Home // GREEN 2023, The Eighth International Conference on Green Communications, Computing and Technologies // View article

Reducing Carbon Footprint of AI Models Without Compromising Performance

Authors:
Austin Deng
Xingzhi Huang
Michael Lu

Keywords: AI; Energy Efficiency; Carbon Emission.

Abstract:
The widespread adoption of Artificial Intelligence (AI) models, such as ChatGPT, has resulted in a significant increase in energy consumption and carbon emissions associ- ated with their training and inference. However, research on sustainable AI is still nascent. This paper aims to explore efficient approaches that can reduce the carbon footprint of AI models without compromising performance. Through an extensive analysis of four distinct categories of AI models (text to text summary, image classification, text to image, and image to text) across various sizes, our findings challenge the prevailing notion that larger AI models consistently outperform smaller ones. In specific AI tasks, we observe that small models can achieve comparable performance while significantly reducing carbon emissions. Moreover, we propose a carbon-aware solution that strategically directs computationally intensive AI tasks to regions with low carbon intensity, which can effectively reduce the environmental impact without compromising model quality. Our experimental results demonstrate a significant carbon savings while maintaining the desired performance levels.

Pages: 8 to 14

Copyright: Copyright (c) IARIA, 2023

Publication date: September 25, 2023

Published in: conference

ISSN: 2519-8483

ISBN: 978-1-68558-097-1

Location: Porto, Portugal

Dates: from September 25, 2023 to September 29, 2023