Measure footprint of open LLMs

Benchmark Apertus, compare with other models, and find strategies for efficient prompting.


3 54 44

EDITED v. 60

1 month ago ~ loleg

Research

EDITED v. 57

1 month ago ~ loleg

📚 Project Repository Created!

Just published our hackathon documentation: "Measuring LLM Energy Footprints"

What we built:

  • Automated energy measurement scripts
  • 42 experimental runs with real data
  • Analysis tools for LLM efficiency comparison
  • Reproducible framework for energy research

Key finding: Response length drives energy consumption way more than prompt complexity.

Fun discovery: Portuguese uses 20% more energy than other languages... as a Portuguese myself, I'm not sure if I should be proud or concerned! 😅

Repository: https://github.com/luisantoniio1998/Measure-footprint-of-open-LLMs

Big thanks to the swissAI team for organizing such an amazing event! Special shoutout to Oleg and everyone who made this hackathon possible. Great energy (pun intended) and fantastic collaboration throughout the weekend.

Thanks to my teammates Agustín and Stefan for the excellent teamwork! 🙌

#swissAI #LLMenergy #Hackathon

1 month ago ~ luisdbarros

Evaluation of Project

Technical Functionality: 4 (Excellent)
The project has demonstrated a significant step towards the technical goals by proposing a measurement framework and providing data about expected resource consumption. However, it has not implemented a fully detailed and interactive dashboard or prompting strategies. Instead, it tentatively suggests approaches that could be further developed with some coding effort. A higher score would require more ambition in demonstrating concrete, coded functionality.

User Experience: 3 (Good)
The project creators have incorporated design concepts into a visually appealing presentation. However, further development of a user-friendly interface and interactive dashboard implementation are needed. The user flow is not yet clear, because the current mockups are static and do not show full functionality.

Skillful Use of AI: 3 (Good)
The team shows a intent to apply AI through dashboards for measuring and prompting, but only at the idea level. Actual AI use or demonstrations (which can range from machine learning models to natural language processing strategies for prompting) have not been implemented in detail. For a higher rating, they should provide working code, examples, or at least detailed algorithms.

Uniqueness / Creativity / Fun Factor: 2 (Fair)
While the goal of making AI more sustainable is creative and timely, the project’s approach does not entirely break new ground. Some parts may be similar to existing works focused on sustainable computing or AI ethics. The potential for a viral "wow" factor or novel approaches could be improved with more ambitious designs in measurements or impact reduction strategies.

Potential / Market Impact: 4 (Excellent)
The project addresses a critical and increasingly important subject: sustainable AI, which has both high market opportunity and could offer Switzerland a leading edge in responsible technology practices. However, realizing this potential depends on effective implementation of the proposed strategies, which, given the lack of concrete prototypes or fully developed dashboards, remains conceptual.

Key Feedback for Improvement:

  • More detailed and interactive documentation, especially showing real codebase snippets or demo links.
  • A dynamic, interactive dashboard prototype or mockup that simulates measuring and responding to model usage.
  • Examples or demos of AI prompting strategies to show how they could reduce environmental impact.
  • A clearer, step-by-step user flow to demonstrate practical usage.

The project is off to a strong start conceptually but requires more concrete, functional demonstrations and applications to reach the "excellent" level in all categories.


Note: This is a simulated evaluation and not based on an actual project. The scores and feedback are examples of how a judge might evaluate based on the given criteria.

🅰️ℹ️ generated with APERTUS-70B-INSTRUCT

1 month ago ~ loleg

Event finish

back to "old" view of slides, this one works at least

1 month ago ~ stefan

EDITED v. 54

1 month ago ~ stefan

EDITED v. 53

1 month ago ~ stefan

move to readme

1 month ago ~ stefan

presentation

1 month ago ~ stefan

update presentation

1 month ago ~ stefan

1 month ago ~ stefan

EDITED v. 47

1 month ago ~ stefan

formatting

1 month ago ~ stefan

EDITED v. 46

1 month ago ~ stefan

added report and data

1 month ago ~ stefan

1 month ago ~ stefan

we have first results, and lunch!

1 month ago ~ stefan

JOINED

1 month ago ~ luisdbarros

Our script it's working :D, we need the visualization https://codeshare.io/ay7NPe

1 month ago ~ AgustinHerrerapicazo

Fail of our script #24

https://codeshare.io/5Z7LLV

Trying to have some kinda visualization

1 month ago ~ AgustinHerrerapicazo

Fail of our script #23 https://codeshare.io/G6q4pp Getting close and some automatization

1 month ago ~ AgustinHerrerapicazo

Fail of our script #22 https://codeshare.io/2p8A0V

1 month ago ~ AgustinHerrerapicazo

Fail of our script #21 https://codeshare.io/5v7wEK

1 month ago ~ AgustinHerrerapicazo

Fail of our script #20 (We start at 20 because we don't have the first 20 saved.) https://codeshare.io/G8qADE

1 month ago ~ AgustinHerrerapicazo

better slide link

1 month ago ~ stefan

link to presentation

1 month ago ~ stefan

EDITED v. 34

1 month ago ~ stefan

We change the way of measuring to do it more accurately (More info in the documentation)

1 month ago ~ AgustinHerrerapicazo

1 month ago ~ AgustinHerrerapicazo

1 month ago ~ AgustinHerrerapicazo

Started writing the project report and created a google sheet as a data repository.

1 month ago ~ stefan

2 months ago ~ loleg

Project

After a few hours of work we have managed to collect hardware power metrics of a testbed server in the Begasoft cloud. We need to make another script, because we currently only can read the average energy use. Still struggling with getting Apertus to work on a local NVIDIA machine, kinda wish we had a Mac ;-) j/k

2 months ago ~ loleg

JOINED

2 months ago ~ AgustinHerrerapicazo

Start

Greetings from the Energy Data Hackdays, first of the Swiss {ai} Weeks hackathons. Try the 💡 Energy Bites demo here

2 months ago ~ loleg

green prompting paper

2 months ago ~ stefan

EDITED v. 19

2 months ago ~ stefan

rewrote for remote access to Rig (probably no physical measurement possible). improvement primarily via prompting.

2 months ago ~ stefan

EDITED v. 18

2 months ago ~ stefan

EDITED v. 16

2 months ago ~ loleg

EDITED v. 11

2 months ago ~ loleg

EDITED v. 10

2 months ago ~ loleg

EDITED v. 9

2 months ago ~ loleg

Google has just released a technical paper on the energy impact of an AI prompt. (Technology Review, Hacker News)

2 months ago ~ loleg

EDITED v. 7

2 months ago ~ loleg

added measuring guidelines to outcomes

2 months ago ~ stefan

EDITED v. 4

2 months ago ~ stefan

focus on inference, minor corrections

2 months ago ~ stefan

EDITED v. 2

2 months ago ~ stefan
Posted 2 months ago by stefan for Hackathon Bern
Hackathons full of ideas, collaboration, and innovation are based on the premise of keeping the experience safe, inclusive, and respectful for everyone. We follow a clear Code of Conduct and support the Universal Declaration of Human Rights. Harassment or discrimination of any kind won't be tolerated—this applies to all staff, participants, coaches, visitors and sponsors. Please take a moment to review the full guidelines.

The contents of this website, unless otherwise stated, are licensed under a Creative Commons Attribution 4.0 International License. The application that powers this site is available under the MIT license.