In a two-part series MIT News Explore the impact of generative AI on the environment. This article explores why this technology is so resource-intensive. In the second part, we will examine what experts are doing to reduce genAI’s carbon footprint and other impacts.
From improving worker productivity to advancing scientific research, the excitement surrounding the potential benefits of generative AI is hard to ignore. While the explosive growth of this new technology has enabled the rapid deployment of powerful models across many industries, the environmental impacts of this generative AI “gold rush” are difficult to identify, let alone mitigate.
The computational power needed to train generative AI models, such as OpenAI’s GPT-4, which often have billions of parameters, can require enormous amounts of power, which increases carbon dioxide emissions and increases pressure on the power grid.
Additionally, deploying these models in real-world applications, enabling millions of people to use generative AI in their everyday lives, and fine-tuning the models to improve performance, draws a large amount of energy long after the models have been developed.
In addition to electricity demands, large amounts of water are required to cool the hardware used to train, deploy, and fine-tune generative AI models. This can strain city water supplies and disrupt local ecosystems. As the number of generative AI applications increases, it also fuels demand for high-performance computing hardware, adding to the indirect environmental impact of manufacturing and transportation.
“When we think about the impact of generative AI on the environment, it’s not just about the electricity we consume when we plug in our computers. There are much broader consequences that go down to the system level and persist based on the actions we take,” he says. Elsa A. Olivetti is Professor of Materials Science and Engineering and Decarbonization Mission Director for MIT’s New Climate Project.
Olivetti is lead author of the 2024 paper “Climate and Sustainability Impacts of Generative AI.” This paper was co-authored by MIT colleagues in response to an Institute-wide call for papers exploring the transformative potential of generative AI, both positive and negative. Direction for society.
demanding data centers
The power demand of data centers is one of the key factors affecting the environmental impact of generative AI. That’s because our data centers are used to train and run deep learning models that underlie popular tools like ChatGPT and DALL-E.
A data center is a temperature-controlled building that houses computing infrastructure such as servers, data storage drives, and network equipment. For example, Amazon has more than 100 data centers worldwide, each containing approximately 50,000 servers that the company uses to power its cloud computing services.
Although data centers have been around since the 1940s (the first was built at the University of Pennsylvania in 1945 to support ENIAC, the first general-purpose digital computer), the rise of generative AI has dramatically accelerated the pace of data center construction.
“The difference with generative AI is the power density required. “Basically, it’s just computing, but a generative AI training cluster can consume seven to eight times more energy than a typical computing workload,” said Noman Bashir, Computing and Climate Impact Fellow at MIT Climate and lead author of the Impact paper. says: and Postdoctoral Fellow at the Sustainability Consortium (MCSC) and the Computer Science and Artificial Intelligence Laboratory (CSAIL).
Scientists estimated that power requirements for North American data centers increased from 2,688 megawatts at the end of 2022 to 5,341 megawatts at the end of 2023, driven in part by generated AI demand. Globally, data center power consumption will increase to 460 terawatts in 2022. This would make data centers the world’s 11th largest electricity consumer, between the countries Saudi Arabia (371 terawatts) and France (463 terawatts). Organization for Economic Cooperation and Development.
By 2026, power consumption in data centers is expected to reach 1,050 terawatts (this would push data centers to fifth place on the global list, between Japan and Russia).
Although not all data center computations involve generative AI, this technology has been a key driver of increased energy demand.
“The demand for new data centers cannot be met in a sustainable way. The speed at which companies are building new data centers means that much of the electricity needed to power them will have to come from fossil fuel-based power plants,” says Bashir.
The performance required to train and deploy models like OpenAI’s GPT-3 is difficult to determine. In a 2021 research paper, scientists from Google and the University of California, Berkeley, estimated that the training process alone consumes 1,287 megawatt hours (enough to power about 120 average U.S. homes for a year) and produces about 552 tons of carbon dioxide. I did it.
All machine learning models must be trained, but one of the unique challenges of generative AI is the rapid fluctuations in energy use that occur at different stages of the training process, explains Bashir.
Grid operators must have a way to absorb these fluctuations to protect the grid, and they typically use diesel-based generators for that task.
Increased impact of inferences
Energy needs do not disappear once a generative AI model is trained.
Every time an individual uses a model, such as requesting an email summary from ChatGPT, the computing hardware that performs that task consumes energy. Researchers estimated that ChatGPT queries consume about five times more power than a simple web search.
“But everyday users don’t think too much about it,” says Bashir. “The ease of use of the generative AI interface and the lack of information about the impact of my actions on the environment mean that as a user, I have little incentive to reduce my use of generative AI.”
For traditional AI, energy usage is split fairly evenly between data processing, model training, and inference (the process of using a trained model to make predictions on new data). However, Bashir says the power demands of generative AI inference are such that these models are becoming more common in a very large number of applications, and he expects the power required for inference to increase as future versions of the models become larger and more complex.
Additionally, generative AI models have a particularly short lifespan due to the increasing demand for new AI applications. Companies release new models every few weeks, so the energy used to train older versions is wasted, Bashir adds. Because new models typically have more parameters than previous models, they often consume more energy to train.
Although the power demands of data centers receive most attention in the research literature, the amount of water these facilities consume also has environmental impacts.
Coolant is used to cool data centers by absorbing heat from computing equipment. Bashir said it is estimated that for every kWh of energy a data center consumes, it requires two liters of water for cooling.
“Just because this is ‘cloud computing’ doesn’t mean the hardware is in the cloud. “Data centers exist in our physical world and have direct and indirect impacts on biodiversity due to their water use,” he says.
The computing hardware inside a data center itself has a less direct impact on the environment.
It’s difficult to estimate how much power is needed to manufacture GPUs, a type of powerful processor capable of handling intensive generative AI workloads, but because the manufacturing process is more complex, there is more power than would be needed to produce simple CPUs. You need this. The carbon footprint of GPUs is further complicated by emissions associated with transporting materials and products.
There are also environmental impacts to obtaining the raw materials used to manufacture GPUs, which can include dirty mining procedures and the use of toxic chemicals for processing.
Market research firm TechInsights estimates that the three major manufacturers (NVIDIA, AMD, and Intel) shipped 3.85 million GPUs to data centers in 2023, up from about 2.67 million in 2022. This number is expected to increase at an even greater rate in 2024. .
The industry is heading down an unsustainable path, but there are ways to encourage responsible development of generative AI that supports environmental goals, Bashir says.
He, Olivetti, and their MIT colleagues argue that a comprehensive consideration of all the environmental and social costs of generative AI is needed, as well as a detailed assessment of the value of its perceived benefits.
“We need a more contextualized way to systematically and comprehensively understand the impact of new developments in this field. Because of the speed at which improvements are being made, we haven’t had a chance to catch up on our ability to measure and understand our strengths and weaknesses,” says Olivetti.
