On the packet-switched implementation of a discrete-time CNN

Abstract

Cellular neural networks are widely used with real-time image processing's applications. Such systems can be efficiently realized using macro enriched field-programmable gate-arrays. This paper explores the benefits of packet switching and discusses its advantages over a current design based on circuit switching. The implementation is performed using Xilinx Virtex-II Pro P30 and handles around 500 Mpixels per second using 128 parallel processing nodes.