国内首款光互连光交换GPU超节点——光跃LightSphere X发布｜聚焦WAIC 2025

Core Insights - The rise of trillion-parameter large models and multimodal training is driving computing clusters into the "ten-thousand-card collaboration" era, with supernode architecture becoming a core technological pillar for this evolution [1] Group 1: Product Launch and Technology - The LightSphere X, the first optical interconnect GPU supernode in China, was officially launched at the WAIC 2025 forum, integrating various advanced technologies [1] - The supernode utilizes distributed optical switching technology, which enhances flexibility and scalability while optimizing system cost-performance ratio [3] - The GPU module of LightSphere X is designed with a powerful architecture, supporting high-performance training and inference for large models [7] Group 2: Innovation and Advantages - LightSphere X employs optical interconnect technology to overcome the limitations of traditional copper cabling, allowing for elastic expansion and reduced deployment costs [2] - The distributed optical circuit switch technology allows for real-time topology reconstruction in case of failures, improving training performance and reducing GPU redundancy costs [3] - The system can dynamically adjust the scale of the supernode based on computational needs, supporting deployments of up to 2,000 cards [3] Group 3: Software and Management - The intelligent cloud platform software of LightSphere X enables flexible network topology adjustments and efficient resource allocation, significantly enhancing node scalability [7] - The unified management platform integrates scheduling engines and training frameworks for intelligent lifecycle management of the supernode [7] - The system allows for rapid replacement of faulty nodes and supports continuous training, ensuring stable model operation [7]