Nvidia has called for a new Data Centre Energy Efficiency Metric in a blog post on May 12th. They explain that for the past 17 years, the most widely used metric for data centre energy efficiency has been Power Usage Effectiveness (PUE), however it doesn't take into account the useful output of the facility.
Researchers and experts agree that a suite of benchmarks is needed to accurately measure the energy implications of today's most widely used AI workloads.
To better understand data centre energy efficiency, metrics should focus on energy (measured in kilowatt-hours or joules) rather than power (measured in watts). Additionally, the industry must move away from abstract measures of work, such as processor instructions or maths calculations, and instead focus on the amount of real, useful work being done.
The definition of useful work may vary depending on the data centre's focus. For example, data centres focused on AI may rely on MLPerf benchmarks, while those tackling scientific research or streaming media may use different measures.
The evolving landscape of data centre workloads has made it clear that a new set of energy efficiency metrics is needed. By focusing on energy consumption and useful work, and by taking into account advances in accelerated computing, data centres and supercomputing centres can develop a suite of benchmarks that will hopefully bring about new levels of energy efficiency.