Neuromorphic photonic computing with an electro-optic analog memory.

Area of Science:

• Neuromorphic Engineering
• Photonic Computing
• Machine Learning Hardware

Background:

• Neuromorphic photonic systems rely on analog signals, requiring energy-intensive digital-to-analog converters (DACs) and analog-to-digital converters (ADCs).
• Conventional von Neumann architectures face significant energy costs due to data movement between memory and converters.

Purpose of the Study:

• To propose and demonstrate an analog electronic memory integrated directly with photonic computing units.
• To eliminate energy-consuming data movement and reduce reliance on DACs/ADCs in neuromorphic photonic systems.

Main Methods:

• Monolithic integration of a neuromorphic photonic circuit with on-chip capacitive analog memory.
• Performance evaluation using machine learning for in situ training and inference on the MNIST dataset.

Main Results:

• Achieved over 26x power savings compared to conventional SRAM-DAC architectures.
• Demonstrated >90% inference accuracy with a minimum analog memory retention-to-network-latency ratio of 100.
• Showcased the viability of leaky analog memories without substantial performance degradation.

Conclusions:

• Integrating analog memory into neuromorphic photonic architectures offers a scalable pathway to energy-efficient, high-speed computing.
• This approach significantly minimizes data movement and reliance on converters.
• Enables practical implementation of neuromorphic photonic systems for advanced AI tasks.