The Environmental Price of Intelligence: Evaluating the Social Cost of Carbon in Machine Learning

The rapid expansion of machine learning (ML) technologies has brought with it an increasing concern over the carbon emissions associated with their hardware and computational demands. This study aims to quantify the environmental impact of ML through the lens of the social cost of carbon (SCC), a metric that translates carbon emissions into future economic impacts. Our approach involves a detailed calculation of the SCC for the entire development lifecycle of ML models, offering a monetary perspective on their environmental footprint. By mapping these costs to a sustainable framework, we aim to foster greater awareness and encourage responsible practices in ML development. This paper not only quantifies the often-overlooked environmental costs of ML but also proposes a pathway towards more sustainable model development, navigating the complex balance between technological advancement and environmental responsibility.

I. Introduction

Machine learning (ML) technologies has brought unparalleled capabilities in data processing and analysis. However, this progress comes with a hidden cost: a significant increase in carbon emissions due to the high computational requirements of ML models. The utilization of the resources like PUs, GPUs, and RAM intensively in complex model development, as seen in generative AI, contribute substantially to environmental concerns. For instance, prominent LLMs like T5 [1], Meena [2], and GPT [3] require extensive computations, leading to high carbon emissions during their training and testing phases[4], [5], [6].