A High-Throughput Network On-Chip in Full-Mesh Architecture

Hongyu Meng,Lei Yang,Zijun Liu,Donglin Wang
DOI: https://doi.org/10.1587/elex.15.20180635
2018-01-01
IEICE Electronics Express
Abstract:Just as Moores law dictates, technology scaling still continues in recent years. Thus increasing the number of on-chip cores continues to be the facto strategy to scale performance while interconnection networks remain important in multi-core chip systems. In this paper, we present a brand-new network-on-chip (NoC) based on full-mesh architecture Full-Mesh NoC (FMN) that can provide high connectivity with low-radix routers. We give an example of implementation for a 64-node FMN system in 28 nm process. Results show that this FMN can achieve high frequency of 700MHz while area requirement of each router is less than 0.035mm2. At last, we simulate a no-buffered FMN and a buffered FMN under uniform random traffic. Results show that the throughput of the no-buffered FMN can achieve the level of high-radix crossbar. The buffered FMN can achieve 2X throughput compared with the traditional buffered 2D-mesh NoC.
What problem does this paper attempt to address?