Partial coherence enhances parallelized photonic computing.

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence Hardware

Background:

• Optical coherence control is crucial for applications like communication and sensing.
• Current integrated photonic accelerators rely on highly coherent light sources for optimal performance.
• The prevailing view is that increased light source coherence enhances system functionality.

Purpose of the Study:

• To introduce a photonic convolutional processing system that utilizes partially coherent light.
• To challenge the necessity of high coherence in photonic accelerators.
• To enable larger-size photonic tensor cores and high-throughput computing.

Main Methods:

• Developed a photonic convolutional processing system leveraging partially coherent light.
• Implemented the system on two photonic platforms: phase-change-material photonic memories and silicon photonics with electro-absorption modulators (EAMs).
• Evaluated system performance on Parkinson's disease gait classification and MNIST handwritten digit recognition.

Main Results:

• Achieved high accuracy in classifying Parkinson's disease patient gaits (92.2%) and MNIST digits (92.4%) using partially coherent light.
• Demonstrated parallel convolution operations and high processing speeds (0.108 TOPS).
• Showcased optimized bandwidth utilization through reduced coherence.

Conclusions:

• Partially coherent light can boost computing parallelism in photonic systems without significant accuracy loss.
• This approach challenges the traditional reliance on high coherence in photonic accelerators.
• Enables the use of less complex light sources, reducing system requirements for high-throughput photonic computing.