面向人工智能的数据通信网络发展

摘要：基于人工智能技术与业务对数据通信网络的需求，分析现有网络面向数据入算、智算中心互联、大规模AI训练3类场景时存在的问题，阐述“入算”“算内”“算间”网络关键技术创新情况，包括“入算”网络的业务创新探索，“算内”网络围绕架构以太网技术等多方面的革新，以及“算间”网络从IT、IP、光层开展的技术改进，并提出包含运营层、网络管控层、业务连接层、物理网络层的4层网络架构以优化数据通信网络。认为合理推动产业发展需有序规划标准化研究工作，递进式开展关键技术试点验证。

关键词：人工智能；数据通信网络；入算网络；算间网络；算内网络

Abstract: Based on the requirements of AI technology and business on data communication networks, this paper first analyzes the problems of the existing networks in three scenarios: data access for computing, interconnection of intelligent computing centers, and large-scale AI training. The key technology innovation of "access-artificial intelligence data center", "inter-artificial intelligence data center" and "intra-artificial intelligence data center" networks are then illustrated, including the business innovation exploration of "access-artificial intelligence data center" networks, the architecture innovation of "inter-artificial intelligence data center" networks around the Ethernet technology, and the technical improvement of "intra-artificial intelligence data center" networks from IT, IP, and the optical layer. After that, a 4-layer network architecture is proposed, including the operation layer, network control layer, service connection layer, and physical network layer, to optimize the data communication network. We believe that reasonable promotion of industrial development requires orderly planning of standardization research and progressive pilot verification of key technologies.

Keywords: artificial intelligence; data communication network; access-artificial intelligence data center network; inter-artificial intelligence data center network; intra-artificial intelligence data center network