Quantifying weather emulators to reduce their energy consumption: an initial test bed

A study applies post-training quantification to two global forecasting AI models, DLWP and FourCastNet, to reduce computational load and energy consumption without compromising short-term forecasts.

Weather forecasting emulators based on deep learning now compete with physical models, from a few days to the sub-seasonal scale, for a much lower computational cost. Their inference is mainly based on GPU-accelerated matrix multiplications, an area that can still be optimized.

This work, published on arXiv, tests post-training quantitation—reducing the numerical accuracy of an already trained model—on two global emulators, Deep Learning Weather Prediction (DLWP) and FourCastNet. This technique, which has already been proven in other fields, makes it possible to perform more operations per unit of time while consuming less energy, and to consider deployments on embedded hardware.

The authors systematically examine the effect of several configurations on short-term forecasts. With simulated quantification, the forecasts remain qualitatively exploitable, which provides a first reference for applying these techniques to geophysical fluid dynamics.