Soft DT-CNN core implementation

Abstract

Digital implementations of discrete-time cellular neural networks have steadily been improved, gradually gaining capacity and therefore applicability. It is required that CNN operations can be freely sequenced and iterated to keep the memory bandwidth limited. Therefore the paper introduces a CNN Instruction set architecture. It is shown that this turns a 400 frames per second CNN network into a conventional stream-processing peripheral at a mere 1 - 5% area overhead.