Abstract
Network-on-Chip (NoC)-based Spiking Neural Networks (SNNs) provide a scalable platform for neuromorphic computing, yet efficient global time-step synchronization remains a critical challenge. While event-driven time-step synchronization schemes effectively eliminate idle time, it inevitably generates a massive volume of Time-step Update Packets (TUPs), causing severe network congestion and performance degradation in large-scale meshes. To overcome this traffic bottleneck, this paper proposes a reconfigurable SNN accelerator featuring a novel TUP aggregation mechanism. Unlike traditional schemes, the proposed router intelligently merges TUPs along the feed-forward path. This mechanism significantly suppresses redundant traffic, reduces the communication load caused by TUPs. Experimental results demonstrate that this traffic suppression reduces the TUP transition time overhead by approximately 75% (from 20.51% to 6.06% of total latency). At the core level, we introduce a reconfigurable Neural Processing Element (NPE) designed to support both low-power Leaky Integrate-and-Fire (LIF) and biologically plausible Adaptive LIF (ALIF) models. By logic circuit sharing and mode-dependent clock gating, the design achieves dual-mode flexibility with negligible area overhead compared to a baseline ALIF implementation and a 26.55% area reduction compared to discrete implementations. The 8 × 8 mesh SNN accelerator, implemented in SMIC 55nm CMOS technology, achieves a competitive energy efficiency of 3.58 pJ/SOP, 97.80% classification accuracy with 1.19μJ/sample in LIF mode in image recognition and 93.53% accuracy with 8.23μJ/sample in ALIF mode in keyword spotting(KWS).
Funding Information
  • Scientific and Technological Innovation 2030 Major Project (2021ZD0200300)

This publication has 23 references indexed in Scilit: