A high performance digital neural processor design by Network on Chip architecture

This paper describes a high performance neural processor by using a Network on Chip (NoC) architecture to solve the interconnection and performance problems in hardware neural networks. The proposed NoC-based neural processor is composed of 20 tiles in 4×5 2-D array, and each tile includes a Process Element (PE) and a packet switched router. In each PE, four neurons are implemented to achieve low communication load. The network is 2D torus topology, and it has a 32 G/s bandwidth and asynchronous clocking system. Our proposed neural processor is designed using 90-nm CMOS technology with one Poly and nine metals, and its performance is evaluated. As a result, it can achieve over 3.1 G Connection Per Second (CPS) of performance while power dissipation is 1.1317 W at 1.2 V supply-voltage and 25 mm² chip area. Compared with the other existing hardware neural networks, the proposed processor can achieve low communication load and high performance, and it is reconfigurable and extendable.