Performance and Energy Efficiency of Weight-Quantized Spiking Neural Networks with Resistive Memory Synapses

The surge in computational demand for deep learning systems necessitates energy-efficient hardware. Spiking neural networks (SNNs) implemented with resistive memory devices offer a low-power solution but have constraints regarding weight precision. This study evaluated the trade-off between recognition performance and power dissipation in weight-quantized unsupervised SNNs. Our analysis revealed that performance degradation caused by reduced bit-widths can be effectively compensated for by scaling the network size. Notably, large-scale networks utilizing binary weights demonstrated robust inference capabilities comparable to those of higher-precision models. Furthermore, circuit-level simulations confirmed that minimizing weight precision results in a substantial reduction in power consumption. These findings suggest that large-scale, low-precision architectures provide a compelling design strategy for energy-efficient edge neuromorphic systems.