Integrated Neuromorphic Photonic Computing for AI Acceleration: Emerging Devices, Network Architectures, and Future...

Area of Science:

• Photonic Neuromorphic Computing
• Artificial Intelligence Hardware Acceleration
• Integrated Photonics

Background:

• Electronic hardware faces physical limits (transistor scaling, von Neumann architecture, thermal dissipation) hindering AI computational density and energy efficiency.
• Deep learning, including large language models (LLMs), demands significant computational resources, exacerbating hardware bottlenecks.

Purpose of the Study:

• To review a decade of progress in photonic neural networks (PNNs) as a solution for AI hardware limitations.
• To analyze core PNN components, network architectures, and application-specific requirements for cloud and edge AI.
• To outline pathways for overcoming material and system-level barriers in PNN development.

Main Methods:

• Systematic review and critical analysis of advances in linear synaptic devices, nonlinear neuron devices, and PNN architectures.
• Analysis of application-specific requirements for PNN deployment in cloud-scale and edge/client-side AI.
• Identification of material and system-level barriers and proposed solutions, including topology-optimized devices and advanced packaging.

Main Results:

• Photonic neural networks (PNNs) demonstrate potential for AI acceleration, achieving single-chip integration of inference and in situ training.
• Significant progress has been made in core PNN components, though challenges remain.
• PNNs leverage light's parallelism, low latency, and minimal thermal loss for efficient matrix operations.

Conclusions:

• Photonic neuromorphic computing, particularly PNNs, presents a paradigm shift for post-Moore AI hardware.
• Overcoming material and packaging challenges is crucial for widespread PNN deployment.
• PNNs offer a promising platform for next-generation, energy-efficient AI acceleration.