EECS Colloquium: Interconnect specialization for addressing data movement inefficiencies

About the event

Abstract:
Data explosion is facing grand computing challenges on computation power, memory capacity, and communication bandwidth. In the post-Moore’s Law era, specialization has become a new driving force to tackle these problems for performance and energy efficiency improvement. While computation acceleration and memory advancement have been widely studied, communication specialization is still in its infancy.  As systems scale up and scale out further, communication can soon become a problematic performance bottleneck. In this talk, I introduce interconnect specialization approaches to addressing the data movement inefficiency problem from a communication perspective by exploiting the information context of data being moved. First, I show that awareness of data error tolerance can enable approximation to improve compression for high-throughput communication. Second, I present that computation semantics of data can empower in-network computing for data movement reduction. Third, I demonstrate that the static data communication pattern of distributed DNN training can facilitate message scheduling for high bandwidth utilization. Finally, I conclude my talk by laying out my future research road map towards intelligent and secure computing.

Bio:
Jiayi Huang is currently a postdoctoral researcher in the Department of Electrical and Computer Engineering at UC Santa Barbara. Before that, he received his Ph.D. degree in Computer Engineering from Texas A&M University in 2020 and his bachelor’s degree from Zhejiang University in 2014. His research interest is computer architecture and systems, with a special focus on data-centric architectures, intersection of machine learning and system design, and security. His dissertation advocates for communication specialization for emerging applications. He was a recipient of the Dissertation Fellowship at Texas A&M.