SourceIT之家•Stalecollected in 17m
China Deploys Its First 100,000-GPU AI Supercluster
#ai-supercluster#domestic-chips#liquid-cooling曙光8000（登峰）曙光8000中科曙光國家超算互聯網鄭州核心節點
A fully domestic 100,000-card cluster is now running AI and scientific workloads at unprecedented scale.
30-Second TL;DR
What Changed
The supercluster is China's first fully domestically produced 100,000-card AI deployment.
Why It Matters
The deployment strengthens China's domestic AI infrastructure and could expand access to large-scale training, inference, and scientific computing resources. Its mixed-precision design may also help organizations run AI and high-performance computing workloads on a shared platform.
What To Do Next
Evaluate whether your training or scientific-computing pipeline can use FP64-to-INT8 mixed precision before seeking access to a 100,000-card domestic cluster.
Who should care:Researchers & Academics
Key Points
- •The supercluster is China's first fully domestically produced 100,000-card AI deployment.
- •Peak computing performance is described as equivalent to humanity computing continuously for 200 years.
- •It supports new materials, drug discovery, AI, industrial simulation, and other workloads across 26 fields.
- •曙光8000（登峰） supports precision levels from FP64 to INT8 through a hybrid supercomputing and AI architecture.
- •Immersion phase-change liquid cooling enables megawatt-scale power density per rack and improved energy efficiency.
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •The supercluster utilizes the Sugon (曙光) 'Dengfeng' architecture, which integrates proprietary high-speed interconnects to mitigate the performance bottlenecks typically associated with domestic GPU clusters.
- •The deployment is part of the 'East Data, West Computing' (东数西算) national strategy, specifically designed to alleviate computational resource imbalances between China's coastal and inland regions.
- •The system employs a unified software stack that supports mainstream domestic AI frameworks like MindSpore and PaddlePaddle, ensuring compatibility with the broader Chinese AI ecosystem.
- •The Zhengzhou node serves as a critical hub for the National Supercomputing Internet, acting as a centralized platform to aggregate and distribute computing power to research institutions and commercial enterprises.
- •The project marks a significant shift in China's semiconductor supply chain, demonstrating the viability of large-scale AI training using exclusively domestic silicon, reducing reliance on restricted high-end Western GPUs.
Competitor Analysis
Primary Focus
- Sugon 8000 (Dengfeng)
- Domestic Sovereignty/Hybrid
- NVIDIA DGX SuperPOD
- Global AI Performance
- Cerebras CS-3 Cluster
- Wafer-Scale AI Training
Interconnect
- Sugon 8000 (Dengfeng)
- Proprietary Domestic Fabric
- NVIDIA DGX SuperPOD
- NVLink/InfiniBand
- Cerebras CS-3 Cluster
- SwarmX Fabric
Ecosystem
- Sugon 8000 (Dengfeng)
- MindSpore/PaddlePaddle
- NVIDIA DGX SuperPOD
- CUDA/PyTorch/JAX
- Cerebras CS-3 Cluster
- PyTorch/TensorFlow
Availability
- Sugon 8000 (Dengfeng)
- China-restricted
- NVIDIA DGX SuperPOD
- Global
- Cerebras CS-3 Cluster
- Global
| Feature | Sugon 8000 (Dengfeng) | NVIDIA DGX SuperPOD | Cerebras CS-3 Cluster |
|---|---|---|---|
| Primary Focus | Domestic Sovereignty/Hybrid | Global AI Performance | Wafer-Scale AI Training |
| Interconnect | Proprietary Domestic Fabric | NVLink/InfiniBand | SwarmX Fabric |
| Ecosystem | MindSpore/PaddlePaddle | CUDA/PyTorch/JAX | PyTorch/TensorFlow |
| Availability | China-restricted | Global | Global |
Technical Deep Dive
- Architecture: Hybrid supercomputing and AI cluster utilizing Sugon 8000 series nodes.
- Interconnect: High-bandwidth, low-latency domestic interconnect fabric designed to scale to 100,000+ nodes without significant throughput degradation.
- Precision Support: Native hardware support for FP64 (scientific computing) and INT8/FP8 (AI inference/training) to maximize utilization across diverse workloads.
- Cooling: Advanced immersion phase-change liquid cooling system, achieving a Power Usage Effectiveness (PUE) significantly lower than traditional air-cooled data centers.
- Scalability: Modular rack design allowing for megawatt-scale power density, facilitating high-density GPU packing.
Future ImplicationsAI analysis grounded in cited sources
Domestic AI training costs will decrease by 20-30% for Chinese firms.
The transition to a state-subsidized, domestically produced supercomputing infrastructure reduces dependence on expensive, supply-constrained imported hardware.
China will increase its share of global AI research publications.
Providing researchers with massive, accessible domestic compute resources removes the bottleneck of limited access to high-end GPU clusters.
Timeline
2021-05
China officially launches the 'East Data, West Computing' national strategy.
2023-04
National Supercomputing Internet platform officially begins trial operations.
2025-02
Sugon announces the next-generation 'Dengfeng' architecture for large-scale AI.
2026-07
Completion and commissioning of the 100,000-card supercluster at the Zhengzhou node.
- 2021-05China officially launches the 'East Data, West Computing' national strategy.
- 2023-04National Supercomputing Internet platform officially begins trial operations.
- 2025-02Sugon announces the next-generation 'Dengfeng' architecture for large-scale AI.
- 2026-07Completion and commissioning of the 100,000-card supercluster at the Zhengzhou node.
Weekly AI Recap
Read this week's curated digest of top AI events →
Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: IT之家 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.