Sugon's 100,000-Card Domestic Computing Cluster Completed in Zhengzhou
N.R. Finch
Sugon (中科曙光) on July 10 announced that Sugon 8000, a 100,000-card fully domestic AI supercluster, is now operational in Zhengzhou and connected to China's national computing network — the country's first cluster of this scale built entirely on domestic chips.
100,000 cards — what scale of computing power is this?
Sugon 8000 runs on 100,000 domestically made accelerator cards, with underlying chips supplied by Hygon (海光) and other Chinese vendors. The entire stack — chips, networking, cooling — is domestically developed.
This means → China has completed engineering-scale validation of a fully domestic computing cluster at the 100,000-card level, moving beyond lab prototypes and smaller pilot deployments.
In plain terms = the question has shifted from "can we build it?" to "can we run it at scale?" — this machine proves the answer is yes.
How does one system handle both scientific computing and AI training?
Sugon 8000 uses a "native super-intelligent convergence" architecture: the same hardware supports both FP64 double-precision scientific computing and trillion-parameter large-model training, covering the full precision range from FP64 down to INT8.
This means → research institutions do not need separate systems for scientific computing and AI training — one cluster handles both.
Real-world runs already logged include: 80,000 cards accelerating protein-folding simulation, 88,000 cards completing a 3.28-quadrillion-grid turbulence simulation, and 90,000 cards running a high-precision DFT simulation of 3.16 trillion atoms.
Can the networking and storage keep up with 100,000 cards?
The network layer uses scaleFabric-class IB-native RDMA interconnects — a high-speed networking technology that lets massive numbers of accelerator cards communicate directly — solving the data-transfer bottleneck across 100,000 cards.
Storage is handled by ParaStor distributed storage, which ranked first in both the full-system and 10-node production categories on the 2026 global IO500 list.
Cooling relies on liquid-cooling technology supporting megawatt-class power density, using domestically sourced coolant and year-round natural cooling to boost energy efficiency.
From 30,000 cards to 100,000 — how long did it take?
Timeline: system R&D completed in 2024; 30,000 cards online for trial runs in February 2025; 60,000 cards deployed in April; full 100,000-card deployment completed in July — five months from first trial to full capacity.
The cluster has completed over 300 application adaptations across more than 20 fields including materials science, electromagnetics, quantum physics, biomedicine, and meteorology, with over 70 10,000-card-scale tests logged.
This reflects a shift: domestic computing has moved past the "does it work?" stage into an ecosystem-validation phase — testing whether real-world applications across industries can actually run on it.
Can this be replicated? What comes next?
Sugon 8000 is built on an open AI computing architecture that supports accelerator cards from multiple vendors and is compatible with mainstream software ecosystems, allowing existing models and applications to migrate quickly.
In plain terms = the system is not locked to one chip supplier. Swapping in another vendor's cards is feasible, which lowers the barrier to replicating this cluster in other cities.
During the launch event, Sugon and the Beijing Academy of Scientific Intelligence announced the start of R&D and construction for a second 100,000-card system. Whether this second cluster ships on schedule — and whether the model can be replicated at more sites — will be the key proof point for China's domestic computing scale-up roadmap.
Content is for reference only, not financial advice.