Fearless Forecast: Two Days of Weather Predicted in Just 20 Minutes, Down to a Square Kilometer, in SC18 Demo

Demo shows what's possible with the prediction of weather, down to forecasts for microclimates, when running on a single NVIDIA DGX-2 supercomputer.
by Scott Martin

For a peek at what the future of supercomputer-powered weather prediction could look like, head into the sprawling Dallas convention center where the SC18 supercomputing show has come to life.

A rich map of a million square kilometers of Switzerland and its neighbors — demarcated with high- and low-pressure areas and varying winds — shimmers on a large 4K screen.

The map is the product of a GPU-accelerated COSMO weather prediction model. It offers resolution down to one square kilometer as well as 80 altitude layers — from the ground to as high as 20 kilometers.

The model can predict two days of weather in just 20 minutes running on a single NVIDIA DGX-2 supercomputer. That’s computing muscle equivalent to 75 dual-socket CPU nodes.

The demo was initially shown on the big screen Monday by NVIDIA founder and CEO Jensen Huang in his pre-show keynote to a crowd of more than 700 conference attendees.

It depicts the current state of the atmosphere and predicts its future state by solving equations for fluid flow and atmospheric physical processes. The model’s ultra-fine level of resolution enables the prediction of weather down to microclimates, which has enormous applications for those engaged in transportation, agriculture, aviation and even simple outdoor recreation activities like hiking.

“We’re looking at a future prediction of microclimates,” Huang said.

Behind the scenes, the GPU-accelerated COSMO model was running on a domain from the Swiss Federal Office of Meteorology and Climatology, known as MeteoSwiss, with enormous computing requirements. A 24-hour forecast created using the COSMO-1 model requires more than 4 quadrillion computing operations. COSMO-1 forecasts are initiated eight times a day, every three hours.

The demo’s visualization is further supported by using ParaView, an open-source visualization tool recently enhanced with ray-tracing capabilities using NVIDIA OptiX to take advantage of the latest NVIDIA Turing GPUs, including the Quadro RTX 6000.