| 111 | 1 | 91 |
| 下载次数 | 被引频次 | 阅读次数 |
针对智算中心互联对光网络的新需求，结合当前智算网络发展现状，探讨智算中心互联架构及关键技术，以实现高性能算力互联。同时，针对跨智算中心分布式协同训练场景，搭建基于光传送网（OTN）的跨智算中心现网试验环境，在广域收敛比不低于16∶1的场景下，百亿AI大模型跨域分布式训练性能达到95%以上。该试验验证采用单波800G实现300 km的传输，并验证其超高可靠传输能力。
Abstract：The interconnection architecture and key technologies for intelligent computing centers are explored to address the new demands of optical networks for their interconnection, while considering the current development status of intelligent computing networks, with the aim of achieving high-performance computing power interconnection. Furthermore, focusing on the scenario of distributed collaborative training spanning multiple intelligent computing centers, an optical transport network(OTN)-based experimental testbed for cross-center interconnection is implemented on a live network. Under conditions where the wide-area convergence ratio is no less than 16:1, a performance of over 95% is achieved for cross-domain distributed training of AI large models with 10 billion parameters. Single-wave 800G transmission over 300 km is employed, and its ultra-high reliability and transmission capability are verified.
[1]李国杰.智能计算技术的历史性突破与巨大挑战[J].集成技术,2025, 14(1):1-8
[2]中国信息通信研究院.中国算力发展报告(2024年)[R]. 2024
[3]丁宏庆,张鹏飞,牛红,等.云化的智算中心万卡集群创新与实践[J].电信科学, 2024, 40(12):125-135
[4]TAN Y X, MAN X K, WANG G Q, et al. Field trial of long-distance RDMA lossless transmission for wide-area data center interconnection[EB/OL].(2024-11-05)[2025-11-08]. https://ieeexplore.ieee.org/document/10809882
[5]张德朝,孙将,曹珊,等.面向跨智算集群互联的新型HIC-OTN技术[J].电信科学, 2025, 41(4):53-60
[6]LIU Y Y, ZHANG A X, WANG X S, et al. Field trial of multidatacenter distributed training for LLM based on bandwidth convergence and two parallel strategies over 120km highreliability 800Gbit/s C+L OTN[EB/OL].[2025-11-09]. https://ieeexplore.ieee.org/document/11047207
[7]中国信息通信研究院.算力时代全光网架构研究报告[R]. 2024
[8]中国联通研究院.基于RDMA的长距无损数据搬移技术白皮书[R].2024
[9]易昕昕,张乃晗,刘雅承,等.算力智联网关键技术研究[J].中兴通讯技术, 2025, 31(2):31-38. DOI:10.12142/ZTETJ.202502005
[10]王光全,满祥锟,徐博华,等.确定性光传输支撑广域长距算力互联[J].邮电设计技术，2024(2):7-13
[11]MACARTHUR P, RUSSELL R D. A performance study to guide RDMA programming decisions[EB/OL].[2025-11-09]. https://ieeexplore.ieee.org/document/6332248
[12]唐雄燕,王海军,杨宏博.面向专线业务的光传送网（OTN）关键技术及应用[J].电信科学, 2020, 36(7):18-25
[13]WANG C Y, HU Y K, SHEN S K, et al. Channel power management of 400 G transmission system based on C6T+L6T spectrum and QPSK modulation format[J]. Optics express,2024, 32(11):20279. DOI:10.1364/oe.523644
[14]QU W X, ZHANG Y, LU Y M, et al. Low-cost lightweight-client twin-field quantum key distribution network with wavelength division multiplexing[EB/OL].[2025-11-10]. https://ui. adsabs.harvard.edu/abs/2022OptEn..61a6102Q/abstract
[15]中国联通研究院. AI时代的全光底座白皮书[R]. 2025
基本信息:
中图分类号:TN929.1
引用信息:
[1]谭艳霞,满祥锟,吴绍辉,等.面向智算中心互联的光算协同技术研究[J].中兴通讯技术,2025,31(06):13-19.