AI at the Edge of Precision: Accelerating AI Hardware Efficiency
with Approximate Computing
Abstract
The rapid evolution of Generative AI models has ushered in groundbreaking advancements across industries, but their ever-increasing complexity presents a critical challenge: the need for efficient and scalable hardware accelerators. As model sizes grow exponentially, traditional computing paradigms struggle to meet the demands for higher performance, lower latency, and reduced energy consumption—particularly at the edge, where resource constraints are more pronounced. This talk explores how innovative microarchitectural approaches can unlock new frontiers for Generative AI computing. Central to this vision is the integration of approximate computing techniques, which embrace the trade-off between precision and efficiency to deliver significant gains in hardware efficiency and power savings with relatively negligible compromise in model accuracy. By leveraging decomposition, dynamic scheduling, and value-based computing, we can optimize execution efficiency, prioritize critical computations, and reduce redundancy, paving the way for next-generation AI accelerators. The talk will also highlight the role of these approaches in enabling efficient Generative AI at the edge, where constraints on power, latency, and hardware resources are most acute.
Short Biography
Prof. Freddy Gabbay leads the VLSI Lab at the Hebrew University’s Institute of Electrical Engineering and Applied Physics, focusing on VLSI design, computer architecture, machine learning, and domain-specific accelerators. He earned his B.Sc., M.Sc., and Ph.D. in Electrical Engineering from the Technion and has over two decades of experience in academia and industry. Prof. Gabbay began his career at Intel’s Microprocessor Research Lab in 1998 before joining Mellanox Technologies in 1999, where he contributed to switch architecture and ASIC design. He later served as a senior design manager at Freescale Semiconductor, leading baseband ASIC development, and returned to Mellanox in 2012 as Vice President of Chip Design. Holding 19 patents, Prof. Gabbay is a senior IEEE member, Associate Editor for IEEE Computer Architecture Letters, Guest Editor for Nature Distributed Computing, and Vice Chair of IEEE Computer Society Israel.
