How to estimate carbon footprint when training deep learning models? A guide and review

Area of Science:

• Computer Science
• Environmental Science

Background:

• Machine learning (ML) and deep learning (DL) are integral to artificial intelligence (AI) advancement across various sectors.
• The environmental cost of developing AI models is a growing concern, prompting research into its impact.
• Numerous software and online tools now exist to monitor the energy consumption during ML model training.

Purpose of the Study:

• To provide AI practitioners with a comprehensive guide to tools for assessing the environmental impact of their work.
• To introduce and compare existing tools for tracking energy consumption in ML model development.
• To offer practical advice on selecting appropriate tools and infrastructure for environmental impact estimation.

Main Methods:

• A review of available online and software tools for tracking ML model training energy consumption.
• Detailed examination of the technical requirements and specific vocabulary associated with each tool.
• Comparative analysis of energy consumption estimates from selected tools using two deep neural networks for image processing on various server types.

Main Results:

• The study identifies and categorizes various tools for monitoring AI's environmental footprint.
• Differences in energy consumption estimations were observed across tools and server infrastructures.
• The paper highlights the importance of tool selection and infrastructure in accurately assessing environmental impact.

Conclusions:

• AI practitioners can utilize the findings to make informed decisions about tool and infrastructure choices.
• Accurate estimation of AI's environmental cost is crucial for sustainable AI development.
• This research facilitates a better understanding of the environmental implications of ML and DL model training.