Title: Photonics for Sustainable AI Systems

Presenter: Farbin Fayza

Advisor: Professor Ajay Joshi

Chair: Professor Ayse Coskun

Committee: Professor Ayse Coskun, Professor Sabrina Neuman, Professor Ajay Joshi

Google Scholar Link: https://scholar.google.com/citations?user=El3c-9EAAAAJ&hl=en&authuser=1

Abstract: With the rise in the use of Artificial Intelligence (AI) models and the resulting increase in the computational demand, the carbon emissions from AI infrastructure are skyrocketing. To address this immense computational demand of AI, researchers are focusing on developing energy-efficient hardware for AI through advancements in accelerators optimized for AI workloads. In parallel, recent studies have shown that the carbon emitted from hardware manufacturing and infrastructure (embodied carbon) contributes significantly to the carbon footprint of computing devices, even surpassing the emissions generated during their use (operational carbon). Therefore, we need to design computing systems that are carbon-efficient in both operation and manufacturing. In this research, we explore the potential of photonic computing to reduce the carbon footprint of AI systems, considering both embodied and operational emissions.

Many recently developed electro-photonic accelerators for Deep Neural Networks (DNNs) report orders of magnitude higher throughput and energy efficiency, potentially reducing operational carbon. However, none of them have examined their fabrication cost, i.e., embodied carbon. To address this gap, our first work introduces the first-ever model to estimate the embodied carbon of photonic chips. We then introduce EPiCarbon, an open-source tool to evaluate the carbon footprint of electro-photonic accelerators. Using EPiCarbon, we analyze the carbon footprint of state-of-the-art (SOTA) electro-photonic accelerators, demonstrating their potential as carbon-sustainable solutions for computationally demanding AI applications. One key limitation of photonics is that it struggles to handle the high-precision data of DNN models in an energy-efficient manner. So in our second work, we explore the use of Hyperdimensional Computing (HDC), a brain-inspired, lightweight AI algorithm with lower computational demands than DNNs and high tolerance to low bit precisions (≤8 bits). We design PhotoHDC, an electro-photonic accelerator for HDC, arguing that photonics and HDC form a well-matched pair for energy-efficient computing. PhotoHDC can achieve one to four orders of magnitude lower Energy-Delay-Product (EDP) than leading electro-photonic DNN accelerators for HDC training and inference. Finally, we propose our future work on designing a carbon-sustainable electro-photonic accelerator for HDC+DNN models. While HDC offers a low-precision, lightweight alternative to DNNs, it alone cannot achieve the desired accuracy like DNNs in complex tasks. Our future work builds on all our prior work and aims to reduce AI’s carbon footprint by combining HDC and DNNs to balance accuracy and operational efficiency, with photonic hardware playing the key role to reduce both operational and embodied carbon.