Session II: Mixed-precision for Weather and Climate Predictions

Weather and climate forecasts exist as time constrained problems where predictions must obviously exceed the bare minimum of one forecast day per day. Existing weather forecasts operate at resolutions of ~10km globally, meaning many aspects of physics are either absent or are parameterized. By 1km many of these processes will be resolved, making those parameterizations redundant. Operating at these increased resolutions will require computational gains from a variety of methods, including reduced numerical precision. Double-precision is widely used in the field currently, despite the general acceptance of model inaccuracies and uncertainties. In our work we probe the possibly for using precisions lower than double while still producing accurate forecasts. By emulating reduced precision, we are able to rapidly investigate precision across a range of kernels of weather and climate forecasting. For the vast majority of calculations, half-precision produces accurate forecasts once model uncertainty has been appropriately considered.

Download Presentation

Location: Grand Ballroom C Date: March 27, 2019 Time: 11:10 am - 11:30 am Matthew Chantry, U. of Oxford