More stories

  • in

    Responding to the climate impact of generative AI

    In part 2 of our two-part series on generative artificial intelligence’s environmental impacts, MIT News explores some of the ways experts are working to reduce the technology’s carbon footprint.The energy demands of generative AI are expected to continue increasing dramatically over the next decade.For instance, an April 2025 report from the International Energy Agency predicts that the global electricity demand from data centers, which house the computing infrastructure to train and deploy AI models, will more than double by 2030, to around 945 terawatt-hours. While not all operations performed in a data center are AI-related, this total amount is slightly more than the energy consumption of Japan.Moreover, an August 2025 analysis from Goldman Sachs Research forecasts that about 60 percent of the increasing electricity demands from data centers will be met by burning fossil fuels, increasing global carbon emissions by about 220 million tons. In comparison, driving a gas-powered car for 5,000 miles produces about 1 ton of carbon dioxide.These statistics are staggering, but at the same time, scientists and engineers at MIT and around the world are studying innovations and interventions to mitigate AI’s ballooning carbon footprint, from boosting the efficiency of algorithms to rethinking the design of data centers.Considering carbon emissionsTalk of reducing generative AI’s carbon footprint is typically centered on “operational carbon” — the emissions used by the powerful processors, known as GPUs, inside a data center. It often ignores “embodied carbon,” which are emissions created by building the data center in the first place, says Vijay Gadepally, senior scientist at MIT Lincoln Laboratory, who leads research projects in the Lincoln Laboratory Supercomputing Center.Constructing and retrofitting a data center, built from tons of steel and concrete and filled with air conditioning units, computing hardware, and miles of cable, consumes a huge amount of carbon. In fact, the environmental impact of building data centers is one reason companies like Meta and Google are exploring more sustainable building materials. (Cost is another factor.)Plus, data centers are enormous buildings — the world’s largest, the China Telecomm-Inner Mongolia Information Park, engulfs roughly 10 million square feet — with about 10 to 50 times the energy density of a normal office building, Gadepally adds. “The operational side is only part of the story. Some things we are working on to reduce operational emissions may lend themselves to reducing embodied carbon, too, but we need to do more on that front in the future,” he says.Reducing operational carbon emissionsWhen it comes to reducing operational carbon emissions of AI data centers, there are many parallels with home energy-saving measures. For one, we can simply turn down the lights.“Even if you have the worst lightbulbs in your house from an efficiency standpoint, turning them off or dimming them will always use less energy than leaving them running at full blast,” Gadepally says.In the same fashion, research from the Supercomputing Center has shown that “turning down” the GPUs in a data center so they consume about three-tenths the energy has minimal impacts on the performance of AI models, while also making the hardware easier to cool.Another strategy is to use less energy-intensive computing hardware.Demanding generative AI workloads, such as training new reasoning models like GPT-5, usually need many GPUs working simultaneously. The Goldman Sachs analysis estimates that a state-of-the-art system could soon have as many as 576 connected GPUs operating at once.But engineers can sometimes achieve similar results by reducing the precision of computing hardware, perhaps by switching to less powerful processors that have been tuned to handle a specific AI workload.There are also measures that boost the efficiency of training power-hungry deep-learning models before they are deployed.Gadepally’s group found that about half the electricity used for training an AI model is spent to get the last 2 or 3 percentage points in accuracy. Stopping the training process early can save a lot of that energy.“There might be cases where 70 percent accuracy is good enough for one particular application, like a recommender system for e-commerce,” he says.Researchers can also take advantage of efficiency-boosting measures.For instance, a postdoc in the Supercomputing Center realized the group might run a thousand simulations during the training process to pick the two or three best AI models for their project.By building a tool that allowed them to avoid about 80 percent of those wasted computing cycles, they dramatically reduced the energy demands of training with no reduction in model accuracy, Gadepally says.Leveraging efficiency improvementsConstant innovation in computing hardware, such as denser arrays of transistors on semiconductor chips, is still enabling dramatic improvements in the energy efficiency of AI models.Even though energy efficiency improvements have been slowing for most chips since about 2005, the amount of computation that GPUs can do per joule of energy has been improving by 50 to 60 percent each year, says Neil Thompson, director of the FutureTech Research Project at MIT’s Computer Science and Artificial Intelligence Laboratory and a principal investigator at MIT’s Initiative on the Digital Economy.“The still-ongoing ‘Moore’s Law’ trend of getting more and more transistors on chip still matters for a lot of these AI systems, since running operations in parallel is still very valuable for improving efficiency,” says Thomspon.Even more significant, his group’s research indicates that efficiency gains from new model architectures that can solve complex problems faster, consuming less energy to achieve the same or better results, is doubling every eight or nine months.Thompson coined the term “negaflop” to describe this effect. The same way a “negawatt” represents electricity saved due to energy-saving measures, a “negaflop” is a computing operation that doesn’t need to be performed due to algorithmic improvements.These could be things like “pruning” away unnecessary components of a neural network or employing compression techniques that enable users to do more with less computation.“If you need to use a really powerful model today to complete your task, in just a few years, you might be able to use a significantly smaller model to do the same thing, which would carry much less environmental burden. Making these models more efficient is the single-most important thing you can do to reduce the environmental costs of AI,” Thompson says.Maximizing energy savingsWhile reducing the overall energy use of AI algorithms and computing hardware will cut greenhouse gas emissions, not all energy is the same, Gadepally adds.“The amount of carbon emissions in 1 kilowatt hour varies quite significantly, even just during the day, as well as over the month and year,” he says.Engineers can take advantage of these variations by leveraging the flexibility of AI workloads and data center operations to maximize emissions reductions. For instance, some generative AI workloads don’t need to be performed in their entirety at the same time.Splitting computing operations so some are performed later, when more of the electricity fed into the grid is from renewable sources like solar and wind, can go a long way toward reducing a data center’s carbon footprint, says Deepjyoti Deka, a research scientist in the MIT Energy Initiative.Deka and his team are also studying “smarter” data centers where the AI workloads of multiple companies using the same computing equipment are flexibly adjusted to improve energy efficiency.“By looking at the system as a whole, our hope is to minimize energy use as well as dependence on fossil fuels, while still maintaining reliability standards for AI companies and users,” Deka says.He and others at MITEI are building a flexibility model of a data center that considers the differing energy demands of training a deep-learning model versus deploying that model. Their hope is to uncover the best strategies for scheduling and streamlining computing operations to improve energy efficiency.The researchers are also exploring the use of long-duration energy storage units at data centers, which store excess energy for times when it is needed.With these systems in place, a data center could use stored energy that was generated by renewable sources during a high-demand period, or avoid the use of diesel backup generators if there are fluctuations in the grid.“Long-duration energy storage could be a game-changer here because we can design operations that really change the emission mix of the system to rely more on renewable energy,” Deka says.In addition, researchers at MIT and Princeton University are developing a software tool for investment planning in the power sector, called GenX, which could be used to help companies determine the ideal place to locate a data center to minimize environmental impacts and costs.Location can have a big impact on reducing a data center’s carbon footprint. For instance, Meta operates a data center in Lulea, a city on the coast of northern Sweden where cooler temperatures reduce the amount of electricity needed to cool computing hardware.Thinking farther outside the box (way farther), some governments are even exploring the construction of data centers on the moon where they could potentially be operated with nearly all renewable energy.AI-based solutionsCurrently, the expansion of renewable energy generation here on Earth isn’t keeping pace with the rapid growth of AI, which is one major roadblock to reducing its carbon footprint, says Jennifer Turliuk MBA ’25, a short-term lecturer, former Sloan Fellow, and former practice leader of climate and energy AI at the Martin Trust Center for MIT Entrepreneurship.The local, state, and federal review processes required for a new renewable energy projects can take years.Researchers at MIT and elsewhere are exploring the use of AI to speed up the process of connecting new renewable energy systems to the power grid.For instance, a generative AI model could streamline interconnection studies that determine how a new project will impact the power grid, a step that often takes years to complete.And when it comes to accelerating the development and implementation of clean energy technologies, AI could play a major role.“Machine learning is great for tackling complex situations, and the electrical grid is said to be one of the largest and most complex machines in the world,” Turliuk adds.For instance, AI could help optimize the prediction of solar and wind energy generation or identify ideal locations for new facilities.It could also be used to perform predictive maintenance and fault detection for solar panels or other green energy infrastructure, or to monitor the capacity of transmission wires to maximize efficiency.By helping researchers gather and analyze huge amounts of data, AI could also inform targeted policy interventions aimed at getting the biggest “bang for the buck” from areas such as renewable energy, Turliuk says.To help policymakers, scientists, and enterprises consider the multifaceted costs and benefits of AI systems, she and her collaborators developed the Net Climate Impact Score.The score is a framework that can be used to help determine the net climate impact of AI projects, considering emissions and other environmental costs along with potential environmental benefits in the future.At the end of the day, the most effective solutions will likely result from collaborations among companies, regulators, and researchers, with academia leading the way, Turliuk adds.“Every day counts. We are on a path where the effects of climate change won’t be fully known until it is too late to do anything about it. This is a once-in-a-lifetime opportunity to innovate and make AI systems less carbon-intense,” she says. More

  • in

    Q&A: The climate impact of generative AI

    Vijay Gadepally, a senior staff member at MIT Lincoln Laboratory, leads a number of projects at the Lincoln Laboratory Supercomputing Center (LLSC) to make computing platforms, and the artificial intelligence systems that run on them, more efficient. Here, Gadepally discusses the increasing use of generative AI in everyday tools, its hidden environmental impact, and some of the ways that Lincoln Laboratory and the greater AI community can reduce emissions for a greener future.Q: What trends are you seeing in terms of how generative AI is being used in computing?A: Generative AI uses machine learning (ML) to create new content, like images and text, based on data that is inputted into the ML system. At the LLSC we design and build some of the largest academic computing platforms in the world, and over the past few years we’ve seen an explosion in the number of projects that need access to high-performance computing for generative AI. We’re also seeing how generative AI is changing all sorts of fields and domains — for example, ChatGPT is already influencing the classroom and the workplace faster than regulations can seem to keep up.We can imagine all sorts of uses for generative AI within the next decade or so, like powering highly capable virtual assistants, developing new drugs and materials, and even improving our understanding of basic science. We can’t predict everything that generative AI will be used for, but I can certainly say that with more and more complex algorithms, their compute, energy, and climate impact will continue to grow very quickly.Q: What strategies is the LLSC using to mitigate this climate impact?A: We’re always looking for ways to make computing more efficient, as doing so helps our data center make the most of its resources and allows our scientific colleagues to push their fields forward in as efficient a manner as possible.As one example, we’ve been reducing the amount of power our hardware consumes by making simple changes, similar to dimming or turning off lights when you leave a room. In one experiment, we reduced the energy consumption of a group of graphics processing units by 20 percent to 30 percent, with minimal impact on their performance, by enforcing a power cap. This technique also lowered the hardware operating temperatures, making the GPUs easier to cool and longer lasting.Another strategy is changing our behavior to be more climate-aware. At home, some of us might choose to use renewable energy sources or intelligent scheduling. We are using similar techniques at the LLSC — such as training AI models when temperatures are cooler, or when local grid energy demand is low.We also realized that a lot of the energy spent on computing is often wasted, like how a water leak increases your bill but without any benefits to your home. We developed some new techniques that allow us to monitor computing workloads as they are running and then terminate those that are unlikely to yield good results. Surprisingly, in a number of cases we found that the majority of computations could be terminated early without compromising the end result.Q: What’s an example of a project you’ve done that reduces the energy output of a generative AI program?A: We recently built a climate-aware computer vision tool. Computer vision is a domain that’s focused on applying AI to images; so, differentiating between cats and dogs in an image, correctly labeling objects within an image, or looking for components of interest within an image.In our tool, we included real-time carbon telemetry, which produces information about how much carbon is being emitted by our local grid as a model is running. Depending on this information, our system will automatically switch to a more energy-efficient version of the model, which typically has fewer parameters, in times of high carbon intensity, or a much higher-fidelity version of the model in times of low carbon intensity.By doing this, we saw a nearly 80 percent reduction in carbon emissions over a one- to two-day period. We recently extended this idea to other generative AI tasks such as text summarization and found the same results. Interestingly, the performance sometimes improved after using our technique!Q: What can we do as consumers of generative AI to help mitigate its climate impact?A: As consumers, we can ask our AI providers to offer greater transparency. For example, on Google Flights, I can see a variety of options that indicate a specific flight’s carbon footprint. We should be getting similar kinds of measurements from generative AI tools so that we can make a conscious decision on which product or platform to use based on our priorities.We can also make an effort to be more educated on generative AI emissions in general. Many of us are familiar with vehicle emissions, and it can help to talk about generative AI emissions in comparative terms. People may be surprised to know, for example, that one image-generation task is roughly equivalent to driving four miles in a gas car, or that it takes the same amount of energy to charge an electric car as it does to generate about 1,500 text summarizations.There are many cases where customers would be happy to make a trade-off if they knew the trade-off’s impact.Q: What do you see for the future?A: Mitigating the climate impact of generative AI is one of those problems that people all over the world are working on, and with a similar goal. We’re doing a lot of work here at Lincoln Laboratory, but its only scratching at the surface. In the long term, data centers, AI developers, and energy grids will need to work together to provide “energy audits” to uncover other unique ways that we can improve computing efficiencies. We need more partnerships and more collaboration in order to forge ahead.If you’re interested in learning more, or collaborating with Lincoln Laboratory on these efforts, please contact Vijay Gadepally.

    Play video

    Video: MIT Lincoln Laboratory More

  • in

    Taking a magnifying glass to data center operations

    When the MIT Lincoln Laboratory Supercomputing Center (LLSC) unveiled its TX-GAIA supercomputer in 2019, it provided the MIT community a powerful new resource for applying artificial intelligence to their research. Anyone at MIT can submit a job to the system, which churns through trillions of operations per second to train models for diverse applications, such as spotting tumors in medical images, discovering new drugs, or modeling climate effects. But with this great power comes the great responsibility of managing and operating it in a sustainable manner — and the team is looking for ways to improve.

    “We have these powerful computational tools that let researchers build intricate models to solve problems, but they can essentially be used as black boxes. What gets lost in there is whether we are actually using the hardware as effectively as we can,” says Siddharth Samsi, a research scientist in the LLSC. 

    To gain insight into this challenge, the LLSC has been collecting detailed data on TX-GAIA usage over the past year. More than a million user jobs later, the team has released the dataset open source to the computing community.

    Their goal is to empower computer scientists and data center operators to better understand avenues for data center optimization — an important task as processing needs continue to grow. They also see potential for leveraging AI in the data center itself, by using the data to develop models for predicting failure points, optimizing job scheduling, and improving energy efficiency. While cloud providers are actively working on optimizing their data centers, they do not often make their data or models available for the broader high-performance computing (HPC) community to leverage. The release of this dataset and associated code seeks to fill this space.

    “Data centers are changing. We have an explosion of hardware platforms, the types of workloads are evolving, and the types of people who are using data centers is changing,” says Vijay Gadepally, a senior researcher at the LLSC. “Until now, there hasn’t been a great way to analyze the impact to data centers. We see this research and dataset as a big step toward coming up with a principled approach to understanding how these variables interact with each other and then applying AI for insights and improvements.”

    Papers describing the dataset and potential applications have been accepted to a number of venues, including the IEEE International Symposium on High-Performance Computer Architecture, the IEEE International Parallel and Distributed Processing Symposium, the Annual Conference of the North American Chapter of the Association for Computational Linguistics, the IEEE High-Performance and Embedded Computing Conference, and International Conference for High Performance Computing, Networking, Storage and Analysis. 

    Workload classification

    Among the world’s TOP500 supercomputers, TX-GAIA combines traditional computing hardware (central processing units, or CPUs) with nearly 900 graphics processing unit (GPU) accelerators. These NVIDIA GPUs are specialized for deep learning, the class of AI that has given rise to speech recognition and computer vision.

    The dataset covers CPU, GPU, and memory usage by job; scheduling logs; and physical monitoring data. Compared to similar datasets, such as those from Google and Microsoft, the LLSC dataset offers “labeled data, a variety of known AI workloads, and more detailed time series data compared with prior datasets. To our knowledge, it’s one of the most comprehensive and fine-grained datasets available,” Gadepally says. 

    Notably, the team collected time-series data at an unprecedented level of detail: 100-millisecond intervals on every GPU and 10-second intervals on every CPU, as the machines processed more than 3,000 known deep-learning jobs. One of the first goals is to use this labeled dataset to characterize the workloads that different types of deep-learning jobs place on the system. This process would extract features that reveal differences in how the hardware processes natural language models versus image classification or materials design models, for example.   

    The team has now launched the MIT Datacenter Challenge to mobilize this research. The challenge invites researchers to use AI techniques to identify with 95 percent accuracy the type of job that was run, using their labeled time-series data as ground truth.

    Such insights could enable data centers to better match a user’s job request with the hardware best suited for it, potentially conserving energy and improving system performance. Classifying workloads could also allow operators to quickly notice discrepancies resulting from hardware failures, inefficient data access patterns, or unauthorized usage.

    Too many choices

    Today, the LLSC offers tools that let users submit their job and select the processors they want to use, “but it’s a lot of guesswork on the part of users,” Samsi says. “Somebody might want to use the latest GPU, but maybe their computation doesn’t actually need it and they could get just as impressive results on CPUs, or lower-powered machines.”

    Professor Devesh Tiwari at Northeastern University is working with the LLSC team to develop techniques that can help users match their workloads to appropriate hardware. Tiwari explains that the emergence of different types of AI accelerators, GPUs, and CPUs has left users suffering from too many choices. Without the right tools to take advantage of this heterogeneity, they are missing out on the benefits: better performance, lower costs, and greater productivity.

    “We are fixing this very capability gap — making users more productive and helping users do science better and faster without worrying about managing heterogeneous hardware,” says Tiwari. “My PhD student, Baolin Li, is building new capabilities and tools to help HPC users leverage heterogeneity near-optimally without user intervention, using techniques grounded in Bayesian optimization and other learning-based optimization methods. But, this is just the beginning. We are looking into ways to introduce heterogeneity in our data centers in a principled approach to help our users achieve the maximum advantage of heterogeneity autonomously and cost-effectively.”

    Workload classification is the first of many problems to be posed through the Datacenter Challenge. Others include developing AI techniques to predict job failures, conserve energy, or create job scheduling approaches that improve data center cooling efficiencies.

    Energy conservation 

    To mobilize research into greener computing, the team is also planning to release an environmental dataset of TX-GAIA operations, containing rack temperature, power consumption, and other relevant data.

    According to the researchers, huge opportunities exist to improve the power efficiency of HPC systems being used for AI processing. As one example, recent work in the LLSC determined that simple hardware tuning, such as limiting the amount of power an individual GPU can draw, could reduce the energy cost of training an AI model by 20 percent, with only modest increases in computing time. “This reduction translates to approximately an entire week’s worth of household energy for a mere three-hour time increase,” Gadepally says.

    They have also been developing techniques to predict model accuracy, so that users can quickly terminate experiments that are unlikely to yield meaningful results, saving energy. The Datacenter Challenge will share relevant data to enable researchers to explore other opportunities to conserve energy.

    The team expects that lessons learned from this research can be applied to the thousands of data centers operated by the U.S. Department of Defense. The U.S. Air Force is a sponsor of this work, which is being conducted under the USAF-MIT AI Accelerator.

    Other collaborators include researchers at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). Professor Charles Leiserson’s Supertech Research Group is investigating performance-enhancing techniques for parallel computing, and research scientist Neil Thompson is designing studies on ways to nudge data center users toward climate-friendly behavior.

    Samsi presented this work at the inaugural AI for Datacenter Optimization (ADOPT’22) workshop last spring as part of the IEEE International Parallel and Distributed Processing Symposium. The workshop officially introduced their Datacenter Challenge to the HPC community.

    “We hope this research will allow us and others who run supercomputing centers to be more responsive to user needs while also reducing the energy consumption at the center level,” Samsi says. More