CUHK develops all-optical signal processor to break AI data centre transmission bottleneck

A research team led by The Chinese University of Hong Kong (CUHK) has developed a novel integrated all-optical signal processor (OSP) to address the massive data transmission demands of next-generation AI systems, particularly for high-speed links between multiple data centres.

Built on a silicon photonic chip, the OSP can correct signal distortion in real time while the optical signal is still in the form of light, without first converting it into an electrical signal for processing. In other words, it can handle signal impairments in high-speed transmission faster and more efficiently, helping improve communication efficiency between servers and data centres in large-scale AI systems.

Experimental results show that the OSP can perform real-time processing at an aggregate data rate of 1.6 Tb/s (terabits per second), with a latency of below 60 picoseconds (one-trillionth of a second) and extremely low energy consumption of only tens of femtojoules per bit. This achievement helps overcome the bottlenecks of conventional digital signal processing (DSP) technologies in terms of speed, latency and power consumption, offering a disruptive solution for ultra-low-latency and green AI supercomputing. The findings have recently been published in the leading international academic journal Science.

Rapid AI growth drives demand for high-speed transmission across large-scale distributed data centres

With the rapid development of AI technology, modern AI systems are no longer confined to a single computer or data centre. Many large generative AI models now require thousands of graphics processing units (GPUs) and specialised accelerators working synchronously across multiple locations. For these systems to operate efficiently, computing units must be able to exchange massive volumes of data quickly and reliably. As a result, in addition to computing power itself, communication technologies offering high bandwidth, low latency and low power consumption have become indispensable to AI development.

“Optical fibre communications form the backbone of modern data transmission and data centre interconnect technologies. However, conventional data centre interconnect technologies are increasingly struggling to keep pace with the scale and speed required by modern AI systems,” said Professor Huang Chaoran, Assistant Professor in the Department of Electronic Engineering, CUHK. “As transmission rates continue to rise, signal distortion becomes more severe, and if signal processing still relies on electronic methods, it can also introduce severe latency and significant power consumption.”

Correcting distorted signals directly with light for greater efficiency

To address these challenges, Professor Huang’s team, together with researchers from Huazhong University of Science and Technology (HUST) and Fudan University (FDU), has developed an integrated OSP. Unlike conventional approaches, the OSP can directly correct distorted signals in the optical domain before optical-to-electrical conversion, reducing processing time and energy consumption while improving overall transmission efficiency.

The design of this technology draws inspiration from neuromorphic computing and machine learning. By precisely tuning the optical paths inside the chip, the research team enabled the OSP to analyse the complex temporal features of high-speed optical signals more effectively, allowing more accurate correction of signal impairments.

In addition, the OSP functions as a programmable nonlinear equaliser which can be flexibly adjusted according to different transmission scenarios to compensate for various signal impairments, including fibre chromatic dispersion, the effects of limited receiver and transmitter bandwidth, and nonlinear distortions arising in optical fibres and devices under heavy load. By retaining the full optical field before optical-to-electrical conversion, the OSP can correct distortion more effectively than conventional DSP-based approaches in multiple test scenarios. It also has the potential to expand the usable wavelength-division multiplexing (WDM) bandwidth constrained by chromatic dispersion by a factor of 6.8, further increasing the transmission capacity of each optical fibre. Its programmability also allows the system to dynamically adapt to different impairment scenarios, wavelengths, data rates and modulation formats, demonstrating high flexibility and scalability.

In the experiments, the OSP simultaneously processed signals from eight wavelength channels, with each channel operating at 100 Gbaud PAM4 (equivalent to approximately 200 Gbit/s per channel), yielding an aggregate throughput of 1.6 Tb/s. The study also showed that its processing latency was below 60 picoseconds, even shorter than a single clock cycle in many digital systems, while energy consumption remained stable at the level of tens of femtojoules per bit. These results indicate that the OSP has strong potential for future high-speed data centre interconnects, cross-regional AI training, and other scenarios requiring ultra-high-speed, low-latency data transmission.

Advancing optical signal processing into a new stage

More broadly, the study reflects the continuing evolution of optical communication technology. More than half a century ago, Professor Charles K. Kao, the former vice-chancellor and president of CUHK and widely recognised as the “Father of Fibre Optics”, proposed the use of low-loss optical fibre for long-distance communication, laying the foundation for the modern internet. This research builds on this important legacy, showing that future technologies may not only use light to transmit information, but also to directly process information with light, opening up new possibilities for next-generation communication and computing systems.

Professor Huang Chaoran is the corresponding author. Other contributing authors of this work include Wang Benshan (first author), Xiao Qiarong (co-first author), Xu Tengji, Fan Li and Liu Shaojie from CUHK; and collaborators (Professor Kong Qiuqiang from CUHK, Professor Dong Jianji from HUST, and Professor Zhang Junwen from FDU). The full text can be found at https://www.science.org/doi/10.1126/science.ady5344.