GPU server in an office or small server room: noise, heat output and power requirements

GPU server noise, heat and power in an office

A GPU or AI server can run in an office or a small server room when the circuit, the cooling and the noise fit the room. Up to about 2 kW, which covers a tower with four RTX PRO 6000 Max-Q or a 2U server with two 600 W cards, one 16 A circuit at 230 V carries the load; four 600 W cards need 32 A or three-phase power, and eight need three-phase feeds. Every watt the server draws becomes heat, 3.412 BTU/h per watt, so the room’s cooling has to remove the same kilowatts for as long as the server runs. Noise rules out rack servers for an office, since the GPU rack servers below are declared at 8.5 to 9.1 B of sound power under load and a DGX Spark at 3.5 B.

Racks above 15 kW, containment and floor loading are in our guide to power and cooling for a GPU rack.

Power draw of 2, 4 and 8 GPU configurations

The cards set most of the load. NVIDIA gives the RTX PRO 6000 Blackwell Max-Q a maximum power consumption of 300 W and the Workstation Edition 600 W. NVIDIA lists the Server Edition at “Up to 600W (configurable)”, and the H200 NVL has the same maximum thermal design power. The L4 is a low-profile card at 72 W. A DGX Spark has a 240 W power supply for the whole system.

Card ratings from NVIDIA’s product pages for the RTX PRO 6000 Max-Q, Workstation and Server Edition, H200 NVL, L4 and DGX Spark, read on 9 October 2026; the 2U, four-card and eight-card 600 W rows are Server Edition or H200 NVL cards. Processors assumed at 350 W each. Heat = watts × 3.412.

The third column adds only the rated figures of cards and processors; memory, drives, fans and power-supply losses come on top. We estimate about 1.8 to 2 kW at the wall for the four-card Max-Q tower and for a 2U server with two 600 W cards, and a fully loaded eight-card PCIe server can approach 10 kW, as our article on how many GPUs fit in one server works out. Where a maker’s power calculator covers the exact configuration, plan with its figure.

The Server Edition and the H200 NVL can run at a lower configured limit, which cuts the heat with the power. Our comparison of the Workstation, Max-Q and Server Edition explains which one belongs in a tower and which in a server chassis.

Heat output in BTU/h and cooling for a small server room

Schneider Electric’s White Paper 25 on calculating total cooling requirements, published under its APC brand, states: “The power consumed from the AC power mains is essentially all converted to heat.” The heat output of a GPU server in watts therefore equals its power draw, and NIST’s conversion factor gives 3.412 BTU/h per watt. A server drawing 2 kW at the wall puts 6,824 BTU/h into the room, and an air conditioner labelled 12,000 BTU/h has about 3.5 kW of cooling capacity.

Size the cooling from the full-load wall power, not the idle draw, and add the other heat sources in the room, such as the UPS, which the same white paper covers. Dell’s R760xa guide lists a heat dissipation of 9,000 BTU/hr for its 2,400 W supply, a figure per supply: size from the server’s full-load draw, not the sum of redundant supplies.

Office air conditioning that runs only in working hours or only in summer does not fit a GPU server, which needs cooling whenever it runs, including nights, weekends and winter. Check the unit’s operating range at low outdoor temperatures, its cooling capacity in kW on the hottest days and its sensible share, since IT equipment gives off dry heat. A second unit keeps the room cooled during service.

Circuits, sockets and UPS for a GPU server outside a data centre

A 16 A circuit at 230 V carries 3,680 W, and we plan at 80 per cent of a feed. That leaves room for one machine of about 2 kW, not for two together with the printers on the same circuit, so ask the electrician which outlets share the breaker. A tower plugs into an ordinary 230 V socket.

Rack servers take larger cords. Dell lists a C19 cord for the 2,400 W supply of the R760xa, and Lenovo states for its SR650a V4 that “All AC power supplies support 230V power”, while only some also accept 115 V. Where a rack PDU for 32 A single-phase or three-phase power has an industrial plug, plug and socket-outlet follow IEC 60309, the standard for accessories “primarily intended for industrial use”, and an electrician installs the socket-outlet. Our rack guide covers the C13 and C19 outlet ratings and the PDU check.

Size the UPS above the full-load wall power with headroom, in watts as well as in VA, and decide whether it bridges only short interruptions or carries the server through an orderly shutdown.

We check the rack, power and airflow before we quote. Send us the circuit rating, the room size and its cooling capacity through the form below.

GPU server noise: declared values for towers and rack servers

Sound power, LwA,m, is a property of the machine, given in bels, where 1 B equals 10 dB. Sound pressure, LpA,m, is declared at a defined bystander position; what a person hears at a desk also depends on the distance and the room. Lenovo and Dell base their figures on ISO 7779 measurements and ISO 9296; NVIDIA declares the DGX Spark according to ECMA-109.

Declared values read on 9 October 2026: NVIDIA DGX Spark product page (ECMA-109, June 2025), Lenovo SR650a V4 environmental specifications, Dell PowerEdge XE7745 Technical Guide Rev. A09 (September 2026), all at 25 °C ambient. Dell lists higher values at 28 and 35 °C.

A difference of 10 dB in sound power is a factor of ten in acoustic power, so the 2U server emits 100,000 times the sound power of the Spark. The eight-card server with 72 W L4 cards declares 8.6 B both at idle and in operation, 0.5 B below the H100 NVL configuration under load, so a rack GPU server stays loud with low-power cards and at idle. Dell writes that the XE7745 “is not recommended for deployment in acoustically sensitive environments, such as work areas lacking acoustic insulation”, and calls the R760xa “a rack server appropriate for attended data center environment”. The declared levels rise with the room temperature as well.

As of October 2026 we found no maker’s declaration for a tower with RTX PRO 6000 cards under GPU load. Before a tower goes next to a desk, ask for the acoustic data of the configured system and point its exhaust towards open space. Lenovo’s documentation notes that European Community directives may govern noise exposure in the workplace; what Directive 2003/10/EC on workers’ exposure to noise requires for a given workplace is an assessment for the company’s occupational safety and legal functions.

Inlet temperature: ASHRAE ranges and vendor limits

ASHRAE’s 2021 Equipment Thermal Guidelines recommend 18 to 27 °C at the server inlet for classes A1 to A4. The allowable range is 15 to 32 °C for class A1 and 10 to 35 °C for A2. GPU configurations often allow less than the class: for an ASHRAE A2 environment, Dell’s R760xa guide sets a “Maximum 30° C (86°F) for GPU > 350 W With Processor > 225 W”, and Lenovo states that the SR650a V4 complies with A2 in most configurations and with A3, A4 or H1 depending on the configuration. A small room that reaches 30 °C on a summer afternoon is already at the limit Dell sets for these GPU configurations.

Read the inlet temperature at the front of the server under full load rather than at the thermostat. In a small room the hot exhaust returns to the intake unless the cooling unit draws it off, so place server and air conditioner with that path in mind.

Acoustic racks and separate rooms

Several manufacturers make sound-insulated rack cabinets with fans or a cooling unit built in, for one or two lower-power servers in a room where people work. Compare the cabinet’s rated heat removal in kW with the wall power from the first table, and its declared attenuation with the server’s declared sound power. Unless the cabinet’s cooling unit takes the heat out of the room, through a split unit or a duct, the heat stays in the room, together with the power of the cabinet’s own fans or compressor, and the air conditioning still has to remove it.

For a 4 × 600 W server or more, a separate room with a door, its own cooling and a dedicated circuit is the usual answer. With remote management through the server’s BMC, routine work needs no visit to the room.

Room checklist before a GPU server moves in

Ranges from ASHRAE’s 2021 Equipment Thermal Guidelines and the makers’ documents cited above.

We build GPU workstations and short-depth servers for labs, pilots and branch sites, as well as rack servers. Describe the room and who works next to it in the form below, and we reply with a configuration and quote within one business day.

When a data centre or colocation is the better place

The office stops being the right place when the load passes one circuit, when the cooling cannot run all year, or when people work within earshot of a rack server. An eight-card server needs three-phase feeds and cooling for up to about 10 kW at full load, and with a declared 9.1 B it needs a closed room away from desks. A colocation rack in a data centre provides the feed, the cooling and the distance from desks; above about 50 kW per rack, our guide to liquid cooling for GPU racks takes over.

Where the server room itself is ageing, the systems around the AI server can move as well. Our EU Cloud hosts guaranteed compute or a private cloud in Tier-3 data centres in the EU, with migration in agreed maintenance windows with a rollback plan.

What we supply

We build AI servers to order, from GPU workstations and short-depth servers for labs and branch sites to 2U and 4U GPU chassis with redundant power supplies, with air or liquid cooling chosen by GPU count and site. The cards include the RTX PRO 6000 Workstation, Max-Q and Server Edition, the H200 NVL, the L40S and the L4. Each server is assembled and burn-in tested and comes with manufacturer warranty on one EU contract and invoice. We check the rack, power and airflow before we quote, and the configuration and quote follow within one business day.

FAQ