When it comes to weather forecasting, ordinary people only care about whether it will rain tomorrow, while meteorological engineers face a massive scientific computation. To predict the weather for the next 10 days, one cannot merely look at today's temperature, humidity, and wind speed; instead, the global atmospheric state must be fed into a numerical model and projected step by step into the future according to physical equations. The finer the grid, the more weather details can be observed, but the computational workload also expands rapidly.
On September 15, Sugon announced a major achievement: Relying on Sugon 8000, the first entirely domestically produced 100,000-card-level AI supercluster, China's independently developed MCV next-generation atmospheric model has achieved the completion of global 5km resolution and 10-day forecast calculations within 1 hour, meeting the necessary timeliness requirements for operational running. What truly deserves attention here is not just the two numbers, 5km and 1 hour, but the fact that domestic computing power is transitioning from running scientific research programs to supporting national-level operational systems.
What exactly makes the 5km resolution so difficult? To clarify what 5km resolution means, one can imagine the global atmosphere as a giant three-dimensional grid. In each grid cell, states such as temperature, pressure, wind field, water vapor, clouds, and precipitation must be calculated and continuously updated. If the horizontal grid is reduced from 10km to 5km, the number of grids in a single direction will increase by about 4 times. The number of global two-dimensional grids increases by approximately 4 times, but the weather model does not just calculate a single 2D plane; it also superimposes vertical layers, time steps, and various physical processes.
At a 5km resolution, to ensure numerical stability, the time step usually needs to be further shortened. Assuming the spatial resolution is doubled, the horizontal grids increase by 4 times, and the time step is shortened by about half, the dynamic computational workload alone could approach 8 times the original. If communication, physical parameterization, interpolation, data reading and writing, and intermediate result exchange are also considered, the overall computational pressure will continue to rise.
MCV is not simply about switching to a set of GPUs, but rather adopting a new computational approach. The core model applied this time is MCV, which stands for the Multi-Moment Constrained Finite Volume method. It is not an AI weather model trained purely on historical data, but is based on the governing equations of atmospheric motion, simulating how the atmosphere will evolve in the future through numerical discretization and time integration.
To put it simply, the process is roughly as follows: observational data → data assimilation → current initial atmospheric state → physical equations → numerical solution → future weather results.
MCV has completed the GPU heterogeneous transformation of key parts such as dynamics, physics, and interpolation, and has verified the consistency between GPU and CPU results through unit testing, full-model joint debugging, verification and comparison, and 10-day forecast plotting validation. This step is actually much more difficult than simply migrating the program to run on GPUs, because in scientific computing, the biggest fear is not program errors, but programs running without errors while yielding biased results.
So why are traditional CPUs no longer sufficient?
When many people think of supercomputing, their first reaction is to add more CPUs. This was indeed effective in the past, but at the stage of global kilometer-level or even hectometer-level meteorological simulation, simply stacking CPUs encounters several issues.
First, as the number of computing units steadily increases, system power consumption and cooling pressure rise simultaneously.
Second, CPUs excel at complex control and general-purpose computing, but when faced with a massive amount of structurally similar grid operations that can be executed in parallel, GPUs often have the advantage.
Third, a weather model is not a completely independent computation. Adjacent grids require frequent exchange of boundary data, and a massive amount of communication must occur between different computing nodes.
If the computing speed is improved but communication cannot keep up, GPUs will end up waiting for data. Many projects merely strive to boost the peak computing power of GPUs, resulting in high theoretical chip computing power without a synchronous increase in real-world application speed.
The joint optimization of MCV and Sugon 8000 this time precisely did not stop at the hardware replacement level, but instead underwent multiple rounds of adjustments around device memory management, communication algorithms, and I/O paths. The result is a 10% increase in computing resource utilization, a 3.7x acceleration in the communication module algorithm, and an over 30x improvement in I/O performance. Behind these three numbers lies the fact that the system bottleneck is not solely in the computing chips themselves. 240 trillion calculations are less important than getting results within 1 hour. Sugon 8000 is positioned as the first entirely domestically produced 100,000-card-level AI supercluster, adopting a CPU + GPU heterogeneous architecture, supporting multiple precisions from FP64 to INT8, and covering scientific computing, model training, inference, and industrial simulation.
There is an easily overlooked key point here: meteorological numerical forecasting and large model inference have different requirements for computational precision. Large model inference may focus more on the efficiency of low-precision computing such as INT8 and FP16, while some core computations in traditional meteorological numerical models have higher requirements for FP64 or high-precision numerical stability.
If a system is only proficient in low-precision AI, it may not be able to directly undertake high-precision scientific computing. Sugon 8000 emphasizes the integration of supercomputing and AI, aiming to support both high-precision scientific computing and low-precision intelligent computing within the same set of infrastructure. For future meteorological operations, this integration offers even greater potential. For example, using physical models to generate high-quality simulated data, using AI models for bias correction, using machine learning to accelerate certain physical parameterizations, using AI to assist in identifying extreme weather risks, and using data-driven methods to improve data assimilation and post-processing.
Future weather forecasting is highly likely not a choice between physical models and AI, but rather physical equations ensuring reliability while AI handles partial computational acceleration and data analysis, with both working together. What does this breakthrough truly change? The most critical significance of this achievement is not just that China now has a faster weather forecasting program, but that it validates three things.
First, domestically developed independent numerical models can achieve large-scale operation on domestic heterogeneous computing power.
Second, domestically developed ultra-large-scale computing power systems can undertake real national-level scientific computing operations.
Third, the competition in supercomputing has shifted from merely comparing peak chip performance to comparing the overall efficiency of algorithms, compilers, communication, storage, scheduling, and applications.
Just one month after going online, Sugon 8000 supported MCV in completing the operational timeliness validation for global 5km resolution and 10-day forecasting. This demonstrates that after supercomputing infrastructure is built, the ability to quickly undertake real tasks and form sustainable applications is equally an important metric for evaluating system value.
Of course, 5km, 10 days, and 1 hour do not mean that weather forecasting will be error-free from now on, nor does it mean that all extreme weather can be accurately predicted. The weather system itself possesses complexity and uncertainty, and model precision, initial field quality, physical parameterization, and observational data will all affect the final results.
Weather forecasting may seem like just a map, but behind it lies a joint contest of global atmospheric physics, numerical algorithms, supercomputers, and systems engineering. The results delivered by Sugon 8000 this time indicate that domestic computing power is further transforming "capable of computing" into "capable of computing on time, computing stably, and supporting real operational computing."