New arXiv study reveals when SNNs beat QNNs on energy efficiency

Spiking Neural Networks (SNNs) have long been promoted as more energy-efficient alternatives to traditional Quantized Artificial Neural Networks (QNNs) because they use event-driven, spike-based computation. However, most prior energy comparisons oversimplify by ignoring the costs of data movement and memory accesses. A new paper by Zhanglu Yan, Zhenyu Bai, Kaiwen Tang, and Weng-Fai Wong addresses this gap with a rigorous analytical energy model. They first establish a fair baseline by constructing representation capacity-matched twins: a rate-encoded SNN with T timesteps is compared to a QNN with ⌈log₂(T+1)⌉ bits, ensuring both models have similar expressive power and hardware requirements. The model covers core computation, memory system, and network-on-chip characteristics, then explores a wide parameter space including spike rate, time window, sparsity, weight bits, and model size.

The results reveal that SNNs are not universally more efficient. Under typical neuromorphic hardware, an SNN with a moderate time window of T=5 requires an average spike rate below 5.7% to beat its matched QNN. For larger windows, the spike rate threshold becomes even stricter. The paper also identifies other regimes where SNNs genuinely excel, such as very sparse activation patterns or when memory access is minimized. These insights are critical for hardware designers and algorithm engineers who want to build truly energy-efficient neural network solutions. The paper has been accepted by IEEE Transactions on Computer-Aided Design (TCAD).