With the arrival of chatbots ChatGPT, Microsoft Bing and soon Google Bard, the question of the carbon footprint of artificial intelligence technologies is resurfacing. Bloomberg estimates that updating a simple AI model could consume the equivalent of 100 US homes over a year. An estimate that grows from year to year, as artificial intelligence models improve, accumulate knowledge and take up more space on machines and servers.
Power-hungry chatbots
Training the GPT-3 model, which is just a general-purpose AI program, would have consumed 1,287 gigawatt hours, according to a research paper published in 2021, or about as much electricity as 120 American homes. over a year. This training would also have generated 502 tons of carbon emissions, as much as 110 American cars emit in a year. According to the researchers, this initial learning would represent only about 40% of the energy consumed by the models. And these are multiplying at high speed.
In addition, these AI models are improving, growing, and logically consuming more than when they started. For example, OpenAI’s GPT-3 model now uses 175 billion parameters, or variables, learned during its training, compared to just 1.5 billion for its predecessor. The models must be updated regularly to be aware of the latest world events, for example, and we know that OpenAI is already working on GPT-4.
Google’s AI projects would consume as much as Atlanta
Another relative estimate comes from Google, whose researchers found that artificial intelligence—across projects, not just chatbots—accounted for 10-15% of the company’s total electricity consumption, which was 18 .3 terawatt hours in 2021. That would mean Google’s AI is burning about 2.3 terawatt hours per year, or about as much electricity each year as a city like Atlanta. But here again, the discretion of large companies on the subject does not allow very precise estimates to be made in terms of energy consumption. And this consumption is not likely to stop given the unbridled evolution of AI.
As for the solutions intended to make all this greener, they seem at first glance quite futile. An energy consultant quoted by Bloomberg limits itself to mentioning the possibility of scheduling updates for chatbots and other AI models at times when energy is cheaper, or in surplus…