loading page

Climate Variability Due to Selection of Computational Platform
  • Thomas Robinson,
  • Jessica Liptak
Thomas Robinson
Geophysical Fluid Dynamics Laboratory

Corresponding Author:thomas.robinson@noaa.gov

Author Profile
Jessica Liptak
Author Profile


Climate models generally require that results between runs are bit-for-bit reproducible. This becomes impossible when switching the computational platform or compiler that the model is run on. An ensemble of the Geophysical Fluid Dynamics Laboratory (GFDL) Atmosphere Model 4.0 (AM4, Zhao et al. 2018a,b)) is created by perturbing the temperature at a random point on the order of 10-13 (in the rounding error of the system). Previous results show that three different compilers on the same computing platform results in a spread of the global mean temperature of 0.14 K (Robinson et al. 2018). The current ensembles are run on three different computing platforms with different processors: the main production computer of GFDL with Intel broadwell/haswell, one with Intel knights landing, and the other with Intel skylake. The ensemble means and standard deviations for global surface temperature are compared in order to see if the spread of rounding error in the model is platform dependent. The means are also compared to see if they lie within the spread of each modeling platform.