Mistral AI's recent report reveals the environmental impact of its Mistral Large 2 LLM, focusing on greenhouse gas emissions, water consumption, and materials use. Training and running the model contributed significantly to its carbon footprint, with 85.5 percent of GHG emissions and 91 percent of water consumption attributed to these activities. The training phase alone produced about 20 kilotons of CO2 equivalents and utilized roughly 281,000 cubic meters of water. Notably, 29 percent of materials consumption occurred during the training process, indicating hardware failures, while running the model generated lesser emissions and required minimal water per request.
Training the 123 billion parameter model produced approximately 20 kilotons of CO2 equivalents and consumed 281,000 cubic meters of water, equivalent to roughly 112 Olympic-sized swimming pools.
In the 18 months since Mistral started work on the model, training and running it accounted for 85.5 percent of GHG emissions and 91 percent of water consumption.
Less than two-thirds of materials consumption was attributed to manufacturing, transportation, and end-of-life, while 29 percent occurred during the training and inference stage.
Running the completed model generated far fewer CO2 equivalents and consumed a fraction of the water for each request.
Collection
[
|
...
]