All-Optical Signal Processor Breaks AI Data Center Transmission Bottleneck
Researchers have developed an integrated all-optical signal processor on a silicon photonic chip that corrects signal distortion in real-time, addressing limitations in current AI data center communication.
A collaborative team from The Chinese University of Hong Kong (CUHK), Huazhong University of Science and Technology (HUST), and Fudan University (FDU) has introduced a novel integrated all-optical signal processor (OSP). This advancement aims to tackle the immense data transmission demands of next-generation AI systems, particularly for high-speed connections between multiple data centers.
The OSP is built on a silicon photonic chip and operates by correcting signal distortion while the data is still in its optical form, bypassing the need for conversion to an electrical signal. This real-time optical domain processing allows for faster and more efficient handling of signal impairments, thereby enhancing communication efficiency within large-scale AI systems that rely on synchronous operation across numerous servers and data centers.
Experimental validation demonstrated the OSP's capability to process signals at an aggregate data rate of 1.6 terabits per second (Tb/s). The system achieved a processing latency of less than 60 picoseconds and exhibited remarkably low energy consumption, in the range of tens of femtojoules per bit. These performance metrics offer a disruptive alternative to conventional digital signal processing (DSP) technologies, which often face limitations in speed, latency, and power consumption, paving the way for more energy-efficient AI supercomputing.
Inspired by neuromorphic computing and machine learning principles, the OSP's design allows for precise tuning of optical paths. This enables the processor to effectively analyze complex temporal features of high-speed optical signals, leading to more accurate correction of distortions. The processor functions as a programmable nonlinear equalizer, adaptable to various transmission impairments such as chromatic dispersion and bandwidth limitations. It can also expand usable wavelength-division multiplexing (WDM) bandwidth, potentially increasing fiber transmission capacity significantly.
This development represents a significant stride towards overcoming the data transmission bottlenecks in AI infrastructure. By performing signal processing optically rather than electronically, the OSP dramatically reduces latency and power consumption, crucial for the high-bandwidth, low-latency demands of distributed AI training and inference. This innovation aligns with the broader additive manufacturing push for integrated photonic solutions, potentially enabling more scalable and efficient data center interconnects and future high-performance computing architectures.
Edited by the news editor with AI from the original report — please refer to the original source.