An all-optical signal processor enabling terabit-per-second real-time equalization.
Saved in:
| Title: | An all-optical signal processor enabling terabit-per-second real-time equalization. |
|---|---|
| Authors: | Wang, Benshan (AUTHOR), Xiao, Qiarong (AUTHOR), Xu, Tengji (AUTHOR), Fan, Li (AUTHOR), Liu, Shaojie (AUTHOR), Kong, Qiuqiang (AUTHOR), Dong, Jianji (AUTHOR), Zhang, Junwen (AUTHOR), Huang, Chaoran (AUTHOR) |
| Source: | Science. 6/11/2026, Vol. 392 Issue 6803, p1-12. 12p. |
| Subjects: | Wavelength division multiplexing, Optical dispersion, Data transmission systems, Neuromorphics, Optical information processing, Integrated circuit interconnections, Data centers |
| Abstract: | Large-scale artificial intelligence training demands ultralow-latency, energy-efficient interconnects for massive graphics processing unit clusters. In intensity-modulation/direct-detection links, digital signal processing (DSP) equalization is limited by nonideal equalization caused by phase loss as well as tight power and latency budgets. We present an integrated, programmable optical signal processor (OSP) that functions as a nonlinear universal equalizer and performs all-optical, DSP-free, real-time equalization. A deep reservoir with all-optical readout enables a Vernier scheme with ~1-picosecond (ps) sampling resolution and a tunable memory window. The OSP simultaneously equalizes eight wavelength-division–multiplexing (WDM) channels, delivering 1.6-terabits/second aggregate throughput with <60-picoseconds latency and tens of femtojoules/bit energy consumption. Operating before detection, it provides superior chromatic dispersion compensation, mitigates transceiver bandwidth limits and fiber nonlinearity, and expands the usable WDM window by a factor of 6.8. Editor's summary: Optical interconnects can enable high-bandwidth connectivity for the scale of artificial intelligence (AI) data centers, but they need improvements in latency (the delay before data transfer) and digital processing to equalize signals. Wang et al. designed an all-optical integrated signal-processing chip that can equalize signals from multiple channels and deliver 1.6 terabits per second of data. This chip overcomes the bandwidth-dispersion limitations of digital processing, enables real-time compensation of both linear and nonlinear impairments in transceivers and optical fibers, and delivers an energy efficiency of 67.5 femtojoules per bit and a latency of 55 picoseconds. —Phil Szuromi INTRODUCTION: The rapid growth of generative artificial intelligence (AI) is placing unprecedented demands on computing infrastructure. Large-scale models require vast numbers of graphics processing units to operate together, often across multiple data centers. For this distributed training to work efficiently, data must move between sites with extremely low and consistent delay; otherwise, valuable computing resources sit idle and costs rise. At the same time, the energy used to transfer data between processors must be reduced to keep future systems sustainable. Optical data interconnects are central to meeting these needs, but current solutions struggle to simultaneously deliver low delay, low energy use, and high scalability. RATIONALE: Optical interconnect can carry very large amounts of data, but their performance in short-distance data center connections is limited by heavy reliance on digital signal processing (DSP). In direct-detection systems, phase information in the optical signal is lost, making compensation of chromatic dispersion (CD) and modulation-induced chirp ill-posed. Even in wavelength regions designed to minimize dispersion, residual dispersion creates faded frequency notches that restrict usable bandwidth. As transmission speeds continue to increase to 200 gigabaud (GBaud), this usable window shrinks rapidly, forcing the use of more parallel fibers and increasing hardware complexity. Moreover, the strict energy and delay constraints in data centers make complex DSP-based correction increasingly impractical. Optical signal processing (OSP), particularly neuromorphic photonic processing, offers a promising alternative by combining intrinsic low latency and energy efficiency with emerging programmability and scalability enabled by integrated photonics and machine learning. RESULTS: This work proposes and demonstrates an integrated OSP platform that performs all-optical, DSP-free, real-time equalization. The architecture employs a deep optical reservoir with an all-optical readout, where intentionally detuned sampling periods create a Vernier-like effect that achieves effective sampling resolution down to 1 picosecond (ps) and provides a long, tunable memory window. By preserving the full optical field prior to detection, the system enables superior chromatic dispersion compensation compared with advanced DSP methods, effectively expanding the usable wavelength-division multiplexing (WDM) bandwidth and increasing per-fiber capacity and shoreline density. Experimentally, the programmable OSP compensates both linear and nonlinear impairments, including transceiver bandwidth limitations and fiber nonlinearities, and dynamically adapts to varying wavelengths, data rates, and modulation formats. The system demonstrates simultaneous real-time equalization of eight WDM channels at 100 GBaud each, achieving 1.6 terabits per second aggregate throughput. Measured processing latency is below 60 ps, and energy consumption reaches the tens of femtojoules per bit (fJ/bit) regime. CONCLUSION: This integrated optical processor enables real-time signal equalization with extremely low delay and energy use while expanding usable transmission bandwidth. Its programmable and broadly applicable compensation capability supports scalable, low-latency optical interconnect for future computing systems. These results identify neuromorphic photonic processing as a promising path toward tightly synchronized and energy-efficient connections for large, distributed AI infrastructure. Distributed data center (DC) interconnection supports large-scale AI. Optical interconnects face scaling challenges. The optical signal processor (OSP) acts as a universal nonlinear equalizer, delivering dispersion compensation while mitigating transceiver bandwidth limits and nonlinear distortions with under 60 ps latency and tens of fJ/bit energy. By expanding the usable wavelength-division multiplexing (WDM) window, OSP increases transmission capacity and accelerates AI training across multiscale data centers. [ABSTRACT FROM AUTHOR] |
| Copyright of Science is the property of American Association for the Advancement of Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Psychology and Behavioral Sciences Collection |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Large-scale artificial intelligence training demands ultralow-latency, energy-efficient interconnects for massive graphics processing unit clusters. In intensity-modulation/direct-detection links, digital signal processing (DSP) equalization is limited by nonideal equalization caused by phase loss as well as tight power and latency budgets. We present an integrated, programmable optical signal processor (OSP) that functions as a nonlinear universal equalizer and performs all-optical, DSP-free, real-time equalization. A deep reservoir with all-optical readout enables a Vernier scheme with ~1-picosecond (ps) sampling resolution and a tunable memory window. The OSP simultaneously equalizes eight wavelength-division–multiplexing (WDM) channels, delivering 1.6-terabits/second aggregate throughput with <60-picoseconds latency and tens of femtojoules/bit energy consumption. Operating before detection, it provides superior chromatic dispersion compensation, mitigates transceiver bandwidth limits and fiber nonlinearity, and expands the usable WDM window by a factor of 6.8. Editor's summary: Optical interconnects can enable high-bandwidth connectivity for the scale of artificial intelligence (AI) data centers, but they need improvements in latency (the delay before data transfer) and digital processing to equalize signals. Wang et al. designed an all-optical integrated signal-processing chip that can equalize signals from multiple channels and deliver 1.6 terabits per second of data. This chip overcomes the bandwidth-dispersion limitations of digital processing, enables real-time compensation of both linear and nonlinear impairments in transceivers and optical fibers, and delivers an energy efficiency of 67.5 femtojoules per bit and a latency of 55 picoseconds. —Phil Szuromi INTRODUCTION: The rapid growth of generative artificial intelligence (AI) is placing unprecedented demands on computing infrastructure. Large-scale models require vast numbers of graphics processing units to operate together, often across multiple data centers. For this distributed training to work efficiently, data must move between sites with extremely low and consistent delay; otherwise, valuable computing resources sit idle and costs rise. At the same time, the energy used to transfer data between processors must be reduced to keep future systems sustainable. Optical data interconnects are central to meeting these needs, but current solutions struggle to simultaneously deliver low delay, low energy use, and high scalability. RATIONALE: Optical interconnect can carry very large amounts of data, but their performance in short-distance data center connections is limited by heavy reliance on digital signal processing (DSP). In direct-detection systems, phase information in the optical signal is lost, making compensation of chromatic dispersion (CD) and modulation-induced chirp ill-posed. Even in wavelength regions designed to minimize dispersion, residual dispersion creates faded frequency notches that restrict usable bandwidth. As transmission speeds continue to increase to 200 gigabaud (GBaud), this usable window shrinks rapidly, forcing the use of more parallel fibers and increasing hardware complexity. Moreover, the strict energy and delay constraints in data centers make complex DSP-based correction increasingly impractical. Optical signal processing (OSP), particularly neuromorphic photonic processing, offers a promising alternative by combining intrinsic low latency and energy efficiency with emerging programmability and scalability enabled by integrated photonics and machine learning. RESULTS: This work proposes and demonstrates an integrated OSP platform that performs all-optical, DSP-free, real-time equalization. The architecture employs a deep optical reservoir with an all-optical readout, where intentionally detuned sampling periods create a Vernier-like effect that achieves effective sampling resolution down to 1 picosecond (ps) and provides a long, tunable memory window. By preserving the full optical field prior to detection, the system enables superior chromatic dispersion compensation compared with advanced DSP methods, effectively expanding the usable wavelength-division multiplexing (WDM) bandwidth and increasing per-fiber capacity and shoreline density. Experimentally, the programmable OSP compensates both linear and nonlinear impairments, including transceiver bandwidth limitations and fiber nonlinearities, and dynamically adapts to varying wavelengths, data rates, and modulation formats. The system demonstrates simultaneous real-time equalization of eight WDM channels at 100 GBaud each, achieving 1.6 terabits per second aggregate throughput. Measured processing latency is below 60 ps, and energy consumption reaches the tens of femtojoules per bit (fJ/bit) regime. CONCLUSION: This integrated optical processor enables real-time signal equalization with extremely low delay and energy use while expanding usable transmission bandwidth. Its programmable and broadly applicable compensation capability supports scalable, low-latency optical interconnect for future computing systems. These results identify neuromorphic photonic processing as a promising path toward tightly synchronized and energy-efficient connections for large, distributed AI infrastructure. Distributed data center (DC) interconnection supports large-scale AI. Optical interconnects face scaling challenges. The optical signal processor (OSP) acts as a universal nonlinear equalizer, delivering dispersion compensation while mitigating transceiver bandwidth limits and nonlinear distortions with under 60 ps latency and tens of fJ/bit energy. By expanding the usable wavelength-division multiplexing (WDM) window, OSP increases transmission capacity and accelerates AI training across multiscale data centers. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 00368075 |
| DOI: | 10.1126/science.ady5344 |
