Reproduce “Spiker: an FPGA-optimized Hardware acceleration for Spiking Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Alessio Carpegna, Alessandro Savino, Stefano Di Carlo

Spiker, an FPGA-optimized hardware accelerator developed by researchers at Politecnico di Torino, enhances Spiking Neural Network (SNN) inference for edge devices by applying hardware-aware simplifications. The system achieved an inference time of 215 µs/image on the MNIST dataset, demonstrating approximately 14 times faster processing compared to other FPGA-based SNN accelerators with a minimal reduction in accuracy.

Key insights

  • Carefully chosen hardware-aware simplifications and architectural optimizations can drastically reduce the hardware complexity and footprint of SNN accelerators with only an acceptable impact on inference accuracy.
  • Exploiting the inherent sparsity of SNN spike activity by dynamically skipping computation steps when no active spikes are present can lead to substantial reductions in execution time and power consumption.
  • Even with a clock-driven approach, maximizing parallelism and optimizing memory access for synaptic weights (using BRAMs) can yield superior inference speeds and competitive energy efficiency compared to other FPGA-based SNN designs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.