Bits & Watts Initiative funds six sustainable AI research projects

Stanford University’s Bits & Watts Initiative selected six new research projects on campus to fund as part of its new flagship research program, Sustaining AI.

Artificial intelligence offers enormous potential for a sustainable future. In the energy sector, AI can leverage the vast amounts of ever-changing data to optimize for lower greenhouse gas emissions and other environmental impacts, as well as cost and reliability. However, this intellectual power requires a lot of electricity, which globally is still primarily powered by coal, natural gas, and oil. As for electric system reliability, data centers – especially processing generative AI and cryptocurrency mining – are already straining existing power grids. By the end of next year, power use by data centers globally is projected by the International Energy Agency to be twice what it was in 2022, homing in on 5% of all electricity use. While many large technology companies aim for sustainable operations, generating and transmitting enough clean electricity to meet this demand will be an uphill battle.

“Data centers are new challenging customers for the electric grid. Support is needed to understand their load characteristics, optimize their location, flexibility, power quality, and even the use of their back up power supplies,” said Larry Bekkedahl, senior vice president for strategy and advanced energy delivery at Portland General Electric, which is a corporate supporter of the Bits & Watts Initiative.

To address these complex challenges, the Bits & Watts Initiative – an industrial affiliate program launched by Stanford’s Precourt Institute for Energy in 2016 – is working with electric utilities worldwide, large IT companies, and data center developers.

“We look forward to working with Stanford AI and energy experts and forward-thinking industry members on this new flagship program,” said Liang Min, the initiative’s managing director.

In 2019, the initiative launched a similar research program focused on preparing utilities and grid operators for a future when 50% of cars are EVs. That program led to many valuable insights for the power and EV industries, like a study showing that EVs should be charged at work during the day, not at home at night. The new flagship program is focused on monitoring and reducing AI’s GHG emissions, as well as affordability, grid reliability, and grid interconnections.

“AI is a global game changer with even deeper impacts on the energy sector, as it can provide solutions for the energy transition but is also energy greedy,” said Gabriel Bareux, director of research and development at RTE, the operator of France’s electricity transmission system and a member of Bits & Watts. “This new program on sustainable AI is targeting the key questions of the coming years for the energy sector.”

Funded research projects

Each of the six $100,000 seed grants supports the exploration of new, potentially transformative ideas for sustainable AI. The grants enable Stanford faculty members and their students to demonstrate their concepts through initial experiments and study. Successful proof-of-concept research results in such programs usually garner significant support to advance the work. Bits & Watts has been making seed grants since 2017.

The six projects below were selected on a competitive basis by a review team including internal Stanford reviewers and external reviewers from the utility and IT sectors.

An analysis framework for data center microgrid systems

Principal investigator: Sara Achour, Computer Science and Electrical Engineering

Self-contained microgrids serving data centers provide reliable electric service and stop microgrid failures from causing blackouts on the main grid, but failure-inducing inputs to microgrids are easy to miss and can be catastrophic if accidentally triggered after deployment or discovered by adversarial actors. Achour and her student will develop an analysis framework that identifies “needle-in-a-haystack,” failure-inducing inputs in microgrid systems that supply data centers.

Optimizing energy storage and transmission for data centers

PI: Hunt Allcott, Environmental Social Sciences

The most likely way to add low-carbon generation to satisfy AI energy demand in less than a decade is through wind and solar power. However, these resources are intermittent and not always located in the same places where companies want to build data centers. Thus, key to powering AI sustainably are energy storage and more transmission to get the power from windy and sunny places to data centers often far away. Allcott and his team will conduct empirical estimations of the value of transmission expansion and storage to power data centers. They will also analyze transmission cost allocation and suggest beneficial tariff structures.

A power supply to remove machine learning transients from the grid

PIs: Philip Levis, Computer Science and Electrical Engineering; Juan Rivas-Davila, Electrical Engineering; and Ram Rajagopal, Civil & Environmental Engineering and Electrical Engineering

Electricity used for today’s machine learning training workloads can peak at 10 megawatts or even multiples of that. Then, when the computers stop processing to communicate their results, their power demand drops to about a fifth of their peak load in microseconds. These huge, synchronized load swings introduce enormous problems to the electric grid. This project seeks to research and build a power supply mechanism that ensures a computing rack’s load does not change quickly due to its energy storage. Data center operators have told the researchers that such a power supply would be much better than their current strategies to deal with the problem.

Optimal siting and flexible operation of data centers to minimize costs, carbon intensity, and cooling water impacts

PI: Meagan S. Mauter, Civil & Environmental Engineering

Besides electricity consumption, data centers also require a lot of cooling water, imposing significant environmental and economic impacts on U.S. surface water supplies. For several reasons, data centers are good candidates for dynamic operation, which responds to changes in electricity prices, emissions, water supply, and utility incentive programs for lowering demand when needed. Unfortunately, U.S. data center operators lack critical datasets to inform siting and dynamic operation. This project will build a national, detailed dataset, including electric rates, revenue potential, GHG emissions, and water availability, to inform data center developers and operators for siting and flexibility.

Moving away from large clusters: Grid-aware AI data center planning

PI: Ram Rajagopal

This research project seeks to reduce large swings in data center electricity use by rethinking their integration with existing electricity infrastructure. The team, including a senior researcher at Microsoft, will model the effects of AI workloads at the substation and distribution network level. The researchers will then develop tools for data center developers to build infrastructure based on the local grid, and for operators to manage workload by distributing it across multiple sites in coordination.

Socio-psychological perspectives in AI model selection for performance and sustainability

PIs: Khalid Osman, Civil & Environmental Engineering, and Tobias Gerstenberg, Psychology

This project, by combining insights from social psychology and AI model development, aims to provide AI developers with contextual information about carbon emissions and energy consumption in relatable terms, making environmental impacts more salient and encouraging more sustainable choices. The products of this work will include a validated framework for sustainable AI practices; a dataset on developer perspectives on selecting AI models for training, fine-tuning and serving AI models; guidelines for developers; and contributions to policy discussions on aligning AI with climate goals.

The Precourt Institute for Energy is part of the Stanford Doerr School of Sustainability.