⚡ electrical engineering

Reconfigurable Photonic Quasi-Unitary Computing Hardware for Optical Signal Processing

This paper presents a compact, scalable, and energy-efficient silicon photonic quasi-unitary device that overcomes traditional trade-offs in on-chip matrix implementation to enable high-performance optical signal processing and recurrent neural network applications.

Original authors: Huan Li, Xiaowan Shen, Chun Gao, Mingyu Zhu, Zejie Yu, Yiwei Xie, Daoxin Dai, Xinxiang Niu, Xiaowen Dong, Jian-Jun He

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Huan Li, Xiaowan Shen, Chun Gao, Mingyu Zhu, Zejie Yu, Yiwei Xie, Daoxin Dai, Xinxiang Niu, Xiaowen Dong, Jian-Jun He

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The world of computing is hitting a wall. For decades, the speed of our electronic devices has doubled roughly every two years, a trend known as Moore's Law. But as the tiny transistors inside our chips become smaller and smaller, they are running into physical limits. They are becoming too crowded to fit more components, and they are generating too much heat to run faster without melting down. To keep the digital world moving, scientists are looking for a new way to process information that does not rely on electricity. They are turning to light. Light travels faster than electricity, carries more data at once, and generates almost no heat. This field, called integrated photonic computing, aims to build computer chips that use beams of light instead of electric currents to perform calculations. The goal is to create machines that can handle the massive data demands of modern artificial intelligence without the energy costs and speed limits of today's electronics.

The challenge, however, is that light is notoriously difficult to control inside a tiny chip. To do useful math with light, researchers need to perform a specific type of operation called a matrix multiplication. In the world of light, this means rearranging beams so they interfere with each other in a precise way to solve a problem. The most reliable way to do this has been to use a grid of tiny interferometers, which are devices that split and recombine light beams. But building these grids is like trying to build a skyscraper out of individual bricks one by one; as the grid gets bigger to handle more complex tasks, the chip becomes impossibly large and difficult to manufacture. The components take up too much space, and the more components you add, the more likely it is that a tiny error in manufacturing will ruin the calculation.

A team of researchers at Zhejiang University and Huawei Technologies has found a new way to solve this problem. Instead of building a massive grid of individual components, they designed a single, compact device that uses the natural behavior of light diffraction to perform these complex calculations. They call their invention a photonic quasi-unitary computing unit. The device is built on a standard silicon chip and looks somewhat like a miniature version of the gratings found in old spectrometers, but with a crucial twist. It uses a special arrangement of waveguides, which are tiny channels that guide light, to lift the data from a smaller number of input channels into a larger, intermediate space, and then bring it back down. This process allows the light to mix and interfere in a way that performs the necessary math without needing thousands of separate components.

The researchers built a physical version of this device on a silicon chip. The core of the chip is incredibly small, measuring just 0.65 millimeters by 2.15 millimeters. Inside this tiny area, they managed to create a system that can handle 32 different input signals at once. The device works by splitting the incoming light into 32 separate paths, passing them through a region where they diffract and mix, and then recombining them. By adjusting the phase of the light in the middle section using tiny heaters, the researchers can reprogram the device to perform different mathematical tasks. This flexibility is key. While the device does not follow the strict mathematical rules of a perfect "unitary" system in every single detail, it is close enough to be highly accurate while offering far more freedom to be reconfigured for different jobs.

To test if this new hardware actually works, the team used it to build a photonic recurrent neural network. This is a type of artificial intelligence that is particularly good at handling data that changes over time, like speech or stock market trends. They put the chip through two difficult tests. First, they asked it to predict a chaotic time-series signal known as the Mackey-Glass series. This is a standard test for AI because the signal is unpredictable and sensitive to tiny changes. The photonic chip managed to predict the next step in the sequence with an error rate of about 0.0023, which is a level of accuracy that matches or beats other photonic systems. Second, they used the chip to fix distorted signals traveling through fiber optic cables. In real-world communication, signals often get scrambled by the glass fibers they travel through. The researchers simulated a scenario where 16 different channels of data were heavily distorted, with a bit error rate of 0.9, meaning the data was almost completely garbled. After the light passed through their new chip, the error rate dropped to 0.0018, effectively restoring the original message.

The implications of this work go beyond just these two tests. The researchers calculated that if they were to replace the slow heating elements in their current chip with faster, high-speed modulators, the device could theoretically perform 102.4 trillion operations per second. Furthermore, because the device uses light and very little power for its calculations, it could achieve an energy efficiency of 42.1 trillion operations per watt. This is one order of magnitude higher than the NVIDIA B100, a state-of-the-art commercial electronic computing chip. The design is also compatible with the existing manufacturing processes used to make silicon chips, which means it could be mass-produced without needing entirely new factories.

This research represents a significant step forward in the race to build optical computers. By moving away from the traditional, bulky grid of components and embracing a design that uses the natural spreading of light, the team has shown that it is possible to build powerful, reconfigurable computing units on a tiny chip. They have demonstrated that light can be used not just to transmit data, but to actively process it with high speed and low energy. While the current version of the device is limited by the speed of its heating controls, the underlying architecture proves that the path to scalable, energy-efficient photonic computing is open. The work suggests a future where the heavy lifting of artificial intelligence could be done by light, freeing our digital world from the heat and speed limits that currently hold it back.

Based on the provided article, here is a detailed technical summary of the work:

Problem Statement

Integrated photonic computing is positioned as a post-Moore solution to the bandwidth and power-efficiency bottlenecks of electronic computing, particularly for deep neural networks. However, implementing on-chip unitary matrices—which are essential for energy-conserving, lossless linear optical transformations—faces a critical trade-off among footprint, scalability, and unitary fidelity.

Current Limitations: Conventional implementations using cascaded Mach-Zehnder interferometer (MZI) meshes (Reck or Clements architectures) suffer from a hardware footprint that scales as $O(N^2)$ , making large-scale integration (e.g., $128 \times 128$ ) physically infeasible in standard silicon photonics. Furthermore, these architectures require stringent fabrication uniformity; small phase deviations lead to unpredictable errors.
Alternative Bottlenecks: Diffractive deep neural networks offer an alternative but suffer from unconfined diffraction and mode mismatch, causing energy leakage that breaks the energy-conservation condition required for strict unitarity.
The Core Challenge: There is a need for a scalable, compact architecture that balances the strict physical constraints of unitarity with the reconfigurability required for diverse matrix computing tasks.

Methodology

The authors propose a Photonic Quasi-Unitary (PQU) computing device based on an Arrayed Waveguide Grating (AWG) architecture. The core innovation lies in relaxing the strict definition of unitarity to enhance reconfigurability while preserving near-unitary physical properties.

Architecture Design:
- The PQU device utilizes non-square free-propagation regions (FPRs) to perform dimension lifting (32 inputs to 81 intermediate channels) and dimension reduction (81 intermediate channels to 32 outputs).
- An intermediate arrayed waveguide region contains a tunable phase-shifter array.
- The overall transfer matrix is modeled as $M = B \Phi B^T$ , where $B$ and $B^T$ are the non-square basis transformation matrices (FPRs) and $\Phi$ is the programmable diagonal phase matrix.
- By using non-square matrices for $B$ and $B^T$ , the design breaks the dimensional limitation of the intermediate phase matrix, creating a "matrix superset" that contains standard unitary matrices. The extra degrees of freedom (49 auxiliary channels) allow for the offsetting of off-diagonal components through precise phase tuning.
Fabrication and Integration:
- A $32 \times 32$ silicon photonic PQU chip was fabricated on a standard SOI process with a 220 nm thick top silicon layer and a 2 µm buried oxide layer.
- The chip integrates beam splitting (via 2 $\times$ 2 MMI), thermo-optic MZI modulators for input loading, the core PQU unit, and grating couplers for output.
- The core PQU unit occupies a compact area of 0.65 mm $\times$ 2.15 mm.
System Verification (RNN):
- The authors constructed a Photonic Recurrent Neural Network (RNN) to test the device, as RNNs require stable, iterative matrix computations that rigorously test long-term numerical stability and matrix fidelity.
- Training Strategy: Due to the difficulty of characterizing the exact FPR transfer matrix post-fabrication, the authors employed a hybrid training approach:
  - Particle Swarm Optimization (PSO): Used to optimize the phase configurations ( $\Phi$ ) of the arrayed waveguides (a black-box strategy robust to fabrication errors).
  - Ridge Regression: Used to train the output matrix ( $W_{hy}$ ).
  - A coordinated process was used where the faster-converging component ( $\Phi$ ) was fixed before optimizing the other.

Key Contributions

Quasi-Unitary Concept: Introduction of a design that abandons rigorous full unitarity constraints in favor of non-square basis transformations, effectively expanding the range of programmable phase operations while maintaining near-unitary left and right multiplication properties.
Scalable AWG-Based Core: Demonstration of a compact, AWG-based computing core that avoids the $O(N^2)$ scaling of MZI meshes, significantly reducing on-chip integration complexity.
Experimental Validation: Successful fabrication and testing of a $32 \times 32$ PQU chip, validating its reconfigurability and computing performance in a real-world RNN framework.

Experimental Results

The system was benchmarked on two distinct tasks:

Mackey-Glass Time-Series Prediction:
- The PQU-RNN system achieved a Normalized Mean Square Error (NMSE) of ~0.0023 on the test set, representing a state-of-the-art level for photonic recurrent computing systems.
Fiber-Distorted Signal Restoration:
- The system was tested on 16-channel fiber-distorted signals (simulating 1 km fiber transmission with dispersion and inter-channel crosstalk).
- The raw distorted signals had a Bit Error Rate (BER) of 0.9.
- After PQU-RNN processing, the average BER across the 16 channels was reduced to 0.0018, effectively restoring the original pseudo-random binary sequence (PRBS) data.

Performance Metrics:

Footprint: The PQU core is 0.65 mm $\times$ 2.15 mm.
Projected Performance: If integrated with 50 GHz modulators, the system supports a projected 102.4 TOPS (Tera Operations Per Second).
Energy Efficiency: With an estimated power consumption of ~2.4 W (dominated by 81 thermo-optic resistors), the projected energy efficiency is 42.1 TOPS/W. The authors note this is approximately one order of magnitude higher than the NVIDIA B100 electronic chip.

Significance and Claims

The paper claims that the PQU architecture provides a scalable and energy-efficient hardware route for on-chip photonic computing, specifically targeting edge optical communication and AI scenarios.

Feasibility: The work validates the feasibility of using quasi-unitary designs to overcome the scale-reconfigurability trade-off inherent in current photonic neural networks.
Process Compatibility: The design is compatible with commercial standard silicon photonic processes (using single-mode waveguides, titanium nitride heaters, and metal deposition), ensuring low fabrication difficulty and potential for monolithic integration with existing foundry PDKs.
Future Potential: While the current work focuses on RNNs for edge signal processing, the authors suggest the architecture could support multi-wavelength operation via on-chip frequency combs and time-domain scaling via photonic delay structures to further expand computational capacity without massive hardware duplication.

The authors maintain a modest tone regarding the current limitations, acknowledging that the experimental speed is constrained by thermo-optic modulators and that the current focus is on RNN-based signal processing rather than complex image processing, which presents higher spatial feature coupling challenges.

Problem Statement

Methodology

Key Contributions

Experimental Results

Significance and Claims

More like this