Vana, Filip ; Carver, Glenn ; Lang, Simon ; Leutbecher, Martin ; Salmond, Deborah ; Düben, Peter
Année de publication
Since the early days of numerical weather prediction (NWP), the issue of appropriate numerical precision has been the subject of considerable interest. Indeed, ECMWF’s second Technical Report, published in 1976 by Baede at al., was devoted to ‘The effect of arithmetic precision on some meteorological integrations’. Some precision-sensitive operations, such as matrix inversion, often require at least so-called double-precision arithmetic (i.e. a 64-bit representation of real numbers) to deliver acceptable results. With the growth of computer power and especially the availability of 64-bit processors, double precision came to be commonly used for all floating-point computations in numerical modelling, including in NWP. However, such precision may be wasteful since it uses up precious computing resources while not necessarily making much difference to forecast quality. Building on work carried out at the University of Oxford, we have made single-precision arithmetic (a 32-bit representation of real numbers) available for ensemble forecasts in ECMWF’s Integrated Forecasting System (IFS). Experiments show that, at a horizontal resolution of 50 km, single precision brings significant savings in computational cost without degrading forecast quality.