LLMImpact

CO2 emissions calculator for LLMs.

The impact of artificial intelligence, especially large language models (LLMs) on the environment is not minor and is one of the major concerns surrounding the current revolution in the application of AI to all areas of life.

In this project we want to build a calculator of the CO2 expenditure produced by an execution of a Deep Learning model, specifically an LLM. The calculator will take the task as input (if it is a simple conversation, or if it is a search and also includes a RAG in its execution), calculate the number of tokens expected (input + output) and translate them into CO2 with a simple conversion. This token calculation is an inexpensive process, compared to the execution of the complete model, which can be done on CPU.

The student will have to do a research on sources of CO consumption linked to machine learning, such as the one done around a DeepMind paper [1], adapting the calculation to the current models, the state of the art and the hardware used for its execution.

The tool will be implemented as a backend with its API, and then it can be consumed through a frontend, as a webapp that will be custom developed and where the user can select their LLMs, the tasks to be performed and calculate the impact of CO2 emissions. On the other hand, it is also desired that the calculation service (backend) is usable from third party tools and can be integrated into existing frontends, so the web service must be designed with these parameters in mind and not only to serve the developed web application.

  1. https://152334h.github.io/blog/scaling-exponents/

TOP