Wire flash
Huawei unveils Ascend 960 AI cluster with NPO optical interconnect
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
At the Huawei Connect conference in Shanghai on September 17, Huawei released its next-generation AI computing infrastructure, the Ascend 960 supernode. The system supports up to 4,096 compute cards, up from 1,024 in the previous generation, and debuts NPO (near-package optics) technology, a new optical interconnect product called Hi-ONE. Huawei claims the Hi-ONE module, the industry's first mass-produced NPO product with 7.2T capacity and built-in light source, replaces 48,000 800G optical modules with 5,500 Hi-ONE modules, reducing power consumption by over 550 kW and doubling system uptime to 99.8% availability. The Ascend 960 is designed to match training and inference needs for 10-trillion-parameter models. Huawei rotating chairman Wang Tao stated the company's AI strategy centers on computing power, hardware monetization, and open-source ecosystem building. The Ascend 910C supernode has been deployed in over 1,000 units, and the Ascend 950 is commercially available. Experts quoted include Zhu Keli, who noted the system provides a domestic foundation for large model training and drives upstream industries like optical communications and liquid cooling, and Miao Fuyou of the Global Computing Alliance, who highlighted China's full industrial chain advantage in AI infrastructure, particularly in connectivity technology.
Source report
Shanghai, September 17 — At the HUAWEI CONNECT 2024 conference in Shanghai, Huawei launched the Ascend 960 Supernode, a new-generation AI computing infrastructure. The system supports up to 4,096 computing cards, up from 1,024 in the previous generation, and introduces a key technology called NPO (Near-Package Optics) for the first time. The platform is designed to meet the training and inference demands of models with up to 10 trillion parameters.
"If data is the fuel of AI, then computing power is its engine," the company stated. In recent years, computing clusters have scaled from tens of thousands of cards to hundreds of thousands, with inter-card communication becoming a critical challenge.
"We will continue to innovate through system architecture to meet the explosive growth in AI training and inference demand in China," said Wang Tao, Vice Chairman and Rotating Chairman of Huawei. He noted that the development of the Ascend 960 chip has exceeded expectations. The air-cooled and liquid-cooled versions of the Ascend 960 Supernode are scheduled for release in the second and third quarters of next year, respectively.
Wang emphasized that the core of Huawei's AI strategy is computing power, with a focus on hardware monetization. The company aims to build competitiveness through system architecture innovation around the "Supernode + Cluster" model, while fostering an open-source and open computing ecosystem. This includes supporting mainstream large models for native training on the Ascend platform and embracing a diverse range of AI models. To date, over 1,000 sets of Ascend 910C Supernodes have been deployed, and the Ascend 950 Supernode has entered commercial use.
Key Technology: NPO (Near-Package Optics)
NPO is a critical technology in the Ascend 960 Supernode. Traditional large-scale computing systems rely heavily on copper interconnects, which face physical limitations in transmission distance and bandwidth. NPO addresses these issues by reducing interconnect power consumption and latency, while solving cabling and heat dissipation challenges in high-density computing systems.
Huawei has developed a next-generation NPO-based optical interconnect product called Hi-ONE, which is the industry's first mass-produced NPO product. It offers the highest transmission capacity in the industry at 7.2T and is the only NPO product with a built-in light source. The Ascend 960 Supernode uses 5,500 Hi-ONE modules to replace the 48,000 800G optical modules that would otherwise be required. This results in:
- Power reduction of over 550 kW
- Double the system uptime (mean time between failures)
- System availability of 99.8%
The Hi-ONE module combines high bandwidth and high reliability with low latency and low power consumption, and works with the Lingqu high-speed interconnect protocol to build a scalable supernode interconnect system.
Industry Perspectives
Zhu Keli, Founding Dean of the National Research Institute of New Economy and Chief Expert on the Intelligent Economy, noted that competition in computing power is not just about chips, but also about system architecture, software ecosystems, and industrial coordination. He pointed out that high communication latency, high costs, and insufficient reliability remain key pain points.
"Huawei's new-generation AI computing infrastructure provides a domestically producible foundation for large model training in China, driving upgrades in upstream and downstream industries such as optical communication and liquid cooling," Zhu said. "At the same time, by opening up interconnect standards, it helps reduce fragmentation in China's computing ecosystem and promotes the engineering deployment of domestic computing clusters."
Miao Fuyou, Chief Technology Officer of the Global Computing Consortium Secretariat, stated that China's greatest advantage in AI infrastructure lies in its full industrial chain capabilities—covering chips, components, complete systems, power supply, and cooling solutions. "Each area is developing rapidly, thanks to China's manufacturing strengths and strong engineering capabilities," he said. "From a technological perspective, interconnect technology is a standout area for China, playing a key role in the scalability and overall efficiency of supernodes and computing clusters."
Source: Economic Information Daily
Source
东方财富网-媒体头条Eastern
Part of this Story
Huawei unveils Ascend 960 supernode with NPO, accelerating AI chip roadmap to 2027