NERO mastering 300k CNN cells

Abstract

A novel massively-parallel fine-grain architecture featuring a digital emulation of Cellular Nonlinear Networks (CNN) is presented. A virtual cellular network is processed line-by-line by a locally connected linear array of processing elements. The resulting computing system is able to execute complex CNN program code consisting of consecutive operations. Furthermore we present a scalable FPGA implementation of this architecture for currently up to 480 × 640 cells with a precision up to 18 bit.