ENERGY STAR Score for Data Centers in the United States - Technical Reference (November 2014)

Overview

The document sets out the statistical basis of the ENERGY STAR score for spaces designed and equipped for high-density computing. The goal is a fair comparison of a property's energy performance with its peers, taking climate and business activity into account. The model predicts the energy use expected from a building's actual activity, compares it with measured use and yields a 1 to 100 percentile rank against the national population. The score was released in June 2010 and applies to individual buildings only, never to campuses.

The reference data come from an EPA survey run with industry bodies including Uptime Institute, Green Grid, 7x24 Exchange and AFCOM, because no sufficiently robust public dataset existed. Of 120 facilities with complete data, 108 metered IT energy at the UPS output, and 61 of those were stand-alone data centers. Stand-alone sites gave more robust data and more significant regression equations, so the analysis uses only those 61 properties.

The dependent variable is PUE in source kBtu: total energy divided by IT energy. Total energy covers every fuel, including chilled water and natural gas, not just electricity. Source energy use intensity was rejected because it varies widely across data centers and ignores IT equipment density. IT energy must be measured at the UPS output; metering at the PDU or nearer the racks would be more accurate, but it is not yet common enough, so a metric most operators can supply was chosen.

Candidate variables included building and data center floor area, Tier level, rack count, UPS utilisation, annual IT energy, building type, data center type, and heating and cooling degree days. Only annual IT energy proved significant: larger IT loads show lower average PUE, but the economy of scale levels off beyond 0.4 source TBtu. Climate had no significant effect because cooling is dominated by the internal IT load. Tier was excluded as not significant, hard to define, and likely to reward needlessly high Tier levels.

The report provides the regression results, the full score lookup table from 1 to 100 and a worked calculation. Its limits are plain: the model rests on 2010 survey data, the sample is small, and the site-to-source factor for electricity may have changed since. Its value to a designer lies in translating a target PUE into an expected score and in showing that the metering architecture, with IT energy measured at the UPS output, must be planned from the start.

Key points

1. The score was first released in June 2010 and covers individual buildings only; co-located data centers can each score, but a campus cannot.
2. After filtering, the regression dataset holds 61 stand-alone data centers, from 120 properties with complete data and 108 with IT energy metered at the UPS output.
3. PUE is defined as total energy divided by IT energy, both in source kBtu, with IT energy measured at the UPS output.
4. The prediction equation has a constant of 1.924 and a coefficient of -0.9506 on centred annual IT energy in TBtu; the mean is 0.2091 TBtu.
5. The IT energy adjustment is capped at 0.4 source TBtu; larger sites receive the same adjustment.
6. The equation explains 11.38% of PUE variance (R2 = 0.1138) but 92.2% of the variation in data center source energy.
7. No statistically significant relationship was found between heating or cooling degree days and PUE, nor between Tier level and PUE.
8. The worked example uses a site-to-source multiplier of 3.14 for electricity and 1.05 for natural gas; 15,000,000 kWh total with 8,500,000 kWh IT gives an actual source PUE of 1.788.
9. An efficiency ratio (actual PUE / predicted PUE) of 0.8781 earns a score of 71; ratios above 1.3315 score 1.
10. The data date from a 2010 survey and site-to-source factors may have changed, so treat the results as a method reference rather than a current benchmark.

Using it in design

• Design sub-metering for IT energy at the UPS output and feed it into monitoring with monthly records; this is a precondition for the score.
• Use the prediction equation and lookup table to see what score a target PUE would earn; for example, a source PUE about 12% below prediction corresponds to a score near 71.
• Meter non-electric energy inputs such as chilled water and natural gas as well; they count towards total energy.
• Set redundancy level by business need, knowing that a higher Tier gives no scoring advantage and the score does not normalise for it.
• When comparing small and large sites, keep source-energy IT load as the common denominator and remember the scale effect stops at 0.4 TBtu.

Who should read it

Designers, operations teams and sustainability staff who want to benchmark data center energy performance, set up metering or translate a PUE target into a score.

Document contents

1. Overview
2. Reference Data & Filters
3. Variables Analyzed
4. Dependent Variable
5. Independent Variables
6. IT Energy Analysis
7. Testing and Variables not Correlated with PUE
8. Regression Equation Results
9. ENERGY STAR Score Lookup Table
10. Example Calculation

Read the document

Publication details

Authors

U.S. EPA ENERGY STAR

Publisher

U.S. EPA / ENERGY STAR

Year

2014

Document language

English

Source type

Government / public

Access

Open licence — downloadable

Licence

U.S. public domain — U.S. Government work; free to reuse

Licence evidence

www.copyright.gov

Citation