Three years have passed since the first paper on this technology's impact on the environment was published. Making changes can be difficult if you don't have accurate numbers.
Jesse Dodge is a researcher at the Allen Institute for Artificial Intelligence in Seattle. If we want to reduce emissions, we need to get a good measurement.
The Allen Institute collaborated with Microsoft, Hugging Face, and three universities to create a tool that measures the electricity usage of any machine- learning program that runs on Microsoft's cloud service. During every phase of their project, from selecting a model to training it and putting it to use, users can see the total electricity consumed by graphics processing units. The first major cloud provider to give users access to information about the energy impact of their machine- learning programs.
Tools that measure energy use and emissions from machine-learning are already available, but they can't be used in the cloud. The existing tools need those values in order to provide accurate estimates.
The new tool reports energy use instead of emissions. Dodge and other researchers presented a companion paper on how to map energy use to emissions at the FAccT in June. The researchers used a service called Watttime to estimate emissions based on zip codes.
Researchers can reduce emissions if they use specific server locations and times of day. When more renewable electricity is available on the grid, emissions from small machine-learning models can be reduced up to 80%, while emissions from large models can be reduced over 20%.
All Rights Reserved © 2024
