智通财经网·2025-09-18 03:40Core Insights - The report from Guohai Securities highlights the increasing demand for AI computing power driven by large model training, with new architectures like GB300 and Vera Rubin expected to emerge, benefiting various segments of the computing power supply chain [1] Group 1: GPU Developments - The GB300 GPU, based on the Blackwell Ultra architecture, shows a significant performance increase with a floating-point computing capability of 15 PFLOPS, which is 1.5 times that of the B200 [1] - Vera Rubin's NVL144 performance is 3.3 times that of GB300's NVL72, while Rubin Ultra NVL576 is projected to outperform GB300 NVL72 by 14 times, with memory capacity being 8 times greater [1] - NVIDIA's revenue for FY2026 Q2 is expected to reach $46.7 billion, reflecting a year-on-year growth of 56% [1] Group 2: Server Architecture - The GB300 NVL72 system consists of 18 compute trays and 9 switch trays, integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs, with potential performance improvements of up to 50 times compared to Hopper [1] - The GB300 NVL72 integrates ConnectX-8 network cards and BlueField-3 DPUs, while the Rubin Ultra NVL576, set to launch in 2027, will utilize the Kyber architecture for enhanced density [1] Group 3: Networking Innovations - CPO technology replaces traditional pluggable optical modules, enhancing energy efficiency by 3.5 times and deployment speed by 1.3 times [2] - Quantum-X and Spectrum-X switches reduce reliance on traditional optical transceivers, providing up to 400 Tbps throughput for large-scale AI factories [2] - The new NVSwitch 7.0 expands to connect 576 GPUs, enabling non-blocking communication for larger GPU interconnections [2] Group 4: HBM Market Trends - HBM4 is expected to enter mass production in 2026, with SK Hynix leading the market [2] - SK Hynix delivered the world's first 12-layer stacked HBM4 samples to major clients in March 2025, while Samsung plans to provide samples to AMD and NVIDIA by July 2025 [2] - Micron is also showcasing plans for next-generation HBM4 in its AI memory product lineup, with mass production anticipated in 2026 [2] Group 5: Cooling Technologies - The GB300 NVL72 employs a full liquid cooling solution with independent liquid cooling plates for each chip, enhancing heat dissipation efficiency [2] - Liquid cooling technology, including both cold plate and immersion cooling, is noted for its higher efficiency and lower operational costs, with cold plate cooling being more mature and widely adopted [2]