A curated list of resources dedicated to enhancing efficiency in AI systems. This repository covers a wide range of topics essential for optimizing AI models and processes, aiming to make AI faster, cheaper, smaller, and greener!
| Topic | Description | Topics |
|---|---|---|
| Quantization | Reducing precision of AI models without loss |  |
| Pruning | Removing unnecessary model parameters for efficiency |  |
| Caching | Storing computation results for faster reuse |  |
| Distillation | Transferring knowledge from a large model to a smaller one |  |
| Factorization | Breaking down complex models into simpler, efficient components |  |
| Compilation | Optimizing model code for specific hardware and environments |  |
| Parameter-Efficient Fine-tuning | Learning a subset of parameters |  |
| Speculative Decoding | Decoding with batches |  |
| Hardware | Leveraging specialized hardware for faster model execution |  |
| Training | Techniques for making model training faster and more efficient |  |
| Inference | Optimizing the speed and resource usage during model inference |  |
| Sustainability | Strategies to reduce the environmental impact of AI systems |  |
| Scalability | Approaches for scaling AI models and infrastructure efficiently |  |
If you find this list helpful, give it a β on GitHub, share it, and contribute by submitting a pull request or issue!
- Facts π
- Tools π οΈ
- Articles π°
- Reports π
- Research Articles π
- Blogs π°
- Books π
- Lectures π
- People π§βπ»
- Organizations π
- Contributing π€
- License π
- 3-40Wh: Amount of energy consumed for one small to long ChatGPT query (Source, 2025)
- 1L: Estimated amount of water required for 20-100 ChatGPT queries (Source, 2025)
- 2 nuclear plants: Number of nuclear plants to constantly work ot generate enough energy if 80M people generate 5 pages per day (Source, 2025)
- 1 smartphone charge: Amount of energy required to AI generate a couple of images or run a few thousands inference with an LLM (Source, 2024)
- >10s: Time requried to generate 1 HD image with Flux on H100 or to generate 100 tokens with Llama 3 on T4 (Source and Source, 2024)
- 61,848.0x: Difference between the highest and lowest energy use in energy leaderboard for AI models (Source, 2025).
- 1,300MWh: GPT-3, for example, is estimated to use just under 1,300 megawatt hours (MWh) of electricity; about as much power as consumed annually by 130 US homes (Source, 2024)
- 800M users/week: Amount of users using ChatGPT per week in 2025 (Source)
- 1B messages/day: Amount of ChatGPT queries per day in 2025 (Source)
- +160%: Expected increase of data center power consumption by 2030 (Source)
- β€οΈ Pruna β€οΈ: A package to make AI models faster, smaller, faster, greener by combining compression methods (incl. quantization, pruning, caching, compilation, distillation...) on various hardware.
- TensorRT: High-performance deep learning inference library for NVIDIA GPUs.
- ONNX: Open Neural Network Exchange format for interoperability among deep learning frameworks.
- Code Carbon: Library to track energy and carbon efficiency of various hardware.
- LLM Perf: A framework for benchmarking the performance of transformers models with different hardwares, backends and optimizations.
- ML.ENERGY Leaderboard: An initiative to benchmark energy efficiency of AI models.
- AI Energy Score: An initiative to establish comparable energy efficiency ratings for AI models, helping the industry make informed decisions about sustainability in AI development.
- Model Optimization Toolkit: TensorFlow toolkit for optimizing machine learning models for deployment and execution.
- Green Coding: LLM service that you can use to prompt most open source models and see the resource usage.
- EcoLogits: EcoLogits is a python library that tracks the energy consumption and environmental footprint of using generative AI models through APIs.
- Perplexity Kernels: GPU kernels by Perplexity.
- Fast Tokenizer: Fast tokenizer is an efficient and optimized tokenizer engine for llm inference serving.
- WeightWatcher: WeightWatcher (WW) is an open-source, diagnostic tool for analyzing Deep Neural Networks (DNN), without needing access to training or even test data..
- Cockpit: A Practical Debugging Tool for Training Deep Neural Networks.
- Electrictiy Map: A live map showing the origin of the electricity in world regions and their CO2 intensity.
- MLCA: A tool for machine learning life cycle assessment.
- TritonParse: A visualization and analysis tool for Triton IR files, designed to help developers analyze, debug, and understand Triton kernel compilation processes.
- Routing on Random Forests: A framework for training and serving LLM based on random forest-based routers, thus allowing to optimize for costs.
- "Energy and AI Observatory" (2025) - IEA
- "AIβs Impacts, how to limit them, and why" (2025) - Better Tech
- "How much energy does ChatGPT use?" (2025) - Epoch AI
- Data centers et intelligence artificielle : la course au gigantisme (2025) - Le Monde
- "What's the environmental cost of AI?" (2024) - CO2 AI
- "Shrinking the giants: Paving the way for TinyAI" (2024) - Cell Press
- "DeepSeek might not be such good news for energy after all" (2024) - MIT Technology Review
- "AI already uses as much energy as a small country. Itβs only the beginning." (2024) - Vox
- "Quelle contribution du numΓ©rique Γ la dΓ©carbonation ?" (2024) - France StratΓ©gie
- "Les promesses de lβIA grevΓ©es par un lourd bilan carbone" (2024) - Le Monde
- "How much electricity does AI consume?" (2024) - The Verge
- "How do I track the direct environmental impact of my own inference and training when working with AI?" (2024) - Blog
- "Data center emissions probably 662% higher than big tech claims. Can it keep up the ruse?" (2024) - The Guardian
- "Light bulbs have energy ratings β so why canβt AI chatbots?" (2024) - Nature
- "The Environmental Impacts of AI -- Primer" (2024) - Hugging Face
- "The Climate and Sustainability Implications of Generative AI" (2024) - MIT
- "AI's "eye-watering" use of resources could be a hurdle to achieving climate goals, argue experts" (2023) - dezeen
- "How coders can help save the planet?" (2023) - Blog
- "Reducing the Carbon Footprint of Generative AI" (2023) - Blog
- "The MPG of LLMs: Exploring the Energy Efficiency of Generative AI" (2023) - Blog
- "Ecologie numΓ©rique: LβIA durable, entre vΕu pieux et opportunitΓ© de marchΓ©" (2025) - LibΓ©ration
- "A General Framework for Frugal AI" (2025) - AFNOR
- "The 2025 AI Index Report" (2025) - Stanford Human-centered Artificial Intelligence
- "Energy and AI" (2025) - International Energy Agency
- "Key challenges for the environmental performance of AI" (2025) - French Ministry
- "Artificial Intelligence and electricity: A system dynamics approach" (2024) - Schneider
- "Notable AI Models" (2025) - Epoch AI
- "Powering Artificial Intelligence" (2024) - Deloitte
- "Google Sustainability Reports" (2024) - Google
- "How much water does AI consume? The public deserves to know" (2023) - OECD
- "Measuring the environmental impacts of artificial intelligence compute and applications" (2022) - OECD
- "The LLM Engineer Almanac" (2025) - Modal
- "Enhance Your Models in 5 Minutes with the Hugging Face Kernel Hub" (2025) - Hugging Face
- "Reduce, Reuse, Recycle: Why Open Source is a Win for Sustainability" (2025) - Hugging Face
- "Mixture of Experts: When Does It Really Deliver Energy Efficiency?" (2025) - Neural Watt
- "Efficient and Portable Mixture-of-Experts Communication" (2025) - Perplexity
- "Optimizing Tokenization for Faster and Efficient LLM Processing" (2025) - Medium
- "Tensor Parallelism with CUDA - Multi-GPU Matrix Multiplication" (2025) - Substack
- "Automating GPU Kernel Generation with DeepSeek-R1 and Inference Time Scaling" (2025) - Nvidia Developer
- "AI CUDA Engineer" (2025) - Sakana AI
- "The ML/AI Engineer's starter guide to GPU Programming" (2025) - Neural Bits
- "Understanding Quantization for LLMs" (2024) - Medium
- "Don't Merge Your LoRA Adapter Into a 4-bit LLM" (2023) - Substack
- "Matrix Multiplication Background User's Guide" (2023) - Nvidia Developer
- "GPU Performance Background User's Guide" (2023) - Nvidia Developer
- Programming Massively Parallel Processors: A Hands-on Approach (2022), Wen-mei W. Hwu, David B. Kirk, Izzat El Hajj
- Efficient Deep Learning (2022), Gaurav Menghani, Naresh Singh
- AI Efficiency Courses: Slides, Exercises (2025) - Bertrand Charpentier
- Data Compression, Theory and Applications: YouTube, Slides (2024) - Stanford
- MIT Han's Lab (2024) - MIT Lecture by Han's lab
- GPU Mode (2020) - Tutorials by GPU mode community
| Name | Affiliation | Research Interests | Social Media |
|---|---|---|---|
| James Martin | Better Tech | AI Sustainability |  |
| Saleh Ashkboos | ETH Zurich | Quantization | |
| Dan Alistarh | IST Austria | AI Compression | |
| Elias Frantar | OpenAI | Quantization | |
| Tim Dettmers | CMU | Quantization | |
| Song Han | MIT | AI Efficiency | |
| Scott Chamberlin | TBD | AI Efficiency | |
| Benoit Petit | Boavista | Data Center Efficiency | |
| Samuel RincΓ© | Gen AI Impact | AI Efficiency, Sustainability | |
| ThΓ©o Alves Da Costa | Ekimetrics | AI Efficiency, Sustainability | |
| Sasha Luccioni | Hugging Face | AI Sustainability | |
| Anne-Laure Ligozat | ENSIEE | AI Sustainability | |
| Organization | Description | Website |
|---|---|---|
| Data4Good | A platform that connects data scientists with social impact projects to address global challenges using data. | data4good.org |
| Make.org | A global platform that empowers citizens to propose and take action on social and environmental issues through collective projects. | make.org |
| CodeCarbon | A tool that helps track the carbon emissions of machine learning models and optimizes them for sustainability. | codecarbon.io |
| Sustainable AI Coalition | An organization dedicated to advancing sustainability in AI technologies and promoting best practices for green AI. | sustainableaicoalition.org |
| FruitPunch AI | A community that solves AI solutions for impact organizations that contribute to the SDG's. | fruitpunch.ai |
Contributions are welcome! Please follow our contribution guidelines to add new resources or suggest improvements that promote AI efficiency.
This project is licensed under the MIT License. Feel free to share and use the resources as needed.
