Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SparrowSNN: A Hardware/software Co-design for Energy Efficient ECG Classification

Zhanglu Yan, Zhenyu Bai, Tulika Mitra, Weng-Fai Wong

TL;DR

The paper tackles the challenge of energy-efficient real-time ECG classification on wearables by revealing hardware overheads often ignored in SNN energy estimates and proposing a hardware-software co-design, sparrowSNN. It introduces Sum-Spikes-Fire (SSF), a hardware-friendly spike activation that decouples memory access from firing, and demonstrates a co-designed SNN architecture with an ultra-low-power 22 nm ASIC. Through post-training 8-bit quantization and ANN-to-SNN conversion, SparrowSNN achieves 98.29% accuracy on MIT-BIH while consuming only 31.39 nJ per inference at 4 MHz (≈6.1 μW), outperforming comparable systems in energy efficiency. The work also analyzes the energy-accuracy trade-offs between SNNs and quantized ANNs and shows patient-wise online training can further boost performance, highlighting practical implications for continuous ECG monitoring on edge devices.

Abstract

Heart disease is one of the leading causes of death worldwide. Given its high risk and often asymptomatic nature, real-time continuous monitoring is essential. Unlike traditional artificial neural networks (ANNs), spiking neural networks (SNNs) are well-known for their energy efficiency, making them ideal for wearable devices and energy-constrained edge computing platforms. However, current energy measurement of SNN implementations for detecting heart diseases typically rely on empirical values, often overlooking hardware overhead. Additionally, the integer and fire activations in SNNs require multiple memory accesses and repeated computations, which can further compromise energy efficiency. In this paper, we propose sparrowSNN, a redesign of the standard SNN workflow from a hardware perspective, and present a dedicated ASIC design for SNNs, optimized for ultra-low power wearable devices used in heartbeat classification. Using the MIT-BIH dataset, our SNN achieves a state-of-the-art accuracy of 98.29% for SNNs, with energy consumption of 31.39nJ per inference and power usage of 6.1uW, making sparrowSNN the highest accuracy with the lowest energy use among comparable systems. We also compare the energy-to-accuracy trade-offs between SNNs and quantized ANNs, offering recommendations on insights on how best to use SNNs.

SparrowSNN: A Hardware/software Co-design for Energy Efficient ECG Classification

TL;DR

The paper tackles the challenge of energy-efficient real-time ECG classification on wearables by revealing hardware overheads often ignored in SNN energy estimates and proposing a hardware-software co-design, sparrowSNN. It introduces Sum-Spikes-Fire (SSF), a hardware-friendly spike activation that decouples memory access from firing, and demonstrates a co-designed SNN architecture with an ultra-low-power 22 nm ASIC. Through post-training 8-bit quantization and ANN-to-SNN conversion, SparrowSNN achieves 98.29% accuracy on MIT-BIH while consuming only 31.39 nJ per inference at 4 MHz (≈6.1 μW), outperforming comparable systems in energy efficiency. The work also analyzes the energy-accuracy trade-offs between SNNs and quantized ANNs and shows patient-wise online training can further boost performance, highlighting practical implications for continuous ECG monitoring on edge devices.

Abstract

Heart disease is one of the leading causes of death worldwide. Given its high risk and often asymptomatic nature, real-time continuous monitoring is essential. Unlike traditional artificial neural networks (ANNs), spiking neural networks (SNNs) are well-known for their energy efficiency, making them ideal for wearable devices and energy-constrained edge computing platforms. However, current energy measurement of SNN implementations for detecting heart diseases typically rely on empirical values, often overlooking hardware overhead. Additionally, the integer and fire activations in SNNs require multiple memory accesses and repeated computations, which can further compromise energy efficiency. In this paper, we propose sparrowSNN, a redesign of the standard SNN workflow from a hardware perspective, and present a dedicated ASIC design for SNNs, optimized for ultra-low power wearable devices used in heartbeat classification. Using the MIT-BIH dataset, our SNN achieves a state-of-the-art accuracy of 98.29% for SNNs, with energy consumption of 31.39nJ per inference and power usage of 6.1uW, making sparrowSNN the highest accuracy with the lowest energy use among comparable systems. We also compare the energy-to-accuracy trade-offs between SNNs and quantized ANNs, offering recommendations on insights on how best to use SNNs.
Paper Structure (27 sections, 10 equations, 8 figures, 10 tables, 4 algorithms)

This paper contains 27 sections, 10 equations, 8 figures, 10 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Spiking neural network
  4. MIT-BIH Arrhythmia Database
  5. Methods
  6. Sum-Spikes-Fire model vs Integrate-and-Fire model
  7. Network design
  8. Post-training Static Quantization for SNN
  9. Workflow of SparrowSNN
  10. Ultra-Low Power ASIC Designs for SNN
  11. Dynamic and Static Power Consumption
  12. Memory Design
  13. Improving the energy efficiency with the SSF mechanism
  14. ASIC design
  15. Performance
  16. ...and 12 more sections

Figures (8)

  • Figure 1: Workflow of SparrowSNN
  • Figure 2: Energy Consumption per bit of SRAM block for different bus width
  • Figure 3: The FSM. TODO: make it more vertical to save place
  • Figure 4: The design of the CU and the scheduler associated.
  • Figure 5: N, SVEB, VEB and F heartbeats
  • ...and 3 more figures