Huawei unveils the Atlas 960 SuperPoD, an AI computing cluster equipped with up to 15,488 Ascend 960 chips, and also introduces the SuperCluster, which connects multiple SuperPoD units.
Chinese technology giant Huawei has unveiled a new AI infrastructure to boost its computing power and better compete with rival chipmaker Nvidia.
Groundbreaking SuperPoD Interconnect: Leading a New Paradigm for AI Infrastructure - Huawei
Huawei Launches Open-Access SuperPoD Architecture for All-Scenario Computing - Huawei
https://www.huawei.com/en/news/2025/9/hc-superpod-innovation
Key products in Huawei's AI chips and computing power roadmap | Reuters
https://www.reuters.com/world/china/key-products-huaweis-ai-chips-computing-power-roadmap-2025-09-18/
Huawei announces new AI infrastructure as Nvidia gets locked out of China | TechCrunch
https://techcrunch.com/2025/09/18/huawei-announces-new-ai-infrastructure-as-nvidia-gets-locked-out-of-china/
On September 18, 2025, Huawei delivered a keynote speech at Huawei Connect 2025, during which the company announced its new AI strategy, which focuses on ensuring sustainable computing power in response to growing demand for AI.
Huawei is developing a large-scale AI computing cluster called 'SuperPoD.' SuperPoD is an AI computing cluster similar to NVIDIA's DGX SuperPOD , but it uses Huawei's proprietary Ascend series AI chips.
To prevent communication bottlenecks caused by connecting thousands of NPUs, Huawei is developing a proprietary high-bandwidth, low-latency network called UnifiedBus (UB), which allows multiple NPUs to function as a single computer.
One of the SuperPoDs, the Atlas 900 A3 SuperPoD, announced by Huawei in 2024, is equipped with 384 Ascend 910C NPUs, achieving a processing power of 300 PFLOPS .
The newly announced 'Atlas 950 SuperPoD' is equipped with 8,192 Ascend 950 NPUs, Huawei's new NPU, and is scheduled to be released in the fourth quarter of 2026. Furthermore, the 'Atlas 960 SuperPoD,' which further improves the performance of the 'Atlas 950 SuperPoD,' is scheduled to be released in the fourth quarter of 2027. The 'Atlas 960 SuperPoD' will be equipped with 15,488 Ascend 960 NPUs.
Additionally, the company also announced the Atlas 850 SuperPoD, the industry's first air-cooled SuperPoD designed for enterprises. The Atlas 850 SuperPoD is equipped with eight Ascend NPUs, meeting the model post-training processing and multi-scenario inference needs of most enterprises.
In addition, the company has also announced 'SuperCluster,' a large-scale cluster consisting of multiple SuperPoD units. The 'Atlas 950 SuperCluster' will be equipped with more than 500,000 NPUs, and the 'Atlas 960 SuperCluster' will be equipped with more than 1 million NPUs.
The Ascend 950PR processors used in SuperPoD and SuperCluster are equipped with high-bandwidth memory (HBM) called 'HiBL 1.0,' which is more cost-effective than industry-leading HBM3E and HBM4E, achieving 128GB capacity and 1.6TB/s memory bandwidth.
The Ascend 950DT also features a higher-performance HBM called 'HiZQ2.0,' which provides 144GB of memory and a memory bandwidth of 4TB/s.
Related Posts:
in Hardware, Posted by logu_ii