Joint Seminar: Shall we sacrifice exactness for performance in weather and climate models?

We investigate two approaches to reduce accuracy in order to gain performance in atmospheric modelling: The use of Stochastic Processors and the use of very low real number accuracy. Stochastic Processors reduce the accuracy since they show faulty calculations, but they offer a significant reduction of the power consumption. The use of lower accuracy for real numbers can significantly increase the performance of the model simulations, since less storage is needed and more data is fitting into memory and cache.
We provide tests with a dynamical core of a model for the global atmosphere that is based on spectral discretisation methods (IGCM). Spectral discretisation methods allow us to calculate the dynamics of the large scales with a different accuracy than the dynamics of the small scales. While we use an exact processor and double precision to calculate the large scale dynamics, for which exactness is crucially important, we emulate inexact stochastic processors and very low real number accuracy for the small scale dynamics. The description of the small scales already possess an inherent uncertainty due to limited grid resolution and parametrisation schemes. We emulate stochastic processors and very low real number accuracy by randomly flipping bits at a prescribed error rate, after all real number operation.
The results show that both approaches posses high potential for weather and climate models. Scale separation is a necessary but sufficient tool to allow a significant reduction of accuracy in large parts of the model.




13:30 Uhr


Bundesstr. 53, room 022/023
Seminar Room 022/023, Ground Floor, Bundesstrasse 53, 20146 Hamburg, Hamburg


Peter Dueben, Oxford University


Kenji Shimizu

Zur Übersicht