Gordon Bell Prize returns to China: USTC quantum chemistry win runs on all-domestic Sugon 8000 100k-GPU cluster
量子位 · wechat · 2026-11-21
At SC 2026, a USTC-led quantum chemistry project on the FeMo cofactor won the Gordon Bell Prize, China's first in five years, while CAS Tianji's SD3 1.5 km-resolution global weather simulation was shortlisted for the climate modeling prize — a first for a Chinese company team. Both ran on Sugon 8000, an all-domestic 100k-GPU AI supercluster put into production only in July.
The two workloads demand opposite precision profiles: the chemistry application uses neural-network quantum states, juggling low-precision matrix throughput with high-precision solvers, while the weather model requires FP64 as a floor with 99%+ of compute on accelerators. Sugon 8000 answers with a heterogeneous CPU+GPU architecture natively optimized for mixed precision — heterogeneous nodes deliver 3x the performance of CPU-only nodes, BF16 weak scaling sustains 94.9% efficiency up to 45,000 GPUs, and the prize-winning run used only 10k GPUs to match a full CPU system.
Over 300 key applications across 30+ domains have been optimized on the cluster. The article argues Chinese HPC's next edge lies in reproducible scientific results and heterogeneous 'AI+HPC' system design rather than peak benchmark scores.
More from Infra
- Inkling 2 Mini quantized to 5.4GB/8.9GB, keeping 91-97% performance — teortaxesTex · 2026-11-21
- Open-weight models are less efficient, so their rise may increase chip sales — aronchick · 2026-11-21
- Score centering, rumored OpenAI RL trick, tolerates staleness and may rescue slow chips — teortaxesTex · 2026-11-21
- $31/month quad-channel EPYC + 128GB runs 3-bit Qwen at 128k context — Ne00n · 2026-11-21
- Ternary Bonsai-2 vs Qwen3.5-2B on A13/A14/M3: 32-55% faster decode, and why thermal logging matters — leo-k7v · 2026-11-21
- Essay 'The Organic Web' weaves 62 citations to dissect data's physical weight — round · 2026-11-21