Assistant Professor Tianyu Wang (ECE) works to make AI Data Centers more Energy Efficient
AI is already using an enormous amount of energy, and the problem is about to get much worse. According to Energy Digital, data centers, which are used to host and run large-scale AI optimisations, are estimated to have consumed 1-1.5% of global electricity in 2022, and we are expected to see electricity demands double by 2026.
In Virgina, which is considered the global hub for data centers — with at least 70% of the world’s internet traffic running through Northern Virginia alone — the presence of data centers is rapidly expanding. Dominion Energy, a main supplier of power in Virginia, estimates that power demand will grow by about 85% over the next 15 years.
Why is this an issue? For one, the new hyperscale data centers required to keep up with the rapid growth of generative AI will exceed the ability of utility providers to expand their capacity fast enough, according to a VP Analyst at Gartner. This threatens to disrupt energy availability and could lead to energy shortages worldwide.
Further, the demand of AI on energy resources will have an environmental impact, especially through the increase in CO₂ emissions. In Virginia, these impacts have already been seen — there has been a spike in projected energy use for the region, causing concern for environmental advocates about Virginia’s ability to meet renewable energy goals. Additionally, residential energy bills will increase to fund renewable energy and natural gas projects.
Assistant Professor Tianyu Wang (ECE), who joined BU in early 2024, combines physics and optical principles with computer science to work on building more energy efficient hardware that can be used in applications ranging from these large AI data centers to locally deployed autonomous agents.
Wang’s research concerns finding alternative physical platforms to carry out neural network computation, which is essential to every powerful AI model today, in a more energy efficient manner. “I’m really interested in using physical processes, like optical phenomena, for computing. Compared to digital computers, these analog systems are likely to be more efficient at neuromorphic computing which is exactly what is needed to make today’s gigantic artificial intelligence models more efficient and accessible.” Wang said.
Wang said that he and his group, the AI Photonics Lab, have identified photonics as a promising solution for its high communication bandwidth and parallelism, making it a natural choice for neuromorphic computing, a computing approach that mimics the human brain. To pave the way for this new form of computing hardware, he and his team use experiments to tackle questions such as: what is the most promising architecture to scale these unconventional optical computers? How can these analog computers be trained and aligned in a scalable way? What are the most promising near-term applications for optical computing?
Currently, Wang is working on two projects in his lab, one that deals with processing, and the other with computational sensing. “On the processor side, we are working on integrat[ing] photonics to large language models, like ChatGPT,” he said. “We hope to be able to train these models and to run such practically useful models on people’s hardware to demonstrate its potential.” On the computation sensing side, where he and his team find an application is cell sorting for cytometry, which is used to characterize and count blood cells, useful in both cell biology research and in medical diagnostics for diseases such as early cancer diagnosis. In both projects, Wang’s goal is to use optical systems to do computation, thus making processing more energy efficient and faster as compared to digital electronics, which use much more energy and thus have significant environmental impacts.
Wang’s research is unique in that it brings together both physics and computing, disciplines that are currently “really divided.” According to him, the current curriculum on computational science was designed mainly for computer science and computer engineering professionals, with almost no mention on how to apply these methods to natural science and engineering research, whereas computational sciences and physical sciences were historically much more closely connected (see the Nobel Prize in Physics in 2025), and we will witness more interdisciplinary work across these disciplines in future (see the Nobel Prize in Chemistry in 2025). . To address this gap, he is starting a graduate level course on physics and computing.
“I will continue to research these new forms of computing based on physical systems, particularly optical systems. I’m interested in understanding what their advantages are, where their advantages come from, and what they are useful for. ” Wang said.
Tianyu Wang is an Associate Professor at Boston University, whose primary appointment is in the Department of Electrical and Computer Engineering. He was awarded the 2024 Peter J. Levine Career Development Professorship at Boston University, and was previously awarded Schmidt AI in Science Postdoctoral Fellowship from 2022-2023 prior to joining BU.