Design space exploration for the integrated digital CNN camera

Abstract

The main thrust of CNN technology has been in analog implementations. Early attempts for digital implementations have a limited potential, confining the usage to for instance linear algorithms for B/W pictures. The paper looks into the prospects of successfully integrating the full CNN potential. This requires a carefully tuning of memory and network-on-chip bandwidth to the temporal/spatial aspects of the neural nodes. The paper introduces an exploration model and discusses architecture with application to streaming grey-scale images.