Research manuscript published by Emma Strubell, Ananya Ganesh, and Andrew McCallumin in June 2019 and available on arxiv.org (https://arxiv.org/pdf/1906.02243.pdf) develops an AI environmental cost framework and assesses several algorithms. Popular NLP approaches exhibit wide variation in power, hours, and emissions. Recommendation:
Authors should report training time and sensitivity to hyperparameters. Our experiments suggest that it would be beneficial to directly compare different models to perform a cost-benefit (accuracy) analysis.