华为AI芯片计划全盘托出！全球最强超节点+超级集群，未来2年全面领先

Core Viewpoint - Huawei's chip development has entered a new phase, focusing on AI computing power and advanced chip architecture to compete with global leaders like NVIDIA [1][2][3]. Group 1: Chip Development and Innovations - Huawei has introduced the Atlas 950 SuperPoD and Atlas 960 SuperPoD, which support 8192 and 15488 Ascend cards respectively, aiming to maintain the world's leading computing power [3]. - The company plans to release the Ascend 950PR in Q1 2024, adhering to a "one generation per year, doubling computing power" strategy [7]. - The Ascend 950 series, 960 series, and 970 series have been outlined for future development, with significant enhancements in performance and capabilities [8][21][24]. Group 2: Performance Metrics - The Atlas 950 SuperNode is expected to achieve 8 EFlops in FP8 computing power and 16 EFlops in FP4, with a memory capacity of 1152 TB and interconnect bandwidth of 16.3 PB/s [35]. - The Atlas 960 SuperNode will have a total throughput of 4.91 million TPS for training and 19.6 million TPS for inference, with FP8 computing power reaching 30 EFlops [42]. - The Atlas 950 SuperPlus cluster will integrate 64 Atlas 950 SuperNodes, achieving a total computing power of 524 EFlops [58]. Group 3: Strategic Positioning - Huawei acknowledges a short-term performance gap in single-chip capabilities compared to NVIDIA but aims to leverage system architecture to create supercomputers that outperform at the cluster level [5][30]. - The company emphasizes the importance of interconnect technology for large-scale supernodes, introducing the UnifiedBus interconnect protocol to enhance reliability and bandwidth [54]. - Huawei's strategy includes the development of general-purpose computing supernodes, with the TaiShan 950 supernode set to replace traditional database servers [49][50].