Trading at the speed of light with scalable photonic neurons

Light Publishing Center, Changchun Institute of Optics, Fine Mechanics And Physics, CAS

FacebookXLinkedInWeChatBlueskyMessageWhatsAppEmail

 

image: 

Figure | Working principle of the photonic neuron based on modulation-and-weighting MRR banks. a, A typical MRR-based photonic neuron, comprising a bank of multiple lasers at different wavelengths, a ring modulator bank, an MRR-based weight bank, and a balanced PD. b, The proposed approach, which consolidates both modulation and weighting in a single MRR bank and employs a single-ended PD, thereby reducing complexity and footprint. c-e, Three possible configurations of the proposed MRR bank, enabling neuron types in feedforward, recurrent (short-term memory), and combined long- and short-term memory modes. 

view more 

Credit: Weipeng Zhang et al.

In stock markets, trading faster often means winning more, making speed the single most important factor in high-frequency trading. Today, the fastest trading systems are built on FPGA-based electronic processors, which offer the lowest latency achievable with conventional electronics. Despite extensive optimization, these systems remain fundamentally constrained by clock speeds, data-conversion overhead, and electronic signal routing. As a result, further reductions in latency using electronic technologies are becoming increasingly difficult to achieve.

 

In a new paper published in eLight, scientists report a photonic neuromorphic computing architecture designed to enable real-time data processing at the speed of light. The researchers demonstrate a scalable photonic neuron that performs weighted summation and nonlinear activation directly in the optical domain, allowing information to be processed continuously as light propagates through the device, rather than sequentially under clock control.

 

A central motivation for this work is the long-standing scalability challenge of photonic processors. Many existing photonic computing approaches, such as those based on Mach–Zehnder interferometers, rely on architectures with large footprints, limiting the achievable size of photonic neurons. More compact implementations based on microring resonators (MRRs) offer higher integration density but are constrained by stringent spectral alignment requirements that worsen as system size grows. These challenges have hindered the realization of large-scale photonic neurons capable of processing complex, real-world data. The architecture reported here directly addresses these limitations, making scalable photonic computing practical.

 

The proposed photonic neuron integrates modulation and weighting within a single microring resonator, rather than implementing these functions using separate photonic elements. This functional unification directly reduces the number of components that must be spectrally aligned, substantially relaxing a major scalability barrier in microring-based photonic processors. Beyond compactness, the architecture is inherently reconfigurable: by introducing simple electrical feedback paths, the same photonic neuron can be configured to support short- and long-term memory. This capability enables effective temporal processing, allowing the neuron to capture both recent and historical information, which is essential for analyzing real-world time-series data.

 

As a proof of concept, the researchers apply the scalable photonic neuron to high-frequency trading tasks, demonstrating real-time processing of financial time-series data using a single neuron. Experiments on several representative stock symbols show generally positive cumulative gains, highlighting the suitability of the photonic architecture for latency-critical trading applications. The neuron is further configured into multiple operating modes, including feedforward processing as well as recurrent configurations with short- and long-term historical feedback. Incorporating temporal memory consistently improves performance and stability, offering insight into how historical information can enhance trading decisions. These results illustrate how the reconfigurable photonic neuron can adapt to different temporal dynamics, while maintaining intrinsic processing latencies on the order of tens of picoseconds, being far below those of state-of-the-art FPGA-based electronic trading systems.

 

Beyond high-frequency trading, the significance of the proposed photonic neuron architecture lies in its ability to enable scalable neuromorphic photonic systems capable of processing complex, real-world data. By directly addressing long-standing limitations in footprint, spectral alignment, and functional integration, this architecture provides a practical pathway toward building larger photonic neurons and, ultimately, large-scale photonic neural networks. Its compactness, reconfigurability, and compatibility with standard photonic integration processes make the realization of usable neuromorphic photonic computers increasingly realistic for industrial deployment. As a result, this approach is well positioned to extend the intrinsic advantages of photonic computing, including ultra-low latency, high parallelism, and energy efficiency, to a broad range of real-world applications, including real-time signal processing, communications, and adaptive control systems.

---

Journal

eLight

DOI

10.1186/s43593-026-00122-3 

Article Title

Compact, reconfigurable, and scalable photonic neurons by modulation-and-weighting microring resonators

Figure | Application of high-frequency trading using the proposed photonic neuron. Schematic of the photonic high-frequency trading processor.  the top row shows the stock price over one trading day. The photonic neuron can process 10 consecutive price points and generate indices for making decisions among buy, sell, and hold.

Credit

Weipeng Zhang et al.