Huawei Unveils Ascend 960 Super Node With NPO Optical Interconnect

Huawei Connect 2026 opened in Shanghai on September 17, 2026. At the event, Huawei released the industry's first Ascend 960 super node to adopt near-packaged optics (NPO) technology, upgraded its Kunpeng super node, introduced the OceanStor M900 AI memory storage system, and built an Agentic super node cluster based on Lingqu UnifiedBus unified interconnect. Wang Tao, Huawei's deputy chairman and rotating chairman, said the core of Huawei's AI strategy is computing power, and that the company is building a computing foundation around "super node plus cluster" through system architecture innovation, developing an open-source and open computing ecosystem, and providing customers with flexible offline and online computing solutions. To date, more than 1,000 Ascend 910C super nodes have been deployed, and Ascend 950 super nodes have begun commercial-scale deployment. A super node is a computing system in which multiple computing nodes are tightly connected through a high-efficiency interconnect protocol, with unified memory addressing across physical nodes, functioning logically as a single computer.

The Ascend 960DT will be launched in the first quarter of 2027, three quarters earlier than originally planned; the Ascend 960PR will be launched in the third quarter of 2027, one quarter earlier than originally planned. Huawei will maintain an annual generation cadence, with the Ascend 970 planned for 2028 and the Ascend 980 planned for 2029, continuing to improve memory access bandwidth, memory access capacity and interconnect bandwidth.

Huawei introduced Hi-ONE, a new-generation NPO-format optical interconnect product with a single-engine transmission capacity of 7.2T, which the company described as the industry's first mass-produced NPO product. Based on the Ascend 960 chip and Hi-ONE, Huawei built the industry's first Ascend 960 super node using NPO technology. A single super node has a scale of 4,096 cards, can provide up to 8E FP8 compute, and has an HBM capacity of 1PB. The super node uses 5,500 Hi-ONE units to replace the 48,000 800G optical modules otherwise required, reducing power consumption by more than 550 kilowatts, doubling system mean time between failures, and achieving 99.8% system availability. Huawei data shows that traditional server architecture causes intra-cluster communication to exceed 40% of training time; according to simulation results from Huawei's Markov Lab, a 100,000-card cluster composed of 4K super nodes improves model floating-point compute utilization by 2.75 times compared with a 100,000-card cluster composed of 8-card servers.

As state-of-the-art models reach the 10T parameter scale, their training and inference systems include intelligent computing super nodes, general computing super nodes and protocol-conversion-free interconnect systems, while inference systems also need to deploy PB-level KV cache clusters. The Kunpeng super node has been upgraded based on Lingqu all-optical networking, supporting up to 4,096 nodes and forming a 256TB unified memory pool. Based on Lingqu, Huawei launched the OceanStor M900 cluster, an L3.5-layer PB-level KV cache supporting "one-hop direct connection." The product uses hybrid media plus an optimized retention algorithm to extend SSD read-write life by 16 times. Huawei also launched the Agentic super node cluster, which uses Lingqu UnifiedBus unified interconnect to unify multiple interconnect protocols into the Lingqu protocol, enabling equal interconnection among subsystems including Ascend super nodes, Kunpeng super nodes and KV cache clusters. Through two-layer CLOS four-plane networking, the maximum cluster scale can reach 512,000 cards, and combined with multi-track topology technology, it can support an Ascend super node cluster of up to 1 million cards.

On the ecosystem side, Kunpeng now has more than 4.16 million global developers, more than 7,200 global ecosystem partners, support for more than 560 global open-source projects, and more than 20 million openEuler installations, giving it the largest share of China's server operating system market. The CANN community has more than 5,200 monthly active developers, and more than 40 native training models have been built on Ascend plus CANN.