Sugon 8000 Powers 100,000-Card Cluster
量子位 · wechat · 2026-07-11
中科曙光 announced the official completion of its fully domestically-produced AI supercluster, 曙光8000（登峰）, in Zhengzhou. Connected to the National Supercomputing Internet and the national integrated computing network, it scales to 100,000 domestic accelerator cards.
The article emphasizes that this isn't merely about stacking cards, but rather building a native supercomputing-intelligent computing integrated system. The single architecture handles both FP64 scientific computing and trillion-parameter model training, covering full-precision computing from FP64 down to INT8.
It also highlights full-stack domestic R&D capabilities spanning chips, computing, storage, networking, cooling, and management. Specifically:
- A scaleFabric, IB-like native RDMA high-speed network supports the 100,000-card interconnect.
- ParaStor distributed storage handles data read/write for large models and scientific computing.
- Liquid cooling and free cooling mitigate high power density deployment pressures.
In terms of applications, 曙光8000 is already deployed in protein folding, turbulence simulation, and high-precision DFT simulation, with over 300 adapted applications. Furthermore, 中科曙光 and the Institute for AI Industry Research (AIR) have initiated the development of a second fully domestic 100,000-card integrated supercomputing system.
Related event: China's Domestic 100,000-Card Shuguang 8000 Cluster Goes Live(3 posts)→
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27