SUSTAINABLE AI

This document presents a curated library of recent proposals aimed at optimising the energy consumption of AI systems across four key domains: “Hardware – Type”, “Hardware - Location/Timing”, “Software – Model”, and “Software – Method”.

These innovative approaches are the result of an extensive exercise conducted by the Master in AI students at Vlerick Business School. By exploring cutting-edge techniques and practical solutions, this collection serves as a comprehensive guide for improving the sustainability of AI technologies.

Each domain is explored with a focus on actionable insights and real-world applications. The aim is to contribute to the growing need for energy-efficient and environmentally responsible AI systems.

The following pages describe the proposals whicch can be applied across four levels. To each proposal, a definition, some metrics (when available) and the source of the information is added:

1. Hardware - Type:

   • Example: Selecting energy-efficient hardware like GPUs, TPUs, or custom AI accelerators designed for AI workloads reduces energy consumption.

2. Hardware - Location/Timing:

   • Example: Optimising the location of training (e.g., data centers powered by renewable energy) and scheduling training during off-peak hours can minimize energy impact.

3. Software - Model:

   • Example: Choosing the appropriate model architecture and size for the task (e.g., using smaller or task-specific models) reduces computational demand.

4. Software - Method:

   • Example: Techniques like model pruning, quantization, efficient training algorithms, and checkpointing during training can reduce energy use in the training phase.

At the end of the document, an appendix provides additional links to valuable resources and information for further exploration. By sharing these insights, we hope to inspire more energy-efficient AI solutions in the future, contributing to a more sustainable technological landscape. 

Please find here the link to the document. 

The document is written in full collaboration with the following students: Santino Agnello, Ishaq Ahmed, Roxanne Baeten, Lisse Belis, Anouk Bergez, Romy Cavelier, Trisha Chatterjee, Wolfgang Constandt, Xander De Smet, Lucas Dewaele, Xander Fontaine, Marie Gabriëls, Luna Geens, Yash Ghai, Jeewat Kumar Harejin, Lothar Henne, Tanja Impens, Reda Karim, Khalil Ahmad Khan, Sein Him Law, Robbe Mannens, Lyn Mbaabu, Bharat Lal Meghwar, Thi Thanh Huong Nguyen, Stijn Oosterlinck, Anna Pakhardymova, Emile Peeters, Marthe Spriet, Lucca Van Hover, Jianing (Giner) Wang