Temperature Optimisation in Data Centres

Raising server inlet temperatures tends to increase the server power consumption due to heightened server fan activity needed to compensate for the warmer air, and to decrease cooling infrastructure power consumption due to less intense and possibly less frequent chiller activity. Relying on primary data from two data centres of a colocation provider in Switzerland, this study succeeds in confirming and quantifying the first effect: In the upper half of the ASHRAE recommended temperature range and slightly above it, i.e., between 23 - 30 {\deg}C, the server power consumption correlated positively with the temperature, with a server room-wide temperature sensitivity of 0.35 - 0.5 %/{\deg}C. Based on the available data and due to ambivalent results, the study did not succeed in answering which effect is stronger for the entire data centre: the increase in server consumption or the savings in cooling consumption. Whether raising inlet temperatures beyond the ASHRAE upper limit (i.e., 27 {\deg}C) yields additional overall benefits thus remains an open question that would be best addressed through controlled experiments. A few pieces of evidence indicate nevertheless that the upper part of the current ASHRAE recommendation (i.e., around 25 - 27 {\deg}C) might be the current sweet spot for colocation DCs. The widely used power usage effectiveness (PUE) metric is not at all useful in this context. For practical reasons but semantically wrong, the server fan power consumption is included in the consumption of servers (and thus, of the IT infrastructure) and is thus on the wrong side of the PUE fraction. An inlet temperature raise, which will lead to increased fan consumption, will thus always yield a PUE decrease, irrespective of what happens to the overall data centre power consumption.

💡 Research Summary

The paper investigates the trade‑off between raising server inlet temperatures and overall data‑centre power consumption, using real‑world measurements from two Swiss colocation facilities. The authors begin by outlining the conventional view that higher inlet temperatures increase server fan speeds, thereby raising server power draw, while simultaneously reducing the load on cooling infrastructure such as chillers and air‑handling units. They note that the industry standard metric, Power Usage Effectiveness (PUE), can be misleading in this context because fan power is usually counted as part of the “IT” load; consequently, any temperature‑induced increase in fan power will artificially lower PUE even if total facility power does not improve.

Data were collected over several months to years, including timestamps for inlet temperature, server power consumption, cooling‑system power, and ambient weather conditions. After cleaning the data and aligning the time series, the authors performed correlation and regression analyses to quantify the relationship between temperature and power. Their key finding is that, in the temperature band from 23 °C to 30 °C—covering the upper half of the ASHRAE recommended range and extending slightly beyond—the server‑side power consumption rises linearly with temperature at a rate of 0.35 % to 0.5 % per degree Celsius. This increase is attributed primarily to higher fan speeds inside the servers, which consume additional electricity.

The cooling‑side response, however, proved ambiguous. In some periods the higher inlet temperature corresponded with a measurable drop in chiller and air‑handling power, but in other periods the reduction was negligible or even reversed, likely due to seasonal weather swings, varying load profiles, and differing operational strategies of the cooling plant. Because the dataset did not provide a consistent cooling‑power trend, the authors could not definitively state whether the cooling‑savings outweigh the server‑fan penalty for the whole data centre.

The authors discuss the practical implications of these results. First, the observed temperature sensitivity suggests that operating at the upper end of the ASHRAE range (around 25 °C–27 °C) may be the “sweet spot” for colocation data centres: fan power increases are modest, while cooling demand is already reduced compared with lower temperatures. Second, the analysis highlights a fundamental limitation of PUE: because fan power is placed on the IT side of the ratio, any temperature‑driven fan increase will make PUE appear better, even if total power consumption stays the same or rises. The paper therefore recommends separating fan power from core IT power when evaluating efficiency.

Finally, the study acknowledges its limitations. The lack of controlled experiments means that causality cannot be fully established, and the data do not capture long‑term reliability or hardware‑lifetime effects of operating at higher temperatures. The authors call for future work that includes controlled temperature ramps, broader climate zones, and explicit measurement of hardware degradation. Such studies would clarify whether pushing inlet temperatures beyond the ASHRAE upper limit (27 °C) can deliver net energy savings without compromising equipment reliability.

In summary, the paper confirms a positive correlation between inlet temperature and server power (0.35 %–0.5 % per °C) in the 23 °C–30 °C range, but it leaves the overall data‑centre energy balance unresolved due to inconsistent cooling‑side data. It cautions against relying solely on PUE for decision‑making and suggests that the current ASHRAE upper range may already represent an optimal operating point for many colocation facilities.