Full Breakdown
Huawei Unveils Atlas 960 SuperPoD and Peerium Architecture to Accelerate Domestic AI Model Training
By Drooid · · How we work
Core Announcement: New SuperPoD Cluster and Unified Interconnect
At Huawei Connect 2026 in Shanghai on September 17, 2026, Huawei Technologies introduced the Atlas 960 SuperPoD cluster, powered by the Ascend 960 AI chip. The system uses the Peerium Computing Architecture and an upgraded UnifiedBus interconnect with Near-Packaged Optics (NPO) via the Hi-ONE engine. Huawei says the Atlas 960 can link up to 4,096 NPU cards, delivering roughly 8 EFLOPS of FP8 performance and 1 PB of HBM for models with up to 10 trillion parameters. The Ascend 960DT chip is slated for the first quarter of 2027, three quarters ahead of schedule.
Background & Context: U.S. Export Controls and the Drive for Self-Sufficiency
U.S. restrictions on advanced semiconductor exports have limited Chinese access to Nvidia’s latest AI processors. Huawei is expanding its domestic AI stack, positioning Ascend chips as a home-grown alternative and emphasizing “full chip self-sufficiency” for China’s AI ecosystem.
Technical Details and Performance Metrics
- UnifiedBus + Hi-ONE NPO: The interconnect merges more than ten protocols, raising bandwidth to terabytes per second and cutting latency from 7 µs to 2 µs. Hi-ONE provides 7.2 TB/s per optical engine, replacing roughly 5,500 units and eliminating about 48,000 800 G optical modules, saving over 550 kW of power.
- Scaling Capability: Peerium Architecture can interconnect up to 1 million processors as a single logical computer. Current deployments include Atlas 950 SuperPoDs, with the largest SuperCluster supporting 256,000 accelerator cards.
- Performance Gains: Huawei claims a 2.3-fold training improvement and 2.5-fold inference improvement for 10-trillion-parameter models versus the Atlas 950 generation. Expected mean time between failures is to double, with availability projected at 99.8 %.
Official Statements & Responses
Yang Chaobin, executive director of the board and CEO of the ICT Business Group, described UnifiedBus as a solution to the “resource-utilization decline” that occurs when clusters scale beyond 100 k NPU cards.
Conflicting Reports & Gaps
TechCrunch reported that Huawei originally planned the Atlas 960 SuperPoD to scale to 15,488 Ascend 960 chips, while the September 17 announcement described a system with 4,096 chips. No clarification has been provided. Huawei’s claim of domestic dominance over Nvidia lacks independent market-share data.
Verbatim Quotes
- “The timing — less than two weeks before President Xi Jinping’s visit to Washington — underscores Beijing’s confidence and ambition in technology and innovation,” — George Chen, partner and chair of digital practice at The Asia Group.
What’s Next
Huawei projects global AI computing supply and demand will reach equilibrium around 2029, with China balancing slightly later in 2030. The company plans to launch the Ascend 970 and Ascend 980 series in 2028 and 2029, respectively, and to continue expanding Peerium-based SuperClusters toward the million-processor goal.