Google’s efficient datacenter design [/ˈguːgle’s ɪˈfɪʃənt ˈdatasenter dɪˈzaɪn/] n - Google has published efficiency data since 2008, and the worth of their labor lies chiefly in its being public. Their datacenters are not perfect; they are something rarer, a consistent and measurable benchmark against which smaller operators may reckon themselves.

Of Their Singular Practices

• PUE as a discipline. [/ˈpʌɛ ˈæs ˈæ ˈdɪskɪplɪnɛ/] n Google reports an annual fleet-wide PUE, not the best site alone. That forces honesty about averages and not the flagship; the number is driven by relentless measurement of every subcomponent.
• Machine learning for cooling. [/ˈmækhɪnɛ ˈlɪːrnɪng ˈfɒr ˈkʌːlɪng/] n The DeepMind collaboration showed that a neural network could commend cooling setpoints beyond what human operators achieved. The true victory was not the algorithm; it was the possession of the sensor data and the control authority to act upon it.
• Custom servers. [/ˈkʌstɒm ˈsɛrvɛrs/] n Google contrives its own motherboards and power supplies. A server built for the exact workload and the exact facility will run at higher temperatures and lower cost than a generic box.
• Location matters. [/ˈlɒkætɪɒn ˈmættɛrs/] n Sites are chosen near cheap renewable power and cold water; the cooling system is designed for the local climate, not copied from the last build.
• Waste heat reuse. [/ˈwæstɛ ˈhɪːt ˈrɛʌsɛ/] n Some locations put rejected heat to district heating or industrial processes. The economics are site-specific, yet the principle is the same everywhere: heat is a byproduct, not a waste stream.

What Lesser Houses May Borrow

• Begin with honest measurement. You cannot copy what you do not measure.
• Set sensors where they belong. See On Early Detection.
• Design servers for the actual environment; a well-chosen inlet temperature can save more than a bespoke cooling plant.
• Use outside air when the climate allows. The cheapest cooling is the air you need not chill.

Google datacenter in The Dalles, Oregon, showing the cooling and electrical infrastructure outside the building. Photo by Visitor7, licensed CC BY-SA 3.0 via Wikimedia Commons.

See Efficient datacenter case studies for the other cases, and Power usage effectiveness for the metric that governs the comparison.