our estimate: board power from NVIDIA's product brief, two AMD EPYC 9355 at 280 W, plus 15 per cent for memory, drives, fans and conversion losses; feeds planned at 80 per cent, 2.94 kW at 16 A and 5.89 kW at 32 A.

Two things follow by our arithmetic. First, at 2.7 kW one 3,200 W supply carries the capped 2U server alone, so two such supplies give 1+1 redundancy. At 3.4 kW, two supplies rated at up to 2,200 W, as in the HPE kit, together carry the load and leave no spare; ask the maker how its power management handles the loss of one supply in that configuration. Second, the cords: a 4U server with five supplies needs five C19 outlets on the rack PDUs, against two for a 2U server, and with A and B feeds each side has to carry the whole server. Feed types and PDUs are in our guide to GPU server power on 16 A, 32 A and three-phase.

Noise

GPU servers of either height belong in a server room. Few makers publish noise figures for GPU configurations; of the documents we read, only Dell's technical guide for the 2U R7725 does. For its GPU configuration, two double-width cards of up to 450 W with two 3,200 W supplies, Dell gives a sound pressure of 60 dB(A) in operation at 25 °C and 76 dB(A) at maximum load at 35 °C, with a sound power of 7.4 and 9.2 bels. For the 4U eight-card XE7745, Dell's guide gives 87 dB(A) at maximum load at 35 °C for its H100 NVL and L4 configurations.

We found no maker that publishes figures for the same four cards in 2U and in 4U, so the difference between them is not measured. Larger fans move the same air at lower speed, which usually lowers noise, and a cooler inlet lets every fan run slower; both point towards the 4U server in a warm room. For a room shared with people, our guide to a GPU server in an office covers noise, heat and power.

Rack density: height units against kilowatts

A 2U server with four cards gives two GPUs per height unit, a 4U server one. In most racks the feed decides before the height does. By our estimate, a 10 kW air-cooled rack takes three four-card servers at the 450 W cap, 8.1 kW in 6U of 2U servers, or two at 600 W, 6.8 kW in 8U of 4U servers, as our guide to the air-cooled AI server in a 10 kW rack shows. Dell's datasheet for its AI Factory with RTX PRO 6000 notes that customers typically air-cool racks with a heat load of up to 30 kW. At that load, eight four-card servers at 600 W draw about 27 kW and take 32U in 4U or 16U in 2U, so even then a 42U rack has room for the 4U version.

Height matters where the rack is already full, or in a colocation contract priced per height unit. Depth and weight matter too: Lenovo gives the SR650a V4 a depth of 924 mm and a maximum weight of 38.8 kg, Dell the R7725 802.4 mm with bezel, and Supermicro the SYS-322GA-NR 800 mm. Check the rack's depth and rails before ordering.

Growing from four to eight cards

The 2U servers we read stop at four double-width cards: the SR650a V4 at four capped cards, the DL385 Gen11 at four. A 2U server bought with two 600 W cards grows to four only with capped cards on Lenovo's platform, as our comparison of two, four or eight Server Edition cards explains. Going to eight means a second server or a new chassis.

A 4U chassis can grow in place. HPE's DL380a Gen12 allows two, four or eight RTX PRO 6000 Server Edition; eight need eight power supplies instead of five, and HPE's front chassis fan module kit and front panel kit. Supermicro's SYS-422GA-NRT takes up to eight 600 W cards behind dual-root PCIe switches. Plan the feed with it: eight cards at 600 W need about 6.3 kW at the wall by our estimate and a three-phase feed per side. Whether one eight-card server or two four-card servers fits better is the subject of one 8-GPU server or two 4-GPU servers.

Other cards in 2U and 4U

The same reasoning holds for other cards. The RTX PRO 4500 Server Edition is a single-slot card of 165 W, and Supermicro lists four of them in 2U servers such as the SYS-222C-TN, as our page on four RTX PRO 4500 Server Edition in 2U shows. The H200 NVL draws up to 600 W and supports only that power level; Lenovo's SR650a V4 takes two of them, and we recommend 4U for four, as on our page to buy a 4× H200 NVL server. The RTX PRO 6000 Max-Q has its own short list of 2U servers: Lenovo lists the Max-Q for its SR655 V3 and SR665 V3; its guide gives the value 3 for both without a legend, so we confirm the count in Lenovo's configuration before we quote, and our page on buying a Max-Q rack server covers the others. For continuous operation in a rack we recommend the Server Edition.

Which height for which situation

• A cool server room, a 230 V 16 A feed per side and no plan beyond four cards: 2U with the 450 W cap. It runs inference within a few per cent of the throughput at 600 W and fits the smallest feed.
• Long contexts, batch jobs or fine-tuning: 3U or 4U at 600 W per card, with a 32 A or three-phase feed per side.
• A room that reaches 27 °C or more at the rack front: 3U or 4U with a 35 °C rating.
• Eight cards within two or three years: a 4U chassis that lists eight cards, with the feed planned for about 6.3 kW.
• A full rack or colocation priced per height unit: 2U, if the room and the feed allow it.

Our inference node, four Server Edition cards, two AMD EPYC processors, 768 GB of DDR5 ECC and four NVMe drives, can be built in either height, and we choose the chassis with you. You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build it to order, assembled and tested before delivery, with manufacturer warranty on every component on one EU contract and invoice, as described on our page to buy a 4× RTX PRO 6000 Server Edition server. Send us the rack depth, free height units and feeds, and we propose the chassis.

FAQ