How to choose a UPS for a server and server room: calculating power and runtime
Servix technical team · Updated
In brief. A server room needs an online double-conversion UPS. Power is calculated not from power supply nameplates but from the actual peak consumption of each device (HPE Power Advisor, Dell EIPT, the iLO or iDRAC log): add the figures up, add 20–30%, and check both watts and volt-amperes. Two HPE DL380 Gen11 servers, an MSA 2060 storage system and a switch in our example come to about 1.64 kW: a 3 kVA / 2.7 kW UPS is enough, whereas a calculation from nameplates would lead to the 6 kVA class. Runtime is set by the time needed for a proper shutdown with margin, and it is extended with external battery modules.
A UPS is often chosen by eye: add up the power supply ratings and round up. The result is either a UPS that costs twice as much as needed, or one that holds for three minutes instead of fifteen. We go step by step through how to calculate power and runtime, where to get real consumption figures and how to set up virtual machine shutdown. This article is for administrators and IT managers building a server room of one or two racks. For a quick estimate there is a server room UPS calculator, and here is what stands behind its figures.
Volt-amperes and watts
A UPS has two power figures. Volt-amperes (VA) are apparent power; watts (W) are active power that actually goes to the load. They are related by the power factor: W = VA × cos φ. According to the Schneider Electric specification, the 3000 VA APC Smart-UPS SRT delivers 2.7 kW, i.e. a factor of 0.9. Other models may have a higher one, so look at the “VA/W” line in the specification, not just the figure in the name.
Server equipment has a power factor close to one: its power supplies have active correction. The HPE MSA 2060 QuickSpecs give 0.976–0.999 for its power supplies, and HPE Power Advisor gives 533 W and 536 VA for the server in our example. In practice the limit is watts, but the load must not exceed either of the two values.
A 20–30% margin is needed for three reasons: servers get additional memory and drives; battery capacity drops with age, and at a load near 100% the runtime is very short; there are modes above the calculated ones, for example fans at full speed in hot weather. In the example below we take 25%.
How much a server actually consumes
The power supply rating is a limit, not consumption. A server with two 800 W power supplies does not consume 1600 W: the second power supply is redundant, and each is designed to carry the machine alone. The Dell PowerEdge R760 12LFF from the catalog has two 1100 W power supplies — that is not a 2200 W load either. The real figure depends on the processors, memory, drives, cards and workload profile.
Manufacturer calculators
HPE has HPE Power Advisor, Dell has the Enterprise Infrastructure Planning Tool (EIPT, dell.com/calc). You build your configuration and get consumption at idle, at a given utilization and at maximum. In HPE Power Advisor (online version 16.11.0, 230 V mains) we built a DL380 Gen11 8SFF server: 2 × Xeon Silver 4410Y, 256 GB (8 × 32 GB DDR5), 8 SAS 2.4 TB 10K drives, MR408i-o controller, 10GbE SFP+ adapter, two 800 W power supplies. Result: 184 W at idle, 358 W at 50% utilization and 533 W at maximum load (Max Load Input Power).
A separate “Include Fan Max Power” checkbox shows consumption with the fans at 100% — the tool itself calls this the worst case, not normal operation. For the same configuration it comes to 775 W. For a server room that may overheat in the Tashkent summer, this mode is worth keeping in mind.
Measurement on a running server
If the server is already installed, the best source is the management controller. In iLO 6 the power meter keeps a 24-hour history: average, maximum and minimum; Redfish returns the same data (the Power resource, the PowerConsumedWatts and PowerMetrics fields). iDRAC9 has consumption monitoring, a warning threshold and power capping. Take the maximum over a week that includes nightly backups and peak days: the average is not suitable for UPS sizing.
The workload profile changes the figures noticeably. In the SPECpower_ssj2008 database, the same DL380 Gen11 with two Xeon Platinum 8480+ and 256 GB showed 594 W at 100% utilization under Windows Server 2022 and 689 W under SUSE Linux Enterprise Server 15 SP4. So take the peak of your own workload.
Equipment power guidelines
Figures for specific configurations from manufacturer tools and documentation. This is an order of magnitude; for your own calculation, build your configuration or take the peak from iLO and iDRAC.
Equipment | Configuration | Idle, W | Maximum, W | Source |
|---|---|---|---|---|
HPE DL380 Gen11 | 1 × Xeon Bronze 3408U, 32 GB, no drives, 1 × 800 W power supply | 91 | 251 | HPE Power Advisor |
HPE DL380 Gen11 | 2 × Xeon Silver 4410Y, 256 GB, 8 × SAS 2.4 TB 10K, 10GbE, 2 × 800 W power supplies | 184 | 533 (775 with fans at 100%) | HPE Power Advisor |
HPE DL380 Gen11 | 2 × Xeon Gold 5418Y, 512 GB, 8 × SAS 2.4 TB 10K, 10GbE | 200 | 674 | HPE Power Advisor |
HPE DL380 Gen11 | 2 × Xeon Platinum 8480+, 256 GB, 1 SSD | 158–201 | 594–689 at 100% utilization | SPECpower_ssj2008, two runs |
Dell PowerEdge R760 | 2 × Xeon Gold 6430, 256 GB, 2 HDD | 192 | not published | ENERGY STAR |
HPE MSA 2060 | SFF or LFF controller enclosure | — | 289–352 in operation; ≈ 518 by heat dissipation of 1766 BTU/h | HPE QuickSpecs |
24 × 1G switch without PoE | — | — | per datasheet; in the example we allow 50 | assumption |
Switch with PoE | Aruba Instant On 1930 24G PoE | — | plus a PoE budget of 195 or 370 W | model range |
Offline, line-interactive or online
The IEC 62040-3 standard classifies UPSs by how much the output depends on the mains: VFD, VI and VFI. In practice these are three familiar types.
Type | How it works | Transfer to battery | Where it fits |
|---|---|---|---|
Offline (VFD) | The load is powered directly from the mains; on failure a relay switches it to the inverter | With a break, on every failure | Workstations |
Line-interactive (VI) | The same, plus an autotransformer evens out the voltage in steps | With a break, on large deviations | A single server in an office, network equipment |
Online, double conversion (VFI) | Mains → rectifier → inverter: the load is always powered from the inverter | No break: the Eaton 9PX has a transfer time of 0 ms | Server room, storage systems |
Why online for a server room. The output is formed by the inverter, and the load sees neither sags, nor surges, nor frequency drift. There is no transfer moment that the server power supply has to ride through on its capacitors. The batteries are not engaged at every sag, as happens with a line-interactive UPS on an unstable grid. And an online UPS works fine from a diesel generator, whose frequency fluctuates more than that of the city grid. A caveat: many online UPSs have an economy mode in which the load bypasses the inverter most of the time; for a server room, keep double conversion.
Power quality: what matters for Tashkent
Sags, surges and short voltage dropouts do happen in office buildings, and their frequency depends on the specific building and line — there are no city-wide average figures here. A practical approach: install a UPS with an event log and after two or three weeks see how many times it switched to battery and what input voltage it saw. In the specification, look at the input voltage range within which the UPS works without the battery. A dedicated line with its own circuit breaker and proper grounding matter more than any UPS — this is covered in the office server room checklist. If there is a generator, the UPS must hold the load while it starts.
Runtime and external batteries
A UPS in an office server room does not need to run for hours — that is what a generator is for. Its job is to ride through short outages and, if power does not come back, give time to shut everything down in order. Calculation: delay before shutdown + shutdown of virtual machines + hosts + storage system + margin for battery aging and heat.
Runtime depends on the load non-linearly, so calculate it using the manufacturer's tables. An example from the specification of the Eaton 9PX Li-Ion 3000 VA / 2700 W (model 9PX3000GRT-L): at 2700 W — 9.1 minutes on internal batteries, at 1620 W — 14.9 minutes, at 810 W — 30 minutes. With one external battery module (EBM) at 1620 W — 40.3 minutes, with two — 63.3. Up to four 1U modules can be connected to this model, and up to ten SRT96RMBP units to the APC Smart-UPS SRT 3000.
Battery capacity drops with age, and faster in heat, so take the calculated time with margin and check the real runtime once a year. After an outage the batteries need time to charge: they may not handle a repeat failure half an hour later.
Proper shutdown: network card and software
A USB cable from the UPS serves one computer. If there are two servers or a hypervisor, you need a UPS network card: for APC — the Network Management Card in the SmartSlot, for Eaton — the Network-M2. The card reports status over the network, including via SNMP, keeps a log and sends out the shutdown signal.
The signal is received by an agent on the servers: for APC — PowerChute Network Shutdown, for Eaton — Intelligent Power Manager (IPM). Both have versions for VMware vSphere that shut down machines and hosts via vCenter; check the supported versions in the manufacturer's compatibility matrix. An independent alternative is Network UPS Tools (NUT): it works with UPSs over USB, a serial port and over the network via SNMP. One machine in primary mode communicates with the UPS, and the others in secondary mode receive commands from it.
The Proxmox VE documentation recommends installing a UPS, but there is no built-in agent for it. Proxmox VE 9 is built on Debian 13, and NUT is installed with the standard nut package. When the host shuts down, Proxmox itself shuts down the guest machines, by default waiting up to 180 seconds for each — include these timeouts in the runtime. In a cluster with HA, check shutdown_policy: in a full outage, migration to neighboring nodes, which are also on batteries, only eats up time. On choosing a platform, see the article Proxmox or VMware in 2026.
The order: virtual machines, hosts, then the storage system. Test the whole scenario during a maintenance window by pulling the UPS plug from the outlet: errors in the agent settings can only be found this way.
Rack-mount design
For a server room, rack or universal models (rack/tower) with rails are used. A 3 kVA UPS usually takes 2U, and a battery module another 1–2U. It is the heaviest item in the rack and goes at the bottom. Check the outlets: the APC SRT 3000 has 8 × C13 and 2 × C19; if there are not enough outlets, PDUs are installed. To replace a UPS without stopping the servers, manufacturers offer external service bypass panels.
Two server power supplies — two UPSs
Two power supplies protect against a power supply failure, but not against a power failure if both are plugged into one UPS. The correct scheme is two feeds: power supply A of each server and the storage system via PDU A to UPS 1, power supply B via PDU B to UPS 2. If one UPS fails, the equipment runs from the second.
Hence the rule: each UPS must carry the entire load alone. In normal operation each carries about half and holds longer, but you need to calculate for the worst case. Devices with a single power supply, like many access switches, are connected via an automatic transfer switch for two feeds (rack ATS), or two switches are connected to different UPSs. And do not forget the second power supply in the server itself: the base DL380 Gen11 8SFF listing has only one, and for a production system a second one is bought.
Calculation example: two 2U servers, a storage system and a switch
A typical server room for virtualization: two HPE ProLiant DL380 Gen11 servers, a shared HPE MSA 2060 storage system with direct SAS connection to both servers, and a 24-port switch. Assumptions are marked.
Step 1. List and consumption. We calculate the servers in the configuration from the Power Advisor section: it is heavier than the base listing, and the processors, memory and drives are selected for the task. Other versions are in the DL380 Gen11 line.
Device | Catalog listing, price as of November 2026 | For calculation, W | Source of the figure |
|---|---|---|---|
Server × 2 | HPE DL380 Gen11 8SFF — price on request (new; base: 1 × Xeon Bronze 3408U, 32 GB, one 800 W power supply) | 2 × 533 | HPE Power Advisor: 2 × Silver 4410Y, 256 GB, 8 × SAS 2.4 TB |
Storage system | HPE MSA 2060 SAS 24SFF — price on request (new, without drives) | 520 | QuickSpecs: 289–352 W in operation, 1766 BTU/h ≈ 518 W; we take the upper estimate |
Switch | Aruba Instant On 1930 24G 4SFP+ (JL682A) — 5,500,000 UZS (new) | 50 | assumption with margin |
Step 2. Total. 2 × 533 + 520 + 50 ≈ 1637 W. In volt-amperes: servers 2 × 536 VA per Power Advisor, storage system — 520 W at a power factor of 0.976 from QuickSpecs ≈ 533 VA, switch — an assumed 60 VA. Total ≈ 1666 VA.
Step 3. 25% margin. 1637 × 1.25 ≈ 2046 W and 1666 × 1.25 ≈ 2083 VA.
Step 4. UPS class. An online 3000 VA / 2700 W UPS fits: the calculated load is about 61% by watts, 76% with margin. The worst case with fans at 100% (2 × 775 + 520 + 50 = 2120 W) also fits, at about 79%.
Step 5. Runtime. Assumptions: 3 minutes waiting for power to return, up to 3 minutes to shut down the virtual machines (Proxmox timeout), about 3 minutes for the hosts and 1–2 minutes for the storage system. That is about 10 minutes, or 15 with a one-and-a-half-fold margin. According to the Eaton 9PX Li-Ion 3000 VA table, at 1620 W the internal batteries give 14.9 minutes — just barely enough, so we add one external module: 40.3 minutes. At a real load of about 1.1 kW (servers at 50%), the internal batteries would last about 22 minutes, but the calculation must be based on the peak.
Step 6. Redundancy. For a dual-feed scheme you need two such UPSs, each sized for the full load. In normal operation each carries about 820 W, and the internal batteries last about 30 minutes.
For comparison — a calculation from nameplates: 4 × 800 W power supplies on the servers, 2 × 580 W on the MSA 2060 and the switch give 4410 W, or 5.5 kW with margin. That is a 6 kVA class UPS for a real load more than two times smaller. Ready-made builds for such a project are in the selection guide server for virtualization.
Frequently asked questions
Can a UPS be sized by power supply ratings?
No: that is the power supply limit, not consumption, and the second power supply is redundant. In our example, such a calculation gave 4.4 kW instead of 1.64 kW. Use Power Advisor, EIPT or the peak from iLO and iDRAC.
How many minutes of runtime does a server room need?
As much as it takes for the delay and a proper shutdown of all equipment, plus a margin for battery aging. For a small server room it is usually 10–20 minutes at the calculated load. If services must keep running throughout the outage, you need a generator.
Do you need an online UPS if you have a generator?
Yes. A generator does not start instantly, and its frequency and voltage fluctuate more than those of the grid. An online UPS covers the startup pause and delivers stable voltage to the load.
Can both server power supplies be plugged into one UPS?
They can, but then a UPS failure will shut down the server. Redundancy requires two UPSs or at least two PDUs from different feeds, and each UPS must carry the entire load.
How will the UPS shut down virtual machines in Proxmox?
Via the UPS network card and NUT on each node: one node in primary mode, the others in secondary. When a host shuts down, Proxmox itself properly shuts down the guest machines.
Send to [email protected] the list of equipment for the rack: models, configurations, number of power supplies, and, if the servers are already running, the peak consumption from iLO or iDRAC. Add how many minutes the server room must hold and whether there is a generator. We will calculate the power and runtime, suggest the UPS class and connection scheme, select servers, storage systems and switches from the Servix catalog and send a quote. Prices include 12% VAT, warranty 12 months on new and refurbished equipment (serviced by Servix), free delivery in Tashkent; made to order — 2–3 weeks, the price is fixed in the invoice.
See also: SAN, NAS or all-flash: how to choose a storage system, DL360 Gen11 or DL380 Gen11, HPE MSA 2060 and 2062: configurations and prices, Rack or tower server: which to choose.
Related reading
An office server room: a checklist before buying a rack server
Power, cooling and a rack for a server in the office: UPS sizing, air conditioning in the Tashkent heat, rack depth, noise, iLO and iDRAC. A 12-point checklist.
Rack or tower server: which to choose
Rack or tower server: a 19″ cabinet or an office, noise from 22 to 49 dBA, 1U and 2U, which HPE and Dell towers fit into a rack and what is cheaper with the same components.
iLO and iDRAC: remote server management and which license you need
What iLO 5/6/7 and iDRAC9/10 can do without an OS, how iLO Standard differs from Advanced and iDRAC Express from Enterprise, and how to check the license when buying.
How much an office server costs in 2026: builds for 10, 30 and 50+ users
An office server for 10 people — from 14.3 million UZS, for 30 — from 20.9 million, for 50+ — from 45.8 million. Three builds from the Servix catalog, prices including VAT as of November 2026.




