Global AI and Data Science

 View Only

Estimating ChatGPT's Carbon Footprint

By Rachana Vishwanathula posted Thu May 04, 2023 05:17 PM


Carbon Footprint Calculation for ML Models:

I have explained in my earlier blog about how to calculate carbon footprint for ML models. Here's the link. Let's use a similar approach and try estimating the carbon footprint of ChatGPT. The carbon footprint is calculated using two things - 

1. The amount of electricity consumed by ML models: This depends on the hardware that is used by the model and utilization rate of the hardware
2. The carbon intensity of this ML models: This depends on how the electricity is produced. For example, solar and wind energies are greener than coal. To quantify this, we look at the location where the hardware is situated at and use the average carbon intensity of electricity in the grid in that location. I have explained this in detail in my previous blog


Now let's understand what hardware ChatGPT used and accordingly calculate the carbon footprint. Here are some of the assumptions I will make to calculate this.
1. Let's assume that ChatGPT uses NIVIDIA A100 40/80 GB GPUs since the similar LLM models like Google Bloom used the same. ChatGPT uses GPT-3 and it's believed that it uses 30,000 GPUs. Here's a link that estimates that. 
2. OpenAI is headquartered in San Fransisco so lets  assume that they are running these servers in the same region.
3. Let's assume that ChatGPT uses Azure cloud since OpenAI has a partnership with Microsoft. On Azure, the data center that's used becomes West US. 

ML CO2 Impact is a tool which calculates carbon footprint. So, if I calculate carbon footprint using above assumptions with ML CO2 Impact, here's what I get for 24 hours for one GPU. 

If it's 1.44 kg co2e for one GPU, and since GPT3 is estimated to use 30000 GPUs, 1.44*30000 i.e., 43200 kg co2e for 24 hours. 
ChatGPT was launched on November 30, 2023. It's 157 days from then making the carbon footprint 157 * 43200 kg co2e i.e., 6782400 kg co2e. 

In conclusion, I estimate that ChatGPT's carbon footprint is 6782.4 tons co2e. This value is just the number for running and maintaining ChatGPT but training GPT3 would have different number. Co2e is one measure within sustainability but there are many other factors like water consumption to be taken into consideration if executive sustainability must be calculated.