The physics (and mathematics) of Artificial Intelligence
What is the environmental impact of using an AI chatbot? It is possible to model this from first principles by considering the GPU energy usage and the water required for cooling data centres. This project aims to estimate the joules of energy, litres of water and grams of carbon dioxide released from each LLM chatbot response.
Project status
Content navigation
About
Synopsis
What is the environmental impact of using an AI chatbot? It is possible to model this from first principles by considering the GPU energy usage and the water required for cooling data centres. This project aims to estimate the joules of energy, litres of water and grams of carbon dioxide released from each LLM chatbot response.
Description
Every prompt sent to an AI chatbot results in a chain of physical processes: electrons are moved across silicon, heat is dispersed into water or air and energy was produced somewhere in the power grid to power it. This project focuses on determining the computation cost and therefore power consumption and environmental impact of Large Language Model (LLM) usage.
Students will conduct research, Fermi estimation, and uncertainty propagation to create a first-principles model of the end-to-end cost of LLM inference.
Required background
Students should have a strong foundation in a range of physics topics including thermodynamics. Experience in solving Fermi programs, order-of-magnitude estimation and uncertainty analysis are also essential.
Computer science knowledge is desirable. Experience with Python is also desirable.