An Asynchronous Mixed-Signal Resonate-and-Fire Neuron

TL;DR

This work tackles low-power, real-time temporal processing at the edge by implementing a CMOS mixed-signal Resonate-and-Fire neuron with asynchronous handshake. The circuit combines a resonator, spike generator, handshake logic, and synapses to realize event-driven, sub-threshold dynamics inspired by biological resonator neurons, building on prior analog implementations. Experimental results demonstrate tunable resonance, Class II excitability, and frequency-selective spiking, supported by die-to-die variability and power analyses that indicate feasibility for large-scale neuromorphic deployment. The approach advocates integration with neuromorphic transceivers for efficient edge processing of temporal signals such as audio, offering a path toward scalable, bio-inspired, energy-efficient hardware.

Abstract

Analog computing at the edge is an emerging strategy to limit data storage and transmission requirements, as well as energy consumption, and its practical implementation is in its initial stages of development. Translating properties of biological neurons into hardware offers a pathway towards low-power, real-time edge processing. Specifically, resonator neurons offer selectivity to specific frequencies as a potential solution for temporal signal processing. Here, we show a fabricated Complementary Metal-Oxide-Semiconductor (CMOS) mixed-signal Resonate-and-Fire (R&F) neuron circuit implementation that emulates the behavior of these neural cells responsible for controlling oscillations within the central nervous system. We integrate the design with asynchronous handshake capabilities, perform comprehensive variability analyses, and characterize its frequency detection functionality. Our results demonstrate the feasibility of large-scale integration within neuromorphic systems, thereby advancing the exploitation of bio-inspired circuits for efficient edge temporal signal processing.

Figures (6)

• Figure 1: Photograph of the fabricated wafer containing multiple dies. The dashed white rectangle highlights one die and the solid lined rectangle delimits the test structure that includes the neuron, the I/O circuits, and the pads (122$\times$1772). The neuron has an area of 35$\times$155.
• Figure 2: Schematic of the circuit. The four functional blocks are highlighted in the circuit: in red, the two analog synapses, in orange, the resonator, in green, the handshake logic, and in blue, the spike generator.
• Figure 3: Measured membrane voltage dynamics for spikes and constant input. (a) Time-domain (up to 0.15) and (b) phase-plane dynamics (up to 0.3) after one inhibitory input. (c) Rhythmic firing in response to step excitatory stimulation, with a baseline value of 0 and a step value of 0.5. (d) Response to an up-chirp consisting of 10 spikes per frequency across a total of 13 frequencies. It shows frequency-selective spiking around 150. The spike input in a) and d) has a time width of 100 and an amplitude of 0.5. We define the inhibitory input voltage as $V_{\text{inh}}$ and the excitatory input voltage as $V_{\text{DD}} - V_{\text{exc}}$.
• Figure 4: Measured die-to-die variability (100 dies). (a) Baseline voltage distributions for $U$ (mean value 724) and $V$ (mean value 758). (b) First peak amplitude distributions for $U$ (mean value 757) and $V$ (mean value 786). (c) Resonant frequency distributions (mean value 170). (d) Q-factor distributions (mean value 129). (e) Distribution of the power consumption (mean value 571).
• Figure 5: (a) The resonant frequency depends linearly on the bias current (post-layout simulation results), ranging from 6 at 10 to 2000 at 2.51. (b) curve typical of Class ii neurons (experimental results). The neuron is silent for inputs below 420 and exhibits high-frequency rhythmic firing for larger inputs. The empty circle in the graph represents this discontinuous behavior. Each point corresponds to the mean calculated from 100 spikes, with bars showing the standard deviation.