Domestic AI Chips: The Iteration Roadmap of Huawei Ascend

[Image] The roadmap for Huawei’s Ascend AI chips sparked heated discussions at the 2025 Huawei Connect Conference. From Ascend 910 to 950, 960, and then to 970, Huawei has iterated annually, increasing the single-chip computing power from 300 TFLOPS (FP16) to over 1 PFLOPS, doubling the HBM bandwidth, with the goal aimed at training trillion-parameter large models and edge inference. Ascend focuses on computing, interconnectivity, and ecosystem…

Leave a Comment