Server Rack Power, PSU Sizing & Heat (BTU) Calculator

Describe one server's parts and how many such servers share the rack, and the calculator works out the rack's energy and cost each month, its full-load draw, the heat the cooling must remove, the UPS it needs and the rating each power supply needs. The defaults describe four two-socket servers, each with sixteen DDR5 sticks, four hard drives and two NVMe SSDs. Together they use 1,312.4 kWh a month, 196.86 at 0.15 per kWh, and at full load they draw 3,341 W and give off 11,400 BTU/h.

How the estimate works

Every part has an idle and a full-load figure. CPUs and GPUs draw their TDP at full load and a share of it at idle, 20 percent by default. Memory, hard drives and NVMe SSDs use datasheet figures, and the board, fans and network cards one allowance for both states. The sum is the DC power the supplies must deliver.

Converting mains power to DC loses some of it, more or less depending on how hard a supply works, so the calculator reads efficiency off the 80 PLUS curve at the share of rated output the supplies carry. Two redundant supplies split the load:

W = P ÷ η(P ÷ (n × R))

Here P is the DC load, n the number of supplies, R the rating of each and η the efficiency at that share. Idle and full load each go through the curve, and the duty cycle d, the share of time at full load, mixes them over a 730-hour month, a year's 8,760 hours divided by twelve:

kWh = servers × ((1 − d) × W_idle + d × W_full) × 730 ÷ 1000

Every watt taken from the wall ends up as heat, so heat is the full-load draw in BTU per hour. The UPS keeps that draw within 80 percent of its rating and covers the brief peak while drives spin up:

BTU/h = W_full ÷ 0.2930711

VA = max(W_full ÷ 0.8, W_peak) ÷ PF

PF is the lowest power factor 80 PLUS allows the level. Each power supply must also carry the server's peak on its own, because a redundant pair has to keep running when one fails.

A worked example

Each default server has two 250 W CPUs: 500 W at full load and 100 W at idle. Sixteen 64 GB DDR5 sticks add 68.35 W at idle and 153.22 W at full load, four hard drives 22.45 W and 37.19 W, two NVMe SSDs 10 W and 28 W, and the board, fans and network 50 W. That makes 250.80 W at idle and 768.40 W at full load.

Two 1,100 W Platinum supplies share it. At idle each carries 125.40 W, 11.40 percent of its rating, where the curve gives 88.28 percent efficiency, so the server takes 284.10 W from the wall. At full load each carries 384.20 W, 34.93 percent of its rating, at 91.99 percent, for 835.31 W. With 30 percent of the time at full load the average is 449.46 W, and four servers over 730 hours use 1,312.43 kWh, shown as 1,312.4.

At full load the rack draws 3,341 W, which is 11,400 BTU/h. Divided by 0.8 that is 4,177 W, more than the 3,697.97 W peak while the drives start, and Platinum's 0.95 power factor makes it 4,397 VA. Starting all four drives at once lifts each server's DC load to 855.31 W, so each supply needs at least 856 W, and 1,100 W leaves room to spare.

Where the part figures come from

Intel defines TDP as "the power consumption under the maximum theoretical load" and "the maximum power that one should be designing the system for", though turbo "can exceed the maximum TDP but only for a limited time." The memory figures are datasheet currents times their voltages, active standby for idle and burst reads for full load. Kingston's DDR4 currents are the memory chips', with no word on the register, while Micron's DDR5 figures include the module's power chip and register, so the two are not like for like. Smaller sticks with fewer chips draw less, which is what the custom option is for. NVIDIA rates the L40S at 350 W, the GPU default, and the LLM VRAM calculator estimates how many GPUs a model needs.

Seagate lists the Exos X20's start current as 2.197 A at 12 V and 0.932 A at 5 V, about 31 W against 9.30 W while reading. Drives can wait to spin up one at a time, which the staggered option models, but Seagate's manual lists that feature, power-up in standby, as supported and not enabled, so the default starts every drive at once.

The 80 PLUS curve

The efficiency figures come from 80 PLUS, which Ecos Consulting announced in March 2004 at ACEEE's Market Transformation Symposium. Utilities paid computer makers $5 for each desktop and $10 for each desktop-derived server whose supply was at least 80 percent efficient at 20, 50 and 100 percent of rated load, and Southern California Edison, the Northwest Energy Efficiency Alliance and Pacific Gas & Electric were among the early sponsors. ENERGY STAR required the same of computers from 20 July 2007. Bronze, Silver, Gold and Platinum followed between 2008 and 2010, Titanium for servers in 2011, and the table now tops out at Ruby.

These are the minimums 80 PLUS sets for internal redundant supplies at 230 and 277 V. Below Titanium the program calls its 10 percent figures suggestions rather than requirements, and the calculator uses them anyway, because a redundant pair often runs that lightly. Between published points efficiency follows a straight line. Below the lowest one the power lost stays at its value there, since a supply's fixed losses do not shrink with the load.

Heat and the UPS

Schneider Electric's White Paper 25 says the power IT equipment draws from the mains "is essentially all converted to heat", so its thermal output "is simply the same as the power input." NIST puts a Btu per hour at 0.2930711 W. For rooms under 4,000 square feet the same paper rates cooling at 1.3 times the IT load plus any redundancy, and for an ordinary room the BTU calculator does the sizing.

A UPS has two ratings, and Schneider's White Paper 15 is plain that "neither the watt nor the VA rating of a UPS may be exceeded." Servers made after about 1996 have power-factor-corrected supplies, it notes, and 80 PLUS holds them to at least 0.90, or 0.95 from Platinum up. Schneider suggests loading a UPS "preferably around 80%" of its capacity, and Vertiv says it should "not be loaded to more than 80% capacity." The UPS runtime calculator shows how long its batteries would last.

Why idle power matters

In 2007 Luiz André Barroso and Urs Hölzle of Google found that servers "operate most of the time at between 10 and 50 percent of their maximum utilization levels", and that "even an energy-efficient server still consumes about half its full power when doing virtually no work." The 30 percent duty cycle sits in that range: if power rises in a straight line with load, average utilization and time at full load are the same number. Processors, they wrote, could already drop below a third of their peak power at very low activity, and the 20 percent idle share sits under that. That year the EPA put the electricity used by American servers and data centers in 2006 at about 61 billion kWh, 1.5 percent of the national total. The Department of Energy puts 2023 at 176 TWh, about 4.4 percent.

How far to trust the estimate

Treat the result as a planning figure. TDP and burst-read currents are ratings few workloads sustain, and most supplies beat the 80 PLUS minimums, so full-load figures lean high. Idle depends on sleep states: TU Dresden researchers measured a two-socket EPYC server at 99.1 W idle and 180.3 W with one core kept out of its deepest sleep. A straight line between idle and full load also overstates middling loads. In a 2025 SPECpower run, Fujitsu's PRIMERGY RX2530 M8 drew 214 W idle, 361 W at 30 percent load and 959 W at full load, where the line gives 437.5 W. The page counts the servers alone; the cloud vs. self-hosted calculator adds a facility's PUE, and the electricity cost calculator prices other loads.

These are planning estimates from datasheet figures and 80 PLUS minimums, not measurements of your hardware. Size circuits, UPS units and cooling with a qualified electrician or engineer and your vendor's figures. See the site disclaimer.

Frequently asked questions

How much power does a server use?

It depends on its parts and on how busy it is. The defaults, two 250 W CPUs, sixteen 64 GB DDR5 sticks, four hard drives and two NVMe SSDs on a pair of 1,100 W Platinum supplies, take 284 W from the wall at idle and 835 W at full load. Spending 30 percent of the time at full load, a server averages 449 W.

How many BTU per hour does a server rack give off?

Every watt a server takes from the wall ends up as heat in the room, so divide the watts by 0.2930711, or multiply by about 3.412. The default rack of four servers draws 3,341 W at full load, which is 11,400 BTU/h, just under one ton of cooling at 12,000 BTU/h a ton.

What size UPS do I need for a server rack?

One whose watt and VA ratings both leave the full load at no more than about 80 percent. The default rack draws 3,341 W at full load, so the UPS needs at least 4,177 W, and with the 0.95 power factor 80 PLUS requires of Platinum supplies, 4,397 VA. The calculator also checks that the UPS covers the brief peak while hard drives spin up.

Is a Titanium power supply worth it?

It saves about 3 percent. Swapping the default rack's Platinum supplies for Titanium cuts its energy from 1,312.4 to 1,274.7 kWh a month, 37.7 kWh, which is 5.65 a month at 0.15 per kWh. Whether that pays for dearer supplies depends on your electricity price and how many years the servers run.

Do redundant power supplies use more electricity?

A little. Two supplies sharing a load each run at a smaller share of their rating, where efficiency is lower. With one 1,100 W Platinum supply per server instead of two, the default rack would use 1,292.5 kWh a month instead of 1,312.4, so redundancy costs 19.9 kWh a month, or 2.98 at 0.15 per kWh.

Why do hard drives need more power when they start?

Spinning the platters up to speed takes a surge of current. Seagate's manual for the 20 TB Exos X20 lists a start current of 2.197 A at 12 V and 0.932 A at 5 V, which is 31.02 W, against 9.30 W while reading. Four drives starting at once add 86.91 W to each default server, so each supply needs at least 856 W rather than 769 W. Starting them one at a time brings that to 791 W.

Sources

1. What do the different PSU (power supply units) ratings mean?, CLEAResult 80 PLUS
2. 80 PLUS program details and history, CLEAResult 80 PLUS
3. 80 PLUS: Market Impacts and Lessons Learned, ACEEE Summer Study on Energy Efficiency in Buildings, 2008
4. ENERGY STAR Program Requirements for Computers, Version 4.0, U.S. Environmental Protection Agency
5. Thermal Design Power (TDP) in Intel Processors, Intel
6. The Case for Energy-Proportional Computing, IEEE Computer, 2007
7. Report to Congress on Server and Data Center Energy Efficiency, U.S. Environmental Protection Agency, ENERGY STAR, 2007
8. DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers, U.S. Department of Energy, 2024
9. Calculating Total Cooling Requirements for Data Centers (White Paper 25), Schneider Electric
10. Watts and Volt-Amps: Powerful Confusion (White Paper 15), Schneider Electric
11. Things to consider in sizing a UPS and load, Schneider Electric
12. Uninterruptible power supply (UPS) FAQs, Vertiv
13. NIST Guide to the SI, Appendix B.8: Factors for units listed alphabetically, National Institute of Standards and Technology
14. Exos X20 SATA Product Manual, Seagate
15. KSM32RD4/32HDR memory module specifications, Kingston Technology
16. DDR5 SDRAM RDIMM Addendum, MTC40F2046S1RC 64GB, Micron Technology (hosted by DigiKey)
17. KIOXIA CD8-R Series product brief, KIOXIA
18. NVIDIA L40S GPU, NVIDIA
19. SPECpower_ssj2008 result: Fujitsu PRIMERGY RX2530 M8, Standard Performance Evaluation Corporation, 2025
20. Energy Efficiency Aspects of the AMD Zen 2 Architecture, TU Dresden, via arXiv, 2021