IBM announced the release of IBM GRAF, which stands for the Global High-Resolution Atmospheric Forecasting System, a global-scale, high-resolution model for weather forecasting, in collaboration with the National Center for Atmospheric Research (NCAR).
IBM touts GRAF as the first “operational high-resolution, hourly-updating model that covers the entire globe,” in contrast to regional models which may update on hourly cadences but cover only North America, parts of Western Europe, or Japan, or global models such as GFS, which are run every six hours. GRAF provides forecasts “down to 3 kilometers (1.9 miles),” whereas competing models cover “10-15 square kilometers (6.2-9.3 miles),” the company claims.
GRAF is run on top of IBM’s POWER9 systems, to which NVIDIA V100 Tensor Core GPUs are connected, as compute accelerators—much in the same way GPGPU compute is used for training artificial intelligence (AI) models. GRAF is built in Fortran, using OpenACC for GPU acceleration, and Message Passing Interface (MPI) used for communication across nodes.
Forecasts derived from this model are available to the general public through IBM-owned weather services, including Weather.com, the digital operations of The Weather Channel, that IBM purchased in 2016, as well as Weather Underground, which IBM acquired as part of that sale.
“What you’re going to see is improvements in the forecast, particularly in areas of the world that haven’t had this level of granularity, or this resolution in terms of updating as frequently. That will be transparent to [consumers],” said Kevin Petty, director of science and forecast operations and public-private partnerships at The Weather Company. “On the enterprise side, the offerings that we provide to the aviation, energy, or agricultural industries, we’re going to be able to provide them with enhanced forecasts for their operations, whether that is via an API or other means, as a service.”
While GRAF represents a significant advance in forecast modeling, this is not sufficient to counteract potentially negative effects that accompany the use of 24 GHz frequencies in commercial 5G mobile network deployments, which has resulted in objections being raised by NASA, NOAA, and the US Navy, alongside lawmakers, seeking to prevent the FCC from opening that frequency range to commercial use.
“Any numerical model around the world, whether it’s IBM GRAF, GFS, ECMWF, etc., there is a strong dependence on good observations—a combination of space based, terrestrial observations from the ground, and observations from the air, such as those coming from aircraft,” Petty said. “Having good observations is necessary for any forecast system.”