GPU server inlet temperature: 25, 30 or 35 °C for RTX PRO 6000 Server Edition, H200 NVL

Which inlet temperature a GPU server needs

A GPU server with the NVIDIA RTX PRO 6000 Blackwell Server Edition or the NVIDIA H200 NVL needs inlet air of at most 25, 30 or 35 °C, depending on the chassis it sits in. Dell sets 30 °C for both cards in its eight-card PowerEdge XE7740 and XE7745, Lenovo 25 °C for two Server Edition cards in its 2U SR650a V4, and Supermicro 35 °C for eight Server Edition cards in its 4U SYS-422GA-NRT. ASHRAE's recommended range for the server room is 18 to 27 °C, so a room designed to that range suits all three, but only with little margin at its upper end.

The limits differ because the card does not set them. NVIDIA specifies how much air its passive cards need at a given temperature, and the server maker decides, with its fans, ducts and slot layout, up to which room temperature its chassis delivers that air. The same 600 W card therefore gets a different limit in each server. For a company that wants to put one or two GPU servers into an existing room, the inlet temperature at the rack position is the figure to measure before choosing the chassis.

ASHRAE classes A1 to A4: recommended and allowable

ASHRAE's Technical Committee 9.9 publishes the thermal guidelines that server makers rate their equipment against. The reference card for the fifth edition of its Thermal Guidelines for Data Processing Environments (2021) gives one recommended range for all air-cooled classes and a wider allowable range per class.

ASHRAE TC 9.9 reference card "2021 Equipment Thermal Guidelines for Data Processing Environments", from the fifth edition, read on 24 October 2026. Humidity limits differ by class and are on the card.

The two ranges answer different questions. In ASHRAE's words, facilities "should be designed and operated to target the recommended range", while the allowable range is where "IT manufacturers test their equipment to verify it will function". The card adds that the allowable range describes the functionality of the equipment and makes no statement about its reliability. The card also limits how fast the temperature may change: 20 °C in an hour and no more than 5 °C in any 15 minutes for equipment other than tape.

Dell rates the XE7740 and XE7745 to "Continuous Operation Specifications for ASHRAE A2", 10 to 35 °C below 900 m, with the maximum reduced by 1 °C per 300 m above that. Supermicro gives 10 to 35 °C for the SYS-422GA-NRT, the same range as A2. For these servers A2 is the general range only: the GPU configurations come with their own, lower limits.

Inlet limits per chassis and GPU

The table collects the limits from the makers' own documents. Dell writes "RTX Pro 6000" for the passive 600 W card, which is the Server Edition.

Dell PowerEdge XE7740 Technical Guide Rev. A04 (table 51, April 2026) and XE7745 Technical Guide Rev. A09 (September 2026); Lenovo ThinkSystem SR650a V4 thermal rules (pubs.lenovo.com) and product guide LP2128 (updated 20 October 2026); Supermicro SYS-422GA-NRT datasheet; 2CRSi Godì 1.8SP5-D product page (undated). All read on 24 October 2026.

Three things stand out. The first is that Lenovo's 450 W cap does not raise its limit: four capped Server Edition cards in the 2U SR650a V4 still need 25 °C, the same as two cards at full power. The second is that the network cards have limits of their own. Dell limits the DAC cable of a BlueField-3 B3220 DPU in the XE7745 to 30 °C, and Lenovo lowers the H200 NVL configuration from 30 to 25 °C when a B3220 is installed. The third is that the 4U chassis do better than the 2U one: Supermicro and Dell carry eight 600 W cards at 35 and 30 °C, where Lenovo's 2U server carries two at 25 °C.

The 2CRSi server, a French maker's 5U chassis, lists the RTX PRO 6000 Max-Q and the RTX PRO 5000 at up to 300 to 350 W per card and gives an operating range of only 10 to 25 °C. The Max-Q is NVIDIA's workstation card with a blower fan, and many server makers do not list it for their GPU servers. For continuous operation in a rack we recommend the RTX PRO 6000 Server Edition in a chassis whose maker lists it, as our comparison of Max-Q and Server Edition in a four-GPU server explains.

We build servers with two, four or eight RTX PRO 6000 Server Edition or H200 NVL for the room they go into; send us the inlet temperature measured at the rack position, and we reply with the chassis and card count that fit.

What the card needs: NVIDIA's airflow table

NVIDIA's product brief for the RTX PRO 6000 Server Edition, SP-12355-001_v02 of June 2025, gives the card an ambient operating range of 10 to 45 °C and a minimum airflow per heat-sink inlet temperature. The temperature in the table is that of the air reaching the card, after it has passed through the front of the server, and the airflow assumes the air is fully ducted through the heat sink.

NVIDIA RTX PRO 6000 Blackwell Server Edition product brief SP-12355-001_v02, table 3-1, read on 24 October 2026. Eight-card column: minimum per card × 8, by our arithmetic, without bypass air.

From 25 to 35 °C the airflow a card needs rises by about half, 65 against 43 CFM, and from 35 to 45 °C it nearly doubles again. The server's fans make up the difference by running faster, which costs fan power and raises the noise. Dell declares 87 dB(A) for its eight-card servers at maximum load and 35 °C and lists lower values at lower room temperatures. The brief also notes that 43 CFM is "the minimum required airflow whenever input power is provided to the card", at idle too.

When the air is not enough, the card protects itself. The brief describes a hardware slowdown that cuts the core clocks by 50 per cent, with "a significant impact to performance", and a hardware shutdown that should only occur if the cooling fails, for example with a fan fault. The H200 NVL is a passive card with a bidirectional heat sink and the same ambient range of 10 to 45 °C in its product brief PB-12128-001_v01. Its 30 °C limits at Dell and Lenovo show that the chassis decides what the room may reach; the card's own range is wider.

What this means for a company server room

A room designed for ASHRAE's recommended range runs at 18 to 27 °C at the server inlets. At 27 °C it has 3 °C left to Dell's 30 °C for the RTX PRO 6000 Server Edition and the H200 NVL, 8 °C to Supermicro's 35 °C, and none to Lenovo's 25 °C for its two-card configuration. A room set to 22 °C has room to spare for all of them. The question for the IT team is where in that band the room is on the hottest afternoon of the year, at the height in the rack where the GPU server will sit.

Three situations move the inlet above the setpoint of the cooling unit. Hot exhaust can return to the front of the rack through empty rack units, which blanking panels prevent, as our guide to air-cooled AI servers in a 10 kW rack describes. A GPU server adds its full load as heat, 6.3 kW for eight 600 W cards by our estimate in our GPU server power guide, and a room sized for less warms up. And when a cooling unit fails, the temperature rises until the remaining units or a shutdown stop it. ASHRAE's limit of 5 °C per 15 minutes then shows how little time a room with a small margin has.

For a director the decision comes down to two numbers: the measured inlet temperature at the planned rack position under the current load, and the limit of the configuration that the server maker documents. If the first stays below the second with a few degrees to spare in summer, the room is fit for the server; if not, the room needs more cooling capacity, or the configuration needs a chassis with a higher limit.

Our reading of the limits in the table above. Margins in the room are our recommendation, not a maker's figure.

Altitude, network cards and other limits that apply

The general range of Dell's A2 rating falls by 1 °C per 300 m above 900 m. A site at 1,500 m, by our arithmetic, has its maximum lowered by 2 °C, from 35 to 33 °C. Dell's tables of GPU limits do not say whether that derating also applies to the 30 °C GPU figure; plan as if it does, or ask the maker for the exact configuration.

Network and storage parts carry limits of their own. Dell requires 85 °C optical cables or DACs for OCP cards of 100 Gb and more, and limits the DAC cable of a BlueField-3 B3220 to 30 °C ambient in the XE7745. Lenovo's thermal rules lower the SR650a V4 limit to 25 °C for an H200 NVL configuration with a B3220, and with the GPU air baffle they limit memory modules to 128 GB. Dell's XE7745 limits apply to its configuration with eight E3.S NVMe drives and no rear drives. A configuration that adds a fast network card, more memory or more drives can therefore end up with a lower limit than the GPU alone suggests, which is why we check the whole bill of materials against the thermal tables, not only the cards.

Checking the inlet and the cards in operation

Measure the inlet temperature at the front of the server, not at the thermostat of the room. The server's management controller reports its own inlet sensor, and its log shows the highest values over a summer week better than a single reading. For the cards, nvidia-smi -q -d TEMPERATURE lists the slowdown and shutdown temperatures of each GPU, and nvidia-smi -q -d PERFORMANCE lists the clock event reasons. NVIDIA's documentation describes "HW Thermal Slowdown" as reducing the core clocks "by a factor of 2 or more due to temperature being too high" and "SW Thermal Slowdown" as clocks held below the requested value because the GPU is above its maximum operating temperature.

If either reason appears under load, the cards lack air: the inlet is too warm for the chassis, the fans are not in the required configuration, or blanks and fillers are missing. A run of the planned workload during your acceptance test, and again in the first warm weeks, shows whether the room holds. For a machine outside a data centre, our guide to a GPU server in an office covers noise and heat as well, and GPU rack power and cooling the airflow per kilowatt. Our checklist for preparing a server room for a GPU server puts cooling, power and floor loading in order, and our guide to a UPS for a GPU server covers what happens when the power fails.

We check room temperature, power feed and rack position before we quote; describe your server room and its cooling, and we return the configuration that runs there within the maker's limits.

What we supply

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 AI servers to order with two, four or eight RTX PRO 6000 Server Edition or H200 NVL cards, air-cooled, with chassis, fans and network cards chosen for the inlet temperature of your room. Our entry server has two RTX PRO 6000 Server Edition, the inference node four, and the training node eight H200 NVL with NVLink bridges. Each server is assembled and tested before delivery, and every component carries the manufacturer warranty on one EU contract and invoice. Commissioning on site is available on request.

FAQ