The carbon and water footprints of generative AI
Some estimates of the footprints.
➤ The causes for the footprints include:
👉Electricity to run the servers
👉Water to cool down the servers
👉Both are needed during the training of the models and when users interact with it
➤Estimated electricity consumption
🌏Training GPT3 is estimated to have emitted about 502 tonnes of CO2. That's equivalent to 27 US lives. [1]
🌏User interactions with ChatGPT in January 2023 is estimated to have consumed as much electricity as 175,000 people in January 2023. [2]
➤Estimated water consumption
🌏 If GPT3 was trained in the US, the estimated water consumption for training is 3.5 million liters. If it was trained in Asia, the water consumption is 4.9 million liters [3]
🌏Each interaction with the ChatGPT consumes water. For example, for GPT3, an interaction with 20-30 questions consumes 500ml water. [3]
---
References:
[1] Carbon emissions for GPT3 training
(page 120)
[2] Electricity usage of ChatGPT interactions in Jan 2023
[3] Water consumption