HERO_TEXT = """

🧮 EcoLogits Calculator

EcoLogits is a python library that tracks the energy consumption and environmental footprint of using generative AI models through APIs.

This tool is developed and maintained by [GenAI Impact](https://genai-impact.org/) non-profit. Learn more about 🌱 EcoLogits by reading the documentation on [ecologits.ai](https://ecologits.ai). 🩷 Support us by giving a ⭐️ on our [GitHub repository](https://github.com/genai-impact/ecologits) and by following our [LinkedIn page](https://www.linkedin.com/company/genai-impact/). """ ABOUT_TEXT = r""" ## 🎯 Our goal **The main goal of the EcoLogits Calculator is to raise awareness on the environmental impacts of LLM inference.** The rapid evolution of generative AI is reshaping numerous industries and aspects of our daily lives. While these advancements offer some benefits, they also **pose substantial environmental challenges that cannot be overlooked**. Plus the issue of AI's environmental footprint has been mainly discussed at training stage but rarely at the inference stage. That is an issue because **inference impacts for LLMs can largely overcome the training impacts when deployed at large scales**. At **[GenAI Impact](https://genai-impact.org/) we are dedicated to understanding and mitigating the environmental impacts of generative AI** through rigorous research, innovative tools, and community engagement. Especially, in early 2024 we have launched an new open-source tool called [EcoLogits](https://github.com/genai-impact/ecologits) that tracks the energy consumption and environmental footprint of using generative AI models through APIs. ## 🙋 FAQ **How we assess the impacts of closed-source models?** Environmental impacts are calculated based on model architecture and parameter count. For closed-source models, we lack transparency from providers, so we estimate parameter counts using available information. For GPT models, we based our estimates on leaked GPT-4 architecture and scaled parameters count for GPT-4-Turbo and GPT-4o based on pricing differences. For other proprietary models like Anthropic's Claude, we assume similar impacts for models released around the same time with similar performance on public benchmarks. Please note that these estimates are based on assumptions and may not be exact. Our methods are open-source and transparent so you can always see the hypotheses we use. **Which generative AI models or providers are supported?** To see the full list of **generative AI providers** currently supported by EcoLogits, see the following [documentation page](https://ecologits.ai/providers/). As of today we only support LLMs but we plan to add support for embeddings, image generation, multi-modal models and more. If you are interested don't hesitate to [join us](https://genai-impact.org/contact/) and accelerate our work! **How to reduce AI environmental impacts?** * Look at **indirect impacts** of your project. Does the finality of your project is impacting negatively the environment? * **Be frugal** and question your usage or need of AI * Do you really need AI to solve your problem? * Do you really need GenAI to solve your problem? (you can read this [paper](https://aclanthology.org/2023.emnlp-industry.39.pdf)) * Use small and specialized models to solve your problem. * Evaluate before, during and after the development of your project the environmental impacts with tools like 🌱 [EcoLogits](https://github.com/genai-impact/ecologits) or [CodeCarbon](https://github.com/mlco2/codecarbon) (see [more tools](https://github.com/samuelrince/awesome-green-ai)) * Restrict the use case and limit the usage of your tool or feature to the desired purpose. * Do NOT buy new GPUs / hardware * Hardware manufacturing for data centers is around 50% of the impact. * Use cloud instances that are located in low emissions / high energy efficiency data centers (see [electricitymaps.com](https://app.electricitymaps.com/map)) * Optimize your models for production * Quantize your models. * Use inference optimization tricks. * Prefer fine-tuning of small and existing models over generalist models. **What is the difference between **EcoLogits** and [CodeCarbon](https://github.com/mlco2/codecarbon)?** EcoLogits is focused on estimating the environmental impacts of generative AI (only LLMs for now) used **through API providers (such as OpenAI, Anthropic, Cloud APIs...)** whereas CodeCarbon is more general tool to measure energy consumption and estimate GHG emissions measurement. If you deploy LLMs locally we encourage you to use CodeCarbon to get real numbers of your energy consumption. ## 🤗 Contributing We are eager to get feedback from the community, don't hesitate to engage the discussion with us on this [GitHub thread](https://github.com/genai-impact/ecologits/discussions/45) or message us on [LinkedIn](https://www.linkedin.com/company/genai-impact/). We also welcome any open-source contributions on 🌱 **[EcoLogits](https://github.com/genai-impact/ecologits)** or on 🧮 **EcoLogits Calculator**. ## ⚖️ License

This work is licensed under CC BY-SA 4.0

## 🙌 Acknowledgement We thank [Data For Good](https://dataforgood.fr/) and [Boavizta](https://boavizta.org/en) for supporting the development of this project. Their contributions of tools, best practices, and expertise in environmental impact assessment have been invaluable. We also extend our gratitude to the open-source contributions of 🤗 [Hugging Face](huggingface.com) on the LLM-Perf Leaderboard. ## 🤝 Contact For general question on the project, please use the [GitHub thread](https://github.com/genai-impact/ecologits/discussions/45). Otherwise use our contact form on [genai-impact.org/contact](https://genai-impact.org/contact/). """ METHODOLOGY_TEXT = r""" ## 📖 Methodology We have developed a methodology to **estimate the energy consumption and environmental impacts for an LLM inference** based on request parameters and hypotheses on the data center location, the hardware used, the model architecture and more. In this section we will only cover the principles of the methodology related to the 🧮 **EcoLogits Calculator**. If you wish to learn more on the environmental impacts modeling of an LLM request checkout the 🌱 [EcoLogits documentation page](https://ecologits.ai/methodology/). ### Modeling impacts of an LLM request The environmental impacts of an LLM inference are split into the **usage impacts** $I_{request}^u$ to account for electricity consumption and the **embodied impacts** $I_{request}^e$ that relates to resource extraction, hardware manufacturing and transportation. In general terms it can be expressed as follow: $$ I_{request} = I_{request}^u + I_{request}^e $$ $$ I_{request} = E_{request}*F_{em}+\frac{\Delta T}{\Delta L}*I_{server}^e $$ With, * $E_{request}$ the estimated energy consumption of the server and its cooling system. * $F_{em}$ the electricity mix that depends on the country and time. * $\frac{\Delta T}{\Delta L}$ the hardware usage ratio i.e. the computation time over the lifetime of the hardware. * $I_{server}^e$ the embodied impacts of the server. Additionally, to ⚡️ **direct energy consumption** the environmental impacts are expressed in **three dimensions (multi-criteria impacts)** that are: * 🌍 **Global Warming Potential** (GWP): Potential impact on global warming in kgCO2eq (commonly known as GHG/carbon emissions). * 🪨 **Abiotic Depletion Potential for Elements** (ADPe): Impact on the depletion of non-living resources such as minerals or metals in kgSbeq. * ⛽️ **Primary Energy** (PE): Total energy consumed from primary sources in MJ. ### Principles, Data and Hypotheses We use a **bottom-up methodology** to model impacts, meaning that we will estimate the impacts of low-level physical components to then estimate the impacts at software level (in that case an LLM inference). We also rely on **Life Cycle Approach (LCA) proxies and approach** to model both usage and embodied phases with multi-criteria impacts. If you are interested in this approach we recommend you to read the following [Boavizta](https://boavizta.org/) resources. * [Digital & environment: How to evaluate server manufacturing footprint, beyond greenhouse gas emissions?](https://boavizta.org/en/blog/empreinte-de-la-fabrication-d-un-serveur) * [Boavizta API automated evaluation of environmental impacts of ICT services and equipments](https://boavizta.org/en/blog/boavizta-api-automated-evaluation-of-ict-impacts-on-the-environment) * [Boavizta API documentation](https://doc.api.boavizta.org/) We leverage **open data to estimate the environmental impacts**, here is an exhaustive list of our data providers. * [LLM-Perf Leaderboard](https://huggingface.co/spaces/optimum/llm-perf-leaderboard) to estimate GPU energy consumption and latency based on the model architecture and number of output tokens. * [Boavizta API](https://github.com/Boavizta/boaviztapi) to estimate server embodied impacts and base energy consumption. * [ADEME Base Empreinte®](https://base-empreinte.ademe.fr/) for electricity mix impacts per country. Finally here are the **main hypotheses** we have made to compute the impacts. * ⚠️ **We *"guesstimate"* the model architecture of proprietary LLMs when not disclosed by the provider.** * Production setup: quantized models running on data center grade servers and GPUs such as A100. * Electricity mix does not depend on time (help us enhance EcoLogits and work on this [issue](https://github.com/genai-impact/ecologits/issues/42)) * Ignore the following impacts: unused cloud resources, data center building, network and end-user devices... (for now) ## Equivalents We have integrated impact equivalents to help people better understand the impacts and have reference points for standard use cases and everyday activities. ### Request impacts These equivalents are computed based on the request impacts only. #### 🚶‍♂️‍➡️ Walking or 🏃‍♂️‍➡️ running distance We compare the ⚡️ direct energy consumption with the energy consumption of someone 🚶‍♂️‍➡️ walking or 🏃‍♂️‍➡️ running. From [runningtools.com](https://www.runningtools.com/energyusage.htm) we consider the following energy values per physical activity (for someone weighing 70kg): * 🚶‍♂️‍➡️ walking: $ 196\ kJ/km $ (speed of $ 3\ km/h $) * 🏃‍♂️‍➡️ running: $ 294\ kJ/km $ (speed of $ 10\ km/h $) We divide the request energy consumption by these values to compute the distance traveled. #### 🔋 Electric Vehicle distance We compare the ⚡️ direct energy consumption with the energy consumer by a EV car. From [selectra.info](https://selectra.info/energie/actualites/insolite/consommation-vehicules-electriques-france-2040) or [tesla.com](https://www.tesla.com/fr_fr/support/power-consumption) we consider an average value of energy consumed per kilometer of: $ 0.17\ kWh/km $. We divide the request energy consumption by this value to compute the distance driven by an EV. #### ⏯️ Streaming time We compare the 🌍 GHG emissions of the request and of streaming a video. From [impactco2.fr](https://impactco2.fr/outils/comparateur?value=1&comparisons=streamingvideo), we consider that $ 1\ kgCO2eq $ is equivalent to $ 15.6\ h $ of streaming. We multiply that value by the GHG emissions of the request to get an equivalent in hours of video streaming. ### Scaled impacts These equivalents are computed based on the request impacts scaled to a worldwide adoption use case. We imply that the same request is done 1% of the planet everyday for 1 year, and then compute impact equivalents. $$ I_{scaled} = I_{request} * [1 \\% \ \text{of}\ 8B\ \text{people on earth}] * 365\ \text{days} $$ #### Number of 💨 wind turbines or ☢️ nuclear plants We compare the ⚡️ direct energy consumption (scaled) by the energy production of wind turbines and nuclear power plants. From [ecologie.gouv.fr](https://www.ecologie.gouv.fr/eolien-terrestre) we consider that a $ 2\ MW $ wind turbine produces $ 4.2\ GWh $ a year. And from [edf.fr](https://www.edf.fr/groupe-edf/espaces-dedies/jeunes-enseignants/pour-les-jeunes/lenergie-de-a-a-z/produire-de-lelectricite/le-nucleaire-en-chiffres) we learn that a $ 900\ MW $ nuclear power plant produces $ 6\ TWh $ a year. We divide the scaled energy consumption by these values to get the number of wind turbines or nuclear power plants needed. #### Multiplier of 🇮🇪 Ireland electricity consumption We compare the ⚡️ direct energy consumption (scaled) by the electricity consumption of Ireland per year. From [wikipedia.org](https://en.wikipedia.org/wiki/List_of_countries_by_electricity_consumption) we consider the Ireland electricity consumption to be $ 33\ TWh $ a year for a population of 5M. We divide the scaled energy consumption by this value to get the equivalent number of "Ireland countries". #### Number of ✈️ Paris ↔ New York City flights We compare the 🌍 GHG emissions (scaled) of the request and of a return flight Paris ↔ New York City. From [impactco2.fr](https://impactco2.fr/outils/comparateur?value=1&comparisons=&equivalent=avion-pny) we consider that a return flight Paris → New York City → Paris for one passenger emits $ 1,770\ kgCO2eq $ and we consider an overall average load of 100 passengers per flight. We divide the scaled GHG emissions by this value to get the equivalent number of return flights. **If you are motivated to help us test and enhance this methodology [contact us](https://genai-impact.org/contact/)!** 💪 """ CITATION_LABEL = "BibTeX citation for EcoLogits Calculator and the EcoLogits library:" CITATION_TEXT = r"""@misc{ecologits-calculator, author={Samuel Rincé, Adrien Banse and Valentin Defour}, title={EcoLogits Calculator}, year={2024}, howpublished= {\url{https://huggingface.co/spaces/genai-impact/ecologits-calculator}}, } @software{ecologits, author = {Samuel Rincé, Adrien Banse, Vinh Nguyen and Luc Berton}, publisher = {GenAI Impact}, title = {EcoLogits: track the energy consumption and environmental footprint of using generative AI models through APIs.}, }""" LICENCE_TEXT = """

This work is licensed under CC BY-SA 4.0

"""