Reconfigurable versatile integrated photonic computing chip

The rapid development of photonic computing chips has been driven by the increasing demand for high data-throughput, low-power computation across various application fields. Researchers have aimed to develop efficient integrated photonic computing chips capable of supporting a wide range of application scenarios in both static and dynamic temporal domains. Progress has been made with single or hybrid configurations for individual network models, which primarily focus on feedforward neural networks for static tasks. However, until now, due to the challenges in improving the scalability of integrated photonic hardware while maintaining the computing performance, it remains a significant challenge to achieve a scalable and flexible solutions for multi-model photonic chips in multi-tasking applications.

 

In a new paper published in eLight, a team of scientists, led by Professor Xiaoyong Hu from State Key Laboratory for Mesoscopic Physics & Department of Physics, Collaborative Innovation Center of Quantum Matter & Frontiers Science Center for Nano-optoelectronics, Peking University, China, and co-author have developed a scalable versatile integrated photonic chip that facilitates the implementation of diverse neural network within a unified structure capable of handling both static and dynamic temporal tasks with high efficiency. Fully reconfigurable microring resonator (MRR) and Mach-Zehnder interferometers (MZI) arrays are utilized to handle diverse processing demands at different stages of large-scale tasks, including feature extraction, temporal reasoning, and output decision-making. Each MRR is based on a cross-waveguide coupling design, allowing each MRR unit to switch between handling static and temporal tasks by inputting information through one or both ports of the cross-waveguide, without requiring additional structures. As a proof of concept, they experimentally integrated a turnkey soliton microcomb with a free spectral range (FSR) of 100 GHz as the light source, fully leveraging the wavelength freedom to demonstrating the realization of fully connected neural network (FCNN), convolutional neural network (CNN), and photonic gated recurrent neural network (PGRNN) models with a tunable area efficiency of up to 2.45 TOPS/mm² at a typical frequency of 10 GHz.

 

Three neural network tasks were demonstrated to validate the proposed chip’s adaptability and performance, particularly for multimodal tasks, which are in high demand but rarely addressed by photonic neural networks. First, for image classification, a combination of CNN and FCNN was used to construct an Inception model, achieving test accuracies of 92.93% and 56.57% on the MNIST and CIFAR-10 datasets, respectively. Second, PGRNN and FCNN were employed for sentiment analysis, achieving test accuracy of 80.81% on the IMDB dataset. Third, a combination of CNN, PGRNN, and FCNN with a scaled-up architecture was used for speech recognition. These scientists summarize the computing strategy of their photonic chip:

 

“The photonic computing chip leverages the multi-wavelength channels of an optical frequency comb and the dual-input-port structure of MRR to enable flexible and efficient all-optical processing. In the FCNN implementation, weights are encoded on the resonant wavelength and tuned electrically, while biases are added via the detuning wavelength. This structure allows simultaneous multiplication and bias addition within a single MRR, with results directly detected at the output. For CNN, the MRR array serves as a photonic convolution kernel, enabling scalable multi-channel convolutions. For PGRNN, cross-waveguide MRRs receive inputs from current and previous time steps at separate ports, and output a combined signal using distinct FSRs to avoid crosstalk.”

 

“By enabling dual-path computation within a single MRR, this architecture doubles the processing throughput over traditional MRR-based schemes, reaching a record-high area efficiency.” they added.

---

---