As artificial intelligence (AI) continues to advance, it has become imperative to consider the environmental impact of AI systems. One prominent AI model that has gained widespread usage is ChatGPT, a language model developed by OpenAI. ChatGPT, like many other AI models, has a carbon footprint associated with its computational processes, which includes training, deployment, and maintenance. This article will provide a factual overview of ChatGPT’s carbon footprint, including its energy consumption, emissions, and efforts to mitigate its environmental impact.
Understanding the carbon footprint of ChatGPT is crucial for evaluating the sustainability of AI models and making informed decisions about their usage in the context of environmental sustainability. So first, let us learn more about ChatGPT.
What is ChatGPT?
The San Francisco-based startup OpenAI created the AI chatbot ChatGPT. Elon Musk and Sam Altman co-founded OpenAI in 2015, and well-known investors, most notably Microsoft, support it. ChatGPT uses deep learning techniques to generate text-based responses to prompts or queries. It is part of the GPT (Generative Pre-trained Transformer) family of models, which are trained on vast amounts of text data to learn patterns and generate coherent and contextually relevant responses.
The LLM of ChatGPT is known as GPT-3.5. The GPT-3 language model from OpenAI has been upgraded. GPT-3 is one of the most significant and most potent language-processing AI models available today, with a massive 175 billion parameters.
ChatGPT is designed to engage in conversations with users, simulating human-like interactions. It can be used for a wide range of applications, including customer service, virtual assistants, content generation, and more. ChatGPT is capable of understanding and generating text in multiple languages, making it a versatile tool for natural language processing tasks.
It’s important to note that ChatGPT is an AI language model and does not possess real-time cognitive abilities or consciousness. It generates responses based solely on the patterns it has learned from the data it was trained on and does not have its own opinions, emotions, or awareness of the world.
What is Carbon Footprint?
The total carbon dioxide (CO2) emissions caused by a person’s or an organization’s actions are known as their “carbon footprint” (e.g., building, corporation, country, etc.). It comprises direct emissions from producing electricity used to power the consumption of products and services and indirect emissions from burning fossil fuels in manufacturing, heating, and transportation. Moreover, the idea of a carbon footprint frequently considers other greenhouse gas emissions, like methane, nitrous oxide, or chlorofluorocarbons (CFCs).
What Is The Carbon Footprint of ChatGPT?
When calculating a machine learning model ChatGPT’s carbon footprint, there are three categories to consider: the carbon footprint from training the model, the carbon footprint from using the model to make inferences after deployment, and a model’s entire life cycle carbon footprint.
No matter the scale, two things are necessary to determine any model’s carbon footprint:
1. The quantity of electricity it uses depends a lot on the hardware it uses and how much of that hardware is being used.
2. The amount of carbon this electricity contains largely relies on how the electricity is generated; solar and wind energy are obviously far greener than coal. The typical carbon intensity of the power in the grid where the hardware is located is frequently used to quantify this.
The Carbon Footprint From The Training of ChatGPT
If understood correctly, a GPT-3 variant serves as the foundation for ChatGPT. According to estimates, training GPT-3 used 1,287 MWh and produced 552 tonnes of CO2.
But, it is only sometimes evident how one would attribute some of these emissions to ChatGPT. These should be assigned to ChatGPT as well. Also, ChatGPT was trained using reinforcement learning, although we are not aware of any reasonable substitutes or information on this training process that is pertinent.
The Carbon Footprint From Running ChatGPT
We can calculate ChatGPT’s daily carbon footprint to be 23.04 kgCO2 using the ML CO2 Impact calculator. The daily carbon footprint of ChatGPT is around 0.2 percent of the yearly carbon footprint of a Dane, who emits 11 tonnes of CO2 on average per year. The carbon footprint of ChatGPT, if it ran for a year, would be 365 * 23.04 kg = 8.4 tonnes, or nearly 76% of the carbon footprint of a Dane.
The estimated 384 GPU hours per day, or 16.4 GPUs multiplied by 24 hours, yielded daily CO2 emissions of 23.04 kg. Although it’s not immediately evident, one believes ML CO2 Impact assumes constant 100% hardware utilization, which may be a reasonable assumption in this case, considering the allegedly high load the service is under.
The daily emissions of 23.04 kgCO2 over 18 days would total 414 kgCO2 for ChatGPT. BLOOM, in contrast, released 360 kg over 18 days. Given how closely the two estimates match up, 23.04 kgCO2 is okay. The disparity between the two emission estimates can be attributed to various factors, such as the power produced by BLOOM and ChatGPT having different carbon intensities.
ChatGPT and BLOOM
It is also important to note that BLOOM processed 230,768 requests over 18 days, or 12,820 on average daily. ChatGPT would receive the same number of requests per day as BLOOM did during that time if 1.2% of its 1 million users sent one request every day. ChatGPT has a more significant carbon footprint because it processes many more daily requests, at least based on the amount of conversation about it in traditional and social media outlets. Although there is a lot of uncertainty surrounding this estimate because it is built on some shaky assumptions, it seems fair compared to detailed calculations of the carbon footprint of BLOOM, a comparable language model.
How Can ChatGPT Be Used Sustainably?
We must use AI environmentally and sustainably as technology gets increasingly ingrained in our daily lives.
Fortunately, there are several strategies for using ChatGPT sustainably:
1. It may be used to examine enormous amounts of environmental data and offer insights for environmental studies, including discovering climatic patterns and forecasting natural disasters.
2. Automating time-consuming and repetitive operations like monitoring and reporting on environmental conditions can free up human resources for more crucial work.
3. By eliminating the need for printing and mailing physical copies, which have a high carbon footprint, it can be used to create reports and presentations that are environmentally friendly.
4. It can be used to create content that encourages eco-friendly behaviour and raises awareness of environmental issues, such as articles, social media posts, and educational materials.
The Bottom Line
One of the most comprehensive language models now in use, ChatGPT has over 175 billion parameters. This indicates that ChatGPT uses a significant amount of energy. OpenAI is making great efforts to ensure that ChatGPT has a negligible effect on the environment. Using renewable energy to power the data centres where it is stored and processed lessens ChatGPT’s carbon footprint. Moreover, ChatGPT’s energy usage is optimized through several methods, including shrinking the model’s size and utilizing more energy-efficient technology.
In addition to CO2 emissions, it’s essential to consider other environmental effects, such as water use, air pollution, soil contamination, etc. The AI sector and the larger IT community must think about the environmental impact of their work and take action to reduce any unfavourable effects. By doing this, we can contribute to ensuring that future generations may profit from AI sustainably.