On the Importance of Precision for Short-Range Forecasting and Climate Simulation

Abstract

This paper presents results of experiments designed to test the importance of arithmetic precision in short-term forecasts and long-term climate simulations with the NCAR global circulation model. It is expected that the next-generation computers will have a sizable speed gain when using lower-precision arithmetic. To determine how precision affects the model results, we compare several short- and long-term experiments using 48-bit mantissa arithmetic (normal for the CDC 6600 and 7600 computers) with corresponding experiments using 24- and 21-bit mantissa arithmetic. The errors due to the lower precision are much smaller than typical observational errors. In addition, it appears that in the short-term experiments the rapid error growth of the model dominates the round-off error accumulation resulting from the lower-precision arithmetic. Therefore, the lower precision used by the next-generation computers should not have a detrimental effect on short-term forecast accuracy. The long-term climate simulation experiments indicated a very similar conclusion. Even though there were some differences in the results of the experiments, climate indicators such as zonal wind, zonal temperature or eddy transport are quite similar.