H3C's Full-Stack Super-Node: 128 GPUs/Cabinet, 96% Net Utilization, 30-Day Stable Training

Introduction

At the GACS 2026 Super-Node Technology Forum, H3C Cloud & Compute Storage Product Line Solution Department General Manager Wang Feng presented "Full-Stack Collaborative Architecture Practice for Ultra-Large-Scale Super-Node Systems" . H3C aims to build a next-generation super-node system that is ultra-large-scale,极致效率, secure and trusted, and business-continuous to address four critical challenges in the agent era where token consumption grows exponentially:

Difficulty scaling cluster size

Low compute resource utilization

Insufficient hardware security protection

Slow operations and fault diagnosis

The solution spans six domains: compute, network, storage, cloud, security, and operations.

Compute: Ultra-High-Density + Full Liquid Cooling

Ultra-High-Density Compute Architecture

H3C shrinks motherboard volume by 25% , removes local PSUs and fans, and uses next-gen liquid cooling with integrated components to minimize space. The cabinet adopts 1U 38.4 kW high-density power supply (first in China) with native 800V HVDC support. This enables 128 high-performance GPUs per cabinet and supports future higher density evolution.

Full Liquid-Cooled (Fanless) System

All heat-generating components on the motherboard are liquid-cooled; cold plates use higher-density flow channels improving heat dissipation by 10%+ . Cabinet-level centralized PSU and busbar are also fully liquid-cooled. H3C has mass-produced a fanless full liquid-cooled whole-system . Next step: phase-change cold plates targeting 60%+ further cooling improvement .

Network: Optical-Electrical Parallel + Compute-Network Co-Optimization

Optical-Electrical Parallel Evolution

Intra-cabinet SerDes reaches 224G ; electrical signaling on PCB is distance-limited, so H3C uses high-speed material selection and high-density manufacturing for reliable 224G transmission. Inter-cabinet, optical is the superior choice ; CPO technology improves distance and bandwidth density but brings manufacturing precision, cooling, and maintenance challenges. H3C does not simply "replace electrical with optical" but engineers both to extremes for high-density reliable interconnect.

Compute-Network Multi-Dimensional Tuning

Untuned physical bandwidth utilization in Scale-out is only ~ 70% . H3C's network scheduling raised utilization to 92.5%+ in one super-node cluster and 96%+ GPU bandwidth utilization in an internet customer project. Future work: millisecond micro-burst detection, rapid network state notification, congestion control.

Unified Scale-up/Scale-out Architecture

A three-layer fused architecture: "Level-1 Scale-up + Level-2 Scale-out + Scale-Across" achieves seamless interconnect from chip to cabinet to cluster.

Storage: Eliminating Compute Wait for Data

High-Performance Distributed Storage Engine

Full zero-copy, intelligent data read, multi-node bandwidth aggregation deliver 200 GB/s per node bandwidth, 3 million IOPS, <100 μs latency , ranking #1 on global benchmarks.

KV Cache Inference Acceleration (XCache)

Long contexts and high concurrency exceed GPU VRAM. H3C's "storage-for-compute" XCache builds a multi-tier cache: shared memory → CXL memory pool → external storage, expanding cache capacity and dramatically improving long-context response speed.

Cloud: Making Compute Available On-Demand

Intelligent Compute Fusion Architecture ("One Cloud, Multiple Chips")

Breaks heterogeneous compute barriers via unified management of diverse chips, affinity-aware topology scheduling, fine-grained compute partitioning, and global resource pooling. Elastic scaling and high fault tolerance enable 30+ days stable training at 100k-node scale .

Intelligent Compute Acceleration Engine

Core: communication-domain topology-aware optimization ( cloud-compute-network collaboration + low-latency collective communication library ). Network topology awareness and path navigation enable global load balancing; communication operator offload and compute-communication overlap reduce compute overhead; second-level fault detection and minute-level self-healing ensure reliability. Results: 50% reduction in cross-node communication time , breaking traditional AI server communication limits.

Second core: operator fusion and memory pooling. Heterogeneous operator fusion and multi-granularity parallelism yield 20-50%+ operator performance gains . Heterogeneous memory pooling (GPU VRAM + CXL global memory) breaks single-card VRAM bottleneck, achieving 90%+ memory utilization and 10%+ inference efficiency improvement .

Security: Three-Layer End-to-End Protection

Three typical threats: compute theft (illegal GPU occupancy), data security risks (tampering, unauthorized access, leaks), missing boundary protection (DDoS, malicious high-risk operations). H3C builds a layered, full-chain security system across compute runtime, model, and token output layers, covering attack-defense detection, data leak prevention, AI security audit. Combined with data security and supply-chain security policies (compliant model training, data classification, server security scanning) to eliminate risks and ensure optimal ROI.

Operations: Unified AIOps Agent for Business Continuity

Pain points: massive scale creates blind spots, alarm storms overwhelm handlers, risks hard to predict. H3C's Unified Operations Intelligent Agent integrates 50 million real operations cases with AI agent technology. Multi-agent collaboration (troubleshooting, inspection, optimization) covers the full "plan, build, maintain, optimize" lifecycle. Outcomes: 60% deployment time reduction, 90% duplicate alarm compression, 90% fault diagnosis efficiency improvement .

Product: UniPoD S80000 Series

The UniPoD S80000 series spans entry to ultra-large scale. 1024 super-nodes scale via Scale-Up to 16,384 GPUs , also supporting 73,728 GPU ultra-large clusters . The 256 super-node (64 GPUs/cabinet, 256 GPU full interconnect) delivers 4x cluster interconnect bandwidth, 2x single-card inference performance, 1.5x single-card training performance vs traditional AI servers.

Open Ecosystem: Fully Compatible with Domestic Ecosystem

H3C adheres to "technology without boundaries, choice without limits." Open decoupled design via three ecosystems: heterogeneous compute, compute interconnect, and fusion software. Supports smooth multi-vendor GPU onboarding, deep compute-network integration, delivering optimal super-node solutions.

Conclusion

Wang Feng emphasized: "Future super-node competition is not about single products or technologies, but a comprehensive contest of full-stack capabilities and system efficiency." H3C will continue open, compatible design, leveraging full-stack strength to continuously optimize super-node systems, helping users achieve the best token price-performance ratio.