4U vs 8U GPU servers: PCIe cards or an HGX baseboard, by density, cooling and service

4U and 8U GPU servers: two ways to house eight accelerators

A 4U GPU server and an 8U GPU server can both hold eight accelerators, and they are different machines. The 4U class, which includes some 5U models, takes eight double-width PCIe cards of up to 600 W each, such as the NVIDIA RTX PRO 6000 Blackwell Server Edition or the NVIDIA H200 NVL. The 8U and 10U class carries eight GPU modules on NVIDIA's HGX baseboard, joined by NVLink switches; NVIDIA's own DGX B300 uses the same GPUs. Of the air-cooled B300 systems we read, Supermicro's and Gigabyte's are 8U, and NVIDIA's, HPE's and Dell's are 10U. Published system power for that class is 14.5 kW for the DGX B300 and 14.3 kW continuous for HPE's Compute XD690.

makers' documents listed under the two tables below, read 12 November 2026; card memory from NVIDIA's datasheets for the RTX PRO 6000 Blackwell Server Edition and the H200 NVL; HGX B300 figures from NVIDIA's HGX page of 30 September 2026. GPUs per rack unit is our division.

Every figure here keeps the basis its maker uses: net or gross chassis weight, a typical or a maximum system weight, continuous or peak power. The bases differ from maker to maker, so the tables describe each model, and two rows do not compare like with like.

We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project and build servers with two to eight cards to order; send us the rack depth, the feed and the inlet temperature and we reply with a configuration and quote within one business day.

The 4U 8-GPU server: eight PCIe cards in 4U or 5U

In this class each accelerator is a separate PCIe card in its own x16 slot. Supermicro calls its SYS-422GA-NRT a "DP Intel 4U Dual-Root PCIe GPU System with up to 8 GPUs" and puts the slots behind PCIe switches; ASUS connects the GPUs of the ESC8000-E12 directly to the processors. The RTX PRO 6000 Blackwell Server Edition has no NVLink, so traffic between cards crosses PCIe. The H200 NVL can be bridged: Gigabyte lists eight of them for its G494-ZB4-AAP2 "arranged as two 4-GPU sets, each with a 4-way NVLink bridge".

read 12 November 2026: Supermicro product pages for the SYS-422GA-NRT and SYS-522GA-NRT (undated); Gigabyte product page and datasheet for the G494-ZB4-AAP2 (datasheet last updated 29 October 2026); Dell PowerEdge XE7745 Technical Guide, Rev. A09, September 2026, and Installation and Service Manual; HPE ProLiant Compute DL380a Gen12 QuickSpecs, version 21, 2 November 2026; ASUS datasheet for the ESC8000-E12 (undated).

The fourth column shows how much "up to eight GPUs" depends on the maker. Supermicro states for the SYS-422GA-NRT: "Support up to 8 double-width 600W PCIe GPU accelerator cards at 35ºC". Gigabyte lists eight RTX PRO 6000 Blackwell Server Edition or eight H200 NVL for the G494-ZB4-AAP2 "at 25°C ambient" and writes: "The system supports up to 8 x 600 W GPUs with an optional fan kit." With that kit the chassis is 1,000.5 mm deep instead of 880 mm. Dell's thermal matrix for the XE7745 gives a maximum ambient of 30 °C for the H200 NVL and for the card it calls "NVIDIA RTX Pro 6000", and 35 °C for the L40S. ASUS adds to its range of 10 to 35 °C: "For operating temperatures above 30°C, please contact your local sales representative". HPE's QuickSpecs require a front chassis fan module kit and a front panel kit for eight RTX PRO 6000 cards.

The weight column needs its qualifiers. Supermicro prints the same net and gross weight for its 4U and its 5U chassis, ASUS states "CPU, DRAM & HDD not included", Gigabyte names no basis, Dell gives a maximum, HPE a minimum and a maximum.

These chassis also take single-slot cards. Supermicro lists the NVIDIA RTX PRO 4500 Blackwell Server Edition, a passive single-slot card of 165 W with 32 GB, for both models in the table; our pages cover buying the RTX PRO 4500 Server Edition and its specifications. The models are compared in Supermicro vs Gigabyte vs ASUS: 8-GPU servers, and the step below in 2U or 4U for a four-GPU server.

8U and 10U: HGX B300 systems and the DGX B300

An HGX system has no GPU cards. NVIDIA's HGX page of 30 September 2026 lists the HGX B300 as "8x NVIDIA Blackwell Ultra SXM" on one baseboard, with fifth-generation NVLink through an NVLink switch, 1.8 TB/s from GPU to GPU and 2.1 TB of GPU memory in total. The server maker builds processors, memory, power and cooling around it. NVIDIA's own system with these GPUs is the DGX B300, whose user guide prints "8 x 288 GB = 2.3 TB total"; our pages plan with the 2.1 TB of NVIDIA's HGX page.

read 12 November 2026: NVIDIA DGX B300 System User Guide, "Introduction to NVIDIA DGX B300 System"; HPE Compute XD690 QuickSpecs, a50009220enw, version 7, 2 November 2026; Dell PowerEdge XE9780 Installation and Service Manual and product page (data refreshed 3 November 2026); Supermicro product page for the SYS-822GS-NB3RT (undated); Gigabyte product page and datasheet for the G894-SD3-AAX7 (datasheet last updated 24 August 2026). "Not published" refers to these documents.

HPE writes: "The standard power draw is 14.3 kW continuous and 16.4 kW peak." NVIDIA's user guide gives a power consumption of 14.5 kW and a system maximum of 15 kW. Supermicro, Gigabyte and Dell state no system power in the documents read. Dell's manual lists the B300 modules at 1,100 W each, which by our arithmetic is 8.8 kW for eight before processors, memory, fans and conversion losses. More on each: Supermicro SYS-822GS-NB3RT, Dell PowerEdge XE servers and HPE ProLiant servers for AI.

Density: GPUs per rack unit and per rack

Counted per rack unit the PCIe class is denser: two accelerators per rack unit in 4U against one in 8U and 0.8 in 10U. The makers' rack plans count differently. Supermicro's datasheet for its B300 systems (2025) plans four air-cooled 8U systems, 32 GPUs, in one rack of 750 mm by 1,400 mm, and Dell's Technical Guide for the XE9780 (Rev. A03, August 2026) supports up to four of the 10U servers in its IR5048 and IR5052 racks. Four such systems fill 32U or 40U; the feed and the cooling remain to be checked. The same Supermicro datasheet places eight liquid-cooled 4U HGX B300 systems, 64 GPUs, in a rack of 800 mm by 1,470 mm. For PCIe servers we found no maker's rack plan; our guide to the air-cooled AI server in a 10 kW rack counts servers per rack from the power side.

Power: supplies, feeds and what the makers publish

The number of power supplies is the number of cords and outlets at the rack: four to eight in the PCIe class, twelve of 3.0 to 3.2 kW in the HGX class, or six of 6.6 kW at Supermicro. Supermicro and ASUS state 3+1 or 4+2 for their PCIe servers and Dell N+N per zone; for the HGX systems NVIDIA and HPE state N+N, Dell 6+6 and Supermicro 3+3. NVIDIA adds for the DGX B300: "The minimum number of required PSUs is six." Our article on PSU redundancy in GPU servers explains the schemes.

A supply reaches its label rating only at the upper end of the voltage range. Supermicro's 3,200 W unit delivers 2,900 W at 200 to 207 V and 3,200 W from 219.1 V. Dell's 3,200 W unit in the XE7745 gives 2,900 W at 200 to 220 V. Supermicro's 6,600 W unit gives 5,900 W at 200 to 207.9 V, 6,500 W at 230 to 238.9 V and 6,600 W only at 239 to 240 V.

For the 4U and 5U servers no product page or manual we read states a maximum system power. A Dell datasheet for eight-card nodes with the RTX PRO 6000 Blackwell Server Edition gives "~8 kW" at full load and calls the figure "indicative only". Our own estimate for eight 600 W cards with two AMD EPYC 9555 processors is about 6.3 kW at the wall. That is more than the 5.89 kW a 230 V 32 A feed carries at 80 per cent, so an eight-card server needs a three-phase feed per side, as our guide to GPU server power on 16 A, 32 A and three-phase shows. For the DGX B300 the user guide prints 12 × 3.2 kW in its physical table and, in its power table, 3.3 kW at 200 to 240 V and 16 A per supply.

Cooling: inlet limits on air, liquid in 4U

The models in both tables are air-cooled. In the PCIe class the stated limits for eight 600 W cards run from 25 °C to 35 °C. In the HGX class Gigabyte gives the G894-SD3-AAX7 an operating temperature of "10°C to 30°C", Supermicro and NVIDIA give 10 to 35 °C, and HPE gives a system inlet of 5 to 35 °C with the note "The upper limit and rate of change may be limited by the type and number of options installed." NVIDIA's data-centre guide for the DGX B300 plans supply air of at most 30 °C.

The DGX B300 user guide gives an airflow of 1,500 CFM at 70 per cent fan speed for the version with power supplies. Gigabyte's 8U system has 27 fans, fifteen of them 80 mm fans for the GPU tray; Supermicro's 4U PCIe server has ten 80 mm fans. Where to measure the inlet is covered in our guide to GPU server inlet temperature.

An HGX B300 system in 4U is a liquid-cooled system. Supermicro's SYS-422GS-NB3RT-LCC is a "DP Intel 4U Liquid-Cooled System with NVIDIA HGX B300 8-GPU (complete rack and onsite service are required)" with direct-to-chip cold plates, and Gigabyte's G4L4-SD3-LAX7 cools GPUs, processors and memory with liquid. In the PCIe class Gigabyte lists the 6U XLS4-SX2-LAS1 for eight liquid-cooled RTX PRO 6000 Blackwell Server Edition cards. For racks of air-cooled XD690 systems HPE lists its Adaptive Rack Cooling System and a rear door heat exchanger. See our overview of liquid cooling for GPU racks.

Weight, depth and lifting

The HGX systems are not always deeper: HPE's XD690 measures 800 mm, less than Dell's 4U XE7745 with its bezel, while Dell's XE9780 reaches 1,044.7 mm and Dell warns that its "Chassis length is longer than usual". Gigabyte's 4U PCIe server grows to 1,000.5 mm with the fan kit that eight 600 W cards need.

Weights can be compared within one maker. Dell gives a maximum of 68.5 kg for the 4U XE7745 and of 159.65 kg for the 10U XE9780 with B300. Across makers the bases differ: NVIDIA states 168 kg for the DGX B300 with power supplies and 207.6 kg packaged, HPE a typical 226 kg for the XD690, Supermicro 119 kg net for the 8U chassis. HPE requires "a customer provided Server-Lift" for break-fix work on the XD690 and writes that "A mechanized lift will be required to accompany the order"; Dell's manual states that "a lift is required for loads over 120 pounds". Floor load, doors and rails belong in a site survey for GPU servers; the rack side is in our guide to GPU rack power and cooling.

Service access: cards, trays and sleds

In a PCIe server a failed accelerator is one card in one slot. Dell notes for the XE7745 "Support for full extension of the system out of the rack to allow serviceability of key internal components."

In an HGX system the eight GPUs sit on one tray. NVIDIA's user guide says: "The motherboard and GPU trays are installed on sliding rails for easy access." Its service manual lists the parts a customer may replace on site, among them power supplies, fan modules, DIMMs and drives; the GPU tray and the GPUs are not in that list. Dell describes for the XE9780 a GPU sled that is released with levers and pulled out after the GPU fan cage is removed. Supermicro's manual for the SYS-822GS-NB3RT names two field-replaceable units, one for the motherboard and one named GPU-FRU-B300-6U-AC-NODE, and the product page carries the note "(onsite service required)". What makers require for service on site is the subject of on-site service for GPU servers.

Which GPU server form factor for which job

• Each model copy fits one card. Inference for several teams or mixed inference and rendering: a 4U or 5U server with eight PCIe cards.
• One model split across GPUs at low latency, or training. NVLink between all eight GPUs is the reason for an HGX system; the published measurements are in DGX B300 vs an eight-card RTX PRO 6000 server.
• Bridged cards in 4U. Eight H200 NVL in two bridged sets of four stay in the PCIe class; see eight H200 NVL server vs DGX B300.
• A site without 14 kW rack positions or a server lift. A PCIe server, or a hosted position as in colocation for GPU servers in the EU.

We build servers with two to eight RTX PRO 6000 Blackwell Server Edition or H200 NVL cards to order and check rack, power and airflow before we quote. The DGX B300 and HGX B300 servers we supply per project; the distribution asks for the company, the project and the intended use before a quote. Tell us the models you run and the rack position you have, and we answer with the form factor that fits.

FAQ