It's raining, then? Hallelujah. Big Blue super 'puter sharpens forecasts

As legal storm over weather app location data brews, its new year's resolution is: 3km

IBM's embattled Weather Company subsidiary has said it is building a GPU-tastic supercomputer to model global weather conditions faster and more accurately.

The firm – under fire in the US for allegedly not having the proper permissions to collect app location data – is aiming for 3km resolution and hourly updates for 40 per cent of the Earth's surface.

This is impressive as most weather forecasting operates in 10km resolution and does not provide hyper-local or fast refreshes. While the work would be possible on CPU-only supercomputers, they just would not be nearly as fast or accurate - as our sister publication, The Next Platform, has pointed out.

Outside the US, Europe, Japan and Korea, which have a 3km forecasting resolution and hourly updates, the rest of the world has 12-15km resolution and six- to 12-hour update timing. This restricts the accuracy of local forecasts in the rest of the world.

Big Blue will endevour to accomplish this using a GRAF (Global High-Resolution Atmospheric Forecasting System) with IBM components:

  • 84 x AC922 server nodes, each containing two Power9 CPUs and four Nvidia V100 GPUs
  • 3.5PB of Spectrum Scale parallel access file storage.

The software uses NCAR's global weather Model for Prediction Across Scales (MPAS), which it developed in partnership with the Los Alamos National Laboratory. It also uses IBM's Watson AI.

Weather data comes in from hundreds of thousands of weather stations around the globe, some amateur, as well as airplane temperature data, collected every five seconds during flights, and crowd-sourced smartphone barometric (pressure) data, if users – er – permit.

GRAF should be able to process up to 10TB of weather data a day and will output day-ahead forecasts. It may look at 10-day forecasts in the future.

The increased resolution – hyperlocality – and timing accuracy should permit better recognition of thunderstorms, which can be too small to be captured by a 12-15km resolution model.

A weather model of the Indian monsoon using the best current resolution for the continent (13km) looks like this:


Click to enlarge

A 3km model shows much finer detail and will enable more accurate region and locality forecasts:


Click to enlarge

IBM has said Weather Company customers could benefit by having, for example, minimised airplane flight disruption from storms or turbulence – if they're lucky enough to live close to an airport that's free of rogue drone reports.

The GRAF system could encourage other weather forecasting and modelling organisations and enterprises to start down the GPU trail to add more hyperlocal services.

We might imagine a future step in weather forecasting accuracy could be micro-locality and sub-hour update capability with, for example, tornado modelling and prediction. ®

Biting the hand that feeds IT © 1998–2019