White Paper: Sustainability Metrics Considerations - PUE, Revision 1.0 (October 2024)

Overview

The executive summary notes that PUE is widely used to report data-centre infrastructure efficiency, but variability in infrastructure, measurement and reporting makes PUE figures hard to compare. The global annual average PUE reported for 2022 was 1.55, yet many data centres already run below 1.10 (over 90% infrastructure efficiency); the PUE number alone does not show what to change next. The authors are from Intel, with reviewers from Microsoft and the OCP Foundation. Part of the motivation is preparing for upcoming data-centre efficiency regulation.

On history and definition, PUE was defined in 2007 by The Green Grid and adopted as ISO/IEC 30134-2:2016. By measurement point it has categories PUE1 (UPS output, basic resolution), PUE2 (PDU output, intermediate) and PUE3 (IT equipment input, advanced). PUE is total data-centre energy divided by total IT energy, that is 1 plus non-IT energy over IT energy. Partial PUE (pPUE) is defined per zone, and the site PUE can be built from zone values; the two routes must agree.

The measurement section separates static and dynamic PUE. Static PUE (sPUE) is computed from demand, used in design, and can be evaluated at maximum (max TDP), idle and optimal load points. Dynamic PUE is measured consumption and, when PUE is quoted without a prefix or suffix, dynamic is assumed (for example PUE@wue). The paper shows PUE varies with IT load: values measured at low load (<10%) and high load (>50%) differ, so without load factor and sampling period (15 minutes up to annual) comparisons are difficult.

The core argument sits in the power-delivery and cooling sections. In a traditional rack each server's PSU and DC-DC converters create losses; in a DC busbar rack the PSU loss is shared. ORV1 supports 12 V, ORV2 12 V and 48 V, ORV3 48 V; a 12 V design means higher current and IR loss for the same copper. Yet all of these losses count as IT energy. In cooling, server fans use about 10% of server power in traditional air, rear-door HX and in-row designs, under 5% with DTC (fans below 50% duty) and none in immersion, yet fans and rack CDUs are counted as IT energy.

The paper also covers climate and water: identical data centres running the same workloads in Ireland and Singapore run at different PUE, so PUE should be tagged with location, WUE and load (PUE@geo, @wue, @load). Colder water can improve PUE but may be less sustainable overall. Planned future work includes better classification of IT and non-IT energy buckets, new metric definitions, measurement and modelling frameworks, an OCP sustainability profile, a possible performance-per-watt metric (for example SERT) and a GWP/CUE review. The paper is licensed CC BY-SA 4.0.

Key points

1. PUE is ISO/IEC 30134-2:2016, with categories PUE1 (UPS output), PUE2 (PDU output) and PUE3 (IT equipment input).
2. The global annual average PUE reported for 2022 was 1.55, while many sites already operate below 1.10.
3. Partial PUE (pPUE) is defined per zone and the site PUE can be built from the zone values.
4. Static PUE is computed from demand at defined load points (max TDP, idle, optimal); dynamic PUE is measured and is assumed when PUE is quoted without qualifiers.
5. PUE at low load (<10%) and high load (>50%) differs, so values quoted without load factor and sampling period are hard to compare.
6. Server fans take about 10% of server power in air, rear-door HX and in-row designs, under 5% with DTC and zero in immersion.
7. In-server fans, rack CDUs, immersion pumps and PSU/DC-DC conversion losses are counted as IT energy in conventional PUE.
8. A 12 V busbar design means higher current and IR losses than 48 V for the same power (ORV3 is 48 V).
9. Identical sites in Ireland and Singapore give different PUE; PUE should be tagged with location, WUE and load (@geo, @wue, @load).
10. Colder water can improve PUE yet be less sustainable; PUE should be read together with WUE.

Using it in design

• The energy model should report PUE with its category (1/2/3), load point and sampling period.
• Support zone-level partial PUE for mixed air/liquid halls.
• Show an adjusted metric that moves server fan and rack-CDU power from IT to facility overhead so air-cooled and liquid-cooled options compare fairly (for example about 10% fan power for air versus under 5% for DLC).
• Warn users that part of a PUE gain from moving to DLC reflects lower in-IT fan energy.
• Report PUE tagged with location (climate data), WUE and load factor, and evaluate chilled-water temperature choices together with WUE.
• Compute design-stage sPUE for max TDP, idle and optimal load points.

Who should read it

Data-centre engineers and sustainability leads who report PUE, compare cooling architectures or prepare for efficiency regulation.

Document contents

1. Executive Summary
2. List of Abbreviations
3. Introduction
4. Power Usage Effectiveness (PUE)
5. PUE History and Definition
6. Partial PUE
7. PUE Measurement
8. Static PUE
9. Dynamic PUE
10. IT Load Variability
11. Sampling period
12. Measurement and Reporting
13. Power Delivery Architecture
14. Data Center Power Converters
15. Traditional Rack Server
16. DC Bus Bar
17. DC Cooling Architecture
18. Traditional Cooling
19. DTC and Hex
20. Immersion Cooling
21. Emerging Cooling and its impact on PUE summary
22. IT Load Consideration
23. Geo Impact
24. Water Impact
25. Future work (Outlook for 2024-2025)
26. Conclusion

Related standards

• ISO/IEC 30134-2:2016
• EN 50600-4-2
• The Green Grid PUE white papers
• OCP DCF Sustainability Metrics

Read the document

Publication details

Authors

Nachimuthu, M.; Kumar, M.J.; Dahlen, E.J. (Intel)

Publisher

Open Compute Project Foundation

Year

2024

Document number

Rev 1.0

Document language

English

Source type

Industry body

Access

Open licence — downloadable

Licence

CC BY-SA 4.0 — share with attribution under the same licence

Licence evidence

www.opencompute.org

Citation