MIT Researchers Tackle AI’s Growing Climate Impact with Innovative Solutions

MIT researchers have unveiled comprehensive strategies to address generative AI’s escalating environmental impact, as data centre electricity demand is projected to more than double by 2030. The International Energy Agency forecasts global data centre consumption will reach 945 terawatt-hours, whilst Goldman Sachs Research estimates 60 percent of this increased demand will be met through fossil fuels, adding 220 million tons to global carbon emissions.

Context and Background

Scientists and engineers at MIT are developing interventions across multiple fronts, from algorithmic efficiency to data centre design. Vijay Gadepally, senior scientist at MIT Lincoln Laboratory, emphasises that discussions typically focus on “operational carbon” from GPU processors whilst overlooking “embodied carbon” from data centre construction itself.

Research from the Lincoln Laboratory Supercomputing Center demonstrates that reducing GPU energy consumption to three-tenths normal levels produces minimal impact on AI model performance whilst significantly improving cooling efficiency. The team discovered that approximately half the electricity used in AI model training achieves only the final 2-3 percentage points of accuracy, suggesting substantial energy savings through early training termination.

Neil Thompson, director of MIT’s FutureTech Research Project, reports that GPU computational efficiency per joule improves by 50-60 percent annually, whilst new model architectures delivering equivalent results with less energy double efficiency every eight to nine months.

Looking Forward

MIT Energy Initiative researchers are developing “smarter” data centres that flexibly adjust multiple companies’ AI workloads to maximise energy efficiency. Deepjyoti Deka’s team is exploring long-duration energy storage systems that could store renewable energy for high-demand periods, potentially eliminating diesel backup generators.

Jennifer Turliuk, former Sloan Fellow and climate AI expert, advocates for AI-powered solutions to accelerate renewable energy grid integration, noting that interconnection studies currently requiring years could be streamlined through generative AI models. Her team developed the Net Climate Impact Score framework to help organisations evaluate AI projects’ comprehensive environmental costs and benefits.

Source Attribution:

Share this article