At the Huawei Connect event in Shanghai on September 17, 2026, Huawei announced that it plans to launch the Ascend 960DT AI chip in the first quarter of 2027. This represents an acceleration from the previously expected third quarter of 2027. A Huawei spokesperson stated that development of the Ascend 960 series is progressing faster than anticipated. The company also plans to release the Ascend 960PR in the third quarter of 2027, with subsequent models Ascend 970 and 980 planned for 2028 and 2029, respectively.
Huawei Unveils Peerium Architecture to Link One Million Processors and Accelerates Ascend 960DT Launch to Q1 2027
To support large-scale AI workloads, Huawei introduced the Peerium Computing Architecture. This architecture aims to allow up to one million processors to function as a single, unified computer through nested parallelism, unified memory addressing, and peer interconnect.
A core component of this architecture is UnifiedBus (UB), a high-speed interconnect technology based on a single open protocol. UB enables scalable connections across CPUs, NPUs, memory, SSDs, network interface cards (NICs), and switches, facilitating peer interconnect between compute, storage, and networking.
Huawei is currently deploying the Atlas 950 SuperCluster, which utilizes 256,000 cards and is built on the Peerium architecture. The company is also testing the Atlas 960 system, which is expected to use near-packaged optics (NPO) for connectivity. Additionally, Huawei announced the Atlas 960E SuperPoD, which can connect up to 4,096 AI processors in a single system.
Sources
- Huaweiが次世代AIチップ「Ascend 960DT」を2027年初めに前倒し投入、最大100万個のプロセッサーを連携させる構想も発表 （GIGAZINE, 2026-09-18）
- Huawei
3 more sourcesHide sources
- Z.aiが中国製AIインフラで「GLM-5.3-Flash」の本番サービスを提供したノウハウを共有、AIエージェントでインフラを管理してNVIDIA GPUと同等まで効率化 （GIGAZINE, 2026-09-18）
- nvidia/GLM-5.3-NVFP4 （HF: NVIDIA, 2026-09-14）
- Model Optimizer