PUE explained: why the brochure number tells you less than you think

Photo: Jemimus via wikimedia (CC BY 2.0)
1.1. It’s on the brochure, it’s in the tender response, it’s the first number a sales engineer says when you ask how efficient a facility is, and it’s true in roughly the way a car’s advertised fuel economy is true. Power usage effectiveness, PUE, is the most useful single figure the data centre industry has and also the one I’ve watched be stretched furthest in the direction of whoever is quoting it. This is a guide to what it measures and what it doesn’t, from someone who has calculated it on real meters and knows where the bodies are buried.
What the number is #
PUE is a ratio. Take the total energy a facility consumes over a period, and divide it by the energy that reached the IT equipment: the servers, storage and network gear that do the work the building exists for. A PUE of 2.0 means that for every kilowatt-hour the computers used, another kilowatt-hour went to cooling, power conversion losses, lighting and everything else. A PUE of 1.1 means the overhead is ten per cent. A PUE of exactly 1.0 is impossible, since transformers and fans are not free, and anyone quoting below about 1.05 is either running an unusual facility in an unusual climate or has redefined the denominator.
The definition is formalised in an international standard, ISO/IEC 30134-2, and the industry body that popularised the metric, The Green Grid, published guidance on where to take the measurements. That matters because the answer depends enormously on where you put the meter. Measure the IT load at the output of the uninterruptible power supply and you’ve included the UPS losses in the overhead, which is honest. Measure it at the rack power strip and you’ve included the losses in the distribution as well, which is more honest and gives a worse number. Measure it at the UPS input and you’ve hidden the UPS entirely, which gives a flattering number and is, in my experience, more common than it should be.

Annual, design, or best afternoon in July #
The second thing to ask about any PUE claim is the period. A facility in western Sydney has a very different overhead at 3pm on a 41-degree January afternoon than at 3am in July, because the cooling plant is working hard in one case and barely at all in the other. An annualised PUE, calculated from twelve months of real metered data, is the only version that means much. A “design PUE” is a modelled figure for a building that may not be full, may not be commissioned, and may never be operated the way the model assumed. A “partial PUE” for one hall, or a figure quoted for the facility’s best month, is marketing.
When I was commissioning plant the design figure and the first-year annual figure were routinely a tenth apart, and sometimes more, because a half-empty building runs its chillers and fans well below their efficient range. Cooling plant is sized for the full load. At 30 per cent occupancy it’s oversized by a factor of three and it shows in the ratio. This is why a brand-new facility’s PUE usually gets better over its first few years as tenants move in, and why the claim on the day it opens should be read as a promise rather than a measurement.
Climate does a lot of the work #
A facility in Canberra, Hobart or Christchurch can use outside air for free cooling for a large part of the year and post a genuinely low annual PUE with no particular cleverness. The same design in Darwin cannot. Comparing a Melbourne PUE with a Singapore PUE tells you more about the weather than the engineering, and the honest operators say so. I’d argue this is the single most misunderstood thing about the metric among the people who buy on it: they’re comparing climates and think they’re comparing competence.
What PUE doesn’t count #
Four things, mainly.
It doesn’t count the efficiency of the IT equipment itself. A hall full of ten-year-old servers idling at 15 per cent utilisation can sit inside a building with a superb PUE. The building is efficient; the computing isn’t. The ratio rewards the landlord and says nothing about the tenant, which is why a hyperscaler’s reported PUE and a colocation operator’s are measuring different things even when the numbers match.
It doesn’t count water. An evaporative cooling system can produce an excellent PUE by trading electricity for water, and in a dry Australian summer that trade has a cost the ratio never shows. The companion metric is water usage effectiveness, litres per kilowatt-hour of IT energy, and I’d like to see it quoted alongside PUE far more often than it is. It rarely is, because the number is often unflattering.
It doesn’t count where the electricity came from. A PUE of 1.2 on a coal-heavy grid and a PUE of 1.3 on a facility contracted entirely to wind and solar are not equivalent, and carbon usage effectiveness exists to make the point. Australia’s grid mix varies by state and by hour, so a facility’s real emissions depend on when it’s busy as much as where it is.
And it doesn’t count the load that liquid cooling moves out of the building’s accounting. When a rack’s heat is carried away by a coolant loop and rejected through a dry cooler, the pumps and the cooler are part of the overhead. If those are metered on the tenant’s side of the line, they vanish from the landlord’s PUE. This is a new problem, it’s growing with every AI hall, and the standards bodies are still arguing about the right boundary.
What a good claim looks like #
The version I trust is annual, metered, states where the meters are, states the occupancy over the period and is independently rated. In Australia that last part means the NABERS Energy rating for data centres, which is administered by the NSW government on behalf of all states, requires a year of real data and an accredited assessor, and produces a star rating you can check on a public register. A facility that has a NABERS rating and quotes a PUE has, at minimum, been audited. One that quotes a PUE and doesn’t have a rating may be equally good, but you’re taking its word for it.
The federal energy department and the Clean Energy Regulator both publish material on facility-level reporting, and large operators report their energy use under the National Greenhouse and Energy Reporting scheme, which is another place to cross-check a claim. It takes an afternoon. If you’re signing a ten-year lease it’s an afternoon well spent.
A worked example from a real building #
Numbers help. Take a facility with a 10-megawatt IT capacity that is 40 per cent occupied, so the IT load is 4 megawatts averaged over the year. Its chillers, pumps, fans, UPS losses, lighting and office load add up to 1.4 megawatts averaged across the same year. Total facility load is 5.4 megawatts. PUE is 5.4 divided by 4, which is 1.35. Not bad, not a brochure number.
Now fill the building. IT load rises to 9 megawatts. The cooling plant, which was sized for 10, runs closer to its efficient range, and the fixed overheads (lighting, the office, the security system, the losses in an oversized transformer) don’t grow at all. Overhead rises to perhaps 2.1 megawatts. Total is 11.1. PUE is 1.23. Same building, same engineering, a tenth better on the ratio because more tenants moved in. That’s the occupancy effect, and it’s why the first-year number and the fifth-year number are different documents.
Now move the meter. If the operator measured IT load at the UPS input instead of its output, the UPS losses, call them 3 per cent of the load, move from the numerator to the denominator. The ratio drops by a few hundredths with no change to a single piece of equipment. Do the same trick with the distribution losses and the lighting in the halls and you can shave another few hundredths. None of it is lying, exactly. All of it is why “where’s the meter” is the first question.
Why I still like the metric #
For all the above, PUE did something the industry badly needed, which was to make cooling and power overhead visible and comparable. When I started, a PUE of 2.0 was ordinary and nobody thought about it. The number gave engineers a target, gave tenants a question to ask, and pushed a generation of facilities from 2.0 to 1.3 in about a decade. That’s a real achievement and the metric deserves the credit.
What it doesn’t deserve is to be the only number on the brochure. Ask for the water figure. Ask for the meter locations. Ask what the occupancy was. And if the person across the table says “design PUE” with a straight face, ask them what it was last February. I was up at five this morning fixing a puncture, so I’m in no mood to be polite about it; the annual figure is the figure, and the rest is weather.
— Declan Ferris, Infrastructure Correspondent, Technology Spectator