Do you know...

how much electricity one ChatGPT request consumes ?

Do you know...

how much electricity one ChatGPT request consumes ?

~ 0.001-0.01 kWh / query[1]

10W lamp turned on for one hour !

[1] https://limited.systems/articles/google-search-vs-chatgpt-emissions/

Do you also know...

how much electricity you consume when using GitHub Copilot for one hour?

[shrug.jpg]

Green My LLM: Studying the key factors affecting the energy consumption of code assistants

Coignion Tristan, Quinton Clément, Rouvoy Romain

GT GL-IA 2025 - Rennes

Find the paper online :

https://arxiv.org/abs/2411.11892

AI is everywhere

AI is everywhere

Code Assistant

GitHub Copilot

Demo of GitHub Copilot

How much energy am I consuming when using a code assistant like GitHub Copilot?

Sends and receives
generation requests

Inference Server in a datacenter somewhere

Makes code

suggestions

Code assistant

How code assistant usually work

Sends and receives
generation requests

Inference Server in a datacenter somewhere

Makes code

suggestions

Code assistant

This is where we want to measure !

Our method

Phase 1 : Build a dataset of development traces

Phase 2 : Use this dataset to simulate the usage of GitHub Copilot and measure the consumption of the inferences

generation requests

Normal GitHub Copilot's inference server

Makes code

suggestions

Code assistant

GitHub's

telemetry server

telemetry data

20 participants

developing for one hour

Phase 1 : Collecting development traces

generation requests

Normal GitHub Copilot's inference server

Makes code

suggestions

Code assistant

GitHub's

telemetry server

telemetry data

20 participants

developing for one hour

Phase 1 : Collecting development traces

generation requests

Normal GitHub Copilot's inference server

Makes code

suggestions

Code assistant

telemetry data

20 participants

developing for one hour

GitHub's

telemetry server

Our telemetry server

Phase 1 : Collecting development traces

generation requests

Normal GitHub Copilot's inference server

Makes code

suggestions

Code assistant

telemetry data

20 participants

developing for one hour

GitHub's

telemetry server

Our telemetry server

Phase 1 : Collecting development traces

Phase 2

Simulating the

code assistant

Our inference server (hosted on G5K)

Our telemetry server

Code assistant simulator

Simulates GitHub Copilot generation requests

`perf` + `nvidia-smi`

telemetry

data

records energy consumption

generation

requests

Configuration options for the simulations

  • Number of concurrent developers - [1-500]
  • Request Streaming - [Yes, No]
  • Manual triggering of the code assistant* - [Yes, No]
  • Large Language Model - [StarCoder, StarCoder2-{7,15}B]
  • Quantization method  - [EETQ, BitsAndBytes, None]
  • Number of GPUS - [1-4]

* emulation of a manual triggering by the user.

4,896 possible configurations.

829 simulations with 314 unique configurations

Before the results...

Some limitations

  • The measures are only an estimation of the energy GitHub Copilot would consume if it was using the same hardware as us.
     
  • Does not take into account the PUE of the datacenter.
     
  • Only considers the environmental impact from the energy standpoint.
     
  • Needs more diverse configurations (hardware, models, etc.).
     
  • Does not take into account code quality or validity.

Less generation requests == Less energy consumed

The least you use GitHub Copilot, the lower you energy consumption is.

A few stats about our participants' usage of GitHub Copilot

Depending on users, between 2 and 20 requests per minute.

Average of 9 requests per minute

Students accepted more often the suggestions from GitHub Copilot than the professional developers

Final state of suggestions made by GitHub Copilot.

Final state of suggestions made by GitHub Copilot.

  • Suggestions happening when the user is typing are often cancelled before the completion of the generation.
     
  • Empty results happen when a suggestion generation is triggered at the end of sentences or lines.
     
  • Suggestions that were not accepted typically happen because users did not ask for them specifically.*

* and sometimes because they were simply bad

Automatic suggestions can lead to a great waste of computing power.

Saturation point

Energy usage of the whole server and by developer, and latency of the requests depending on the number of concurrent developers

Impact on the energy consumption of switching from one configuration to another
(e.g. adding a GPU, or using another LLM)

Impact on the energy consumption of switching from one configuration to another
(e.g. adding a GPU, or using another LLM)

Impact on the energy consumption of switching from one configuration to another
(e.g. adding a GPU, or using another LLM)

Impact on the energy consumption of switching from one configuration to another
(e.g. adding a GPU, or using another LLM)

Impact on the energy consumption of switching from one configuration to another
(e.g. adding a GPU, or using another LLM)

Impact on the latency of the requests of switching from one configuration to another

Scenarios

  • Small team : A dedicated server for 5 concurrent developers
     
  • Medium team : A dedicated server for 20 concurrent developers
     
  • Distributed service : A server whose load is maximized (e.g. commercial code assistants)

Scenarios

  • Small team : A dedicated server for 5 concurrent developers
  • Medium team : A dedicated server for 20 concurrent developers
  • Distributed service : A server whose load is maximized (e.g. commercial code assistants)

Objectives

  • Frugal objective : Minimize energy consumption
  • Performance objective : Minimize latency and user discomfort (automatic suggestions)

Average consumption of one developer :

Average consumption of one developer :

Users of GitHub Copilot = 28.7 millions developers[1] x 41%[2]

[1] https://www.statista.com/statistics/627312/worldwide-developer-population/

[2] 2024 Stack Overflow Developer Survey

= 11.77 million users

x  11,770,000

=
235 MW

Users of GitHub Copilot = 28.7 millions developers[1] x 41%[2]

[1] https://www.statista.com/statistics/627312/worldwide-developer-population/

[2] 2024 Stack Overflow Developer Survey

= 11.77 million users

x  11,770,000

235 MW

Users of GitHub Copilot = 28.7 millions developers[1] x 41%[2]

[1] https://www.statista.com/statistics/627312/worldwide-developer-population/

[2] 2024 Stack Overflow Developer Survey

= 11.77 million users

25% x

235 MW

[1] https://www.edf.fr/groupe-edf/comprendre/production/nucleaire/nucleaire-en-chiffres

25% x

235 MW

[1] https://www.edf.fr/groupe-edf/comprendre/production/nucleaire/nucleaire-en-chiffres

235 MW

Also...

So... What can we do about it ?

Code assistant providers

  • Use smaller and more specialized models
  • Quantize the models
  • Maximize the usage of each server.
  • Allow user to easily manually trigger code assistants

So... What can we do about it ?

Code assistant providers

  • Use smaller and more specialized models
  • Quantize the models
  • Maximize the usage of each server.
  • Allow user to easily manually trigger code assistants

Users

  • Use assistants mindfully
  • Don't make them the default
  • Disable the automatic suggestions
  • Think before prompting. Only use AI when you need it.

So... What can we do about it ?

Code assistant providers

  • Use smaller and more specialized models
  • Quantize the models
  • Maximize the usage of each server.
  • Allow user to easily manually trigger code assistants

Users

  • Use assistants mindfully
  • Don't make them the default
  • Disable the automatic suggestions
  • Think before prompting. Only use AI when you need it.

 

Researchers

  • Measure the environmental impact of AI.
  • Make the measuring process easier.
  • Talk about it around you

Thank you for listening !

Green My LLM - GT GL-IA 2025

By Tristan Coignion

Made with Slides.com

Green My LLM - GT GL-IA 2025

  • 31

More from Tristan Coignion