N. Shukla, S. Datta, A. Parihar and A. Raychowdhury
Department of Electrical Engineering, The Pennsylvania State University, University Park,, PA, 16802, USA; Department of Electrical Engineering, University of Notre Dame, Notre Dame,, IN, 46556, USA
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta,, GA, 30332, USA
Complementary metal-oxide-semiconductor (CMOS) transistors supporting the Boolean computational framework have been the backbone of modern computation, and have fueled the information technology revolution for the past four decades. However, there is a class of computing tasks, such as associative processing, which involves the determination of a degree of match (or association) between two quantities in multidimensional space, for which the Boolean approach consumes a significant amount of computing resources. Mathematically, the quantification of the degree of match between two quantities requires the calculation of a distance (using a suitable distance norm) between two points that represents the two quantities in a high-dimensional space. It turns out that the so-called “Boolean bottleneck” arises from the requirement of a significant number of power-intensive multiply–accumulate (MAC) operations involved in the calculation of the distance norm.
Associative computing-based applications such as pattern recognition, visual saliency, and image segmentation are ...