Rounding errors may be beneficial for simulations of atmospheric flow: results from the forced 1D Burgers equation

摘要: Inexact hardware can reduce computational cost, due to a reduced energy demand and an increase in performance, therefore allow higher-resolution simulations of the atmosphere within same budget for computation. We investigate use emulated inexact model randomly forced 1D Burgers equation with stochastic sub-grid-scale parametrisation. Results show that numerical precision be only 12 bits significand floating-point numbers—instead 52 double precision—with no serious degradation results all diagnostics considered. Simulations on grid higher spatial resolution are significantly better compared coarser at similar estimated computing cost. In second half paper, we compare forcing rounding errors parametrisation scheme is used represent variability standard setup. argue forcings schemes provide first guess upper limit magnitude tolerated by suggest hidden distribution forcing. present idealised setup replaces expensive engineered error provides quality. The create forecast ensemble spread based conclude not necessarily degrading quality simulations. Instead, they beneficial representation variability.