September 2025 Huawei announced large Supernodes, enhancing scale and efficiency through connectivity
September 2025
Â
Abstract
by Carol Cao and Bob Wheeler
September 25, 2025
The 10th Huawei CONNECT 2025 Conference, themed “Accelerate Industrial Intelligence” was held at the Shanghai World Expo Exhibition Center from September 18 to 20, 2025.
The conference brought together thought leaders, business elites, technology experts, developers and other influential figures from different countries and industries. The conference presented Huawei’s comprehensive intelligentization strategy through a three-dimensional lens of “strategic panorama – industrial technology – ecological development.” It unveiled new digital and intelligent infrastructure products, industry-specific solutions, and development tools. Experts from various countries and industries shared AI application case studies.
On the first day of the conference, Huawei unveiled its three-year Ascend chip plan and launched two new super nodes: Atlas 950 SuperPoD and Atlas 960 SuperPoD, supporting 8,192 and 15,488 Ascend cards respectively. Additionally, Huawei introduced the Atlas 950 SuperCluster and Atlas 960 SuperCluster clusters based on these super nodes, with computing power exceeding 500,000 NPUs and reaching one million NPUs respectively.
Eric Xu delivered a keynote speech, announcing the Ascend chip plan for the next three years and the world’s largest computing power super-node.
In early 2024, Huawei unveiled the 910C, its most advanced AI processor. During the conference, Rotating Chairman Eric Xu announced an ambitious strategy featuring three groundbreaking AI chips in a single announcement. The Ascend 950 series, the Ascend 960 series, and the Ascend 970 series.
Ascend 950 series: The Ascend 950 series includes the Ascend 950PR (optimized for prefill and recommendation) and Ascend 950DT (optimized for decode and training). The Ascend 950PR and Ascend 950DT chips will adopt the same Ascend 950 chip die, but there are two proprietary HBM: HiBL 1.0 and HiZQ 2.0, respectively packaged with the Ascend 950 chip die. The Ascend 950DT chip will be available in the fourth quarter of 2026.
Ascend 960 series: Compared with Ascend 950 chips, the Ascend 960 will have twice the computing power, memory access bandwidth, memory capacity, and number of interconnect ports. It is designed to significantly boost training and inference performance. The Ascend 960 chips will be available in the fourth quarter of 2027.
Ascend 970 series: Currently, this series is in the planning stage. The goal is to double computing power in FP4 and FP8, double interconnect bandwidth compared to the Ascend 960, and increase memory access bandwidth by at least 1.5x. The Ascend 970 is expected to launch in Q4 2028.
The Atlas 950 SuperPoD is built with Ascend 950DT chips. This SuperPoD will have up to 8,192 Ascend 950DT chips, 20 times more NPUs than the Atlas 900 A3 SuperPoD. It will be available in the fourth quarter of 2026. In its full configuration, the Atlas 950 SuperPoD will have 160 cabinets, including 128 compute cabinets and 32 communications cabinets, deployed in 1,000 m2 of floor space. All of these cabinets will be linked with all-optical interconnects.
Full text of the research note is available for LightCounting subscribers at: https://www.lightcounting.com/login