Questa sezione non è ancora tradotta. Il contenuto seguente è mostrato in inglese, come pubblicato. Leggi la versione inglese.
Last reviewed · How we check sources
Data centre cooling and power explained
How a data centre gets rid of heat and where its power comes from, in plain English. Start with the path below, which follows a watt from the chip to the grid, or pick a topic. Each guide cites the standards body, regulator or company that published the facts.
Follow the heat: from the chip to the grid in six steps
- The chipWhere the heat startsGPU and memory power, per chip and per rack
- The rackPower into the rack800 V DC and rack-scale systems such as Kyber
- The cold plateLiquid takes the heatHow direct liquid cooling works
- The CDUThe secondary loopHow the CDU controls the loop, and the rules
- OutsideHeat, water and noiseDry coolers, cooling towers and what 1 GW needs
- The gridWhere the power comes fromGrid, on-site generation, batteries and nuclear
New to the jargon? Terms such as PUE, CDU and UPS are in the glossary, and the headline numbers are on statistics.
The guides at a glance
12
plain-English guides and explainers on this page
6
topics, from cooling and power to cost and resilience
6
steps from the chip to the grid, each linked to its guide
11
official sources cited in the 3 technology explainers alone
All guides, by topic
Cooling
- Cooling explainerDirect liquid cooling, explainedHow direct liquid cooling works: cold plates, the technology cooling loop, CDUs and facility water, ASHRAE W17 to W45 water classes, and single versus two-phase.Evidence 5
- Regulation explainerSecondary cooling loop regulationHow the CDU controls the secondary (TCS) loop and the rules that apply: ASHRAE classes, EU reporting, EnEfG, pressure equipment and F-gas rules.
Power and chips
- Power explainer800 V DC power for AI racks, explainedWhat 800 V DC power for AI racks is, why NVIDIA says 54 V in-rack power hits limits above 200 kW, the copper and efficiency claims, and what is still maturing.Evidence 2
- Rack explainerNVIDIA Kyber rack-scale systems, explainedWhat NVIDIA Kyber is: the MGX rack after Oberon, vertical compute blades, 144 GPUs per rack, 800 VDC power and liquid cooling. NVIDIA's own roadmap and dates.Evidence 4
- Chip guideChip power, rack density and coolingWhy GPU and HBM power keeps rising, what it means per rack, and the specifications behind NVIDIA Vera Rubin and Kyber.
- Energy guideEnergy supply pathwaysGrid connection, on-site generation, batteries and nuclear: the options for powering a data centre and what each needs.
Water and noise
- Power and water guideData centre power and water, explainedHow 1 GW of IT load turns into grid supply, heat rejected and water use, with PUE and WUE by cooling method, a calculator and the noise of each heat rejection system.
Cost and money
- Cost guideConstruction cost per MWPublished cost benchmarks for building a data centre by market, with a worksheet to test your own assumptions.
- Money explainerEuropean electricity prices by countryMonthly wholesale power prices in the UK, Ireland, Germany, the Netherlands, France and the Nordics since 2019, and what 1 MW of constant load pays.
Resilience
- Incident explainerAWS data centre attacks in the UAE and BahrainWhat happened zone by zone, how power, cooling and fire suppression were affected, and the resilience lessons.
Reference
- ReferenceData centre glossaryPUE, CDU, UPS, cooling towers, PPAs and AI terms in plain English, each with a source.
- ReferenceData centre statisticsHeadline figures on electricity, PUE, rack density, water and grid queues, each with its source and date.