华为全联接大会：昇腾芯片路线图公布，灵衢2.0技术规范开放

Core Insights - Huawei's rotating chairman Xu Zhijun delivered a keynote speech at the All-Connect Conference, unveiling the evolution roadmap of the Ascend chip and the Lingqu 2.0 technology specifications for ultra-node interconnection [1][3] Group 1: Ascend Chip Development - Over the next three years, Huawei plans to develop three series of Ascend chips: Ascend 950 series, including Ascend 950PR and Ascend 950DT, with launches scheduled for Q1 2026 and Q4 2026 respectively, followed by Ascend 960 in Q4 2027 and Ascend 970 in Q4 2028 [3] - The Ascend chips will enable Huawei to create computing solutions that meet customer demands, with ultra-nodes becoming the dominant product form in AI infrastructure [3] Group 2: Ultra-node Deployment - Since the launch of the Atlas 900 ultra-node in March, Huawei has deployed over 300 units, serving more than 20 clients across various industries, making it the largest ultra-node in the world [3] - The planned ultra-nodes based on future Ascend chips, Atlas 950 and 960, will support 8192 and 15488 computing cards respectively, achieving computing power of 8 EFLOPS and 30 EFLOPS, significantly surpassing NVIDIA's planned products [5] Group 3: Interconnection Technology - To address the technical challenges of large-scale ultra-nodes, Huawei has introduced high-reliability mechanisms at every layer of the interconnection protocol, enhancing optical interconnection reliability by 100 times [5] - The company has developed the ultra-node interconnection protocol architecture, Lingqu, to meet the requirements of large-scale ultra-nodes [5][7] Group 4: Lingqu 2.0 Specifications - Xu Zhijun announced the opening of the Lingqu 2.0 technology specifications, which include UnifiedBus 2.0 protocol and various firmware and OS reference designs [8][9][10]