[10-3] February 28, 17:30-17:50

Improving Weather Forecasts Through Reduced Precision Data Assimilation

Sam Hatfield (University of Oxford), Peter Dueben (ECMWF) and Tim Palmer (University of Oxford)

We present a new approach to improve the efficiency of data assimilation for numerical weather prediction, by trading numerical precision for computational speed. Data assimilation is inherently uncertain due to the use of relatively long assimilation windows, noisy observations and imperfect models. Therefore, errors incurred from using a precision below double precision may be within the tolerance of the system. Lower precision arithmetic is cheaper, and so by reducing precision in ensemble data assimilation, we can redistribute computational resources towards, for example, a larger ensemble size.

We will present results on how lowering numerical precision affects the performance of an ensemble data assimilation system, consisting of the Lorenz '96 toy atmospheric model and the ensemble square root filter. We compare the performance of the system at half, single and double precision. We estimate that half precision assimilation with a larger ensemble can reduce assimilation error by up to 50%, with respect to double precision assimilation with a smaller ensemble, for no extra computational cost. Additionally, we investigate the sensitivity of these results to the assimilation window length. We hope that the results presented here will encourage further investigation into half precision hardware, which is now becoming available to end-users.