How the Sunway 8000 Achieves Full‑Precision, End‑to‑End, Fully Domestic AI Supercomputing

In July 2026, the Sunway 8000 (also called "Dengfeng") was unveiled at the Shanghai World Expo as China’s first fully domestic 100,000‑GPU AI super‑cluster, earning the official "treasure" title at the World AI Conference (WAIC) 2026.

1. Native Super‑Intelligent Fusion: Redefining the Compute Base

The system abandons the traditional split between high‑precision scientific computing (FP64) and low‑precision AI training (BF16/FP8). Instead, it adopts a native super‑intelligent fusion architecture that integrates chips, hardware clusters, and the scheduling platform into a single, cohesive design. This enables a single platform to run high‑precision scientific workloads (quantum chemistry, fluid simulation) and AI large‑model training/inference across the full precision spectrum (FP64/FP32/BF16/TF32/FP8/INT8), achieving the promised “one system, all scenarios” coverage.

2. Full‑Stack Autonomous Technology: Chip‑to‑Network Breakthroughs

The 100k‑GPU cluster’s challenges lie not only in scale but also in deep system coupling. Sunway 8000 addresses this with a fully domestic, end‑to‑end technology stack:

Six self‑developed chips covering general‑purpose processors, AI accelerators, and interconnect switches, forming a complete domestic chip‑to‑system chain.

High‑density compute units : a novel cabinet design with orthogonal architecture, full‑electric interconnect, immersion phase‑change liquid cooling and high‑efficiency power delivery raises per‑cabinet compute density by 20× over conventional solutions.

scaleFabric network : a domestically‑designed IB‑native RDMA solution supporting up to 114,000 cards per subnet, end‑to‑end latency < 1 µs, switch latency 260 ns, 800 Gb/s per port and 64 Tb/s total capacity, with millisecond‑level fault self‑healing.

Phase‑change immersion liquid cooling : achieves PUE < 1.04, supports megawatt‑level cabinet power, and enables zero‑water consumption heat recovery.

ParaStor distributed storage : ranked #1 globally on the IO500 list for both production‑node and 10‑node categories, matching the massive data throughput of the cluster.

Gridview 7.0 management platform : combines MetaStack and Kubernetes for dual‑intelligence scheduling, unifying scientific simulation and AI model training, and alleviating scheduling congestion, operational complexity, and low utilization in ten‑thousand‑GPU clusters.

3. Application Validation: From Gordon‑Bell‑Level Simulations to Industry‑Scale Models

Over 300 key applications have been deeply optimized on Sunway 8000, with 15 reaching Gordon‑Bell‑class super‑large‑scale simulation standards. Domains include weather forecasting, quantum computing, energy, biology, new materials, and fluid dynamics, covering more than 20 scientific and industrial tracks. Notable scientific results include high‑precision calculations of 152 FeMoco spin‑orbit clusters, a 3.16 trillion‑atom DFT simulation, and a 3.28 trillion‑cell turbulent flow direct simulation. On the AI side, the system has been tuned for major large‑model families across language, speech, multimodal, and autonomous driving scenarios.

The OneScience scientific model platform further lowers the barrier to massive compute: researchers describe requirements in natural language, and the platform automatically schedules resources, generates code, and submits jobs, already serving more than 20 leading research institutions.

4. Supercomputing Internet: Turning 100k‑GPU Power into Public Infrastructure

Sunway 8000 is not confined to a single data center; it connects to the national supercomputing internet, forming the largest domestic single‑vendor AI compute pool. The network now spans 14 provinces and cities, linking over 30 super‑AI centers, aggregating >350 million CPU cores and >250 k GPU cards, with >1.4 million registered users and 2.4 billion completed jobs.

Three ecosystem co‑creation programs accelerate adoption: the “100k‑GPU Co‑Creator Incentive” for academic and startup teams, the “Intelligent Agent Ecosystem” leveraging 276 agent assets for multi‑agent collaboration, and the “Core Partner Recruitment” offering optimization, marketplace, and channel support to the full AI‑hardware supply chain.

5. Outlook

From six self‑designed chips to the scaleFabric network, sub‑1.04 PUE liquid cooling, IO500‑topping storage, Gridview 7.0 scheduling, and the OneScience platform, Sunway 8000 illustrates China’s shift from isolated breakthroughs to a fully autonomous, full‑stack AI compute ecosystem. Under the spotlight of WAIC 2026, the system not only showcases technical prowess but also marks a pivotal transition for China in the global AI compute race.