• Latest Articles
  • Online First
  • Archive
更多
Volume  期 4，2026 2026年第卷第4期

  • Key Technologies for Computing-Network Convergence in Optical Networks for Intelligent Computing Centers

    LIU Yuanhao, XIE Xiaolan, DENG Ning, FAN Minwei, CHEN Cao

    DOI：10.13756/j.gtxyj.2026.260136
    摘要：ObjectiveThe resilience of Artificial Intelligence Data Centers Optical Network (AIDCON) has become a core bottleneck limiting the efficiency of large-scale distributed Artificial Intelligence (AI) training. This paper systematically elaborates on key technologies of high-resilience AIDCON, covering two levels: intra and inter-artificial intelligence data center.MethodsFor the intra-artificial intelligence data center, targeting potential switch and link failures, a resilient Flow Completion Time (rFCT) metric is designed. The resilience is embedded into the network planning and scheduling process, and a network scheduling scheme is proposed. For the inter-artificial intelligence data center, addressing large-scale network failures caused by natural disasters, a multi-band optical network protection method based on disaggregated tunable C+L transceivers and a protection method for space-division multiplexing optical networks based on inter-core switching technology are proposed.ResultsAnalysis demonstrates that embedding resilience into the scheduling process mitigates the adverse impacts of failures. Moreover, the protection scheme combining multi-band and space-division multiplexing enables service restoration under disaster scenarios across various capacity expansion dimensions.ConclusionThe proposed schemes can effectively enhance the resilience of AIDCON, providing a systematic methodology and theoretical basis for designing highly reliable optical networks for large-scale AI training infrastructures.  
    关键词：AIDCON;network resilience;failure recovery;multi-band optical network;space-division multiplexing optical network   
    138
    |
    96
    |
    0
    <HTML>
    <L-PDF><Meta-XML>
    <引用本文> <批量引用> 169917772 false
    更新时间：2026-08-27

    YANG Shuai, YAN Fulong, YANG Hongrui, XUE Zhiwen, LIAN Boyu, ZHANG Jie

    DOI：10.13756/j.gtxyj.2026.260135
    摘要：ObjectiveAs Artificial Intelligence (AI) accelerates, the parameter scale of foundation models continues to expand. The demand for computing resources in training and inference links continues to increase, catalyzing the deployment of massive AI computing infrastructures. In intelligent compute fabrics, inter-node communication overhead is the chief bottleneck stifling compute utilization. Consequently, profiling network performance with discrete-event simulators before build-out is paramount for optimizing design and slashing these overheads. However, the amount of communication traffic is huge and has a time-space dependence. Traditional full simulation faces huge state overhead, and there is a contradiction between network scalability and simulation efficiency.MethodsFirstly, we analyzed the communication characteristics in the distributed training scenario of large models. The set communication primitives of the dense and sparse models of the DeepSeek architecture are modeled as point-to-point traffic with spatio-temporal dependencies, and the traffic characteristics are analyzed. Then, based on the periodicity of the obtained traffic and the convergence of the cluster network during the training of the large model, a steady-state extrapolation acceleration algorithm is proposed. The OMNeT++ platform is used to capture the transient changes of the network during the simulation process, and the packet-level simulation of the communication traffic is carried out at the initial stage of the simulation. After the network enters the steady state, the discrete packet-level simulation is abstracted as a flow level. The flow rate is analyzed according to the principle of maximum and minimum fairness, and the simulation process of the remaining traffic is mathematically extrapolated.ResultsThe above simulation acceleration effect is verified on three network scales. Experiments show that this method achieves a maximum of 4.7 times acceleration of the total simulation time while retaining the underlying congestion characteristics of the network. It can also reduce the memory usage of the card scale simulation by 30%. Subsequently, a small-scale verification was performed on a 100,000-card network. Under the premise that the network characteristics obtained by simulation are close to the benchmark, the simulation time is reduced by 73%, and the memory usage is reduced by 31%.ConclusionIn this paper, collective communication is modeled as fine-grained point-to-point traffic, which provides data support for the evaluation of intelligent computing networks. The proposed steady-state extrapolation acceleration algorithm effectively improves the simulation efficiency of large-scale intelligent computing networks and provides a reference for the performance evaluation of larger-scale intelligent computing networks.  
    关键词：AI computing network;traffic modeling;collective communication;discrete-event simulation   
    107
    |
    103
    |
    0
    <HTML>
    <L-PDF><Meta-XML>
    <引用本文> <批量引用> 169917770 false
    更新时间：2026-08-27

    WANG Wei, LI Zuyu, DONG Kaiyang, LI Fan

    DOI：10.13756/j.gtxyj.2026.260128
    摘要：ObjectiveDue to the multi-dimensional modulation capability, coherent technology is emerging as a core evolutionary direction for high-speed intelligent computing optical interconnects. However, the high cost and power consumption remain key bottlenecks restricting its deployment in short-reach scenarios. To effectively reduce its implementation cost and power consumption while maintaining the advantages of high-capacity and high-performance, simplified coherent technologies for Artificial Intelligence (AI) supernode interconnects have become a hot research topic.MethodsTo tackle the cost and power bottlenecks inherent in conventional coherent architectures, this article systematically reviews existing simplified coherent solutions from three perspectives: low-cost phase/frequency-locking techniques, hardware-efficient Digital Signal Processing (DSP) algorithms, and real-time implementations. In terms of low-cost phase/frequency-locking techniques, a frequency offset locking scheme based on pre-scanning combined with Fast Fourier Transform (FFT) and Chirp Z-Transform (CZT), as well as a digital heterodyne optical phase-locked loop scheme combining coarse frequency offset acquisition with fine phase tracking, are proposed to avoid using expensive narrow-linewidth lasers. As for hardware-efficient DSP algorithms, a single-pilot-tone-based polarization demultiplexing method and a novel baud-rate 3Sign-Timing Error Detector (TED) scheme based on sign operations are proposed to reduce the hardware requirements of the equalization and clock recovery modules. Besides, a low-complexity real-time coherent system is verified based on 32 Gbaud dual-polarization Quadrature Phase Shift Keying (QPSK) signals.ResultsExperimental results demonstrate that the proposed phase/frequency-locking techniques can effectively ensure frequency-locking accuracy and frequency stability, with frequency fluctuation suppressed within 200 kHz. In terms of simplified equalization, the proposed scheme not only reduces the complexity of the equalization module but also improves the polarization tracking speed to 10 Mrad/s. As for the clock recovery, the proposed 3Sign-TED scheme can effectively avoid any multiplication operations, exhibiting extremely low hardware complexity. Furthermore, based on the 32 Gbaud dual-polarization QPSK real-time coherent platform, a low-complexity real-time coherent system with a net data rate exceeding 100 Gbit/s is successfully demonstrated.ConclusionThe above technical breakthroughs provide a reference for future high-speed optical interconnects in AI data centers.  
    关键词：AI data center interconnects;simplified coherent technology;phase/frequency-locking techniques;real-time coherent system;hollow-core fiber transmission   
    133
    |
    128
    |
    0
    <HTML>
    <L-PDF><Meta-XML>
    <引用本文> <批量引用> 169917769 false
    更新时间：2026-08-27

    ZHAO Shixuan, WANG Xin, LI Yongcheng

    DOI：10.13756/j.gtxyj.2026.260126
    摘要：Objective6th Generation Mobile Communication Technology (6G) holographic services are generated by proactive user interactions and possess dynamic and bursty characteristics. To meet the multi-sensory immersive experience requirements of users, holographic services integrate diverse modalities of data, including holographic point cloud videos, audio, and olfaction. Such multi-modal services impose extremely high bandwidth, ultra-low latency, and strict synchronization transmission requirements for the network. Meanwhile, the user's perspective adopts a flexible Six-Degree-of-Freedom (6DoF) interaction method. A slight change in the perspective rotation will cause a change in the number of point clouds within the field of view, leading to a change in the rate of services requests. Holographic point cloud videos have issues of occlusion and perspective rotation limitations. Users do not need to receive all the data each time to maintain an immersive experience. Multi-modal holographic service requests are dynamic, bursty, and subject to strict synchronization constraints. Therefore, how to achieve efficient and dynamic provisioning of holographic services under 6DoF viewpoint rotation prediction poses a significant challenge. The existing methods are unable to meet the efficient provisioning requirements of the aforementioned immersive services. This paper focuses on the transmission process of multi-modal service data in the network, ignoring the holographic communication process such as data collection, compression, decoding, and rendering. It specifically conducts research on the intelligent provisioning algorithm for 6G holographic services in Optical Transport Networks (OTN), with particular emphasis on resource-efficient dynamic routing and wavelength assignment during data transmission.MethodsTo address these issues, this paper proposes an intelligent provisioning algorithm based on Transformer-based holographic viewpoint rotation prediction combined with graph neural network and deep reinforcement learning. This algorithm employs a Transformer to perform 6DoF holographic viewpoint rotation prediction. It establishes a mapping model between the predicted viewpoints and the holographic service transmission bandwidth, converting the dynamic bandwidth demand into a static one. It then uses the graph neural network based on edge-node switching convolution to extract the features of nodes and links in the optical transport network. Subsequently, it realizes the routing and wavelength assignment of multi-modal 6G holographic service data through an intelligent provisioning strategy based on deep reinforcement learning.ResultsThe simulation results show that, compared with the benchmark algorithms, the proposedalgorithmachieves a reduction of approximately 58% in the combined optimization objective of the service blocking rate and the bandwidth resource occupancy rate.ConclusionThe proposedalgorithm in this paper can enhance the resource utilization efficiency of the optical transport network for 6G holographic services in dynamic scenarios. It also reduces the blocking rate, thereby supporting the immersive experience of 6G holographic users.  
    关键词：6G holographic services;OTN;routing and wavelength assignment;holographic viewpoint rotation prediction;deep reinforcement learning   
    75
    |
    120
    |
    0
    <HTML>
    <L-PDF><Meta-XML>
    <引用本文> <批量引用> 169917771 false
    更新时间：2026-08-27
查看更多

Quick Entry

  • Author Login
  • Peer Review
  • Editorial

 

   Study on Optical Communications

 

 

 

 

 

 

 

Bimonthly, started in 1975

Governed by，China Information Communication Technologies Group Corporation (CICT)

Sponsored by：Wuhan Research Institute of Post & Telecommunications Co., Ltd.

Publisher：Editorial Department of Study on Optical Communications

 

Honorary Editor-in-Chief：YU Shao-hua

Editor-in-Chief：LI Han-bing

Executive Editor-in-Chief：XIAO Xi

Director：WANG Xi

 

Address：3rd floor, Building 4, Fenghuo Science Park, Gaoxinsilu, East Lake New Technology Development Zone, Wuhan,  Hubei Province, P.R.China

Zip Code: 430205

Tel:+8627-87691537

ISSN：1005-8788

CN：42-1266/TN

 

0