Product Launchllm traininghigh performance computingsugon400g rdma
Sugon Unveils 10,000-GPU Supercluster for Trillion-Parameter Models
9.0
On December 18, 2025 at HAIC 2025 in Kunshan, Sugon unveiled the scaleX 10,000‑GPU supercluster, a China-made system designed for trillion-parameter foundation models and scientific AI. The cluster comprises 10,240 accelerator cards across 640‑GPU cabinets, delivers over 5 EFLOPS, employs a 400G native RDMA network and achieves PUE 1.04. It already supports over 400 large models and targets model training and industry workloads.
Key Points
- 1Deploys a 10,000‑GPU supercluster with 10,240 accelerators delivering over 5 EFLOPS and 640‑GPU cabinets
- 2Achieves efficiency via 400G native RDMA, sub‑1µs latency, PUE 1.04, and 55% higher accelerator utilization
- 3Enables practitioners to train trillion‑parameter models and supports 400+ leading large models and applications
Scoring Rationale
Major official infrastructure launch with broad industry impact, but limited immediate access details constrain practical adoption.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
