Sugon Completes China’s First 100,000-Card All-Domestic Computing Cluster Sugon 8000

Sugon announced this week the full deployment of the Sugon 8000, China’s first 100,000-card computing cluster powered entirely by domestic hardware. Scaling from 10,000 to 100,000 cards is not a matter of simply adding more nodes — it represents an exponential leap in system complexity, requiring entirely new solutions for network congestion, storage bottlenecks, thermal dissipation, and scheduling efficiency.

The standout feature of Sugon 8000 is its hyper-fusion architecture that integrates supercomputing and AI computing. Traditional AI data centers often face a painful trade-off: clusters built for AI training cannot handle scientific computing, and vice versa. Sugon 8000 was designed from the ground up to handle both, supporting FP64 double-precision workloads for scientific computing alongside trillion-parameter large model training — essentially building a road that handles both trucks and cars with equal efficiency.

In terms of development timeline, Sugon set a domestic record: system R&D began in 2024, construction completed in 2025, a 30,000-card trial run went live in February 2026, a 60,000-card AI4S cluster came online in April, and the full 100,000-card deployment was finalized by July — an unprecedented cadence for domestic computing infrastructure.

On the technology front, Sugon 8000 uses ScaleFabric, an IB-native RDMA high-speed network, for reliable 100K-card interconnect, paired with ParaStor distributed storage. The storage system achieved first place in both production full-node and 10-node performance categories on the 2026 Global IO500 rankings. Combined with Sugon’s liquid-cooling expertise, the system efficiently supports megawatt-level high-density deployments through domestic coolants and natural cooling methods.

The cluster adopts an open AI computing architecture that supports multi-brand AI accelerators and is compatible with mainstream ecosystems. Developers and researchers can migrate existing models and applications without learning proprietary toolchains. The open approach also enables the broader supply chain to participate in collaborative innovation.

Sugon 8000 has already been connected to the National Supercomputing Internet, providing computing services to research institutions and industry. It covers over 20 domains including materials science, electromagnetics, quantum, biomedicine, astronomy, and meteorology, and has completed more than 70 large-scale computing tests at the 10,000-card level. In aerospace manufacturing, 88,000 cards completed a 328-trillion-grid turbulent flow simulation — a world-class result achieved on domestic computing power.

From a strategic perspective, Zhengzhou played a key role as the deployment hub, with its geographic location, energy resources, and network connectivity making it a natural bridge between eastern and western computing resources. As more 100,000-card nodes are deployed across China, the country is positioned to build the world’s largest computing scheduling network.