Energy-Efficient Spiking Neural Networks Using Approximate Neuron Circuits and 3D Stacking Memory

This paper proposes a combination approach of approximate neuron circuits and approximate stacking synaptic memory to evaluate the impact on accuracy and energy consumption. Instead of having accurate adders, this work utilizes approximate versions in the neuron’s circuits. Furthermore, we divide the fixed-point weight into several subsets, and each subset is allocated into a different silicon layer of 3D-IC-based stacking memory. Undervoting or power-gating each subset allows us to have approximate stacking memory that works in conjunction with approximate neurons. The evaluation showed that with the proper neuron circuit approximation, energy consumption improved by up to 24.84% per neuron with a 0.7% accuracy loss. Moreover, by combining with approximate stacking memory, we can reduce energy consumption by 55.9% while losing 1.6% of the average accuracy with the appropriate configuration.