New weather forecasting system predicts 10-day outlook at 3.1-mile resolution

A new weather forecasting model can now predict conditions up to 10 days ahead at a resolution of 3.1 miles (5 kilometers), outdoing most of the world's leading forecast systems. According to a joint announcement from the China Meteorological Administration and Beijing-based computing company Sugon, the system runs on Sugon 8000, a large-scale compute cluster powered entirely by domestically produced AI chips. Most major global forecasting systems operate at resolutions of roughly 3.1 to 6.2 miles (5 to 10 kilometers).

By comparison, the U.S. National Weather Service's flagship Global Forecast System tops out at about 8.1 miles (13 kilometers) in resolution.

System Performance and Compute Capacity

Sugon said the team has already tested runs at a finer global resolution of 1.9 miles (3 kilometers), bringing the work closer to targets set out in a recent five-year plan for meteorology. The goal is to achieve nationwide 0.6-mile (1-kilometer) forecasts by 2030, with resolution reaching 328 feet (100 meters) in critical regions. Numerical weather models work by dividing the atmosphere into three-dimensional grids and then calculating how factors such as wind, temperature, pressure and humidity change within them.

Smaller grid cells reveal weather systems in greater detail, but the computational cost is steep — halving the distance between grid points can require roughly eight times more computing power.

The forecast model, called Mesoscale Convective Vortex, has been specifically tuned for Sugon 8000. The cluster was unveiled in July and links 100,000 powerful compute units that can operate in parallel. According to the company, Sugon 8000 is distinguished by its ability to handle fundamentally different kinds of computation on the same system. Traditional supercomputers are built for the high-precision arithmetic needed for scientific simulations, while AI systems typically trade precision for the speed needed to process large volumes of data.

Sugon 8000 is designed to meet both needs simultaneously, and is built entirely on domestically developed chips, storage and networking technology.

Broad Applications and Technical Advantages

In recent tests the system produced a full 10-day global forecast in roughly an hour. The turnaround itself is not especially fast — the U.S. Global Forecast System completed a comparable 10-day run in about 45 minutes during the same period. Still, Sugon said its one-hour turnaround is sufficient for routine forecasting needs. The speed can give forecasters more time to evaluate the development of approaching typhoons, heavy rain and other severe weather. Beyond meteorology, Sugon 8000 is being applied to large-scale simulations in protein research, molecular dynamics and turbulence modeling.

The company said the system is now compatible with nearly 200 applications and more than 500 mainstream AI models, spanning over 30 research and industrial fields.