Heterogeneous Computing Powers AI for Science: Sugon 8000's 10 Frontier Breakthroughs

Introduction

The 2026 CCF National High Performance Computing Academic Conference (CCF HPC 2026) held in Shenzhen highlighted how "super-intelligence fusion" is reshaping scientific research. Frontier topics such as meteorological and ocean forecasting, cosmic astronomy simulation, trillion-atom molecular dynamics, and aerospace multi-physics computing are no longer single high-precision simulation tasks but composite workloads intertwining mechanism solving with AI training and inference . Traditional architectures struggle with these mixed workloads. Sugon 8000 (Dengfeng), a 100,000-card domestic AI supercluster, adopts a native heterogeneous super-intelligence fusion approach . The conference released ten frontier application results — benchmark scientific cases under real loads — that directly validate the core advantages of heterogeneous architecture in the AI4S era.

Ten Frontier Applications on Sugon 8000

Sugon 8000 is designed from the bottom up with a CPU + domestic accelerator card heterogeneous computing approach . Chips, hardware, high-speed network, parallel storage, and unified scheduling system are integrated and coordinated so that a single compute foundation simultaneously carries high-precision scientific simulation and AI training/inference . The ten applications cover cosmological simulation, weather forecasting, ocean simulation, molecular dynamics, quantum-supercomputing collaborative biological simulation, digital twin brain, aerospace fluid simulation, deep geophysics, industrial thermal simulation, and more. These topics are extremely difficult; under traditional architectures they either cannot reach target scale or take too long to be practical. Sugon 8000's native heterogeneous foundation and full-stack hardware-software synergy successfully support these world-class scientific practices.

1. Next-Generation Weather-Climate Unified Model MCV

Global high-resolution climate forecasting demands both precision and timeliness. Traditional pure CPU clusters would far exceed the operational time window to complete a global 5 km resolution, 10-day forecast. Sugon 8000's heterogeneous architecture achieves global forecast within 1 hour, meeting international operational timeliness standards, and has completed a global 3 km resolution simulation trial. The dynamical core must maintain FP64 high precision to avoid distortion of atmospheric physical processes, while AI parameterization schemes and post-processing require massive low-precision parallel compute. Co-locating high-precision mechanism solving and AI modules in the same compute pool provides the key computational cornerstone for the continuous evolution of next-generation weather forecasting models.

2. MatPL Machine Learning Force Field: Trillion-Atom Molecular Dynamics Simulation

Large-scale molecular dynamics is a hard problem in materials and life sciences. As simulation scale grows, requirements for parallel scalability, video memory capacity, and communication latency become increasingly stringent. Pure CPU architectures see compute speed drop sharply beyond billion-atom scale. On Sugon 8000, MatPL trillion-atom molecular dynamics simulation achieves 100x speedup over global top-tier results and 1000x speedup over previous Gordon Bell prize-winning solutions. In parallel scaling from single card to large cluster, overall weak scaling efficiency reaches 86.9% with communication overhead only 4.26% , meaning nearly linear utilization of any compute scale to challenge arbitrarily large atomic systems.

3. Quantum-Supercomputing Collaborative Biomolecular Dynamics Simulation

Addressing the challenge of balancing high precision and large scale in biomolecular simulation, the research team leveraged Sugon 8000 and the "Jiuzhang" photonic quantum computing platform to establish a quantum-supercomputing collaborative workflow: HPC generates Random Phase Approximation (RPA) level high-precision quantum chemistry data and performs large-scale molecular dynamics simulation, while quantum computing assists force field learning to achieve more accurate predictions with limited data. In downstream simulation performance tests, system scale reached 52.2 billion atoms , maintaining 89% weak scaling efficiency under 8192 MPI tasks; simultaneously completed a stable 100 picosecond simulation of the SARS-CoV-2 spike protein trimer in aqueous solution, providing computational support for protein research and drug design.

4. Multi-Nozzle Supersonic Jet Ultra-Large-Scale High-Precision Parallel Solver

Ultra-large-scale high-fidelity numerical simulation of rocket plumes is a typical strongly coupled fluid simulation . Such tasks have enormous mesh scale and extremely frequent inter-node data exchange, posing extreme tests for cluster high-speed interconnect and parallel synchronization; large-scale expansion easily hits communication bottlenecks. Sugon 8000, relying on scaleFabric high-speed network and other subsystems, achieved a rocket plume simulation mesh scale of 6.774 trillion grid cells, setting an international record for the largest known simulation of its kind. At nearly 100,000-card scale , weak scaling efficiency reached 99.01% , strong scaling efficiency 73% , and mesh update throughput reached 34.029 trillion cells per second .

5. Smart Breeding Digital Twin Platform "Virtual Maize"

Crop digital twins require coupling environmental simulation, growth mechanisms, and AI prediction — mixed workloads with large load fluctuations demanding high scheduling flexibility. Pure CPU clusters face uneven resource busy/idle under mixed loads. Sugon 8000's Gridview cluster scheduling system automatically assigns different sub-task types to CPUs or accelerators, raising overall effective compute utilization to 70%–90% , dramatically shortening breeding simulation iteration cycles and accelerating superior variety screening.

The ten frontier applications are not single HPC tasks but mechanism simulation + AI coupled composite workflows . Sugon 8000 is not merely hardware stacking of CPUs and accelerators; it achieves fine-grained load splitting and collaborative computing between CPUs and domestic accelerators through scaleFabric native RDMA lossless high-speed network, ParaStor parallel storage, Gridview unified cluster scheduling, and immersion phase-change liquid cooling infrastructure full-stack synergy , enabling support for these world-class complex simulation tasks.

Simultaneously the platform realizes full-chain autonomy and control across core chips, hardware systems, and scheduling software. Heterogeneous architecture inherently offers higher compute density; paired with immersion phase-change liquid cooling, system PUE drops as low as 1.04. At equivalent compute output, compared to pure CPU clusters, total system power consumption and data center footprint are significantly reduced, giving 100,000-card clusters long-term stable operational capability.

Four Bottlenecks of Traditional Architectures

1. Compute Type Mismatch, Low Mixed-Task Efficiency

Frontier scientific research is universally mixed-precision. Fluid, weather, and material mechanism solving rely on FP64 double precision for physical rigor; AI parameterization, model training and inference rely on FP16/BF16/INT8 low-precision compute. CPUs excel at complex branch logic and flow scheduling but have low compute density for massive parallelism. Relying entirely on pure CPUs for trillion-atom simulation or 100,000-card large model training leaves hardware resources largely idle, drastically reducing performance per watt.

2. Physical Cluster Separation Breaks AI Workflows

Traditional construction separates scientific simulation and AI training into two independent clusters. Simulation data must be exported, transferred across clusters, and re-imported into the AI platform; massive intermediate files are repeatedly moved, incurring huge I/O overhead. This not only lengthens research iteration cycles but also risks data version inconsistency and transfer loss, making it difficult to support the core AI4S closed-loop mode of "simulation generates data, AI learns and optimizes, then feeds back to simulation" .

3. Large-Scale Expansion Hits a Clear Ceiling

Pure CPU clusters can only increase compute by continuously stacking nodes. Node count explosion brings interconnect congestion, data center space pressure, and soaring power and cooling costs. At 100,000-card scale and beyond, parallel efficiency decays rapidly. For billion- and trillion-atom ultra-large-scale simulations, pure CPU clusters struggle to complete computation within scientifically acceptable time windows.

4. Low Hardware Resource Utilization, Compute Investment Fails to Translate to Output

Under complex scientific loads, pure CPU clusters typically maintain only 20%–40% effective compute resource utilization . Most resources are consumed by I/O wait, cross-node communication, and task scheduling overhead. Hardware investment is high but effective compute for scientific solving is limited; much compute becomes paper peak performance.

Conclusion: Native Super-Intelligence Fusion Based on Heterogeneous Computing Is the Inevitable Direction for the AI4S Era

Global top-tier compute system construction shows a clear trend: the industry is shifting from physically separate supercomputing and intelligent computing construction toward unified super-intelligence fusion platforms . Compute evaluation standards have moved from pure hardware peak scale to effective compute release capability under real scientific tasks .

Traditional architectures remain viable for single high-precision scientific computing scenarios, but against the new paradigm of deep coupling between AI and physical mechanisms, they have hit their capability boundary. The heterogeneous computing approach represented by Sugon 8000 does not eliminate CPUs; rather it lets CPUs and domestic accelerators each do what they do best: CPUs handle flow scheduling and complex logic; massive parallel computing goes to accelerators, with a single foundation simultaneously carrying traditional HPC and AI tasks.

Sugon 8000 uses real scientific loads to prove that native super-intelligence fusion architecture based on heterogeneous computing holds extreme advantages in supporting major AI4S innovations under new-quality productive forces. Looking ahead, as scientific simulation evolves toward larger scale, higher precision, and more complex loads, domestic heterogeneous compute infrastructure is poised to drive more major original scientific discoveries to fruition.