Reliability Analyses of Ultra-Low Voltage Analog Spiking Neurons

G. Brandsteert, L. Van Brandt, D. Flandre (Univ. Catholique de Louvain, Belgium)

We investigate the behaviour of a bio-inspired artificial neuron, implementing a simplified Morris-Lecar model under downscaled supply voltages in the subthreshold regime towards ultra-low-power performance. The neuron spike characteristics, including amplitude, period, and typical shape, are analysed as the supply voltage is reduced. We observe a dramatic loss of the typical linear-exponential behaviour in the rising edge of the spike related to a drastic reduction in spike amplitude and an increase in spike period. These are used to define figures-of-merit of spiking operation for ad hoc neuromorphic computations. Subsequently, through Monte Carlo variability analyses, we highlight that ultra-low-voltage neurons are highly sensitive to process variations and prone to statistically likely failures. Our study evidences the challenges of maintaining stability and performance in low-power neuromorphic circuits and the importance of reliability assessments.