Preparing a server room for a GPU server: a checklist from rack depth to acceptance test

What to prepare before a GPU server arrives

Before a GPU server with two to eight cards arrives, a company prepares seven things: a deep enough rack position, a floor and delivery route for the weight, power feeds with C19 outlets, cooling within the maker's inlet limit, network uplinks, access to the management controller and a plan for the acceptance test. A server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards draws about 3.4 kW by our estimate, one with eight cards 6.3 to 8 kW, and any check left until delivery day can keep the delivered server switched off.

The list below is the one we go through with a customer before we quote, as our AI servers page states: rack, power and airflow are checked before the quote. Each item names who usually owns it, because facilities, network and server teams each hold a part.

Our checklist; the details and the documents behind each row follow below.

Rack depth, rails and free height

GPU servers are deep. Dell's PowerEdge XE7745, an eight-card 4U server for the RTX PRO 6000 Server Edition and the H200 NVL, measures 174.3 mm high, 886.73 mm deep without the bezel and 899.56 mm with it; from the mounting ears to the handles of the power supplies it is 864.73 mm. Supermicro gives 437 × 178 × 737 mm for its four-unit SYS-422GA-NRT, which takes eight 600 W cards. Add the room behind the server for the C19 plugs and the network cables.

The rails decide whether the server fits the rack at all. Dell offers one rail for the XE7745, a tool-less stab-in sliding rail, for 19-inch EIA-310-E four-post racks with square holes or with unthreaded or threaded round holes, and it points to its rail sizing and rack compatibility matrix for the depth range. Measure the distance between the front and rear posts of the rack position, and compare it with the range in the maker's matrix for the exact rail. Dell lists four-post racks only for this rail.

Count the free height as well. A 4U server takes four rack units, a 2U two-card server two; leave the units around it closed with blanking panels so that hot exhaust does not flow back to the intake. Our article on how many GPUs fit in one server compares 2U, 3U and 4U chassis.

Weight, floor loading and the route to the rack

Dell states a maximum of 68.5 kg for the XE7745 with all drives fitted. Its handling guidance in the same technical guide asks for two people from about 18 kg (40 lb), three people from about 32 kg (70 lb), and a server lift above 54.4 kg (120 lb). An eight-card server is therefore a job for a lift. Supermicro lists 29.7 kg net and 45.3 kg gross for the SYS-422GA-NRT in a package of 685 × 675 × 1,041 mm; its datasheet does not say which parts the net weight includes, so ask for the weight of the configured server with its cards.

For the floor, the rack total counts, and so does the load on each foot or castor. Four servers at 68.5 kg already come to 274 kg by our arithmetic, before the rack, its PDUs and the cabling; our guide to GPU rack power and cooling notes that a 64-GPU rack can approach one metric ton. Raised floors are rated per tile, so the facilities team checks the point load at the rack position.

The route matters on delivery day. Check that the package fits the doors, the corridor turns and the goods lift, and that the lift is rated for the package and the people who move it. Plan who receives the delivery, where the server waits before it is racked, and who removes the packaging. We deliver across the EU, and to other destinations through our logistics partner; commissioning on site is available on request.

Power feeds, PDUs and C19 outlets

The load per server follows from the cards. With two AMD EPYC processors sized to the card count and 15 per cent for the rest, we estimate about 1.8 kW for two 600 W cards, about 3.4 kW for four (about 2.7 kW at the 450 W cap) and about 6.3 kW for eight; Dell gives about 8 kW at full load for its eight-card nodes. Our GPU server power guide explains the method.

Our estimates as in our GPU server power guide, planned at 80 per cent of the feed rating: 2.94 kW for 230 V 16 A, 5.89 kW for 32 A, 8.87 kW for 400 V 16 A three-phase. Dell's figure of about 8 kW for its eight-card nodes is from its AI Factory with RTX PRO 6000 datasheet.

The outlets are a separate check. Under IEC 60320 a C13 outlet is rated at 10 A and a C19 at 16 A. Dell lists C19 cords for its 2,400 and 3,200 W supplies, and HPE states that its 2,400 and 3,200 W supplies take C19 or C19-C20 cords only, so a PDU with only C13 outlets cannot feed these servers. Count one C19 outlet per supply cord on each side: an eight-card Dell server has eight supplies.

With A and B feeds, each feed, PDU and UPS is sized for the whole server, because the supplies on one side carry it alone when the other side fails. Dell supports 4+4 grid redundancy for eight 600 W cards in the XE7740 only with its 3,200 W supplies, which deliver their full rating between 220.1 and 240 V. Ask the maker whether the exact configuration supports A/B at all before you lay two feeds. Sizing the UPS for the same load is covered in our guide to a UPS for a GPU server.

We check the feed, the outlet type and the A/B requirement against the configuration before we quote; send us the rating of the feed at your rack position and the PDU model, and our team replies with the power supplies, cords and card power mode that fit.

Cooling and inlet temperature at the rack position

Practically all the power becomes heat: an eight-card server at 6.3 kW puts about 21,500 BTU/h into the room, at 3.412 BTU/h per watt. At the 160 CFM per kW and 11 °C rise our rack guide uses, it needs about 1,000 CFM of cooled air at its front, and a four-card server at 3.4 kW about 540 CFM. Check what the floor tiles or the in-row unit deliver at that position, not the average for the room.

The inlet limit is set per configuration. Dell's technical guides for the XE7740 and XE7745 limit the ambient temperature to 30 °C with the RTX PRO 6000 Server Edition or the H200 NVL, against 35 °C for the L40S, inside a general range of 10 to 35 °C up to 900 m that falls by 1 °C per 300 m above. ASHRAE recommends 18 to 27 °C at the inlet, so a room at the top of that band leaves 3 °C to Dell's limit. Measure the inlet at the height where the server will sit, under today's load.

Close every empty rack unit with a blanking panel and keep the server's own GPU blanks and fillers installed when it ships with fewer cards than slots. Mount it in the same front-to-back direction as the rest of the row. Noise is part of the room too: Dell gives 87 dB(A) for the XE7745 at maximum load and 35 °C and highly recommends ear protection, so plan service work in the rack with that in mind. Our guide to air-cooled AI servers in a 10 kW rack counts servers per rack, and GPU server inlet temperature compares the limits of more servers.

Network uplinks and how models reach the server

Agree the uplinks before the server is built, because the network cards are chosen with the configuration. Dell's XE7745 has one OCP 3.0 slot for cards of up to 100 GbE and takes add-in network cards at 100, 200 or 400 GbE and a 200 GbE DPU; Supermicro's SYS-422GA-NRT has two 10GbE RJ45 ports on board. NVIDIA's certified-systems guide, updated on 30 September 2026, asks for at least 200 Gbit/s for servers that work together across nodes and up to 400 Gbit/s per GPU. The training node on our AI servers page carries 400G networking.

A single inference server needs much less bandwidth for its users, because prompts and answers are text. The uplink matters for the model store and the data. A checkpoint of 482 GB, the size of Qwen3-Coder-480B-A35B in FP8 on Hugging Face, takes about 6.4 minutes at 10 Gbit/s line rate and about 64 minutes at 1 Gbit/s, by our arithmetic; a restore of a vector index or a fine-tuning dataset behaves the same way. Where several servers exchange data during training or serve one model across nodes, our guide to an inference cluster without InfiniBand covers the fabric.

For the network team, the list is short: switch ports of the agreed speed on each side if the server uses two uplinks, the optics or direct-attach cables that match those ports, the VLANs, and the IP addresses. Switches and optics stay on the customer's side; we size the network cards in the server to match them.

BMC access and the management network

Every GPU server has a baseboard management controller with its own port. Dell's XE7745 has an iDRAC10 with a dedicated 10/100/1000 Mbit/s Ethernet port; HPE's DL380a Gen12 has iLO 6 or iLO 7, depending on the model, with a 1 Gb dedicated rear port and a front service port. NVIDIA's certified-systems guide asks for Redfish 1.0 or later and TPM 2.0 with secure boot. Prepare one switch port on the management network, an IP address or a DHCP reservation, and the account that will hold the BMC credentials.

The management network should be isolated. NVIDIA's guidance, as quoted in our guide to securing a GPU server, is to place BMC interfaces on isolated management networks and never expose them to the internet, and an authentication bypass in a widespread BMC firmware, CVE-2024-54085, has been on CISA's list of known exploited vulnerabilities since 25 June 2025. Decide before delivery which jump host reaches the BMC and whether IPMI over LAN stays off.

The BMC is also how the room is monitored. It reads temperatures, power and fan speeds without the operating system, and Dell lists its GPU telemetry under an iDRAC Datacenter licence, as our guide to GPU out-of-band monitoring explains. Order that licence with the server if your monitoring will use it.

Acceptance test in the rack

Plan the acceptance test before delivery, with the tools installed and a person who signs it off. Our guide to GPU server acceptance testing and burn-in has the procedure: every card present with the expected serial number, firmware and PCIe link, NVIDIA's DCGM diagnostics at level 3, a sustained power test while power limits, clocks and temperatures are watched, and on multi-GPU servers nvbandwidth and nccl-tests. Afterwards the ECC counters, the row remapper and the kernel log should show no new errors.

Run it in the rack, at the inlet temperature where the server will work. A test in the delivery bay says little about the 30 °C limit at the rack position, and the room's cooling only reaches a steady state after the server has run at full power for a while. Record the wall power the PDU reports during the test; it replaces the estimates in the table above.

Every server we supply is assembled and tested before delivery, and on request we install the operating system, drivers, CUDA and a container runtime; tell us what your acceptance protocol has to cover, and we prepare the software setup to match it.

What to send us before we quote

A short description of the room is enough for us to check a configuration. Send the rack model or its post depth, the free units at the position, the feed rating and type on each side, the PDU outlet type, the inlet temperature you measure there, the uplink speed of the switch ports and how the server reaches its rack from the loading bay. With these, the configuration we propose already fits the room.

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, with the chassis, power supplies, cords and network cards chosen for your rack. Our entry server has two RTX PRO 6000 Server Edition, the inference node four, and the training node eight H200 NVL with NVLink bridges. Every component carries the manufacturer warranty, on one EU contract and invoice; quantities and lead time are confirmed in writing with the quote.

FAQ