Efficiency metrics: PUE, WUE, CUE and beyond

What gets measured gets managed, and data centre sustainability depends on good metrics. Power usage effectiveness (PUE) became the industry's best-known number in the late 2000s and drove a dramatic fall in infrastructure overhead. It was soon joined by metrics for water, carbon, renewable energy and energy reuse, now standardized in the ISO/IEC 30134 series. Each metric answers a specific question, and each can mislead if it is measured carelessly or used for something it was not designed to show. This lesson defines the main metrics, explains how to measure them properly, shows how they behave with load and weather, and discusses their limits.

Power usage effectiveness

PUE, defined in ISO/IEC 30134-2 (and originally by The Green Grid), is the ratio of the total energy used by the data centre to the energy used by the IT equipment, over a period:

A PUE of 1.0 would mean no overhead at all; a PUE of 2.0 means that for every kilowatt-hour reaching the IT equipment, another is spent on cooling, power conversion, lighting and other support. Typical values are about 1.1 to 1.2 for the best hyperscale facilities, 1.3 to 1.5 for good modern colocation sites, and 1.6 to 2.0 or more for older enterprise data centres and server rooms. Surveys by the Uptime Institute have reported industry averages around 1.5 to 1.6 for several years, reflecting the large stock of older facilities.

PUE must be calculated from energy (kWh) over a full year, not from instantaneous power, because cooling energy varies strongly with the seasons. The standard defines measurement categories that differ in where IT energy is measured: at the UPS output (the simplest, category 1), at the PDU output (category 2) or at the IT equipment input (category 3, the most accurate). Losses in PDUs and cabling count as IT energy in category 1 but as overhead in category 3, so PUE values should always be reported with their category and period.

How PUE behaves

Two factors drive PUE through the year. Weather changes cooling energy: in summer, chillers work harder and economizers run less, raising PUE; in winter, free cooling lowers it. Load also matters: much infrastructure overhead is fixed (transformer no-load losses, UPS losses at light load, minimum fan speeds, lighting, offices), so a lightly loaded facility has a higher PUE than the same facility at full load. New data centres often report poor PUE in their first years simply because they are partly empty.

import numpy as np

# Monthly metered data for one facility (MWh)
it = np.array([3100, 2850, 3150, 3080, 3220, 3200, 3350, 3380, 3250, 3260, 3150, 3240])
facility = np.array([3720, 3390, 3810, 3820, 4090, 4290, 4660, 4700, 4260, 4010, 3780, 3860])
monthly_pue = facility / it
print("Monthly PUE:", " ".join(f"{p:.2f}" for p in monthly_pue))
print(f"Best month {monthly_pue.min():.2f}, worst month {monthly_pue.max():.2f}")
print(f"Average of monthly PUEs: {monthly_pue.mean():.3f}")
print(f"Annual PUE (energy-weighted): {facility.sum() / it.sum():.3f}")

# Other ISO/IEC 30134 metrics for the year
water_m3, grid_g_per_kWh, reused_MWh, renewable_MWh = 52000, 120, 2400, 30000
wue = water_m3 * 1000 / (it.sum() * 1000)          # L per kWh of IT energy
cue = facility.sum() * grid_g_per_kWh / 1000 / it.sum()   # kg CO2 per kWh of IT energy
erf = reused_MWh / facility.sum()
ref = renewable_MWh / facility.sum()
print(f"WUE {wue:.2f} L/kWh, CUE {cue:.3f} kgCO2/kWh, ERF {erf:.3f}, REF {ref:.2f}")
print(f"Energy reuse effectiveness ERE = {(facility.sum() - reused_MWh) / it.sum():.3f}")

Monthly PUE: 1.20 1.19 1.21 1.24 1.27 1.34 1.39 1.39 1.31 1.23 1.20 1.19
Best month 1.19, worst month 1.39
Average of monthly PUEs: 1.264
Annual PUE (energy-weighted): 1.266
WUE 1.36 L/kWh, CUE 0.152 kgCO2/kWh, ERF 0.050, REF 0.62
Energy reuse effectiveness ERE = 1.203

In this example, monthly PUE ranges from 1.19 in winter to 1.39 in summer. Reporting the best month (1.19) would flatter the facility; the honest annual figure, calculated from total energy, is 1.27. The average of monthly PUEs happens to be close, but in general the annual value must be computed from the energy totals, because months with more IT energy should weigh more.

Water usage effectiveness

Water usage effectiveness (WUE, ISO/IEC 30134-9) relates the water used by the data centre to its IT energy, in litres per kilowatt-hour:

Site WUE counts water used on site, mainly for evaporative cooling and humidification. A source-inclusive variant adds the water consumed in generating the electricity the facility buys, which can be substantial for thermal power plants. Facilities with evaporative cooling commonly have site WUE around 1 to 2 L/kWh, while air-cooled or closed-loop facilities can be close to zero. WUE should be read together with local water stress (Lesson 6): a litre in a desert matters more than a litre beside a large lake.

Carbon usage effectiveness

Carbon usage effectiveness (CUE, ISO/IEC 30134-8) expresses the greenhouse gas emissions from the facility's energy use per unit of IT energy, in kilograms of CO₂-equivalent per kilowatt-hour:

CUE combines efficiency (PUE) with the carbon intensity of the energy (Lesson 7). The same facility has very different CUE on different grids: in the example, a grid intensity of 120 g/kWh and PUE 1.27 give a CUE of about 0.15 kg/kWh. Whether market-based instruments such as renewable energy certificates may reduce reported CUE is a matter of accounting rules (Lesson 7), and CUE values should state which method is used.

Renewable energy and energy reuse

Two further metrics capture other parts of the picture. The renewable energy factor (REF, ISO/IEC 30134-3) is the share of total facility energy that comes from renewable sources, whether generated on site or procured through contracts that meet the standard's criteria. The energy reuse factor (ERF, ISO/IEC 30134-6) is the share of total facility energy that is reused outside the data centre, for example as heat delivered to a district heating network (Lesson 9). A related metric, energy reuse effectiveness (ERE), subtracts reused energy from total energy before dividing by IT energy, so that heat reuse can bring ERE below 1.0, unlike PUE.

IT efficiency metrics

PUE says nothing about whether the IT equipment does useful work efficiently. A facility full of idle servers can have an excellent PUE. Metrics that address IT efficiency include IT equipment energy efficiency (ITEE, the rated capacity of servers relative to their rated power) and IT equipment utilization (ITEU), both in the ISO/IEC 30134 series, and server benchmarks such as SPECpower, which measure performance per watt at different loads. For AI workloads, useful measures include tokens or inferences per joule and training energy per model. Lesson 5 develops IT efficiency in detail; the point here is that infrastructure metrics and IT metrics answer different questions, and both are needed.

Measurement and its pitfalls

Metrics are only as good as the metering behind them. Sound practice requires revenue-grade or calibrated meters at the utility intake, at the boundaries between IT and non-IT loads, and at the levels the chosen category requires; continuous logging at intervals of 15 minutes or less; documented boundaries (for example, whether office space and shared building systems count); and a full year of data. Common pitfalls include reporting design PUE instead of measured PUE, quoting the best month or best hall, excluding loads such as generator block heaters or office HVAC, and comparing facilities in very different climates or at different loads without context. Independent verification, increasingly required by regulation (Lesson 11), reduces these problems.

Partial PUE and subsystem metrics

PUE describes a whole facility, but managers often need to know which subsystem is responsible for the overhead. Partial PUE (pPUE) applies the same ratio to a zone or subsystem: for example, the energy used by a data hall's cooling plus its IT energy, divided by its IT energy, gives a cooling pPUE for that hall. Similar ratios for the electrical chain, lighting and support systems let the overhead be broken down: in a facility with PUE 1.30, perhaps 0.18 comes from cooling, 0.09 from electrical losses and 0.03 from other loads. Tracking these components over time shows where improvement efforts have worked and where they are needed, and makes PUE changes explainable rather than mysterious.

Benchmarking and normalization

Comparing facilities is useful but must be done carefully. Climate affects PUE and WUE through free cooling hours and evaporation; load affects PUE through fixed overheads; redundancy level affects electrical losses; and the IT mix (air-cooled general computing or liquid-cooled AI) affects cooling energy. Fair benchmarking compares facilities in similar climates and conditions, or adjusts for them, for example by modelling the expected PUE of a reference design in each climate. Regulators preparing minimum performance standards and rating labels face exactly this problem, which is why they often grade facilities relative to climate zone and size.

Internally, the most useful benchmark is often a facility's own history: the same site's PUE, WUE and pPUE, normalized for load and weather, tracked month by month after each change.