Evaluating the carbon footprint of code implementation

• Main Author: Tar, Sreeja
• Other Authors: Lim Wei Yang Bryan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Large Language Models (LLMs); Carbon footprint; Fine-tuning; Natural Language Processing (NLP); Emissions; Energy consumption
• Online Access: https://hdl.handle.net/10356/181172
• Institution: Nanyang Technological University

This project undertakes the evaluation of the carbon footprint of large language models (LLMs). Three models are evaluated – Meta’s LLaMA-2 (7-billion parameter configuration), Mistral (7-billion parameter configuration), and Google’s Gemma (2-billion and 7-billion parameter configurations). The emissions generated by these models in the fine-tuning phase are evaluated for three tasks – question answering, text summarisation, and sentiment analysis. This report also examines differences in emissions generated across GPUs. The project ultimately explores the impact of model optimisation on emissions and provides corresponding recommendations to reduce the carbon footprint of the selected models.