Domestic 100,000-Card AI Supercluster Debuts
新智元 · wechat · 2026-07-11
The article introduces 中科曙光's newly completed first fully domestic 100,000-card AI supercluster, "曙光8000（登峰）". It champions a "native supercomputing-intelligent computing integration" approach: merging high-precision scientific computing with low-precision AI training/inference into a single system, rather than simply stitching together separate nodes.
It emphasizes that scaling from 10,000 to 100,000 cards isn't a mere expansion, but a systemic re-engineering of chips, networking, storage, and cooling. To tackle efficiency and reliability challenges at massive scales, 曙光8000 utilizes full-stack in-house R&D, an IB-like RDMA high-speed network, distributed storage, and immersion phase-change liquid cooling.
The system has already optimized over 300 key applications across more than 20 scientific and industrial scenarios, including large models, robotics, and innovative drugs. Now connected to the National Supercomputing Internet, it is open to research institutions, enterprises, and individual developers, highlighting a shift from "building well" to "using well."
Related event: China's Domestic 100,000-Card Shuguang 8000 Cluster Goes Live(3 posts)→
More from Infra
- 12 KV Cache Reduction Techniques Every AI Engineer Should Understand, Explained — blaizedsouza · 2026-09-11
- The shadow GPU capacity market is formalizing, with Meta selling excess compute to outside buyers — DavidLinthicum · 2026-09-11
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11