The use of imprecise processing to improve accuracy in weather & climate prediction

The use of stochastic processing hardware and low precision arithmetic in atmospheric models is investigated. Stochastic processors allow hardware-induced faults in calculations, sacrificing bit-reproducibility in exchange for improvements in performance and potentially accuracy and a reduction in power consumption. A similar trade-off is achieved using low precision arithmetic, with improvements in computation and communication speed and savings in storage and memory requirements. As high-performance computing becomes more massively parallel and power intensive, these two approaches may be important stepping stones in the pursuit of global cloud resolving atmospheric modelling. The impact of both hardware induced faults and low precision arithmetic is tested using the Lorenz ’96 model and the dynamical core of a global atmosphere model. In the Lorenz ’96 model there is a natural scale separation, the spectral discretisation used in the dynamical core also allows large and small scale dynamics to be treated separately within the code. Such scale separation allows the impact of lower-accuracy arithmetic to be restricted to components close to the truncation scales, and hence close to the necessarily inexact parametrised representations of unresolved processes. By contrast, the larger scales are calculated using exact arithmetic. Hardware faults from stochastic processors are emulated using a bit-flip model with different fault rates. Our simulations show that both approaches to inexact calculations do not substantially affect the mean behaviour, provided they are restricted to act only on smaller scales. By contrast, results with inexact calculations can be superior to those where smaller scales are parametrised. This suggests that inexact calculations at the small scale could reduce computation and power costs without adversely affecting the quality of the simulations. This would allow higher resolution models to be run at the same computational cost.