United States Data Center Energy Usage Report (LBNL-1005775, June 2016)

Overview

The report rebuilds historical data center electricity use back to 2000 from earlier studies and shipment data, then forecasts to 2020. U.S. data centers used an estimated 70 billion kWh in 2014, about 1.8% of national consumption. Growth was nearly 90% over 2000-2005, about 24% over 2005-2010 and about 4% over 2010-2014, and the report projects another 4% by 2020, to about 73 billion kWh. Earlier reference points are 61 billion kWh in 2006 (1.5%) and 69 billion kWh for 2008 (1.8%). Its method models servers, storage, network equipment and infrastructure by data center space type, using IDC shipment trackers and SPEC and SERT benchmark databases.

The slowdown is explained by supply-side trends. Server shipments grew 15% a year over 2000-2005, 5% over 2005-2010 and 3% after 2010, with nearly all post-2010 growth going to hyperscale data centers that run at high utilization and need fewer servers for the same service. Server power demand rose until 2005 and has been roughly constant since, while power scaling at low utilization improved. Storage drives became more efficient per drive and network port power estimates fell. A counterfactual that freezes utilization, power scaling, disk and port power and PUE at 2010 levels shows more than 600 billion kWh would have been needed across the decade.

Efficiency scenarios for 2020 are defined as improved management, best practices and hyperscale shift, plus two combinations. Relative to current trends they cut server energy by 4%, 18% and 28% respectively, and infrastructure energy by 22%, 64% and 46%. Storage energy is lower by 25% and network energy by 38% only in the best-practices case. Together the scenarios save up to 33 billion kWh a year in 2020, about 45% below current trends. Infrastructure energy is PUE multiplied by IT energy, so IT reductions and PUE gains compound. The assumed 2014 PUE values run from 2.5 for server rooms down to 1.2 for hyperscale.

Further chapters cover federal government servers under the Federal Data Center Consolidation Initiative, indirect energy impacts of ICT, and water: electricity generation consumes on average about 7.6 L (2.0 gal) of water per kWh delivered, counted as indirect footprint. The future-work chapter says the data are limited and that PUE reflects only building infrastructure, not IT efficiency, naming alternatives such as DCeP and CADE. It warns that PUE gains, utilization and power proportionality have limits, so the plateau may not last. The study is dated and its projections end at 2020.

Key points

1. U.S. data centers used an estimated 70 billion kWh in 2014, about 1.8% of national electricity use, with about 73 billion kWh projected for 2020.
2. Electricity use grew about 4% over 2010-2014, versus about 24% over 2005-2010 and nearly 90% over 2000-2005.
3. Server shipment growth fell from 15% a year (2000-2005) to 5% (2005-2010) and 3% after 2010, nearly all of it into hyperscale facilities.
4. Assumed 2014 PUE by space type: 2.0 closets, 2.5 server rooms, 2.0 localized, 1.9 mid-tier, 1.7 high-end, 1.2 hyperscale.
5. In the 2014 hyperscale PUE breakdown, cooling adds 0.16, transformers 0.02 and lighting 0.02 per unit of IT energy; for server rooms cooling alone adds 1.23.
6. Combined efficiency scenarios could cut 2020 electricity use by up to 33 billion kWh a year, about 45% below current trends.
7. Efficiency improvements since 2010 are estimated to save 620 billion kWh over 2010-2020 compared with 2010 efficiency levels.
8. Best practices cut infrastructure energy by 64% and, with hyperscale shift, by 75%, relative to current trends.
9. Electricity generation consumes on average about 7.6 L (2.0 gal) of water per kWh delivered in the U.S., counted in the indirect water footprint.
10. PUE measures only the building infrastructure, so the report points to productivity metrics such as DCeP and CADE.

Using it in design

• Use the space-type PUE table as a sanity check on early-stage PUE assumptions, remembering these are 2014 national averages, not design targets.
• Break PUE into cooling, power distribution and lighting contributions the way the report does when comparing design options.
• Compare a planned facility's PUE against the hyperscale 1.2 and server room 2.5 values to show the efficiency gap when justifying consolidation or an efficient design.
• Include indirect water of grid electricity (about 7.6 L/kWh average) alongside on-site water when comparing cooling approaches.
• Prefer the 2024 LBNL report for current planning and use this one for historical trend context.

Who should read it

Planners, policy analysts and design engineers who need historical U.S. data center energy trends and benchmark PUE by facility type.

Document contents

1. Executive Summary
2. 1 Introduction
3. 1.1 Report Scope
4. 1.2 Report Organization
5. 2 Estimates of U.S. Server and Data Center Energy Use
6. 2.1 Methodology Overview
7. 2.2 Data
8. 2.3 IT and Infrastructure Equipment Estimates
9. 3 Energy Use Associated with Federal Government Servers and Data Centers
10. 4 Expected Energy Savings Opportunities
11. 4.2 Improved Management scenario
12. 4.3 Best practices scenario
13. 4.4 Hyperscale Shift scenario
14. 5 Indirect energy impacts
15. 6 Future Work

Publication details

Authors

Arman Shehabi; Sarah Smith; Dale Sartor; Richard Brown; Magnus Herrlin; Jonathan Koomey; Eric Masanet; Nathaniel Horner; Inês Azevedo; William Lintner

Publisher

Lawrence Berkeley National Laboratory

Year

2016

Document language

English

Source type

Government / public

Access

Citation only

Licence

Licence conditions not established — confirm the conditions with the publisher

Licence evidence

datacenters.lbl.gov

Citation