Sephia Achieves Sub-One-Laser-Per-Neuron Efficiency with Scalable Optical Neuromorphic Computing Architecture

The pursuit of energy-efficient computing increasingly focuses on mimicking the human brain, leading researchers to explore spiking neural networks. Matěj Hejda, Aishwarya Natarajan, Chaerin Hong, and colleagues at Hewlett Packard Enterprise’s Large-scale Integrated Photonics group, now present SEPhIA, a novel architecture that overcomes limitations in power consumption, signal interference, and physical size that have previously hampered progress in this field. This innovative system employs integrated photonics and electronic components to achieve remarkably efficient operation, requiring less than one laser per spiking neuron. Validating their design through detailed simulations and a physics-aware model, the team demonstrates classification accuracies exceeding 90% on complex datasets, paving the way for scalable and expressive neuromorphic computing systems capable of tackling real-world, spike-encoded tasks.

Photonic Spiking Neural Network Architecture Details

Scientists have detailed a novel optoelectronic spiking neural network (SNN) architecture, SEPhIA, designed to overcome limitations in current neuromorphic computing systems and enable scalable, realistic photonic neural networks. This innovative approach minimizes the footprint per neuron by utilizing microring resonator modulators and multi-wavelength light sources, effectively reducing the laser requirement to less than one per spiking neuron. Researchers engineered a hybrid architecture, integrating photonic and electronic components to leverage the strengths of both domains. The team validated SEPhIA at both the device and architectural levels through time-domain co-simulation, coupling excitable CMOS circuits with a physics-aware, trainable optoelectronic SNN model.

This comprehensive modelling framework utilized experimentally derived device parameters, ensuring the accuracy and relevance of the simulations to real-world hardware constraints. Scientists harnessed balanced photodetection to address challenges inherent in all-optical approaches, such as limited fan-in and the absence of spike inhibition functionality. Experiments demonstrate that the multi-layer optoelectronic SNN achieves classification accuracies exceeding 90% on a four-class spike-encoded dataset, closely matching the performance of software models. A detailed design space study quantified the impact of photonic device parameters on SNN performance under constrained signal-to-noise conditions, revealing critical relationships for optimization. This work establishes SEPhIA as a scalable, expressive, and physically grounded solution for neuromorphic photonic computing, capable of tackling spike-encoded tasks with improved efficiency and practicality.

Spiking Photonic Integration For Scalable Neuromorphic Computing

Scientists have developed SEPhIA, a new photonic-electronic spiking neural network architecture designed for practical implementation and scalability. This work addresses limitations in existing spiking neural networks by focusing on realistic constraints such as power consumption, signal interference, and physical size. The team achieved effective efficiency by utilizing microring resonator modulators and multi-wavelength light sources, requiring less than one laser per spiking neuron. Validation of SEPhIA involved both detailed computer simulations of the electronic and photonic components, and the creation of a physics-based, trainable model of the optoelectronic system.

These simulations incorporated experimentally derived parameters to ensure accuracy. The multi-layer optoelectronic spiking neural network successfully classified spike-encoded data with accuracies exceeding 90%, closely matching the performance of software-based models. Detailed analysis of the individual components revealed diverse spiking behaviors, including tonic spiking, where neurons fire continuously at a constant rate, spike-frequency adaptation, where firing rates decrease over time, and bursting dynamics, characterized by clusters of spikes followed by periods of inactivity, all at rates of 1 GSpike/s. This system uniquely combines excitable analog CMOS circuits with compact microring resonator modulators and multi-wavelength lasers, achieving high efficiency by sharing laser resources across multiple spiking neurons. Validating the design through both detailed co-simulations and a physics-aware training model, the team demonstrates classification accuracies exceeding 90% on a benchmark dataset, comparable to performance achieved with software-based SNNs. This work establishes a physically grounded and scalable solution for neuromorphic photonic computing, addressing limitations found in other approaches.

The team quantified the impact of photonic device parameters on SNN performance, revealing how these elements influence overall accuracy under realistic conditions. Furthermore, the architecture exhibits favorable energy efficiency metrics when compared to alternative spiking neuromorphic systems. The authors acknowledge that future work could explore incorporating factors such as limited bit precision, additional noise sources, and more advanced sparsity techniques, as well as applying the system to natively temporal or optical datasets. These advancements promise to further enhance the capabilities and broaden the applications of this innovative optoelectronic SNN architecture.