国泰海通｜计算机：华为公布AI芯片路线图，全球最强超节点2025Q4上市

Core Insights - Huawei is set to launch the new Ascend 950PR chip in Q1 2026 and the Ascend 970 chip in Q4 2028, indicating a strong commitment to advancing its AI chip technology [2] - The global strongest super node, Atlas 950 SuperPoD, is expected to be released in Q4 2025, showcasing Huawei's leadership in AI computing infrastructure [3] - Huawei emphasizes an open hardware and software ecosystem, allowing industry partners to develop products based on its technology [4] Group 1: AI Chip Roadmap - Huawei's roadmap includes the release of the Ascend 910C in Q1 2025, followed by the Ascend 950PR in Q1 2026 and the Ascend 950DT later in 2026 [2] - The Ascend 970 chip will feature a computing power of 4 PFLOPS (FP8) and 8 PFLOPS (FP4), with a memory capacity of 288GB and bandwidth of 14.4TB/s [2] - In comparison, NVIDIA's Blackwell Ultra GB300 offers 15 PFLOPS (FP4) with 288GB HBM3e and 8TB/s bandwidth, highlighting Huawei's competitive positioning [2] Group 2: Super Node Development - As of September 18, 2025, over 300 CloudMatrix 384 super nodes have been deployed, serving more than 20 customers [3] - The Atlas 950 SuperPoD will support up to 8192 Ascend cards, while the Atlas 960 SuperPoD, expected in Q4 2027, will support 15488 cards, establishing a new benchmark in computing power [3] - Huawei's super node clusters, Atlas 950 SuperCluster and Atlas 960 SuperCluster, will exceed 500,000 and reach one million cards respectively, solidifying their status as the world's strongest computing clusters [3] Group 3: Open Ecosystem Strategy - Huawei is fully opening its super node technology, including the Lingqu protocol and reference architecture, to enable industry self-development [4] - The company will also open source its operating system components, allowing users to integrate and maintain versions based on their needs [4] - This strategy aims to foster a collaborative environment for innovation and development within the AI computing sector [4]