Explanation: Environmental Impact of Generated AI | MIT News

In a two-part series MIT News Explore the environmental meaning of generating AI. In this article, we examine why this technology is so numerous. The second piece will investigate what experts do in reducing Genai’s carbon footprint and other impacts.

The excitement from increasing worker productivity to developing scientific research on the potential benefits of generating AI is hard to ignore. While the explosive growth of this new technology has allowed many industries to deploy powerful models quickly, the environmental consequences of this generative AI “gold rush” are still difficult to determine, let alone mitigate.

Training the computational power required to generate AI models, which usually have billions of parameters, such as OpenAI’s GPT-4, may require a lot of electricity, resulting in increased carbon dioxide emissions and pressures on the grid.

Additionally, deploying these models in real-world applications allows millions to use the generated AI in their daily lives and then fine-tune the models to improve their performance, which will attract a lot of energy long after the model is developed.

In addition to electricity demand, a large amount of water is needed to cool the hardware used to train, deploy and fine-tune the generated AI model that could damage municipal water supply and damage local ecosystems. The increasing number of generative AI applications has also stimulated the need for high-performance computing hardware, thereby increasing the indirect environmental impact of their manufacturing and transportation.

“When we consider the environmental impact of generating AI, it’s not just the power consumed when plugging into a computer. Elsa A. Olivetti, a professor in the Department of Materials Science and Engineering, said there are more consequences to reach the system level and stick with the actions we take and persevere.”

Olivetti is senior author of the 2024 paper, “The Climate and Sustainable Meaning of Generative AI,” co-written by colleagues at MIT in response to the institute-wide appeal of papers that explore the transformational potential of generative AI, which is a positive and negative guidance for society.

A demanding data center

The power demand in data centers is a major factor in the environmental impact of generating AI, as data centers are used to train and run deep learning models behind popular tools such as Chatgpt and Dall-E.

A data center is a temperature-controlled building that can accommodate computing infrastructure such as servers, data storage drives, and network devices. For example, Amazon has more than 100 data centers worldwide, each with approximately 50,000 servers to support cloud computing services.

Data centers have existed since the 1940s (the first was built at the University of Pennsylvania in 1945 to support the first universal digital computer, The Eniac), but the rise of generative AI has significantly increased the speed of data center construction.

“What makes generative AI different is the power density it requires. Fundamentally, it’s just computation, but a generative AI training cluster may consume seven or eight times more energy than a typical computational effort.

Scientists estimate that the power requirement for North American data centers has increased from 2,688 MW by the end of 2022 to 5,341 MW by the end of 2023, driven in part by the need to generate AI. Globally, power consumption in data centers has risen to 460 tons in 2022. This will make the data center the 11th largest electricity consumer in the world, between the Saudi Arabian countries (371 Terawatts) and France (463 Terawatts), which will make the data center the 11th largest electricity consumer in the world.

By 2026, the power consumption of data centers is expected to reach close to 1,050 Terawatts (which will give data centers the most fifth on the global list between Japan and Russia).

Although not all data center computing involves generating AI, the technology has been the main driver of increased energy demand.

“The demand for new data centers cannot be met in a sustainable way. The speed at which companies build new data centers means that most of the electricity that powers them must come from fossil fuel-based power plants,” Bashir said.

The power required to train and deploy models such as 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 consumed 1,287 megawatts of electricity (enough to power about 120 average homes in the U.S. in a year), producing about 552 tons of carbon dioxide.

Bashir explains that while all machine learning models must be trained, one problem that is unique to generating AI is that the rapid fluctuations in energy use occur at different stages of the training process.

Grid operators must have a way to absorb these fluctuations to protect the grid, and typically use diesel-based generators to perform the task.

The impact of reasoning increases

Once the generated AI model is trained, the energy demand does not disappear.

Every time you use the model, maybe someone asks Chatgpt to summarize the email, and the computing hardware that performs these operations consumes energy. The researchers estimate that Chatgpt queries consume five times more than simple web searches.

“But every day users don’t think much about it,” Bashir said. “The ease of use interface for generating AI and the lack of information about the environmental impact of my actions means that, as a user, I don’t have much motivation to reduce my use of generating AI.”

With traditional AI, energy usage is quite uniform between data processing, model training and inference, which is the process of predicting new data using trained models. However, Bashir expects that the power demand for generating AI inferences will eventually dominate, as these models become ubiquitous in so many applications, and as future versions of the model become larger and more complex, the power required for inference will increase.

In addition, due to the growing demand for new AI applications, the generated AI models have an especially short shelf life. Bashir added that the company releases new models every few weeks, so energy wasted for training previous versions. New models usually consume more training energy, as they usually have more parameters than their predecessors.

Although the electricity demand in data centers may have attracted the greatest attention in the research literature, the amount of water consumed by these facilities also has environmental impacts.

Refrigerated water is used to cool the data center by absorbing heat from the computing device. It is estimated that for each kWh of energy consumed by a data center, it requires two liters of water to cool, Bashir said.

“Just because this is called ‘cloud computing’ doesn’t mean that hardware exists in the cloud. Data centers exist in our physical world and because of the amount of water they use, they have a direct and indirect impact on biodiversity,” he said.

The computing hardware inside the data center brings its own, less direct environmental impact.

While it is difficult to estimate how much power it takes to build a GPU, a powerful processor that can handle intensive AI workloads, because the manufacturing process is more complex, the capabilities required to produce a simpler CPU go far beyond what it requires to produce a simpler CPU. Emissions related to material and product transport, the GPU’s carbon footprint is even more complex.

Obtaining raw materials used to make GPUs also has environmental impacts, which can involve dirty mining procedures and processing using toxic chemicals.

Market research firm TechInsights estimates that three major producers (NVIDIA, AMD and INTEL) shipped 3.85 million GPUs to data centers in 2023, up from 2.67 million in 2022.

Bashir said the industry is on an unsustainable path, but there are ways to encourage responsible developments in generative AI that support environmental goals.

He, Olivetti and their MIT colleagues believe that this will require a comprehensive consideration of all the environmental and social costs of generating AI, as well as a detailed assessment of the value of its perceived welfare.

“We need a more contextual way to systematically and comprehensively understand what new developments mean in the field. Due to the speed of improvement, we have no chance to catch up with our ability to measure and understand the tradeoffs.”