Fullerene-Inspired Efficient Neuromorphic Network-on-Chip Scheme.

IEEE Trans. Circuits Syst. II Express Briefs(2024)

引用 0|浏览2
暂无评分
摘要
Router-based network-on-chips (NoCs) are widely used for multi-core interconnection in neuromorphic chips. The NoC scheme has a great impact on communication efficiency. Although many efforts have been made to develop high-efficiency NoC schemes, they still suffer from poor topological connectivity and the single function of routing nodes, making it stay on a low efficiency in neuromorphic computing. This work proposes an efficient merge-and-link (MerLin) NoC scheme by reconsidering interconnection topology and bottom communication nodes. At the topology level, an efficient communication structure inspired by the fullerene-60 structure is proposed as the NoC topology; at the node level, an efficient merge-and-link routing scheme is proposed for implementing the bottom routers. The fullerene-60-inspired topology brings improvements of 8∼63% in average degree and 2× in connectivity and a decrease of 9∼27% in the communication distance compared with other topologies. The merge-and-link router supports hybrid transmission modes, achieves a low transmission latency of 2 clock periods, and only occupies 0.004 mm2 area. Eventually, the MerLin NoC achieves a minimal zero-load latency of 6.32 clock cycles, outperforming related works. This work is expected as a promising solution to optimize the interconnection of next-generation multi-core neuromorphic chips.
更多
查看译文
关键词
Neuromorphic computing,network-on-chip (NoC),router,spiking neural network (SNN)
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要