.png){kind=icon}
.png){kind=icon}
Return
Parallel Matrix-Vector Multiplication Using a Scalable Multiport Interferometer
.png){kind=icon}
OA
AI
DOI：10.1002/lpor.202501876.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Integrated photonic matrix-vector multiplication (MVM) is an emerging technology with broad applications in science and engineering, offering inherent advantages in high-speed and low-power consumption computations. Among various architectures explored, coherent multiport Mach-Zehnder interferometer (MZI) mesh networks are widely utilized for implementing reconfigurable complex-valued matrix computations. However, their scalability is fundamentally constrained by accumulated losses and control complexity as the mesh size grows. Here, we introduce multi-wavelength technology to enhance computational parallelism by leveraging the broad bandwidth of MZIs, effectively adding an additional computational dimension to MZI mesh-based MVM networks. The MZI switches are optimized for low phase error and wide bandwidth to alleviate the complexity of mesh programming and support a greater number of wavelengths. We experimentally demonstrate the mesh performing parallel MVM operations across multiple wavelengths through two applications: (1) multi-channel analysis of clinical electrocardiogram (ECG) signals for cardiovascular risk prediction, achieving 93.8% accuracy in identifying sudden cardiac death risk; (2) parallel processing of segmented images enables edge extraction with significantly improved computational speed. This multi-wavelength parallel MVM architecture overcomes key scalability limitations while offering a highly extensible pathway toward energy-efficient, high-throughput optical computing for advanced artificial neural applications.
Keywords:
acceleration processing
mach-zehnder interferometer (MZI) mesh
matrix-vector multiplication
optical computing
parallel processing