Reconsidering the energy efficiency of spiking neural networks