High performance NoC architecture for two hidden layers BP Neural Network

Abstract

Artificial neural networks (ANNs) are widely used in applications of an intelligent system such as pattern recognition, fuzzy system, optimization and control. We have already proposed a novel NoC architecture for different kinds of BPANNs and it was shown that the architecture is a promising hardware implementation for Neural Network. However, some problems to be solved are still remained. One of them is performance. In this paper, we propose another NoC architecture, network topology and routing strategy for higher performance. Experimental results by NoC simulator show that this new architecture and routing strategy reduce the communication load, reduce both latency by 7.7% and dynamic power consumption by 10.3% and also improve throughput by 8.1%, all compared with the previous one.