Performance Analysis of Different Interconnect Networks for Network on Chip

Abstract

Nowadays, every electronic system, ranging from a small mobile phone to a satellite sent into space, has a System-on-Chip (SoC). SoCs have undergone rapid evolution and are still progressing at a swift pace. Due to explosive evolution of semiconductor industry, the devices are scaling down at a rapid rate and hence, the SoCs today have become communication-centric and shared bus system and crossbar system were fail to performed communication in side SoC. Interconnection networks offer an alternate solution to this communication paradigm and are becoming persistent in SoC. A NoC based interconnect network is a well-organized and efficiently use of limited communication channel while maintaining low packet latency, high saturation throughput, high communication bandwidth amongst different IPs core with a minimum area and low power-dissipation. In this thesis we present details performance analysis of four interconnect network mesh, torus, fat tree and butterfly in term of latency and throughput under uniform, tornado, neighbour, bit reversal and bit complement traffic using cycle accurate simulator. We also implement NoC interconnect networks on FPGA and see the effect of NoC parameters(FDW,FBD,VC) on FPGA, and validate their performance through FPGA synthesis . We found that the FDW and buffer depth have the great effect on FPGA resources, Virtual Channels (VCs) with all NoC parameter have considerably effect on buffer size and routing and logic requirements at NoC. We also analysis all interconnect networks in term of power and area at 65 nm technology by using synopsis tool. We found that butterfly interconnect network has highest power and Area efficient interconnect network but it will suffer heavily degradation on performance at high load so fat tree network is efficient network among all interconnect network.