High-resolution and high-speed weather forecast system using GPU

Weather forecasting technology continues to evolve, such as AI development company DeepMind developing an AI capable of forecasting accurate weather at the level of a weather caster. In the midst of this, NCAR, the American Atmospheric Research Center, has developed FastEddy, a system that can forecast local weather more granular than the existing weather using GPUs.

Recently, drone delivery technology has been developed, and medical supplies delivery drones are active in Africa and other regions. According to NCAR, this drone is strongly influenced by wind, so to make the drone active, high-resolution forecasting of the weather, such as wind that changes depending on buildings and terrain, is required.

Existing weather forecasting technology is a relatively high-resolution method, and numerical forecasting is carried out using a 3-4km area, so the location of buildings and delicate land undulations are not taken into account, and weather forecasts for unmanned aerial vehicles do not have sufficient resolution. In addition, NCAR also has a weather model called WRF-LES that enables high-resolution weather forecasting, but the use of WRF-LES requires large-scale computer facilities and a long time to produce forecasts. Therefore, NCAR set out to develop a model that provides higher resolution and faster forecasting than existing weather forecasting models.

As a result, NCAR succeeded in developing a weather forecast model Fast Eddy using a 5m square. Although the CPU is used to calculate the existing weather forecast model, the calculation is performed using the GPU used for calculations such as Fast Eddy games. Therefore, Fast Eddy is said to be able to predict the weather 8 times faster with the same power as a weather forecast model using a CPU.

According to NCAR, the battery consumption speed of a drone can triple depending on the urban wind flow. NCAR is also looking forward to future enhancements to FastEdit so that it can also simulate sound augmentation methods in the future. Related information can be found here.