Scholarlink

Deep Spiking Neural Networks (DSNNs) excel in brain inspired computing by accurately emulating biological neural information processing through precise temporal coding. However, their implementation on von Neumann architectures suffers from low simulation efficiency and high hardware resource consumption. To address these challenges, this paper proposes a novel deep spiking neuromorphic chip based on a Multichannel Time-Multiplexed Architecture (MTMA). By synergistically combining the biological plausibility of the Spike Response Model (SRM) with the parallel computing advantages of field-programmable gate arrays (FPGAs), we have developed a configurable hierarchical system incorporating Spiking Neural Processing Units (SNPUs) and Virtual Spiking Neurons (VSNs). Implementation on the Virtex-7 FPGA platform demonstrates that for a 32-72-48-4 network configuration, the 8-channel MTMA architecture achieves a 75% reduction in system delay (from 2.04ms to 0.84ms) compared to serial implementations, while maintaining resource utilization at merely 4.07% (LUT) and 1.71% (DSP) of a fully parallel architecture. Validation using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset shows classification accuracies of 84.58% (training set) and 83.63% (test set), consistent with software simulation results.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Spiking Neural Networks FPGA Implementation Based on Multichannel Time-Multiplexed Architecture

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xianghong Lin,
  • Chengyang Xie,
  • Ruidong Ma

摘要

Deep Spiking Neural Networks (DSNNs) excel in brain inspired computing by accurately emulating biological neural information processing through precise temporal coding. However, their implementation on von Neumann architectures suffers from low simulation efficiency and high hardware resource consumption. To address these challenges, this paper proposes a novel deep spiking neuromorphic chip based on a Multichannel Time-Multiplexed Architecture (MTMA). By synergistically combining the biological plausibility of the Spike Response Model (SRM) with the parallel computing advantages of field-programmable gate arrays (FPGAs), we have developed a configurable hierarchical system incorporating Spiking Neural Processing Units (SNPUs) and Virtual Spiking Neurons (VSNs). Implementation on the Virtex-7 FPGA platform demonstrates that for a 32-72-48-4 network configuration, the 8-channel MTMA architecture achieves a 75% reduction in system delay (from 2.04ms to 0.84ms) compared to serial implementations, while maintaining resource utilization at merely 4.07% (LUT) and 1.71% (DSP) of a fully parallel architecture. Validation using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset shows classification accuracies of 84.58% (training set) and 83.63% (test set), consistent with software simulation results.