In a two-part series, MIT News Explores the environmental implications of generic AI. In this article, we will look at why this technology is so resource-intensive. The second part will examine what experts are doing to reduce the carbon footprint and other impacts of GenAI.
From improving worker productivity to advancing scientific research, it’s hard to ignore the excitement over the potential benefits of generic AI. While the explosive growth of this new technology has enabled the rapid deployment of powerful models across many industries, it is difficult to even mitigate, let alone mitigate, the environmental consequences of this generative AI “gold rush.”
The computational power required to train generative AI models, which often have billions of parameters, such as OpenAI’s GPT-4, can demand a staggering amount of electricity, leading to carbon dioxide emissions and pressure on the electric grid. goes.
Furthermore, deploying these models in real-world applications enables millions of people to use generic AI in their daily lives, and then fine-tuning the models to improve their performance long after a model has been developed. Draws large amounts of energy over time.
In addition to power demand, generators used to cool the hardware used for training, deploying, and fine-tuning AI models require a lot of water, which can impact municipal water supplies and local energy costs. Can disrupt the ecosystem. The growing number of generic AI applications has also increased the demand for high-performance computing hardware, adding indirect environmental impacts from its manufacturing and transportation.
“When we think about the environmental impact of generative AI, it’s not just the electricity you consume when you plug a computer in. It has much broader consequences that go down to the system level and the things we do.” remain based on actions,” says. Elsa A. Olivetti, professor in the Department of Materials Science and Engineering and head of the decarbonization mission of MIT’s New Climate Project.
Olivetti is the senior author of the 2024 paper, “The Climate and Sustainability Implications of Generative AI,” which was co-authored by MIT colleagues in response to an institute-wide call for papers that explore the transformative potential of generative AI, both positive and negative. Is a writer. Guidelines for society.
demand for data centers
The electricity demand of data centers is a major factor contributing to the environmental impacts of generator AI, as data centers are used to train and run the deep learning models behind 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 around the world, each containing about 50,000 servers that the company uses to support cloud computing services.
While data centers have existed since the 1940s (the first general purpose digital computer, ENIAC, was built at the University of Pennsylvania in 1945 to support ENIAC), the rise of generative AI has dramatically increased the speed of data center construction.
“What is different about generative AI is the power density it requires. Basically, it’s just computing, but a generic AI training cluster can consume seven or eight times more energy than a typical computing workload,” says Noman Bashir, lead author of the impact paper, who is a professor of climate computing at MIT. and Climate Impact Fellow. and Postdoc in the Sustainability Consortium (MCSC) and Computer Science and Artificial Intelligence Laboratory (CSAIL).
Scientists have estimated that the power requirements of data centers in North America will increase from 2,688 MW at the end of 2022 to 5,341 MW at the end of 2023, partly driven by the demands of generic AI. Globally, power consumption of data centers is expected to increase to 460 terawatts in 2022. This will make the data center the world’s 11th largest power consumer, behind Saudi Arabia (371 terawatts) and France (463 terawatts). Organization for Economic Cooperation and Development.
By 2026, power consumption of data centers is expected to reach 1,050 terawatts (which would put data centers in fifth place on the global list between Japan and Russia).
While not all data center compute involves generative AI, the technology has been a major driver of increasing energy demands.
“The demand for new data centers cannot be met in a sustainable manner. The speed at which companies are building new data centers means that the bulk of the electricity to power them must come from fossil fuel-based power plants,” says Bashir.
It is difficult to find the power needed to train and deploy a model like OpenAI’s GPT-3. In a 2021 research paper, Google and scientists at the University of California at Berkeley estimated that the training process alone consumed 1,287 megawatt hours of electricity (enough to power about 120 average American homes for a year), making About 552 tonnes of carbon dioxide is produced.
Bashir explains that while all machine-learning models must be trained, one issue unique to generative AI is the rapid fluctuations in energy usage that occur at different stages of the training process.
Power grid operators must have a way to absorb those fluctuations to protect the grid, and they typically use diesel-based generators for that task.
predictable effect
Once a generative AI model is trained, the energy demand does not go away.
Every time a model is used, perhaps by someone asking them to summarize emails from ChatGPT, the computing hardware performing those tasks consumes energy. Researchers have estimated that a ChatGPT query consumes about five times more power than a simple web search.
“But an everyday user doesn’t think too much about it,” says Bashir. “The ease of use of generative AI interfaces and the lack of information about the environmental impacts of my actions means that, as a user , I don’t have much incentive to cut back on the use of generator AI.”
With traditional AI, energy usage is divided fairly evenly between data processing, model training, and inference, which is the process of using a trained model to make predictions on new data. However, Bashir expects that the power demands of generic AI inference will eventually dominate as these models are becoming ubiquitous in so many applications, and the power required for inference will increase as future versions of the models become larger and more complex. .
Additionally, the shelf-life of generic AI models is typically short due to the increasing demand for new AI applications. Companies release new models every few weeks, Bashir says, so the energy used to train previous versions is wasted. Newer models often consume more energy for training, as they usually have more parameters than their predecessors.
Although the electricity demand of data centers is receiving the most attention in the research literature, the amount of water consumed by these facilities also has an environmental impact.
Chilled water is used to cool a data center by absorbing heat from computing equipment. It’s estimated that for every kilowatt hour of energy consumed by a data center, it will require two liters of water to cool it, Bashir says.
“Just because it’s called ‘cloud computing’ doesn’t mean the hardware lives in the cloud. Data centers exist in our physical world, and they have direct and indirect impacts on biodiversity because of their water use,” he says.
The computing hardware inside data centers brings its own, less direct environmental impacts.
Although it is difficult to estimate how much power is required to produce a GPU, the kind of powerful processor that can handle intensive generative AI workloads would exceed the power required to produce a simple CPU because the construction The process is more complicated. The carbon footprint of a GPU is complicated by emissions related to material and product transportation.
There are also environmental implications of obtaining the raw materials used to manufacture GPUs, which can include dirty mining processes and the use of toxic chemicals for processing.
Market research firm TechInsights estimates that the three major producers (NVIDIA, AMD and Intel) will ship 3.85 million GPUs to data centers in 2023, up from about 2.67 million in 2022. This number is expected to increase by an even higher percentage in 2024. ,
The industry is on an unsustainable path, Bashir says, but there are ways to encourage responsible development of generative AI that supports environmental objectives.
He, Olivetti, and their MIT colleagues argue that this will require comprehensive consideration of all the environmental and social costs of generative AI, as well as a detailed assessment of the value in its purported benefits.
“We need a more relevant way to systematically and comprehensively understand the implications of new developments in this area. “Because of the speed at which improvements have occurred, we have not had a chance to catch up in our abilities to measure and understand the tradeoffs,” says Olivetti.
