智算光互联技术进展及趋势

发布时间:2026-01-06 作者:张平化,王会涛,王勇

摘要:AI大模型的快速迭代,直接推动集群算力需求进入高速增长通道,而互联技术作为破解集群算力瓶颈的核心抓手,其战略价值不断提升。智算集群更是将互联需求从 “单纯扩规模”推向“规模、性能、灵活性三位一体升级”。从技术底层逻辑出发,介绍了智算光互联关键技术进展,认为光电融合的核心是光互联技术的应用边界持续向芯片级、网络级延伸。而OIO/CPO互联技术与OCS,正是精准适配这一发展趋势的关键技术方向——既能满足智算集群对高带宽、低能耗、高可靠的核心需求,又可兼顾多租户场景下资源灵活分割与调配的互联诉求。

关键词:光电融合;CPO;OIO;OCS

 

Abstract: The rapid iteration of AI large models has directly propelled the demand for cluster computing power onto a high-growth trajectory. As a core lever for breaking through the bottlenecks in cluster computing power, interconnection technology is witnessing a steady rise in its strategic value. Particularly for intelligent computing clusters, interconnection requirements have evolved beyond "simple scale expansion" to a trinity upgrade encompassing scale, performance, and flexibility.From the perspective of underlying technical logic, the key technological advancements in intelligent computing optical interconnection are discussed. It is believed that the core of optoelectronic fusion lies in the continuous extension of optical interconnection technology’s application boundaries—expanding both to the chip level (e.g., optical interconnection between chips) and network level (e.g., cluster optical networks). Notably,OIO (Optical I/O)/CPO (Co-packaged Optics) interconnection technologies and OCS (Optical Circuit Switching) are precisely the key research directions that align with this development trend. They not only meet the core requirements of intelligent computing clusters for high bandwidth, low energy consumption, and high reliability, but also accommodate the interconnection needs for flexible resource partitioning and allocation in multi-tenant scenarios.

Keywords: optoelectronic fusion; CPO; OIO; OCS