China uses 100,000 domestic chips in its Sugon 8000 supercomputer

Summary

The Sugon 8000 supercluster in Zhengzhou integrates approximately 100,000 GPU cards designed and manufactured domestically. This hybrid system is capable of running both high-precision scientific simulations and large-scale artificial intelligence workloads. Sugon claims that every layer of the hardware and software is sourced from Chinese suppliers.

Why it matters

Sugon has been on the US Entity List since 2019. This 100,000-card cluster, built entirely with domestic components, is seen as a concrete step in reducing China's reliance on foreign semiconductors for frontier computing. Serving universities and industry with over 13,000 nodes and 6,000 concurrent tasks, the system sits at the heart of China's National Supercomputer Internet. This shifts the global balance in high-performance computing. The system's hybrid design prioritizes throughput over precision in AI mode. This approach indicates that China is targeting large language model (LLM) training workloads, which dominate the AI roadmaps of the US and Europe.

Highlights

• Liquid immersion cooling in a fluorinated liquid (boiling point approx. 50 °C) manages heat output equivalent to burning 12 tons of coal per hour across 1,500 tons of hardware.
• Every network port operates at 800 Gb/s speed. This allows the 100,000 chips to behave as a single machine. The cluster contains approximately 1,600 km of cabling.
• Sugon reports over 300 optimized applications across more than 30 fields. These include a 10-day global weather forecast with 5 km resolution generated in approximately one hour.

Critical take

Nearly all performance and supply chain claims in the article are attributed to Sugon itself or state-backed Chinese media. No independent benchmarks, third-party audits, or peer-reviewed verification are cited. Therefore, the claim of a 'fully domestic' supply chain and the 800 Gb/s connection speed figures remain unverified. The article also interprets the phrase 'designed and manufactured in China' as competitiveness against leading US accelerators. However, the source text never makes this inference.