Power and cooling for AI clusters: airflow, liquid cooling and PUE
AI clusters pack processors densely, generating heat that air conditioning often cannot remove. Liquid cooling techniques such as cold plates, immersion, and door heat exchangers address this. The PUE metric measures efficiency, and programs like ENERGY STAR reduce server power.
Power demand and density
Data centers are extremely energy-hungry buildings. According to the Department of Energy, they use 10 to 50 times more electricity per square foot than a typical office building, and they consume about 2% of total U.S. power.
The same page notes that many efficiency opportunities exist, such as designing new facilities, using benchmarking tools, and applying 12 strategies. For AI clusters, the rising power density makes cooling challenging, often requiring liquid solutions.
Liquid cooling approaches
High-performance computing has used liquid cooling for many years as a cost-effective and efficient way to extract heat, according to the Open Compute Project's cooling page. With the growth of 5G, IoT, VR, and other latency-sensitive applications, data centers are built in regions where cooling is difficult, and increased power density makes warm water cooling an effective alternative.
OCP's cooling project covers five functional areas: cold plates, coolant distribution units, immersion, door heat exchangers, and heat reuse. The cold plate sub-project works on standardization for direct liquid cooling, and the coolant distribution unit sub-project focuses on integrating cooling with facility infrastructure.
Measuring efficiency with PUE
The Green Grid's document titled PUE: a comprehensive examination of the metric provides guidelines for determining Power Usage Effectiveness (PUE) for a dedicated data center, a mixed-use facility, or a specialized facility. It includes results, scalability analyses, and case studies, and discusses Partial PUE and the approach to reuse waste heat.
The document emphasizes that PUE measures energy efficiency, but notes that other metrics are needed for issues such as equipment availability and productivity. It also supersedes prior white papers and harmonizes PUE definitions globally with a primary focus on energy.
Equipment efficiency and ENERGY STAR
According to ENERGY STAR's data center page, IT equipment represents nearly two-thirds of data center energy consumption, yet it has historically not been the focus of energy savings. An easy way to increase savings is to purchase ENERGY STAR certified products.
For example, an ENERGY STAR certified server uses 30% less power than a comparable standard server, saving more than 650 kWh per year when power management is enabled. ENERGY STAR offers product finders for enterprise servers, uninterruptible power supplies, data center storage, and large network equipment.
Research and best practices
Lawrence Berkeley National Laboratory's Center of Expertise for Data Center Energy partners with industry and electric utilities to accelerate AI infrastructure and drive innovative energy systems. Its research areas include power and cooling technologies, grid planning, procurement, and facility performance modeling.
The center provides a comprehensive data center energy toolkit that can be used sequentially to move from a basic understanding of energy use to identifying opportunities and implementing best practices. It also hosts workshops and publishes reports on data center energy demand.
| Area | Role per OCP |
|---|---|
| Cold plate | Sub-project develops standards for direct liquid cooling |
| Coolant distribution unit | Sub-project works on liquid distribution and facility integration |
| Immersion | Listed as a functional area for advanced cooling |
| Door heat exchanger | Listed as a functional area for advanced cooling |
| Heat reuse | Listed as a functional area for advanced cooling |
Frequently asked questions
What does PUE measure?
PUE, or Power Usage Effectiveness, measures the ratio of total data center energy to IT equipment energy, with guidelines provided by The Green Grid. It helps operators see how much overhead energy is used for cooling and other infrastructure.
Why is liquid cooling used for AI clusters?
High power densities in AI clusters generate heat that air cooling cannot efficiently remove. Liquid cooling, especially warm water cooling, is an effective alternative, as noted by the Open Compute Project.
How much electricity do data centers consume in the U.S.?
Data centers account for about 2% of total U.S. electricity use and consume 10 to 50 times more energy per floor space than a typical commercial office building, per the Department of Energy.
What savings do ENERGY STAR servers offer?
An ENERGY STAR certified server uses 30% less energy than a standard server and saves on average over 650 kWh per year when power management is enabled.
What does LBNL's data center research cover?
LBNL researches power and cooling technologies, grid integration, and facility performance to help accelerate AI infrastructure and improve data center energy systems.