光互连光交换

Search documents
国内首个光互连光交换GPU超节点发布,将实现2000卡规模部署
Guan Cha Zhe Wang· 2025-07-31 09:12
Core Viewpoint - The launch of the LightSphere X marks a significant advancement in AI computing infrastructure in China, utilizing innovative optical interconnect technology to enhance GPU performance and scalability [1][6][8] Group 1: Product Launch and Technology - The LightSphere X is the first optical interconnect GPU supernode in China, developed by Shanghai Yidian, Xizhi Technology, Biran Technology, and ZTE [1] - It employs distributed optical switching technology and silicon photonics to create a high-bandwidth, low-latency computing cluster [1][4] - The supernode architecture integrates GPU resources to form a unified computing entity, facilitating the transition to a "ten-thousand card collaborative" era in computing [3][4] Group 2: Technical Advantages - LightSphere X overcomes traditional power limitations of single cabinets by utilizing optical interconnects, allowing for flexible expansion and reduced deployment costs [3][4] - The distributed optical circuit switch (dOCS) technology enhances system flexibility and scalability, enabling real-time topology reconstruction and dynamic adjustment of supernode size based on computational needs [4][5] - The GPU module features a powerful architecture with liquid cooling, supporting high throughput for large model training while ensuring low-latency inference [5] Group 3: Ecosystem and Future Development - LightSphere X aims to build a fully autonomous computing ecosystem, providing a robust foundation for AI development in China [7][8] - The system is designed to be open and compatible with various interconnect protocols, mitigating risks of ecosystem lock-in [7] - The collaborative framework of "open protocols + autonomous technology + self-developed software" establishes a closed-loop system for technological and ecological autonomy [8]
国内首个光互连光交换GPU超节点发布 即将落地上海仪电智算中心
Shang Hai Zheng Quan Bao· 2025-07-29 12:23
上海仪电方面表示,光跃LightSphere X依托上海仪电算力底座,集合国内智算领域头部企业,共建了"光芯片—GPU—服务器—算力集群—智算云平台"闭 环,打造了开放共享的光电融合算力生态体系。未来,光跃LightSphere X将作为新一代智算集群核心架构,支撑全栈自主可控的国产算力池建设,助力中国 AI算力基础设施实现跨代发展。 万亿参数大模型与多模态训练的崛起,正推动算力集群迈入"万卡协同"时代。超节点架构通过深度整合GPU资源,在超节点内构建起低延迟、高带宽的统一 算力实体,已成为支撑这一演进的关键技术路径。 来源:上海证券报·中国证券网 上证报中国证券网讯(记者 刘怡鹤)7月28日,在2025世界人工智能大会(WAIC)"智算云启,共绘生态"论坛上,上海仪电(集团)有限公司(以下简 称"上海仪电")联合上海曦智科技有限公司(以下简称"曦智科技")、上海壁仞科技股份有限公司(以下简称"壁仞科技")、中兴通讯股份有限公司(以下 简称"中兴通讯"),正式发布国内首个光互连光交换GPU超节点光跃LightSphere X。 据介绍,光跃LightSphere X采用光互连技术,通过增加机柜数量构建超节点, ...