UPS for a GPU server: kW, kVA, runtime and clean shutdown for 2, 4 and 8 GPUs

How large a UPS a GPU server needs

A UPS for a GPU server is sized from the server's full-load wall power, in watts and in VA, with headroom: by our estimate about 1.8 kW for two NVIDIA RTX PRO 6000 Blackwell Server Edition or NVIDIA H200 NVL cards, 3.4 kW for four and 6.3 kW for eight, which at an 80 per cent margin means a UPS of about 2.3, 4.3 and 7.9 kW. The runtime only has to cover a clean shutdown, which on the units we read is a matter of minutes.

Dell's figure of about 8 kW for its eight-card nodes takes an eight-card server to a 10 kW unit. The shutdown covers the inference engine, any running training job and the operating system.

The rest of this guide works through the figures behind that answer: VA against watts, what an online double-conversion UPS does, how long a battery carries which load, how to shorten the shutdown and lengthen the runtime, and what A and B feeds change. The UPS models named are examples from one maker's published datasheets, chosen because their figures are complete enough to calculate with; other makers' units are sized in the same way.

Server load for 2, 4 and 8 cards

The load figures come from our GPU server power guide: the board power of the cards, two AMD EPYC processors sized to the card count, and 15 per cent for memory, drives, network cards, fans and conversion losses. A measurement on the delivered system replaces that estimate with a real value, and the measured value is the one to size the UPS on.

Wall power by our estimate as in our GPU server power guide; Dell's "~8 kW" estimated node maximum at full load from its AI Factory with RTX PRO 6000 datasheet (XE7740 and XE7745). UPS column: wall power ÷ 0.8, our arithmetic. Example units: APC Smart-UPS On-Line SRT6KXLI (6 kVA / 6 kW) and SRTG10KXLI (10 kVA / 10 kW), from APC's product page and datasheet.

The 80 per cent margin is the one we use for electrical feeds. It leaves room for the start-up peaks of the power supplies, for a switch or management device on the same UPS, and for the difference between the server's VA and its watts, which none of the server guides we read states. A 6 kVA unit at 80 per cent carries 4.8 kW and covers every row up to four cards at 600 W; a 10 kVA unit carries 8 kW and covers an eight-card server, with Dell's figure exactly at the margin.

VA, watts and power factor

A UPS has two ratings. The rating in watts is the real power it can deliver, the power that the server turns into computation and heat. The rating in VA is the apparent power, voltage times current, which also counts current that flows without doing work. The ratio of the two is the power factor, and because real power can never exceed apparent power, it is at most 1.

Both ratings are limits, and the load has to stay below each. APC rates its current online units at the same value in both, 6 kVA and 6 kW for the SRT6KXLI and 10 kVA and 10 kW for the SRTG10KXLI. On such a unit the VA rating is the one a server reaches first, since a server's VA is never lower than its watts. On a unit rated lower in watts than in VA, the watt rating can be the tighter one. Read both on the datasheet and compare each with the server.

On the server side, the makers state the maximum input current of each supply, not the power factor. Dell's XE7740 and XE7745 guides rate the 3,200 W Titanium supply at 16 A between 220.1 and 240 V, which at 230 V is 3.68 kVA per supply by our arithmetic. That is the supply's rating, not what the server draws: with eight supplies in a 4+4 arrangement the load is shared, and the server draws its wall power, 6.3 to 8 kW for eight cards. Size the UPS on the wall power and keep the input ratings for the electrician who plans the feed.

Online double conversion and what it does for a GPU server

In an online double-conversion UPS, the rectifier turns the incoming AC into DC for the battery and the inverter, and the inverter feeds the load all the time. When the mains fails, the inverter keeps running from the battery, so the server's power supplies see no switchover at the outlet. Both APC units in the table are of this type; APC describes the SRT6KXLI as a "double-conversion on-line pure sine wave tower UPS".

For a GPU server this matters because its work is expensive to repeat. A training run that stops without a checkpoint loses everything since the last one, and an inference service that drops loses its users' sessions and has to load its models again. A unit that bridges short dips without a transfer keeps both running through them, and the battery covers a longer outage until the shutdown. The server's supplies also see the inverter's output rather than the mains: Dell rates its 3,200 W supply at full output only between 220.1 and 240 V, while the SRTG10KXLI accepts 160 to 285 V at its input.

The outlets matter as much as the topology. The SRT6KXLI has six C13 and four C19 outlets, and GPU server supplies of 2,400 W and more need C19 cords, as our power guide shows. HPE's DL380a Gen12 uses five such supplies for two or four double-wide cards, one more than the unit's C19 outlets. The SRTG10KXLI takes a hardwired input at 230 V (160 to 285 V) and offers two C13 and one C19 outlet plus a hardwired output, so an eight-card server with eight C19 cords hangs on a rack PDU fed from that output. The electrician sizes the input feed of a unit like this from the maker's installation documents.

We build servers with two, four or eight cards and choose the number of power supplies and cords to fit your UPS and PDUs; send us the model and outlets of your UPS, and our team replies with the configuration and the load it puts on each side.

Runtime: what the battery carries

A datasheet with two runtime points is more useful than a single figure, because the runtime more than triples when the load halves. APC's datasheet for the SRTG10KXLI, dated 29 August 2022, gives about 3.5 minutes at 10,000 W and about 11 minutes at 5,000 W with its internal lead-acid batteries, and a recharge time of 4 hours.

APC Smart-UPS On-Line SRTG10KXLI product datasheet of 29 August 2022 (internal batteries), read on 24 October 2026. Rows between the two points are our reading; we do not interpolate a figure. The 300 W row is our estimate: 8 × 300 W cards plus 2 × 360 W processors, plus 15 per cent.

For the 6 kVA SRT6KXLI, APC's product page gives no runtime figure; it states that up to ten external battery packs can be added for longer runtime. A lithium-ion variant of the range exists as well. Whatever the unit, ask the maker for the runtime chart at your load, not at full rating, and remember the recharge time: after one outage the battery needs hours before it carries the full runtime again, so a second outage the same afternoon finds it partly empty.

A few minutes are enough when the UPS is there to shut the server down cleanly, which is what we size for here. Riding out an outage of an hour needs a generator or a much larger battery, and that is a building decision rather than a server one.

What a clean shutdown takes

The runtime has to exceed the time between the UPS reporting a low battery and the server being off. That time depends on the workload, so measure it during commissioning rather than guessing.

1. Inference. The serving engine stops taking new requests, finishes or aborts the running ones, and exits. The model weights are on disk already, so nothing large has to be written.
2. Training. A job that runs for days should write a checkpoint before it stops. By the arithmetic of our guide to CPU, RAM and storage for a 4-GPU server, a resumable checkpoint of a full Llama 3.3 70B fine-tune is about 990 GB. Micron lists 10.0 GB/s of sequential writes for its 9550 NVMe drive, so by our arithmetic one such drive needs at least 99 seconds for it and four drives striped at least 25 seconds.
3. The operating system. Services stop, file systems are synced and unmounted, and the server powers off.

Network UPS Tools (NUT), as of version 2.8.5, documents how this is triggered. Its upsmon daemon treats a UPS that is "on battery and has a low battery" as critical and then starts the shutdown sequence. On a server that controls the UPS directly, the primary, it waits for the other servers on the same UPS, the secondaries, to disconnect, for at most HOSTSYNC, 15 seconds by default, then waits FINALDELAY, 5 seconds by default, and runs its shutdown command. NUT's documentation warns that the secondaries have to finish before the primary cuts the shared power. Secondaries monitor the UPS over the network through the primary's NUT server; APC also includes a network management card with the SRT6KXLI.

Lengthening the runtime with a power cap on battery

The cards are most of the load, and they can be capped while the server runs. NVIDIA's product brief SP-12355-001_v02 gives the RTX PRO 6000 Server Edition a minimum power limit of 300 W in both power modes, settable in-band with nvidia-smi -pl, with a typical response of 50 ms. The H200 NVL's brief gives a minimum of 200 W.

Minimum power limits from NVIDIA's product briefs SP-12355-001_v02 (June 2025) and PB-12128-001_v01 (April 2025). Wall power by our estimate: cards at the minimum plus two processors (2 × 280 W for four cards, 2 × 360 W for eight), plus 15 per cent.

NUT's upsmon can run a command when the UPS goes on battery: NOTIFYCMD is "called with the full text of the message", for every event whose NOTIFYFLAG includes EXEC, and ONBATT is the event "UPS is on battery". A small script that caps every card at its minimum on ONBATT, with a second one that restores the cap when the power returns, moves an eight-card server from between 3.5 and 11 minutes to more than 11 minutes on the 10 kVA unit above, by the datasheet's two points. The jobs run slower while capped, which matters little in the minutes before a shutdown. A cap set with nvidia-smi lasts until the next driver load, as our guide to the RTX PRO 6000 power limit explains, so the script has to restore it after the restart too.

A and B feeds: one UPS per side

With redundant power supplies on two feeds, each side has to carry the whole server when the other fails, as our GPU server power guide explains for the feeds. The same applies to the UPS. With one UPS on feed A and one on feed B, each is sized for the full server load, not for half of it. With one UPS on feed A and the mains on feed B, a mains failure drops side B, and the UPS on side A carries the whole server on its own. Either way, the UPS on a side is sized for the full load.

Dell's eight-card XE7740 supports 4+4 grid redundancy for eight 600 W cards only with 3,200 W supplies, so the four supplies on one side carry up to 12.8 kW of supply rating against a load of 6.3 to 8 kW. The UPS on that side has to carry the 6.3 to 8 kW. For a rack of several servers, add the servers on each side and size that side's UPS for the sum.

Inrush and restart after an outage

When the mains returns and the UPS restarts its outputs, or when the servers are switched on together, every power supply charges its input capacitors at once. The Dell technical guides for the XE7740 and XE7745 and HPE's QuickSpecs for the DL380a Gen12, as read on 24 October 2026, list input current ratings and efficiency but no inrush figure. Ask the server maker for the inrush current of the exact supply, and the UPS maker for the overload the inverter accepts, before several servers share one UPS.

Two practical measures reduce the peak even without figures from either maker. Check whether the servers' management controllers offer a delayed power-on after a power loss, and set different delays for the servers on the same UPS so that they do not start at once. And do not let the GPU workloads start automatically the moment the operating system is up: the UPS battery is still recharging, and a second outage would find it partly empty.

Heat and cooling during an outage

A UPS loses part of its power as heat while it runs, and that heat adds to the server's in the room. If the room's cooling units are not on the UPS, the server keeps heating the room during the runtime while nothing removes the heat. For an eight-card server at 6.3 kW that is the full load going into the room for several minutes. Our guide to GPU server inlet temperature shows that 600 W GPU configurations stop at 25 to 35 °C at the inlet, depending on the chassis, so a room with little margin should shut down the GPU workloads early in an outage rather than at the last minute.

For a machine outside a data centre, our guide to a GPU server in an office covers circuits, noise and heat, and our checklist for preparing a server room for a GPU server puts power, UPS and cooling in order. We check the feed, the UPS and the cooling against the configuration before we quote; describe your power and UPS setup, and we return the power supplies, cords and estimated load per side with the configuration.

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, or eight RTX PRO 4500 Server Edition, with power supplies and cords chosen for your feeds and UPS. 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