Today for AI
HOT RADAR
computeHEAT 10.5°

AI CURATED HOT STORY · 11/15/2026

Sugon scaleFabric Expands Single-Subnet Interconnect to 114,000 GPUs

1 reports archived1 independent sourcesupdated 11/15/2026, 14:17:25
Synthesis & Latest Updates
1 Sources Cross-Validated

At the 8th China Supercomputing Conference, Sugon chief engineer Wan Wei argued that high-speed interconnect is the key to breaking through cluster-scale bottlenecks, requiring system-level innovation for efficient collaboration at larger scales. Sugon's fully self-developed scaleFabric 400G lossless network supports up to 114,000 GPUs per subnet, 2.33x that of traditional InfiniBand, cuts networking costs by about 30%, and already underpins a 30,000-GPU intelligent computing cluster at the Zhengzhou core node.

LATEST/At the 8th China Supercomputing Conference, Sugon chief engineer Wan Wei argued that high-speed interconnect is the key to breaking through cluster-scale bottlenecks, requiring system-level innovation for efficient collaboration at larger scales. Sugon's fully self-developed scaleFabric 400G lossless network supports up to 114,000 GPUs per subnet, 2.33x that of traditional InfiniBand, cuts networking costs by about 30%, and already underpins a 30,000-GPU intelligent computing cluster at the Zhengzhou core node.

Hourly heat curve

5 fully observed hours
Now
10.5
Peak
11.516:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first
  1. IT之家 · IT业界T2·72 pts

    Sugon scaleFabric Expands Single-Subnet Interconnect to 114,000 GPUs

    • Single-subnet interconnect scales to 114,000 GPUs, 2.33x traditional InfiniBand, with roughly 30% lower total networking cost
    • 400G NIC delivers 0.9-microsecond end-to-end latency; switches offer 800Gbps per port, ~260ns switching latency and sub-1ms link failover
    • Already deployed at the National Supercomputing Internet Zhengzhou core node, supporting three 10,000-GPU scaleX clusters totaling 30,000 GPUs