Researchers Build Ultra-Fast Optical Chip for Feature Extraction with Record-Low Latency

A research team from Tsinghua University has developed an optical computing system that dramatically reduces latency in feature extraction - one of the most critical stages in real-time data processing tasks like quantitative trading and robotic surgery.

Many modern artificial intelligence (AI) applications, such as surgical robotics and real-time financial trading, depend heavily on the ability to rapidly extract key features from streams of raw data. This crucial step is often held back by the limitations of traditional digital processors. As data volumes grow, the physical constraints of conventional electronics make it increasingly difficult to reduce latency or boost throughput to meet the demands of emerging, data-intensive services.

One promising alternative is optical computing, which uses light to perform high-speed computations. Among its various approaches, optical diffraction operators (thin, plate-like structures that compute as light propagates through them) stand out for their energy efficiency and ability to process data in parallel. However, pushing these systems beyond 10 GHz in practical applications has been a significant technical challenge, mainly due to the difficulty of maintaining stable, coherent light required for reliable optical calculations.

To overcome this, a team led by Professor Hongwei Chen at Tsinghua University has developed a novel solution. As reported in Advanced Photonics Nexus, the team introduced an optical feature extraction engine (OFE²) designed to enable real-time optical feature extraction across a wide range of applications.

A key innovation lies in OFE²’s data preparation module.

Delivering high-speed, parallel optical signals to coherent optical cores is notoriously difficult, especially when using traditional fiber-based components, which often introduce unwanted phase shifts during power splitting and delay.

The Tsinghua team addressed this by creating an integrated on-chip system featuring tunable power splitters and finely controlled delay lines. This setup effectively de-serializes the input signal, converting it into multiple stable, parallel branches. An adjustable phase array further enhances flexibility, allowing OFE2 to be reconfigured as needed.

Once the data is prepared, the optical waves travel through the diffraction operator. This process is mathematically equivalent to a matrix-vector multiplication, which serves as the mechanism for feature extraction. Here, the diffracted light converges into a focused “bright spot” at the output. By adjusting the phase of the parallel input lights, the bright spot can be partially redirected toward a specific output port. The corresponding variations in output power represent extracted features based on the input signal’s changes over time.

OFE² operates at 12.5 GHz, performing each matrix-vector multiplication in under 250.5 picoseconds - the lowest latency reported among comparable optical computing systems.

We firmly believe this work provides a significant benchmark for advancing integrated optical diffraction computing to exceed a 10 GHz rate in real-world applications.

Hongwei Chen, Professor, Tsinghua University

The research team validated OFE²’s performance across a range of tasks, demonstrating its versatility and effectiveness. In image processing, the system successfully extracted edge features from input images, generating two complementary feature maps, which are described as "relief" and "engraving" views. These features enhanced the performance of downstream AI models, improving classification accuracy and boosting pixel-level precision in semantic segmentation tasks, such as identifying organs in CT scans.

Notably, AI networks that incorporated OFE² required fewer electronic parameters compared to baseline models, showing that optical pre-processing can reduce computational load and enable more efficient hybrid AI systems.

The team also demonstrated OFE²’s effectiveness in a digital trading scenario.

In this task, the system processed real-time time-series market data and generated trading decisions based on a pre-optimized strategy. Traders fed live price signals into OFE², which, after initial training, produced output signals that could be directly interpreted as buy or sell actions through a simple decision logic. Because the computation occurs at the speed of light, the system delivered a major latency advantage, allowing trading opportunities to be captured with minimal delay and achieving consistent profitability. 

Collectively, these results highlight a shift toward offloading some of the most computationally demanding tasks from traditional electronics to ultrafast, low-power photonics. As a result, this approach could lead the way for a new generation of AI systems that are capable of making decisions in real time without the energy and speed constraints of conventional hardware.

The advancements presented in our study push integrated diffraction operators to a higher rate, providing support for compute-intensive services in areas such as image recognition, assisted healthcare, and digital finance. We look forward to collaborating with partners who have data-intensive computational needs.

Hongwei Chen, Professor, Tsinghua University

Journal Reference:

Sun, R., et al. (2025) High-speed and low-latency optical feature extraction engine based on diffraction operators. Advanced Photonics Nexus. doi.org/10.1117/1.APN.4.5.056012.