Software-controlled high-dimensional coherent optical tensor computing via wavelength multiplexing.

Area of Science:

• Photonics
• Artificial Intelligence
• Computer Engineering

Background:

• Real-time data processing is crucial for AI, driving demand for high-speed computing solutions.
• Optical computing offers inherent advantages in speed and parallelism for AI tasks.
• Optoelectronic convolutional neural networks (OE-CNNs) benefit from high-dimensional optical tensor computing for efficient inference.

Purpose of the Study:

• To present a software-controlled reconfigurable optical tensor processing unit.
• To leverage optical coherent technology and wavelength-division multiplexing (WDM) for high-dimensional matrix operations and high-throughput computing.
• To demonstrate a high-performance photonic AI processor.

Main Methods:

• Integration of nine Mach-Zehnder interferometers (MZIs) and nine optical delay lines for coherent serial 3x3 convolution operations.
• Software-tunable ternary weights {-1, 0, 1} for flexible computation.
• Implementation of WDM to support high-dimensional tensor computing and parallel data channel processing.

Main Results:

• Experimental execution of tensor computing on 3D images at 29.14 GBaud, achieving 1.57 TOPS peak speed.
• Potential for up to 19.41 TOPS with 37-channel WDM expansion.
• Demonstration of an OE-CNN with four parallel optical convolution kernels achieving 96.66% classification accuracy on MNIST.

Conclusions:

• The proposed architecture significantly enhances photonic computing performance by integrating coherent optical computing and WDM parallelism.
• This work offers a scalable pathway for developing next-generation high-performance photonic AI processors.
• The developed unit enables efficient high-dimensional matrix operations and high-throughput computing for AI applications.