Waveguide-multiplexed photonic matrix-vector multiplication processor using multiport photodetectors

TL;DR

This work introduces a scalable intensity-based photonic MVM processor using waveguide multiplexing and multiport photodetectors to sum optical intensities, avoiding wavelength- or mode-multiplexing. A $16$-port Ge PD demonstration achieves $3$ dB bandwidth of $11.8$ GHz at $-3$ V and scales toward $250$ ports with $6.1$ GHz bandwidth, while a $4 imes4$ SOI circuit performs MVMs for an Iris-classification NN with $93.3\%$ accuracy and Fashion-MNIST CNN simulations reach $90.53\%$ accuracy. The results highlight high fidelity, linear responsivity, and potential for large-scale optical neural networks, aided by PCM modulators to reduce insertion loss and area in large circuits. Overall, the approach provides a simplified, scalable foundation for on-chip photonic matrix-matrix multiplication and multi-dimensional optical neural network computation.

Abstract

The slowing down of Moore's law has driven the development of application-specific processors for deep learning. Analog photonic processors offer a promising solution for accelerating matrix-vector multiplications (MVMs) in deep learning by leveraging parallel computations in the optical domain. Intensity-based photonic MVM processors, which do not utilize the phase information of light, are appealing due to their simplified operations. However, existing intensity-based schemes for such processors often employ wavelength multiplexing or mode multiplexing, both of which have limited scalability due to high insertion loss or wavelength crosstalk. In this work, we present a scalable intensity-based photonic MVM processor based on the concept of waveguide multiplexing. This scheme employs multiport photodetectors (PDs) to sum the intensities of multiple optical signals, eliminating the need for multiple wavelengths or modes. A 16-port Ge PD with a 3 dB bandwidth of 11.8 GHz at a bias voltage of -3 V is demonstrated, and it can be further scaled up to handle 250 ports while maintaining a 6.1 GHz operation bandwidth. A 4 $\times$ 4 circuit fabricated on a Si-on-insulator (SOI) platform is used to perform MVMs in a 3-layer neural network designed for classifying Iris flowers, achieving a classification accuracy of 93.3%. Furthermore, the performance of large-scale circuits in a convolutional neural network (CNN) for Fashion-MNIST is simulated, resulting in a classification accuracy of 90.53%. This work provides a simplified and scalable approach to photonic MVM, laying a foundation for large-scale and multi-dimensional photonic matrix-matrix multiplication in optical neural networks.

Figures (11)

• Figure 1: Schematic structure of a $4 \times 4$ waveguide-multiplexed circuit using a single-waveguide-layer platform. Intensity modulators, such as MZIs, MRRs, or PCM absorbers, are used to generate the vector and matrix elements. The multiport PD sums multiple optical signals at the same wavelength and can be implemented using Ge-on-Si with a vertical p-i-n structure. Redundant waveguide crossings (not shown in the figure) are inserted into paths with fewer crossings to equalize insertion loss. MZI: Mach-Zehnder interferometer. MRR: microring resonator. PCM: phase change material.
• Figure 2: a Schematic diagram of the waveguide-coupled Ge-on-Si multiport PD using a vertical p-i-n structure. The drawing is not to scale, and metal layers are not shown. b A fabricated 16-port PD consisting of a 100--long multiport waveguide coupler region. The 300-nm waveguide spacings are nearly invisible. c Four PD ports are connected to on-chip components for further characterization. d Measured time stability of the photocurrent (blue curves) and the response of the photocurrent when the light phase in one path is tuned from 0 to more than $2\uppi$ (red curves). e Photocurrent as a function of the incident optical power. f Dark current as a function of the bias voltage. g Measured electro-optic frequency response.
• Figure 3: a A $4 \times 4$ circuit fabricated on an SOI platform, with enlarged images of an MZI and a 4-port PD. b The circuit is wire-bonded and packaged with a fiber array. c Experimental setup.
• Figure 4: a Characterizations of an MZI for a vector element ($x_{1}$) and two MZIs for two matrix elements ($w_{12}$ and $w_{22}$) b Photocurrent response of one 4-port PD (PD-1) when tuning the MZI for $w_{11}$ under various conditions. c Implemented matrices. d Expected and measured photocurrents of the 4-port PDs when 500 randomly generated matrices and vectors are implemented.
• Figure 5: a A 3-layer neural network for classifying Iris flowers. b Classification results when a computer alone is used (left) and when the MVMs are performed by the $4 \times 4$ circuit (right).