Illuminating Energy-Efficient AI
by A.J. Kleber
With the meteoric rise of AI technologies over the last few years, and their rapid integration across diverse industries and societal infrastructures, concerns over their immense energy requirements have come into sharp focus for users and researchers alike. Deep Neural Networks (DNNs) and large language models (LLMs), some of the most influential models on the market today, are trained and deployed in data centers powered by specialized silicon-based hardware such as graphics processing units (GPUs), and their booming energy demands have been well-publicized. Experts are scrambling to find options to mitigate or amend this unsustainable drain on global resources.
To address this critical issue, Professor Ajay Joshi is exploring the development of novel electro-photonic computing architectures that could perform as well or better than conventional electronic GPUs. Supported by a $1.5M grant from the NSF, titled ASCENT: Heterogeneously Integrated Electronic Photonic AI Accelerators (HIEPAA), he is investigating the use of light-based computing for energy-efficient AI hardware.
Light is a far more efficient medium for data transmission than electricity, and recent advances in integrated electronic-photonic systems have shown significant promise. Joshi’s approach incorporates an electro-optical metamaterial called thin-film lithium niobate (TFLN) into silicon photonic chip platforms, in order to produce analog optical modulators; devices which will greatly increase the speed of computing from that of traditional silicon-based electronic devices, while also decreasing data loss. He plans to design new architectures and circuit techniques to achieve efficient, highly accurate AI computation using low-precision building blocks, a concept previously explored by his research group and published in Nature Communications.
With these innovations, the future of AI could be bright indeed.
Professor Ajay Joshi leads the Integrated Circuits, Architectures, and Systems Group at Boston University. His accolades include multiple Google Faculty Research Awards, Best Paper Awards, and a 2012 NSF CAREER AWARD. He also serves as CEO and co-founder of CipherSonic Labs, a start-up dedicated to data privacy technology development, and the recipient of a 2024 BU Technology Development Ignition Award.