Ask most people what a data centre sounds like and they picture server fans. Noizend's briefing Data Centre Noise: Physics, Infrastructure and Policy makes the case that the real acoustic load sits elsewhere, in the mechanical and electrical plant that keeps the servers alive, none of which switches off at night the way a factory or a construction site does.
A note on the numbers below: every level in the report is an unweighted-equivalent figure, converted from A-weighted source data. That conversion is deliberate. A-weighting is built around human sensitivity at conversational frequencies and discounts the low end heavily, which is precisely where this equipment is loudest. Reported the conventional way, the problem partly disappears into the metric.
Cooling is the largest contributor, and the choice of technology is a direct trade between water, energy and noise.
Transformers hum at 100 Hz and its harmonics through magnetostriction in the core, a physical property of the steel rather than a defect that can be engineered out without redesigning the transformer, and a tone that sits where barriers work least well.
Backup power is where the cleaner framing of the industry's move away from diesel deserves closer scrutiny. Diesel generators in the two-megawatt class are commonly rated around 99 dB at seven metres under load, and sites deploy them by the dozen. Amazon's Ashburn campus runs 101 units at roughly 3 MW each, more than 300 MW of standby diesel behind one facility.
Battery storage does not remove that noise so much as change its character. Component levels at one metre run to roughly 110 dB for thermal management, 103 dB for the power conversion system and 93 dB for transformers, and because those sources are tonal and low-frequency, assessments routinely add a penalty on top of the measured level.
The report's central warning is about load. Facilities are approved on models that assume equipment runs at a fraction of its real load, then open louder than promised, a pattern traced through Lane Cove in NSW, Hazelmere in Western Australia and West Footscray in Victoria.
Its recommendations follow from that: cumulative precinct-level assessment rather than project-by-project modelling, operational verification against approved models at each construction phase, and unweighted (Z-weighted) and C-weighted reporting alongside A-weighted figures, so low-frequency content cannot be modelled out of view.
The full report is available to download above.
07
Whether you operate a business or just want some peace and quiet in your home, let's talk about noise.
