What is the SEC and How is it Related to the Environment?
9 min
Level
How does the SEC, otherwise known as the U.S. Securities and Exchange Commission, help the United States regulate environmental policies?
The voice of impact
ESG / CSR
Industries
Greenly solutions
US
US
The Environmental Impact of Artificial Intelligence
Level
New generation artificial Intelligence chatbots like ChatGPT hit the mainstream recently, and due to their increasingly advanced capabilities (ChatGPT can produce incredibly accurate and realistic copy), and their ease of access (anyone can use ChatGPT for free), they’re being used for everything from homework to drafting legal briefs.
Some are even predicting that they’ll eventually change the shape of our job market.
“ But their increased use isn’t without a cost, AI technology and chatbots leave behind a significant carbon footprint. ”
At this point in time, ChatGPT is probably the most widely known and used chatbot on the market. Its technology is based on GPT-3 which is the 3rd generation of OpenAI’s GPT language model. Because of this, it is difficult to calculate the precise percentage of GPT3 emissions that should be allocated to ChatGPT.
For this reason, we’ll look at GPT-3’s carbon footprint, and not ChatGPT’s, but since they’re very similar it provides a good estimate.
A few things before starting
Before we dig into the numbers, let's take a look at the difference between GPT-3 and ChatGPT:
- ChatGPT is designed for conversational AI tasks - it generates text based on its trained parameters. Whereas GPT-3 is used for general predictive modelling tasks that do not require real-time responses.
- Another difference is that ChatGPT has an interactive interface that is accessible to all and free of charge (for now at least), while GPT-3 does not.
In order to assess the carbon emissions relating to GPT-3, Greenly has focused on a realistic action that a business might employ GPT-3 for.
In this case we’ll look at the emissions of GPT-3 technology where it is used to “reply automatically to 1 million emails per month, over the course of 1 year”.
There are two stages to the use of GPT-3 in this scenario: training and use. Both stages have their own carbon footprint that must be considered to accurately calculate the emissions for this scenario.
Training GPT-3
GPT-3 first needs to be trained in order to become competent and efficient at writing appropriate replies to emails - something that takes many hours. In this scenario we can assume that the training would take place in a classic data centre, resulting in the following emissions.
- In order to effectively train GPT-3 in a classic data centre, a variety of equipment would need to be used (for example: servers, cooling and lighting). The running of this equipment over a one year period translates into a total of 160268 kgCO2e (kilograms of carbon dioxide equivalent).
- Refrigerant gas leaks over the period of the year can be estimated to produce a total of 9602 kgCO2e.
- The manufacturing of servers used for the purpose of training GPT-3 to carry out the task would result in emissions totalling 68889 kgCO2e (this is based on a lifespan of 4 years and 100% workload).
“ The total emissions resulting from the training of GPT-3 over the course of 1 year total 238 759 kgCO2e. ”
GPT-3 Use
Once GPT-3 has been effectively trained, the technology can then be employed to actually carry out the task of answering 1 million emails per month. This action would result in the following carbon emissions.
- For GPT-3 to respond to 1 million emails this would take around 38.8 hours. Taking into consideration its electricity consumption over this time-frame, as well as cooling and servers manufacturing, this activity will result in emissions totalling 9.3 kgCO2e.
- The next step in the process is then to transfer the results of GPT-3’s calculations. The emissions from this activity works out at 11.8 kgCO2e.
- We also need to consider that data would need to be stored in order to train the GPT-3 model: 85.2 kgCO2e.
“ The total emissions stemming from the use of GPT-3 over a 1 year period therefore totals 1277 kgCO2e. ”
GPT-3's total carbon footprint
“ Based on these calculations, for GPT-3 to answer 1 million emails per month, over a 1 year period, this would result in a total of 240036 kgCO2e (240 tCO2e). This correlates to a total of 136 round trips between Paris and New York! ”
It should also be noted that if GPT-3 were to be used to respond to emails in more than one language, it would be necessary to re-train the technology for each language.
This means that the total emissions would need to be multiplied by the number of languages required. Where a company operates in a number of different languages, the resulting carbon emissions are significant.
Key findings
Training is the most carbon intensive stage. Greenly’s calculation has revealed that the training of GPT-3 results in significantly higher carbon emissions than its actual use - in fact carbon emissions are as much as 230 times higher.
The choice of data centre strongly influences emissions. For example, a classic data centre produces 40% more emissions than an optimised data centre. The country in which the data centre is located will also play a part: Greenly’s calculation is based on a data centre in France, however, if the calculation had instead been calculated based on a data centre in the US, this would have resulted in emissions that were 6 times higher.
GPT-3 produces higher carbon emissions than other comparable AIs. If we compare GPT-3 with other AI technology carrying out the exact same task, it’s clear that GPT-3 results in significantly higher emissions. This is largely because the number of GPU hours (ie. the sum of the duration of each individual GPU that’s been used for deep learning) for pre-training GPT-3 is between 100 and 30000 times higher than for the other AIs, which means more electricity, more servers, etc. This is because GPT-3 has an incredibly high number of parameters (ie. the variables that are input to allow the AI to make predictions) - 175 billion to be exact.
Conclusion
New AI technology such as GPT-3 and ChatGPT are an exciting advancement and something that many believe has the potential to re-shape the future of how we work and how different industries operate.
However, they’re not flawless and further advances in this field can be expected. Already, Chinese firms such as Huawei and Inspur are working on AIs with more than 200 billion parameters.
“ As the technology advances and we see increasing uptake across industries, the resulting carbon footprint will also grow. ”
Plus d'articles
View allShare
Subscribe to the newsletter