Tutors : Romain Rouvoy, Clément Quinton
Studies on Large Language Models' energy consumption and code efficiency
Tristan Coignion
2nd year PhD student
AI and Large Language Models
Large Language Model (LLM) :
AI technology that generates text.
Code assistant : An LLM that integrates into a developer's workflow (e.g. GitHub Copilot)
Helps developers create software
AI and Large Language Models
Study on the performance of LLM-generated code
Studied the performance of LLM-generated solutions to Leetcode algorithmic problems
- No (or very few) differences between LLMs in terms of code speed
- LLM are a priori better than humans on average in terms of code speed (Careful : weak proofs)
Study on the energy consumption of a code assistant
Below is a depiction of an LLM's usage of electricity :
The code assistant's generation is done in the cloud
The electrical consumption is hidden
Study on the energy consumption of a code assistants
How do we measure ?
A human participant develops using GitHub Copilot
We replay the generations using multiple AIs and scenarios on servers we own, and measure the electrical consumption
We collect the telemetry of GitHub Copilot which allows us to "replay" the generations with other AIs
Study on the energy consumption of a code assistant
Next steps
- Performance of software produced by LLMs
- Environmental impact of using a code assistant
- Making the code assistants use less electricity
Thank you for listening !
Team Seminar Autumn 2023
By Tristan Coignion
