Temperature Optimisation in Data Centres (TODC) - Final report

Overview

This study examines how raising server inlet temperature changes server power, mostly fan power, and whether it saves energy overall. It is based on primary data from two Swiss colocation data centres that raised their inlet setpoints, together with a literature review and expert interviews. It was funded by the Swiss Federal Office of Energy (SFOE) and is an open-access preprint, arXiv:2509.26163.

Server-room power consistently rose with inlet temperature in the 23-30 C range, at roughly 0.35-0.57% per C depending on the site and analysis. In data centre 1, 65 temperature changes and 325 windowed analyses gave a mean sensitivity of 0.57% per C (95% CI ±0.11%/C), and the sensitivity hardly depended on the starting temperature over 23-29 C. In data centre 2 (1-minute data, 7 temperature changes), the mean sensitivity was 0.34% per C.

The basic limit is that server airflow increases linearly with fan speed while fan power scales with the cube of fan speed. In the cited RISE experiment, relative to a 22 C inlet, raising to 28 C increased total energy by 0.9-1.4% and lowering to 16 C cut it by 0.6-1.2%, depending on load. Reported hyperscale inlet temperatures are Google 26.6 C, Microsoft 27 C and Meta 29.4 C.

Building-level data were too confounded by changing compute load to show whether cooling savings outweigh the extra fan power. The author argues that PUE is the wrong metric for this trade-off because fan power sits in the IT denominator: raising inlet temperature always lowers PUE, since cooling power falls and fan power is counted as IT load, so a lower PUE does not show that total energy fell. The author suggests 25-27 C as the current practical sweet spot for colocation.

Key points

1. Between 23 and 30 C, room-wide server power rose by 0.35-0.5 percentage points per degree C.
2. In data centre 1, 65 temperature changes and 325 windowed analyses gave a mean sensitivity of 0.57% per C, hardly dependent on the starting temperature over 23-29 C (R2 = 0.005).
3. In data centre 2 (1-minute data, 7 changes), the mean sensitivity was 0.34% per C, and every 1-hour and 2-hour window correlation was positive.
4. Server airflow increases linearly with fan speed but fan power scales with the cube of fan speed, which sets the basic limit on raising inlet temperature.
5. RISE experiment: relative to 22 C inlet, 28 C increased total energy by 0.9-1.4% and 16 C cut it by 0.6-1.2%, depending on load.
6. Reported hyperscale inlet temperatures: Google 26.6 C, Microsoft 27 C, Meta 29.4 C.
7. Raising inlet temperature always lowers PUE, because fan power is counted as IT load, so a lower PUE does not show that total energy fell.

Using it in design

• Add an IT-side penalty of about 0.3-0.6% of IT power per C of inlet rise above about 23 C (a configurable default) to the setpoint optimiser and energy calculator, and model fan power with the cube of speed.
• Show total facility energy alongside PUE and warn that a lower PUE from higher setpoints may not mean lower total energy.
• Use 25-27 C as the default recommended inlet setpoint band, with an allowable-range warning above 27 C.
• Add a server power versus inlet temperature curve as a catalogue field.

Who should read it

Cooling and energy engineers who set inlet temperature setpoints in colocation and enterprise data centres.

Related standards

• ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments
• ISO/IEC 30134-2 (PUE)

Publication details

Authors

Vlad C. Coroamă

Publisher

Swiss Federal Office of Energy SFOE (Energy Research and Cleantech); preprint on arXiv

Year

2025

Document number

arXiv:2509.26163

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2509.26163

Licence

CC BY-NC-ND 4.0 — share unaltered with attribution for noncommercial purposes

Citation