All-Optical Computing Based on Convolutional Neural Networks

Kun Liao, Ye Chen, Zhongcheng Yu, Xiaoyong Hu, Xingyuan Wang, Cuicui Lu, Hongtao Lin, Qingyang Du, Juejun Hu, Qihuang GongResearch 2020 · DOI

Abstract

The rapid development of information technology has fueled an ever-increasing demand for ultrafast and ultralow-energy-consumption computing. Existing computing instruments are predominantly electronic processors. The scaling of computing speed is limited not only by data transfer between memory and processing units, but also by RC delay associated with integrated circuits. Using photons as information carriers is a promising alternative. Here, we report a strategy to realize ultrafast and ultralow-energy-consumption all-optical computing based on convolutional neural networks, leveraging entirely linear optical interactions. The device is constructed from cascaded silicon Y-shaped waveguides with side-coupled silicon waveguide segments to enable complete phase and amplitude control in each waveguide branch. The generic device concept can be used for equation solving, multifunctional logic operation, Fourier transformation, series expanding and encoding, as well as many other mathematical operations. Multiple computing functions were experimentally demonstrated to validate all-optical computing performances. The time-of-flight of light through the network structure corresponds to an ultrafast computing time of the order of several picoseconds with an ultralow energy consumption of dozens of femtojoules per bit. Our approach can be further expanded to fulfill other complex computing tasks based on non-von Neumann architectures and thus paves a new way for on-chip all-optical computing.

Introduction

As global data demands surge, traditional electronic computing is approaching a physical ceiling. The standard von Neumann architecture, which separates memory and processing units, creates a data transfer bottleneck that limits overall speed. Furthermore, as integrated circuits shrink, they suffer from increasing RC delays and significant heat dissipation due to Ohmic losses.

Optical computing—using photons instead of electrons—presents a compelling solution due to the high speed and low thermal footprint of light. However, many previous optical designs relied on third-order optical nonlinearities, which often require high power or suffer from slow response times. This research introduces a novel alternative: an all-optical convolutional neural network (CNN) that operates entirely through linear optical interactions on a silicon-on-insulator (SOI) platform.

Methodology: The "Physically-Fixed" CNN

The researchers developed a non-von Neumann architecture composed of cascaded silicon Y-shaped waveguides. The fundamental innovation lies in the "weight modulator" units integrated into each waveguide branch.

Schematic representation of the all-optical CNN architecture and the basic Y-branch building block used for phase and amplitude control.

These modulators consist of silicon waveguide segments side-coupled to the main transmission lines. By precisely adjusting the length of these segments and the gap between them and the main waveguide, the researchers can independently control both the amplitude and the phase of the propagating light.

Unlike traditional neural networks that learn in real-time, this system uses a "physically-fixed" approach. The network is first trained in a software environment (using PyTorch and stochastic gradient descent) to determine the optimal physical dimensions for a specific task. Once the weights are optimized, the chip is fabricated with those exact geometries, effectively embedding the neural network's intelligence directly into the hardware.

Experimental Results and Applications

The team validated the performance of this all-optical CNN across several complex mathematical and logic tasks:

  • Transcendental Equation Solving: The device successfully solved complex equations such as $\cos(2kx) + 4 = \tan(kx)$. The experimental results showed a maximum deviation of less than 5% from theoretical values, demonstrating high precision for continuous mathematical functions.
  • Multifunctional Logic Gates: By utilizing control bits, a single network structure was configured to perform various logic operations, including OR and XNOR. The system achieved an intensity contrast between "0" and "1" states as high as 12.9 dB.
  • Binary Half-Adder: The CNN successfully performed binary addition with an average intensity contrast of 14.2 dB, proving its capability for fundamental digital arithmetic.

Performance Metrics

The efficiency of the all-optical CNN is remarkable compared to electronic counterparts. The total computing time—defined by the time-of-flight of light through the chip—was measured at approximately 9.4 picoseconds, with the core CNN operation taking only 1.3 ps. Energy consumption was recorded between 50 and 92 fJ/bit. Theoretical analysis suggests that with further optimization of pulse energies, this could potentially drop to the attojoule ($10^{-18}$ J) range.

Conclusion and Outlook

While the current device is an Application-Specific Integrated Circuit (ASIC)—meaning its "weights" are fixed at fabrication and cannot be reprogrammed—it demonstrates a powerful proof-of-concept for ultrafast, low-power computing. The use of linear optical interactions avoids the trade-offs associated with nonlinear materials, while the cascaded Y-branch design offers a scalable building block for more complex architectures. Future work may incorporate wavelength multiplexing to enable parallel processing, further expanding the potential of on-chip all-optical computing for non-von Neumann tasks.

Summary updated

Abstract

<title>Abstract</title> The rapid development of information technology has fueled an ever-increasing demand for ultrafast and ultralow-energy-consumption computing. Existing computing instruments are pre-dominantly electronic processors. The scaling of computing speed is limited not only by data transfer between memory and processing units, but also by RC delay associated with integrated circuits. Using photons as information carriers is a promising alternative. Here, we report a strategy to realize ultrafast and ultralow-energy-consumption all-optical computing based on convolutional neural networks, leveraging entirely linear optical interactions. The device is constructed from cascaded silicon Y-shaped waveguides with side-coupled silicon waveguide segments to enable complete phase and amplitude control in each waveguide branch. The generic device concept can be used for equation solving, multifunctional logic operation, Fourier transformation, series expanding and encoding, as well as many other mathematical operations. Multiple computing functions were experimentally demonstrated to validate all-optical computing performances. The time-of-flight of light through the network structure corresponds to an ultrafast computing time of the order of several picoseconds with an ultralow energy consumption of dozens of femtojoules per bit. Our approach can be further expanded to fulfill other complex computing tasks based on non-von Neumann architectures and thus paves a new way for on-chip all-optical computing.
BibTeX
@inproceedings{liao2020all,
  author = {Kun Liao and Ye Chen and Zhongcheng Yu and Xiaoyong Hu and Xingyuan Wang and Cuicui Lu and Hongtao Lin and Qingyang Du and Juejun Hu and Qihuang Gong},
  title = {{All-Optical Computing Based on Convolutional Neural Networks}},
  booktitle = {Research},
  year = {2020},
  doi = {10.21203/rs.3.rs-59249/v1},
}
Figure