image: Figure1 | Comparison of scaling limitations in existing ONNs and the proposed solution. a, Scaling challenges in current on-chip ONNs, including depth and input size limitations. The depth limitation arises from NAFs lacking net gain or requiring electrical amplification. The input size limitation is constrained by the need for coherent detection in coherent computing and multiple wavelengths in incoherent computing. b, The proposed partially coherent deep optical neural network, which overcomes these limitations by enabling scalable depth and input size expansion.
Credit: Hailong Zhou et al.
As artificial intelligence (AI) technology advances, the inherent limitations of conventional electronic processors in energy consumption and processing latency have become increasingly prominent. Photonic neural networks (ONNs) are considered a competitive alternative due to their significant advantages in bandwidth and energy efficiency. However, despite this promise, achieving end-to-end, large-scale ONNs on a chip faces serious challenges. Network depth is constrained by the weak cascadability of optical nonlinear activation functions, while input size is limited by the scale of the optical matrix and its reliance on coherent light sources.
In a paper recently published in Light: Science & Applications, a research team from Huazhong University of Science and Technology and The Chinese University of Hong Kong has demonstrated a novel solution. They designed and fabricated a monolithically integrated partially coherent deep optical neural network (PDONN) chip that effectively overcomes these limitations.
"Our design's core is an on-chip nonlinear activation function based on opto-electro-opto conversion, which provides a positive net gain, allowing the network to be stacked deeper," explains one of the corresponding authors. "At the same time, we introduced convolutional layers to rapidly reduce data dimensionality, which allows our chip to process larger-sized input images".
One of the study's most notable innovations is the use of a partially coherent optical source. Unlike the expensive and complex narrow-linewidth lasers used in traditional schemes, partially coherent light (such as from an LED or amplified spontaneous emission source) is more accessible and significantly reduces the stringent requirements for system coherence control. This not only simplifies the system but also paves the way for scaling up the on-chip optical matrix.
The team's PDONN chip integrates hundreds of optical devices into a compact area of approximately 17 mm². Its architecture includes a record-breaking 64-unit input layer, two convolutional layers, and two fully connected layers, making it the on-chip ONN with the largest input size and number of optical layers reported to date.
In experimental tests, the chip successfully performed four-class classification of handwritten digits and two-class classification of fashion images, with accuracies of 94% and 96%, respectively. Notably, the network's performance was well-maintained even when using a partially coherent light source, fully demonstrating the architecture's robustness. Furthermore, the chip's single-inference latency is estimated at 4.1 ns, with an energy efficiency of 121.7 pJ/OP, showcasing its potential for low latency and high energy efficiency.
"This work demonstrates that through architectural innovation, we can use more accessible technology to build large-scale, high-performance optical computing systems," the researchers conclude. "It opens up new avenues for the next generation of energy-efficient, widely applicable AI hardware"
Looking ahead, the team plans to continue improving the chip's performance and scalability by enhancing the modulator extinction ratio and further reducing system latency. With its groundbreaking design and demonstrated capabilities, the PDONN architecture is a strong candidate for future large-scale artificial intelligence applications, marking a new milestone in the field of optical computing.
Journal
Light Science & Applications
Article Title
Scaling up for end-to-end on-chip photonic neural network inference