Sugon Unleashes scaleX Supercluster: China's First True 10,000+ Card AI Beast with World-First 640-Card Per Cabinet Density
Category: Tool Dynamics
Excerpt:
On December 18, 2025, Sugon (Dawning Information Industry) unveiled the scaleX supercluster at the HAIC2025 conference in Kunshan — China's first physical showcase of a massive-scale AI computing system. Built from 16 scaleX640 ultra-nodes interconnected via proprietary scaleFabric network, it deploys over 10,240 AI accelerator cards for total compute exceeding 5 EFlops. Breakthroughs include the world's first single-cabinet 640-card node with immersion phase-change liquid cooling, 20x density boost, native RDMA 400Gb/s networking with sub-1μs latency, and tiered data optimizations lifting accelerator utilization by 55%. Early deployments signal China's push beyond overseas 2027 roadmaps.
🖥️ Sugon's scaleX Supercluster: China's Seismic Jolt to the Global AI Infrastructure Race
The global AI infrastructure race just felt a seismic jolt from the East — and it's wearing a red badge. Sugon's scaleX supercluster isn't another spec sheet flex; it's a full-throttle assault on the bottlenecks plaguing exascale AI training, where density, efficiency, and scalability collide head-on. Debuting in real hardware glory today at HAIC2025, this monster aggregates 16 scaleX640 ultra-nodes into a cohesive beast capable of 10,240+ cards and 5+ EFlops raw power — all while sipping power at PUE 1.04.
Fresh off the scaleX640 node's November launch (which already claimed world's densest cabinet), the full cluster rollout crushes modular limitations, leapfrogging projected overseas milestones like NVIDIA's 2027 NVL576 by years in integrated innovation.
1. Density Revolution: 640 Cards in One Cabinet
Forget sprawling farms — scaleX redefines compute density with ultra-dense blades and immersion phase-change liquid cooling:
- Cooling Tech: Servers submerged in non-conductive fluid → heat boils off via phase change, slashing thermal resistance.
- Architecture: "One-to-two" high-density design → doubles systems per rack without airflow issues.
- Results:
- 20x higher compute density per cabinet.
- 30-40% faster training/inference for MoE trillion-parameter models.
- 55% spike in resource utilization via tiered data transfer optimizations.
2. Networking Nerve Center: Native RDMA at Hyperspeed
Sugon's homegrown scaleFabric outperforms InfiniBand with domestic innovation:
| Metric | Details |
|---|---|
| Hardware | 400G-class domestic RDMA NICs and switches |
| Bandwidth & Latency | 400Gb/s bandwidth + <1μs end-to-end latency |
| Scalability | Seamlessly scales to 100,000+ cards |
| Advantage vs. Traditional IB | 2.33x better performance + 30% lower total network costs |
No more bottleneck bloodbaths: full cluster compute unleashes without choking on data pipelines.
3. Smart Ops That Think Ahead
The cluster’s brain combines digital-twin tech and intelligent orchestration:
- Digital Twin: Physical-to-virtual mirroring → visualized fault hunting and auto-recovery.
- Intelligent Scheduling: Manages 10,000+ nodes for 100,000+ users → fires millions of jobs per second.
- Reliability & Compatibility:
- 99.99% uptime SLA.
- Multi-brand accelerator support.
- 400+ optimized mainstream/world models out the gate.
4. Real-World Rampage Already Underway
scaleX’s 算力 (computing power) is already delivering impact across key sectors:
✅ Big Model Training: Trillion-param models converge faster (financial risk models, geo-exploration sims cut weeks to days).
✅ Scientific Intelligence: AI4Science workflows ingest petabytes → breakthroughs in energy prospecting and climate modeling.
✅ Eco Wins: 30% power savings vs. air-cooled rivals → aligns with China's green compute mandates.
5. The Open-Source Undercurrent
Built on Sugon's AI open computing architecture (launched September 2025 with 20+ partners), scaleX embraces:
- Multi-vendor card support.
- Mainstream ecosystem integration.
- A deliberate counter to closed silos → democratizes access while fortifying domestic sovereignty.
6. Competitive Earthquake
This release is a game-changer for the global race:
- While global giants roadmap denser racks for 2027, Sugon’s live cluster already delivers.
- Lower TCO (Total Cost of Ownership) + greener footprints vs. competitors.
- China’s statement: System-level mastery, not single-chip supremacy, wins the exascale war.
Sugon's scaleX supercluster isn't incremental — it's the inflection point where China's AI infrastructure goes from catching up to carving the path. By fusing unprecedented density, blistering interconnects, and intelligent orchestration into a scalable, open beast, it redefines what's possible for trillion-param frontiers and beyond.
As these clusters proliferate, expect a cascade: faster native model breakthroughs, greener mega-datacenters, and a recalibrated global balance where Eastern innovation sets the pace. The supercluster era? It's here — and it's scaled in China.
