Experiment Impact Tracker

Overview

Experiment Impact Tracker is a drop-in Python method for tracking energy usage, carbon emissions and compute utilization of a system. On Linux hosts with Intel chips supporting the RAPL or powergadget interfaces and on NVIDIA GPUs it records CPU and GPU power draw, hardware information, Python package versions and estimated carbon emissions, and in California it can query caiso.com for realtime carbon intensity. Logged data can be turned into an online appendix, and a companion script generates a carbon impact statement for a paper from the tool's log directories. The project accompanies the paper Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning (Henderson et al., 2020).

Use cases

• Log energy and carbon per ML experiment
• Generate a carbon impact statement for a paper
• Build energy leaderboards for experiments
• Report compute utilization alongside emissions

Sources

• Experiment Impact Tracker repository
• Experiment Impact Tracker on PyPI