College - Author 1
College of Engineering
Department - Author 1
Civil and Environmental Engineering Department
College - Author 2
College of Engineering
Department - Author 2
Computer Engineering Department
Advisor
Franz Kurfess, College of Engineering, Computer Science and Software Engineering Department
Funding Source
Bavaria California Technology Center (BaCaTeC) #34 (2025-1) and the California Polytechnic State University San Luis Obispo, College of Engineering Summer Undergraduate Research Program (SURP 2025)
Acknowledgements
Hochschule München MUC.DAI faculty Dr. Gudrun Socher and Maximilian Dauner, Bella Boulais
Date
10-2025
Abstract/Summary
As the size and demand for large language models (LLMs) increase, the environmental impact of computational inference often exceeds training; yet industry lacks a standardized method of calculating this expanding environmental footprint. Complexity arises with task-specific computational demands, infrastructure overhead, and various GPU architectures, making cross-model assessments burdensome. Combining environmental engineering and computer science principles by validating Jegham et al.’s meta-model, we predict the carbon emissions and water consumption during inference, providing metrics to raise user awareness of AI’s growing environmental footprint. Additional work supports integration into a multi-agent conversational system that encourages responsible scheduling and prompting, guiding the user towards sustainable AI solutions. Our results show that this framework can predict carbon emissions and water usage during inference across a variety of LLM architectures and sizes with reasonable accuracy for both locally-running and proprietary cloud-based LLMs.
October 1, 2025.
URL: https://digitalcommons.calpoly.edu/ceng_surp/87