UNIS(SZ:000938) Huan Qiu Wang·2025-12-19 08:29Core Insights - The article discusses the launch of the H3C UniPoD S80000 super node product by Unisplendour Corporation's subsidiary, H3C, aimed at addressing the challenges of communication walls and computing power utilization in large model training and inference [1][5]. Group 1: Product Features and Innovations - The H3C UniPoD S80000 super node utilizes a "computing power × connectivity" technology concept, achieving full interconnection of GPUs through a Scale-up architecture, resulting in an 8-fold increase in inter-card bandwidth compared to traditional 8-card servers and an 80% improvement in single-card inference efficiency [1][5]. - The super node supports liquid cooling for high-density deployment and is compatible with multiple brands of GPUs, addressing the long-term stability requirements for large model training through software and hardware collaborative optimization [1][5][7]. Group 2: Market Context and Demand - As the market for high-performance computing power surges, driven by the increasing prevalence of high-parameter MoE large models like DeepSeek, the ability to efficiently train and infer large models becomes critical for gaining a competitive edge in the rapidly evolving AI landscape [2][3]. - The article highlights the importance of building robust and efficient AI computing infrastructure, with super node products emerging as a key focus in the current computing power sector [2][3]. Group 3: Technical Advantages - The article emphasizes that traditional cross-node communication methods lead to significant communication overhead, reducing computing power utilization. The Scale-up technology enables direct high-speed communication between GPUs, significantly enhancing GPU utilization and reducing idle time [3][4]. - In the inference phase, the super node's support for independent scaling of computing and storage resources allows for efficient resource allocation, particularly in scenarios requiring frequent access to KV Cache, thus minimizing resource waste and ensuring low latency [4][5]. Group 4: Stability and Reliability - The H3C UniPoD S80000 is designed with a focus on stability and maintainability, crucial for preventing training interruptions that could lead to resource waste and model performance degradation. The product incorporates collaborative optimization of software and hardware to ensure uninterrupted long-term training [7][8]. - The company is actively investing in optical interconnection technology to leverage the benefits of high speed, low latency, and low energy consumption while addressing the reliability issues associated with optical components [7][9]. Group 5: Future Outlook - H3C aims to continue developing super node products that support large-scale deployments of 1024 cards and above, enhancing the scale and efficiency of intelligent computing clusters [7][8]. - The company is committed to building a strong, diverse, and continuously evolving computing infrastructure to support the AI industry's growth and transformation [8][9].