Follow Datanami:
January 9, 2019

IBM’s New Global Weather Forecasting System Runs on GPUs

Anyone who has checked a forecast to decide whether or not to pack an umbrella knows that weather prediction can be a mercurial endeavor. It is a Herculean task: the constant modeling of incredibly complex systems to a high degree of accuracy at a local level within very short spans of time. Now, IBM and The Weather Company (an IBM subsidiary) have announced a new weather model that promises “the most accurate local weather forecasts ever seen worldwide.”

The IBM Global High-Resolution Atmospheric Forecasting System, or GRAF, leverages wind and – temperature data captured every five seconds by airplanes – a previously untapped data source that can lend insight into otherwise poorly-monitored areas of the world – and crowdsourced data from smartphone sensors of users who opt to share that data – another resource that can improve forecasting coverage, though one that has proved controversial. The model also, of course, leverages more traditional datasets, such as data from hundreds of thousands of weather stations (including stations run by amateur weather enthusiasts).

These massive datasets will be fed to GRAF’s IBM Power9-based supercomputer, which is being built now and will go into production over the next few months. The system combines 84 AC922 nodes – equipped with four Nvidia V100 GPUs each – and 3.5 petabytes of IBM Spectrum Scale Storage, capable of processing up to 10 terabytes of weather data every day. IBM estimates that without GPUs, the system would have required 500 servers.

An August 2018 monsoon in India, shown at left by the best current weather model that operates at 13-kilometer resolution. At right, the new IBM Global High-Resolution Atmospheric Forecasting System (GRAF) operates at 3-km resolution and updates 6 to 12 times more often. Credit: IBM

All of this generates an impressively high-resolution forecast of global weather – IBM promises a close to 200 percent improvement in forecasting resolution for most of the world, with the system able to forecast at a 3km resolution for over 40 percent of the globe. GRAF also updates hourly, making it the first commercial weather system able to predict small weather events (such as thunderstorms) at a global level with such quick turnaround. IBM is currently focusing GRAF on day-ahead forecasts and is not yet looking at 10-day forecasts.

GRAF’s high-resolution, quick-turnaround global forecasting marks a change from existing models. Previously, much of the world outside of the U.S., Japan, Korea and Western Europe has relied on coarser 12-15km forecasting, and when those models covered large areas of the globe, they typically only updated every 6-12 hours. Events like thunderstorms were often too fine to be captured by most traditional weather models.

IBM anticipates myriad commercial uses for GRAF. It may be useful, for instance, for aviation companies seeking to reschedule flights or minimize disruption from turbulence. Insurers might be able to prepare more effectively for storm recovery efforts. Power companies might be able to more accurately predict electricity demand in different neighborhoods. Wind energy siting and production might improve thanks to higher-resolution awareness of wind speeds in hard-to-monitor areas. Farmers might be able to better prepare for dramatic shifts in weather like snap freezes, improving crop production. The list goes on, with wide applications across a wide variety of sectors.

The forecasts generated by GRAF will also, of course, have applications for end users. Users of The Weather Channel app, weather.com, the Weather Underground app, and wunderground.com will see improvements in their forecasts thanks to GRAF, says IBM, as will any business utilizing IBM offerings from The Weather Company. Ultimately, the GPU code will also be open source, presumably allowing extensive uses in the research community.

GRAF’s powerful capabilities were made possible not just by IBM’s R&D investments, but also by an open-source collaboration between The Weather Company and the National Center for Atmospheric Research (NCAR). GRAF leverages NCAR’s Model for Prediction Across Scales (MPAS), which it developed in partnership with the Los Alamos National Laboratory. MPAS is NCAR’s latest-generation global weather model.

“Today, weather forecasts around the world are not created equal, so we are changing that,” said Cameron Clayton, general manager of Watson Media and Weather for IBM. “Weather influences what people do day-to-day and is arguably the most important external swing factor in business performance. As extreme weather becomes more common, our new weather system will ensure every person and organization around the world has access to more accurate, more finely-tuned weather forecasts.”

GRAF was announced at CES 2019 in Las Vegas and will be available later this year.

Datanami