Measuring the Effectiveness of Carbon-Aware AI Training Strategies in Cloud Instances: A Confirmation Study

While the massive adoption of Artificial Intelligence (AI) is threatening the environment, new research efforts begin to be employed to measure and mitigate the carbon footprint of both training and inference phases. In this domain, two carbon-aware training strategies have been proposed in the lite...

Full description

Saved in:
Bibliographic Details
Published inFuture internet Vol. 16; no. 9; p. 334
Main Authors Vergallo, Roberto, Mainetti, Luca
Format Journal Article
LanguageEnglish
Published Basel MDPI AG 01.09.2024
Subjects
Algorithms
Alternative energy sources
Anomalies
Artificial intelligence
Carbon
carbon awareness
Carbon footprint
cloud
Computational linguistics
Computer vision
Deep learning
Ecological footprint
Effectiveness
Emissions
Emissions (Pollution)
Energy consumption
Energy efficiency
Energy resources
Environmental impact
Environmental sustainability
green AI
green software
Infrastructure
Language processing
Machine learning
Machine vision
Measurement
Natural language interfaces
Natural language processing
Renewable resources
Software industry
State-of-the-art reviews
sustainability
Workloads
North Carolina
Germany
Online AccessGet full text

Cover

More Information
Summary:While the massive adoption of Artificial Intelligence (AI) is threatening the environment, new research efforts begin to be employed to measure and mitigate the carbon footprint of both training and inference phases. In this domain, two carbon-aware training strategies have been proposed in the literature: Flexible Start and Pause & Resume. Such strategies—natively Cloud-based—use the time resource to postpone or pause the training algorithm when the carbon intensity reaches a threshold. While such strategies have proved to achieve interesting results on a benchmark of modern models covering Natural Language Processing (NLP) and computer vision applications and a wide range of model sizes (up to 6.1B parameters), it is still unclear whether such results may hold also with different algorithms and in different geographical regions. In this confirmation study, we use the same methodology as the state-of-the-art strategies to recompute the saving in carbon emissions of Flexible Start and Pause & Resume in the Anomaly Detection (AD) domain. Results confirm their effectiveness in two specific conditions, but the percentage reduction behaves differently compared with what is stated in the existing literature.
ISSN:1999-5903
1999-5903
DOI:10.3390/fi16090334