数据中心基础设施重构
Search documents
国泰海通|计算机:英伟达GTC前瞻:聚焦Rubin落地、Feynman前瞻与基础设施重构
国泰海通证券研究· 2026-03-12 14:03
Core Insights - The main focus of GTC 2026 is not just on individual chip specifications but on whether NVIDIA can successfully mass-produce the Rubin platform, advance the Feynman architecture, and integrate optical interconnects, power supply, and liquid cooling to transition the AI industry from "buying GPUs" to "deploying AI factories" [1][4] Group 1: Event Overview - NVIDIA's GTC will take place from March 16 to 19 in San Jose, California, covering various fields including agent-based AI, AI factories, scientific AI, CUDA, high-performance inference, open models, physical AI, and quantum computing [2] - The Rubin platform is evolving from a single GPU product to an integrated AI supercomputing platform that includes CPU, GPU, interconnects, networking, and system components, enhancing the delivery unit from boards to complete cabinet systems [2] Group 2: Rubin Platform and Feynman Architecture - The Rubin platform is expected to enter mass production, with the potential unveiling of an enhanced version called Rubin Ultra, which will integrate 144 GPUs and achieve a network scale-up of up to 1.5PB/s, with a bidirectional interconnect bandwidth of 10.8TB/s per chip [2] - The Feynman architecture is anticipated to be one of the first chips utilizing TSMC's A16 process, integrating Groq's LPU hardware stack, with production expected to start in 2028 and customer shipments between 2029 and 2030 [3] Group 3: Infrastructure Transformation - The shift towards optical interconnects, high-voltage direct current (HVDC) power supply, and liquid cooling is driving the reconstruction of data center infrastructure, moving from traditional copper interconnects to higher bandwidth and lower loss optical connections [4] - The future of AI systems' scalability will depend not only on chip manufacturing capabilities but also on the efficient and stable delivery of power to each computing node, with liquid cooling becoming a standard configuration for high-power computing platforms [4]