Are you aware of your AI’s carbon footprint?

Capgemini Invent has developed a tool that helps you measure and reduce the carbon footprint of your AI algorithms.

Thanks to reduced computing power costs and increasingly massive datasets, Artificial Intelligence is pervasive in almost all sectors of activity, businesses and governments included.

However, while AI is being praised as the go-to technology for many climate actions, it also carries its own footprint.

At Capgemini Invent, we believe that measuring and monitoring AI impact is now a business imperative.

Starting from the big picture, digital & IT (computing, storage, network, and devices) footprint accounts for 4% of global CO2 emissions and its energy consumption increases by 9% every year. More specifically, the environmental impact of AI activities is widely unappreciated. As the complexity of models and their energy requirements will only increase in the future, a pressing need for eco-design initiatives around AI has arisen.

From a regulatory standpoint, the European Parliament’s special committee on Artificial Intelligence in the Digital Age expressed the need to align the design and development of AI with the goals of the European Green deal.

We decided to tackle the lack of user-friendly tools and develop our own solution to capture the entire spectrum of emissions, thereby precisely monitoring AI carbon emissions at scale.

At Capgemini Invent, we wondered why the carbon footprint of AI was such a blind spot :

1. First, the heterogeneous nature of Machine Learning and Deep Learning tasks makes for a wide range of potential emissions. These emissions mostly depend on the type of use case (Natural Language Processing, computer vision, time series analysis, etc.), project phases (exploratory analysis, training, inference, run, etc.), type of machine used, and the energy mix of the geographic area. Thus, GHG emissions registered for different AI use cases are quite disparate. A group of Google and Berkeley researchers showed that training a single algorithm (GPT3, a powerful language model developed by Open AI) could generate up to 552 T of CO2 — about the amount produced by 3 direct round trips of a full passenger jet between San Francisco and New York (one round trip is 8 324 kilometers).
2. Most common approaches consist in estimating carbon emissions through indirect calculations. However, it has been demonstrated that extrapolations lack in precision compared to direct measurements. Since there is little to no investigation on power consumption breakdown by project phase, the emissions generated by a particular project phase might be underestimated. To illustrate, Nvidia CEO Jensen Huang recently pointed out that an estimated 80 to 90 percent of the cost of machine learning at scale is devoted to AI inference (running phase). As a result, scientists have been unable to select algorithms based on the best performance and carbon efficiency tradeoff.
3. Finally, AI projects are foremost driven by classic model performance metrics (e.g., accuracy, precision, and recall). However, in the process of fine-tuning a model, the performance enhancement is not worth the generated impact. So how can we figure out the best tradeoff between accuracy and efficiency?

To solve this issue, we have developed our own carbon impact measurement tool: Carbon AI

We think monitoring power usage is the right way to track AI’s footprint. Capgemini Invent has built its own python package allowing to precisely measure the energy consumption of any algorithm. Easily deployed on any data science project, the package is already a standard internally.

Our Approach

The idea is to have probes to accurately measure AI energy consumption, throughout its entire life cycle (from exploratory analysis to production) and for every use case within the client system (laptops, on-premise, cloud, on edge).

We leverage hardware manufacturers’ tools (Intel and Nvidia) to record the power consumption of a machine attributed to an AI process. While an algorithm is executing, our package runs in the background and queries those tools to record the energy usage at every step of the process. The data collected is stored on the user’s computer, providing precise knowledge of AI’s carbon footprint.

How can I use it?

The package is publicly available on Github.

In addition to the package documentation, a tutorial will be published in a second article followed by technical details in a scientific publication. Stay tuned!

With only 2 lines of code, the package is user-friendly and noninvasive. It runs in the background without disrupting your running processes. During algorithm execution, computing power is measured and converted into CO2 emissions.

Why should I use it?

This tool has been designed for both data scientists and AI leaders

On the business side, you will be able to…

1. Precisely measure the impact of each project for each Business Units
2. Identify AI energy efficiency levers and best practices to raise awareness among your data scientists
3. Develop innovative AI carbon footprint KPIs to integrate into your company’s ESG report
4. Set new decision criteria related to ROE (Return on Environment) in addition to ROI, before giving the green light to new projects

On the Data science side, you will…

1. Get the exact breakdown of carbon emissions for each component of your Artificial Intelligence project
2. Identify optimal algorithms with the best tradeoff between performance and energy efficiency for each task before deployment
3. Become aware of energy usage and take your algorithm footprint into account from the beginning

What’s next?

This package is only the start of our journey towards greener AI 🌱.

On top of this open-source tool, we are currently developing additional features to enhance the monitoring and identification of carbon emission reduction levers. Beyond measurement, we aim at providing concrete actions and best practices to support data scientist teams moving towards frugal AI.

Be among the first to integrate Carbon AI in your work environment and join our community on Github!

About the Authors

Simon Gosset is a data strategy consultant at Capgemini Invent with hands-on experience in shaping and building data products at scale (from idea to industrialization) alongside with industrial leaders. Simon is driven by making an impact with AI technology. He is focused on building assets at the crossroad of data & sustainability. He is CarbonAI product owner.

Martin Chauvin is a consultant data-scientist at Capgemini Invent. He had the opportunity to work for many industrial clients to develop AI algorithms to improve their operational efficiency. Passionate about data science and trying to tackle climate change, he leads the technical development of the package.

Emilie Debedde is a data strategy consultant at Capgemini Invent. She has coordinated data projects in several industries and strongly believe in data science to accelerate environmental research.

François Lemeille is a Senior Data-Scientist at Capgemini Invent. He was involved in the packaging and tooling of this project to deliver a reliable tool.

Valentin Millery is Managing Consultant at Capgemini Invent within Data Analytics for Intelligent Industry practice. He has experience in developing new products and services at scale leveraging Data & AI. He is convinced that AI can help reduce the effects of climate crisis with CarbonAI tool ahead.

Tanguy Masgnaux is a data strategy consultant at Capgemini Invent. He is passionate about solving data related business challenges and bridging the gap between data specialists and decision makers.