Source-linked AI summary
Toward Fast Neural Computing using All-Photonic Phase Change Spiking Neurons
Indranil Chakraborty, Gobinda Saha, Abhronil Sengupta, Kaushik Roy
TL;DR
CMOS neuromorphic implementations are limited by energy and area inefficiency, while electrical PCM operation faces high write-time constraints. The paper demonstrates a GST-embedded microring integrate-and-fire neuron using photonic phase-change dynamics and proposes integration with photonic synapses. The authors report 1 pJ per neuron per time-step and an all-photonic spiking-neural-network inferencing framework without significant classification-performance loss.
Problem
CMOS neuromorphic implementations are energy- and area-inefficient, while electrical PCM operation has significant write-time energy restrictions.
Method
The paper implements a purely photonic integrate-and-fire neuron using GST phase-change dynamics embedded on a microring resonator.
Results
1 pJ per neuron per time-step is reported, and the proposed neuron can be integrated into an all-photonic spiking-neural-network inferencing framework without significant classification-performance loss.
Takeaways & Limitations
The design supports potential all-photonic spiking-neural-network implementations with wavelength multiplexing that could reduce cross-talk and enable denser networks.
Abstract
from arXiv · showhide
The rapid growth of brain-inspired computing coupled with the inefficiencies in the CMOS implementations of neuromrphic systems has led to intense exploration of efficient hardware implementations of the functional units of the brain, namely, neurons and synapses. However, efforts have largely been invested in implementations in the electrical domain with potential limitations of switching speed, packing density of large integrated systems and interconnect losses. As an alternative, neuromorphic engineering in the photonic domain has recently gained attention. In this work, we demonstrate a purely photonic operation of an Integrate-and-Fire Spiking neuron, based on the phase change dynamics of Ge$_2$Sb$_2$Te$_5$ (GST) embedded on top of a microring resonator, which alleviates the energy constraints of PCMs in electrical domain. We also show that such a neuron can be potentially integrated with on-chip synapses into an all-Photonic Spiking Neural network inferencing framework which promises to be ultrafast and can potentially offer a large operating bandwidth.
Introduction
Neuromorphic hardware seeks brain-inspired efficiency, but CMOS implementations remain energy- and area-inefficient. Photonic phase-change materials offer a route toward faster PCM operation, including GST with sub-ns photonic write speeds.
- Spiking neural networks use event-driven spike-based processing to support bio-plausible hardware implementations targeting higher energy efficiency.
- CMOS neuromorphic implementations do not match the human brain’s energy efficiency and are area-inefficient.
- Phase-change materials face electrical-domain energy restrictions because of high write times, requiring a 10× reduction in exciting current or write-pulse duration to outperform CMOS.
- GST has demonstrated sub-ns write speeds when excited by photonic laser pulses, addressing a key electrical-domain PCM limitation.
GST embedded Ring Resonator as a Integrate-Fire Neuron
The proposed neuron uses GST-controlled microring transmission to integrate bipolar optical inputs and fire when the GST reaches full amorphization and the membrane potential crosses threshold. Opposite transmission responses from paired resonators are combined interferometrically to produce the resultant integration.
- Ring-resonator operation: A GST-covered ring resonator controls light at the THROUGH and DROP ports through GST-state-dependent absorption and resonant coupling.The ring’s resonant condition and coupling parameters determine conditional guidance through the ports.
- Device modeling: The neuron’s photonic design is parameterized using effective-medium estimates for partially crystallized GST and material and device dimensions.The cited formulation uses crystalline and amorphous permittivities together with waveguide and GST parameters.
- GST optical response: GST crystallization increases effective absorption, whereas amorphization decreases it; consequently, transmission increases with crystallization degree at the modeled ports.The theoretical transmission behavior motivates using the GST-ring system for integrate-and-fire operation.
- Bipolar integration: Paired positive and negative ring resonators receive weighted sums of opposite polarity, and an interferometer combines their outputs into the resultant membrane-potential integration.A phase modulator tunes the positive path so the interferometer produces the sum of the incoming pulses.
- Firing: When GST reaches full amorphization, the membrane potential crosses Pthresh and an additional photonic circuit generates the neuron’s spike.The firing unit includes a photonic amplifier, circulator, and crystalline GST-loaded rectangular waveguide; read and write phases alternate.
- GST optical response: The GST dynamics provide state-dependent integration because amorphization growth depends on both the current GST state and incident-pulse amplitude.GST absorption heats the material, and regions above approximately 877 K grow in amorphous extent.
Results
The results establish optical and thermal behavior of GST-ring devices and show how their transmission can support bipolar integration in a photonic spiking-neural-network framework. Simulations further indicate energy advantages over comparable electrical PCM devices.
- Phase-change dynamics: Simulation and experimental benchmarking validate the GST phase-change modeling framework for the proposed photonic neuron.The benchmark compares transient GST temperature responses under matched excitation conditions and reports good agreement with experimental data.
- Phase-change dynamics: −3.71 dB attenuation for c-GST versus −0.26 dB for a-GST demonstrates strong state-dependent optical absorption.This contrast supports progressive amorphization while preserving already amorphized volume over the selected input-power range.
- Optical response: At λread = 1529 nm, GST amorphization decreases ‘THROUGH’ transmission and increases ‘DROP’ transmission, providing opposite-direction integration responses.The simulated port responses are connected through an interferometer to obtain resultant membrane-potential integration.
- Spiking-neural-network framework: The proposed framework computes bipolar weighted sums with separate dot-product engines and ring resonators, then integrates them into an effective membrane potential.The simulated SNN receives spike streams, updates neuron membrane potentials, and produces integrate-and-fire behavior.
Discussion
The proposed GST-based all-photonic integrate-and-fire neuron addresses electrical PCM speed and energy limitations while enabling photonic SNN integration. Wavelength multiplexing could reduce cross-talk and support denser, larger on-chip networks.
- Discussion: Sub-ns write speeds are proposed to overcome the energy-efficiency scaling bottleneck of electrical phase change materials.The paper states that beating CMOS energy efficiency requires a 10× reduction in current pulse amplitude or increased pulse duration.
- Discussion: The work demonstrates a biologically plausible phase-change spiking neuron operating in the photonic domain.
- Discussion: The proposed neuron can potentially integrate with photonic synapses in an all-photonic SNN inferencing framework without significant classification-performance loss.
- Discussion: Wavelength multiplexing could eliminate cross-talk between neighboring neural elements, enabling denser networks and potentially larger networks on the same chip.The resonant wavelength can be modulated by varying device dimensions.