Nvidia(NVDA) 国泰君安·2025-03-19 11:13Investment Rating - The investment rating for the company is "Buy" [1][29] Core Insights - NVIDIA held its annual GTC conference from March 17 to 21, 2025, focusing on the release of Blackwell Ultra and Vera Rubin chips, as well as advancements in Physical AI and Agentic AI [2][7] - The Blackwell Ultra chip is set to achieve a 1.5x performance increase and is expected to enter mass production in the second half of 2025, creating 50 times the revenue opportunities for data centers compared to the previous Hopper architecture [7][10] - The next-generation Vera Rubin chip will begin shipping in the second half of 2026, featuring a memory capacity 4.2 times that of the Grace CPU and a performance increase of 2 times [12][13] - NVIDIA announced a long-term technology roadmap for its AI chips, outlining a progression from Blackwell (2024) to Feynman (2028) [13] Summary by Sections Blackwell Ultra and Rubin Chip Release - The Blackwell Ultra chip will be equipped with up to 288GB of HBM3e memory and enhanced FP4 performance, achieving a 1.5x increase in FP4 inference performance [7][10] - The Blackwell Ultra NVL72 cabinet will include 72 Blackwell Ultra GPUs and 36 Grace CPUs, with a total memory of 20TB and a bandwidth of 576TB/s [10][11] Vera Rubin Chip - The Vera Rubin platform will feature a CPU with 88 cores and a memory bandwidth 2.4 times that of Grace, with overall performance expected to be 3.3 times greater than the previous generation [12][13] - The Vera Rubin Ultra chip is projected to be released in 2027, with performance capabilities reaching 900 times that of the Hopper architecture [12][13] NVIDIA Photonics and CPO System Update - NVIDIA introduced three new switch products under the "NVIDIA Photonics" platform, significantly enhancing performance and deployment efficiency compared to traditional switches [18] - The Quantum 3450-LD switch features 144 ports with a bandwidth of 115TB/s, while the Spectrum SN6800 switch has 512 ports with a bandwidth of 409.6TB/s [18] NVIDIA Dynamo Release - NVIDIA Dynamo is an open-source software designed to enhance inference performance across data centers, claiming to double the performance of standard models and increase token generation by over 30 times for specialized models [19][21]