China's Homegrown "Computing Power Artery" Rivals Nvidia! Dawning Information Releases First Fully Self-Developed IB Network Product, Achieving Support for Ten-Thousand-Card-Scale Intelligent Computing Clusters

Invest in stocks with Golden Kylin Analyst Reports—authoritative, professional, timely, comprehensive—helping you discover potential thematic opportunities!

(Source: Cailian Press)

Science and Technology Innovation Board Daily March 13 — (Reporter: Guo Hui) On March 12, Sugon announced a major breakthrough in domestically developed high-end native RDMA technology in Zhengzhou, officially launching its first full-stack self-developed 400G lossless high-speed network—scaleFabric.

It is understood that Sugon’s scaleFabric is the first domestically produced full-stack self-developed InfiniBand (IB) network product.

In the field of large-scale intelligent computing clusters, RDMA (Remote Direct Memory Access) networks have become a fundamental requirement for computing power centers. Sugon’s scaleFabric is based on native RDMA architecture, with 100%自主研发 from the bottom layer’s 112G SerDes IP and hardware devices to the upper-layer management software, breaking the overseas monopoly on high-end native RDMA network technology.

Wu Hequan, academician of the Chinese Academy of Engineering, stated that high-speed networks, as a core technology of computing infrastructure, are directly related to the security and quality of the country’s computing infrastructure. In the context of large model training and intelligent computing cluster deployment, networks need to have ultra-low latency, ultra-high bandwidth, and lossless transmission capabilities. RDMA high-speed networks are the “arteries of computing power” for intelligent computing clusters.

Currently, InfiniBand and RoCE are the mainstream high-speed network technologies in the market. Among them, InfiniBand is widely used in top supercomputers and AI clusters worldwide due to its low latency and native lossless transmission capabilities. According to the TOP500 list, about 60% of high-performance computing systems globally use InfiniBand architecture.

Sugon’s scaleFabric adopts the InfiniBand native RDMA route, not RoCE. In this regard, Sugon stated that InfiniBand’s technical advantages in AI/HPC are irreplaceable, as it is a true lossless network. The lossless feature benefits RDMA performance and makes network management easier.

Previously, the InfiniBand (IB) industry chain was largely monopolized by overseas vendors.

In 2019, NVIDIA acquired Mellanox for $6.9 billion. Mellanox specializes in InfiniBand network technology. Since the 1990s, as supercomputing technology diverged, InfiniBand gradually became synonymous with high-speed networks. This acquisition was seen as a pivotal turning point in the history of computing power, transforming NVIDIA from a GPU product manufacturer into an AI data center company, and integrating network technology to address communication bottlenecks in AI scaling.

Industry insiders told Science and Technology Innovation Board Daily that Chinese market users face many pain points when using IB networks. Although IB is an open protocol, the technology is effectively monopolized by Mellanox. Additionally, its cost, service, and response speed are difficult to meet domestic customer needs.

In terms of performance, it is claimed that the technical specifications of scaleFabric 400 series network products fully match NVIDIA’s NDR products.

Wang Wei, chief engineer of Sugon’s high-speed network interconnection product division, said that the core of scaleFabric is two自主研发 high-speed network chips. The scaleFabric 400 network card chip achieves 400G high bandwidth, using自主研发 RDMA engine; the scaleFabric 400 switch chip provides 64T bidirectional switching capacity, employing VCT switch architecture, with forwarding latency as low as 260 nanoseconds.

Additionally, scaleFabric switches support up to 80 ports of 400G, a 25% increase compared to NVIDIA’s NDR 64-port switches, significantly reducing network deployment costs. The maximum QP support per card is 256K, a 100% increase over NVIDIA NDR, and the subnet interconnection scale is 2.33 times that of traditional IB, easily supporting clusters of up to 114,000 cards, with a total network cost reduction of 30%.

“Based on actual testing data, scaleFabric can basically match NVIDIA NDR’s performance, with some data even surpassing the NDR series. The technical specifications of the two chips fully align with NVIDIA’s current mainstream NDR products,” Wang Wei said.

Regarding compatibility with existing IB networks, it is introduced that scaleFabric provides native interfaces, fully compatible with various communication libraries, enabling seamless integration with HPC and AI applications. Large model training can be directly migrated to scaleFabric-based systems without code modification. Additionally, network management and maintenance practices are compatible with IB user habits.

For compatibility with domestically produced computing hardware, Sugon’s Senior Vice President Li Bin told Science and Technology Innovation Board Daily that scaleFabric has already achieved interconnection adaptation with different computing chips. Future plans include exploring more efficient methods, such as connecting computing chips to network cards via proprietary collaborative links, and promoting chip interconnection protocol sharing within the ecosystem to facilitate more efficient direct chip-to-chip communication with other vendors’ chips.

In practical applications, scaleFabric has been deployed at the national supercomputing internet core node in Zhengzhou, supporting three scaleX intelligent computing clusters with a total of 30,000 cards.

Li Bin said that as the product is implemented in ultra-large-scale intelligent computing clusters, domestically developed native RDMA technology is gradually maturing, and a high-performance network industry ecosystem centered around it is accelerating.

With the release of this first domestically developed full-stack self-made IB network product, Sugon has partnered with Lenovo Qitian, ZTE, Damo Data, Ourecos, Lexun Technology, Jiuzhang Cloud, and other industry collaborators to establish the Guanghe Organization High-Performance Computing Special Committee AIDC High-Speed Network Working Group.

Li Liu, Vice President of Sugon Information Industry (Beijing) Co., Ltd., stated, “After the launch of scaleFabric, future efforts will rely on standardization organizations, adopting an open stance to collaborate with more domestic partners. From AI data centers, acceleration chips, to network equipment and AI frameworks, we will promote industry collaborative innovation and build an AI network ecosystem.”