On September 15, the Earth System Numerical Prediction Center of the China Meteorological Administration and Sugon jointly unveiled the latest application results from Sugon 8000 in Zhengzhou.

MCV, a multi-moment constrained finite volume numerical model developed in-house by the Earth System Numerical Prediction Center, now completes a global 10-day forecast at 5-kilometer resolution in under an hour on the Sugon 8000 AI supercomputing cluster, meeting mainstream international operational timeliness standards and passing quality verification.

The announcement comes barely a month after Sugon 8000 formally went live, a sign of how quickly domestic high-end computing platforms are being put to work on major research workloads.

Global Forecast Resolution Sharpened to 5 Kilometers

MCV is China's homegrown next-generation integrated weather-climate model. Global 5-kilometer resolution means the atmosphere is covered by a grid with roughly 5-kilometer spacing, with wind, temperature, water vapor and other variables simulated in each cell.

The denser the grid, the better the model can identify and describe small- and medium-scale weather processes such as typhoons and heavy rainfall, but the computational load rises accordingly.

Chen Qiying, chief scientist at the Earth System Numerical Prediction Center, said that raising global forecast resolution from 12.5 kilometers to 5 kilometers has strengthened the model's ability to capture rainfall intensity and typhoon intensity. Coarser resolutions could predict typhoon tracks reasonably well but fell short on simulating storm intensity.

Li Xingliang, researcher and deputy director at the center, said the MCV model's winter predictability on Sugon 8000 is now comparable to advanced international levels. The progress also brings the team close to targets set in the national meteorological development plan for the 15th Five-Year Plan period, which calls for 5-kilometer global weather model resolution and usable forecast lead times of nine days.

Run Time Cut From About 21 Hours to Under One Hour

Kilometer-scale global numerical forecasting demands not only greater computing power but also higher standards for communication, memory management and system stability. Simply adding more compute hardware does not necessarily deliver a proportional performance gain.

According to the center's estimates, running a global 5-kilometer forecast on a traditional CPU architecture would take about 21 hours per 10-day forecast cycle, too slow for the timeliness requirements of an operational system.

The project team ported the full MCV model to the Sugon 8000 heterogeneous acceleration platform and optimized device memory management, communication algorithms and related computing techniques, cutting run time to under an hour and improving overall efficiency by more than 20 times.

Jiang Qingu, a senior engineer at the Earth System Numerical Prediction Center, said the joint team continually cross-checked results from the heterogeneous platform against CPU baselines during migration and optimization to ensure the model output stayed consistent and reliable.

Writing 200GB of Data Now Takes 10 Seconds

Weather forecasting involves both processing large volumes of observational data and continuously generating enormous model output, so storage and I/O performance shape overall system efficiency as much as raw compute does.

Bu Jingde, chief engineer for high-end computing at Sugon, said the company optimized the storage architecture of Sugon 8000 for meteorological workloads, cutting the time to write roughly 200GB of data per forecast cycle from six minutes to 10 seconds.

The system also uses tiered hot and cold storage along with data lifecycle management. Sugon says the architecture can support long-term storage and retrieval of hundreds of petabytes of meteorological data. The performance and capacity figures come from the project team and Sugon.

Numerical Simulation and AI on One Platform

Meteorological research is shifting from a reliance on high-precision numerical computation alone toward a fusion model in which numerical simulation and artificial intelligence work together.

Numerical forecasting projects weather changes by solving the physical equations of atmospheric motion, while AI can learn patterns from historical data and help handle incomplete observations and uncertainty in initial conditions. Combining the two approaches holds promise for further gains in forecast accuracy and operational efficiency.

Sugon 8000 uses a converged supercomputing-AI architecture, supporting numerical computation at varying precisions alongside AI model training and inference on the same system. In future, numerical forecasting modules and AI modules can run together on one platform.

Unlike general scientific computing, weather forecasting is an operational system that must run continuously and reliably. During flood season or major events in particular, any interruption in model computation, communication or storage can delay forecast products and disaster warnings, making system reliability a key measure of operational readiness.

Domestic Computing Power Moves From Infrastructure to Research Applications

The results suggest domestic high-end computing platforms are moving beyond computing power infrastructure buildout and into concrete application areas such as meteorology, life sciences and materials research.

Sugon said application tuning proceeded in parallel with the construction of Sugon 8000. Beyond weather forecasting, the platform now serves drug discovery, new materials, large models and embodied intelligence.

The project team has also completed simulation trials at 3-kilometer global resolution. By its calculations, each doubling of computational precision requires roughly eight times the compute. Pushing to 3 kilometers and beyond will require simultaneous gains in per-card performance, node interconnect efficiency and storage capability.