In-memory photonic dot-product engine with electrically programmable weight banks

Area of Science:

• Optoelectronics and Photonics
• Materials Science
• Computer Engineering

Background:

• The von Neumann bottleneck limits computational efficiency in traditional architectures.
• Hybrid photonic-electronic systems offer a potential solution but face implementation challenges.
• Phase-change materials (PCMs) are promising for non-volatile memory and computing.

Purpose of the Study:

• To demonstrate a successful in-memory photonic-electronic dot-product engine.
• To decouple electronic programming from photonic computation for improved performance.
• To achieve high-accuracy image processing using hybrid computing.

Main Methods:

• Development of non-volatile, electronically reprogrammable PCM memory cells.
• Utilized non-resonant silicon-on-insulator waveguide microheater devices.
• Implemented parallel multiplications for image processing tasks.

Main Results:

• Achieved record 4-bit weight encoding in PCM memory cells.
• Demonstrated lowest energy consumption (1.7 nJ/dB) for Erase operation.
• Obtained high switching contrast (158.5%) and superior image processing accuracy (SNR ≥87.36, σ ≤0.007).

Conclusions:

• Successfully demonstrated an in-memory hybrid computing system for convolutional processing.
• Achieved 86% and 87% inferencing accuracies on the MNIST database.
• This work represents a milestone in hybrid photonic-electronic computing, resolving the von Neumann bottleneck.