Work

Photonic Tensor Processing Unit by Depthwise Separable Convolution

Year: 2026

Type: article

Abstract: We propose and experimentally demonstrate a hardware-efficient photonic tensor processing unit (PTPU) enabled by depthwise separable convolution (DSC) and wavelength division multiplexing (WDM). Fabricated on an 800nm silicon nitride platform, the proposed PTPU implements a$4\times 2 \times 2\times 1$kernel using only four microring resonators and eight Mach-Zehnder interferometers, reducing the number of photonic devices by 25% compared to standard convolution implementation. The tensor convolution capability of the PTPU is validated through image convolution tasks. Furthermore, when deployed as the convolution module in a DSC-based convolutional neural network (CNN) for NoduleMNIST 3D dataset classification, the proposed PTPU achieves 85.42% prediction accuracy, which is comparable to software-based implementations. These results establish the proposed PTPU as a highly efficient photonic neural network architecture, particularly well-suited for resource-constrained applications.

Source: IEEE Photonics Technology Letters

DOI: 10.1109/lpt.2026.3713231

Authors: Weiwei Pan, Heng Chen, Chen Ji

Institution: Zhejiang University

Language: English

Topic: Neural Networks and Reservoir Computing

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $2,645

OpenAlex ID: W7168299173

FWCI:

Cites: 8

Cited by:

Related to: 10