Core Insights - Nvidia's CEO Jensen Huang emphasized that Nvidia is no longer just a GPU company but an essential AI infrastructure company, introducing the semi-custom AI infrastructure NVLink Fusion [2][3] - NVLink Fusion allows for a mix of Nvidia chips with other custom chips, accelerating the time to market for AI factories, targeting large-scale enterprises [2][3] Product Highlights - NVLink Fusion utilizes Chiplet technology with two configurations: one combining Nvidia's Grace CPU with custom AI accelerators, and the other linking custom CPUs with Nvidia GPUs [3] - The DGX Spark, the world's smallest personal AI supercomputer, features the GB10 Grace Blackwell super chip and achieves AI computing power of 1 PFLOPS, set to launch soon [4] - The Grace Blackwell NVL72, a large-scale system, integrates 36 Grace CPUs and 72 Blackwell GPUs, showcasing Nvidia's vertical scaling capabilities [4] Upcoming Innovations - Nvidia plans to launch the GB300 (Blackwell Ultra) in Q3 2025, featuring enhanced performance metrics compared to its predecessor GB200, including a 1.5x increase in inference performance and doubled network connectivity [4] - The RTX 5060 GPU and a new MSI laptop featuring this GPU are expected to be available in May 2025, catering to the x86 platform [5] - The RTX PRO Blackwell server, designed for enterprise AI, supports up to 8 RTX PRO 6000 Blackwell GPUs and significantly outperforms previous systems in specific workloads [5] Strategic Collaborations - Nvidia announced a partnership with TSMC and Foxconn to establish an AI supercomputer in Taiwan, along with the creation of a new office named "NVIDIA Constellation" [5][6] - The business model of establishing AI factories can be replicated by Chinese AI chip companies, particularly Huawei, through collaborations with large enterprises in various verticals [6]
黄仁勋最新演讲:授权 NVLink 技术 支持定制CPU、AI芯片混搭