Energy-Efficient Approximate Computing: From Circuits to Systems
Minimizing energy consumption of computing circuits enables prolonged operation for battery-operated devices and it is key to alleviate the physical constraints on improving performance. Approximate computing is an emerging paradigm where circuits are deliberately designed such that their results are approximate. By giving up some arithmetic accuracy, it is possible to design nano-scale circuits with dramatically lower power dissipation and smaller silicon footprint. Using approximate circuits is attractive for emerging classes of applications that are inherently tolerant to errors, which include signal and image processing, computer graphics, computer vision, machine learning, and neuromorphic computing. In this talk, l I will first describe new methods to design approximate circuits for basic arithmetic building blocks, such as adders, multipliers and dividers. Our circuits use a novel circuit mechanism to dynamically zoom in on the most relevant bits in input operands and approximate the remaining bits. I will show how our approximate arithmetic circuits can provide dramatic reductions in power consumption with negligible reduction in accuracy. I will then describe new automated methods to synthesize arbitrary approximate circuits. Our techniques enable designers to automatically discover large number of approximate circuits from their original circuit, by mutating the original circuit in intelligent ways and retaining the most promising mutants. Our tool identifies the Pareto-optimal mutant circuits that give optimal trade-off between accuracy and power consumption. I will demonstrate results for various circuit applications, including for deep neural networks, which recently emerged as a major success story in machine learning and artificial intelligence. I will describe how our techniques drastically simplify the underlying computation requirements in terms of circuit area cost and power consumption without compromising accuracy.
Associate Professor, School of Engineering, Brown University
Sherief Reda is an Associate Professor at the School of Engineering, Brown University. He joined the Computer Engineering group at Brown in 2006 after receiving his Ph.D. in computer science and engineering from University of California, San Diego. His research interests are in the areas of energy-efficient computing, thermal-power sensing and management, and low-power design techniques. Professor Reda received a number of research awards and acknowledgments, including a best paper award in DATE 2002, a hot article in Operations Research letters in 2004, best paper nominations in ICCAD 2005 and ASPDAC 2008, a NSF CAREER award in 2010, a best paper award in ISLPED 2010 and a best paper nomination ICCAD 2015. He is a senior member of IEEE.
Host: Professor Yousef Mahmoud