Core Insights - The Rubin CPX GPU, designed by Nvidia for ultra-long context AI inference tasks, features an innovative decoupled inference architecture, significantly enhancing hardware efficiency and cost balance [1][2] Hardware Innovations - The Rubin CPX introduces substantial changes in hardware, including a modular tray design with four sub-cards, upgraded cooling from air to liquid, and a wireless cable architecture for connectors and PCBs [2] - The GPU offers 30 Peta FLOPS of computing performance at NV FP4 precision, equipped with 128GB GDDR7 memory and a memory bandwidth of 2TB/s [1] Market Potential - The PCB market for Nvidia's AI products is projected to reach $6.96 billion by 2027, representing a 142% increase from 2026, driven by the expected shipment of 100,000 racks across various models [3] - The value of a single VR200NVL144 cabinet PCB is estimated at approximately 456,000 yuan, with a single GPU corresponding to a PCB value of 6,333 yuan (880 USD), reflecting a 113% increase compared to the GB300 model [3] Related Companies - Relevant companies in the supply chain include Shengyi Technology (生益科技), Shenzhen South Circuit (深南电路), Xingsen Technology (兴森科技), and others, indicating a broad industry impact [4]
中金:英伟达Rubin CPX采用创新解耦式推理架构 或驱动PCB市场迭代升级