rossjillian/energy-data

Energy data

The climate impact of machine learning is profound yet poorly understood. As machine learning researchers, we constantly make choices that affect how carbon intensive our models are. However, there is no data that enables us to proactively understand the carbon impact of choosing a certain architecture or training paradigm. We plan to release the first open dataset mapping energy usage to model metadata, accompanying data analysis and visualizations, and an interactive website for researchers and practitioners. Our goal is to enable researchers and practitioners to understand the costs of using different techniques and to guide future research directions in efficient machine learning.