As AI Use Grows, UMass Amherst Computer Scientists and Engineers Host Symposium on Sustainable Data Center Solutions

Content

What does sustainable computing look like in the age of AI? That was the question approximately 70 scholars and tech industry representatives sought to answer in a collaborative symposium on Sept. 17 and 18 at UMass Amherst.

Organized by Mohammad Hajiesmaili, a CICS associate professor and director of the UMass SOLAR (Sustainability, Optimization, Learning and Algorithms Research) Lab, the Sustainable Compute and AI Infrastructure (SCAI) symposium was a collaboration between the Manning College of Information and Computer Sciences (CICS) and the Riccio College of Engineering. It featured two days of talks, panel discussions, research highlights and poster presentations focused on AI systems, data-center infrastructure, power systems, environmental impacts and public priorities. 

In his keynote address, Prashant Shenoy, distinguished professor of computer science and director of the NSF Computer Decarbonization (CoDec) Expedition, said it’s important that computer scientists and engineers rethink typical approaches to the problem. Increasingly, he said, computing takes place in the cloud, because online and data-intensive workloads demand it. 

But AI has proved to be more resource-hungry than traditional cloud-reliant applications, requiring more data centers using more energy than ever before. This strains the electric grids, raises costs and causes negative environmental impacts. However, resource management via workload shifting could enhance cloud platforms’ efficiency by taking advantage of variabilities in grid demand, energy availability and electricity prices. Shenoy shared some initial methods of making computing systems more grid-friendly, as well as how to approach performance, efficiency and cost tradeoffs. 

“Adaptation is key,” he said. “You can use workload flexibility; you can use energy storage. But whatever it is, you have to rethink how you have been designing our systems and how they interface with the grid. There’s a lot of opportunity for us as academic researchers to contribute to this type of problem and conversation.”

The sessions that followed focused on three facets of the problem – AI and workloads, computing systems, and energy and infrastructure – and incorporated perspectives from academia as well as industry, while also touching on policy, community planning and public infrastructure. 

“Both industry and academics saw important roles for academic research despite – and perhaps in part because of – the fast pace of industrial systems,” said Laura Haas, CICS professor, who moderated a panel on the role of academia in the future of AI infrastructure. 

“One of the most important roles that academia can play is in democratizing research,” said panelist Ramesh Sitaraman, a CICS distinguished professor who leads the Lab for Internet-scale Distributed Systems (LIDS); he is also chief consulting scientist at Akamai Technologies. “If AI is going to influence all people in society, all people in society must be able to contribute to how it evolves and how it’s researched.”

Ideally, common shared infrastructure built by the community would allow anyone to do research, Sitaraman said. “I don’t see that happening in the AI side of things yet, and that’s something that we ought to focus on,” he added. “This means infrastructure, shared infrastructure, shared applications, shared workloads – so that anyone, anywhere, can do this type of research.”

Industry-academia collaboration was a common thread throughout the symposium – and that was by design, Haas said. “Such symposia are really helpful for networking and forming new collaborations,” she noted, adding that industry representatives were especially keen to find potential partners on the academic side.

A separate panel addressed what responsible data center growth looks like. Moderator Golbon Zakeri, professor of mechanical and industrial engineering, opened the panel by reminding us that this question is not new, citing The Limits to Growth, which raised similar issues in 1972. “While ours is not a new question, it’s a new challenge,” she added.

In addition to the points raised about investing in renewables and protecting water resources, Erin Baker, faculty director of the Energy Transition Institute and distinguished professor of mechanical and industrial engineering, brought in the perspective of energy justice. 

“Justice is the recognition that burdens and benefits are distributed unevenly and often these are along lines of race class and political power,” she said. “And we need to treat these inequities not just as incidental outcomes but as matters of structural justice requiring ethical and political redress.”

Baker highlighted how, in a world with many environmentally damaging industries, she found it implausible that the pollution alone would drive communities to the protests we are currently witnessing. Instead, communities are reacting to the concern that AI will worsen inequality by taking jobs. 

“In the US, we already have a lot of inequality and it’s growing and I think people are very worried about inequality and AI,” she says. The data center becomes an “avatar” for the threat of inequality that people are worried AI could bring — taking away jobs, using up resources and degrading local environments. 

When asked how data centers could be scaled sustainably, Baker emphasized the importance that AI should bring prosperity to the communities it enters by investing in education, healthcare and environmental benefits.

The panelists, presenters and attendees all seemed to agree that greening AI data centers is a big job, but despite the daunting nature of the topics at hand, the symposium ended on a hopeful note.

“I have never been this excited about the research and the possibility of making an impact in the real world in such a short time,” Hajiesmaili said. “We, as scientists, have a substantial responsibility to spread the word out there to first raise awareness with the public about both opportunities and risks – to people and planet – of this technology, and also about the scientific solutions to fundamental problems in this space that could be adapted by industry and tech companies when they are building such an infrastructure.”