Weather forecast: the Astra model equals GenCast with ten times less calculation, and turns 50 scenarios in 40 minutes on a consumer graphics card

Driven across eight GPU H200, this quarter-degree overall model beats the entire ECMWF European centre on the 600 combinations of variables, levels and deadlines tested.

Producing high-resolution probabilistic weather forecasts remains very expensive. Two Korean researchers present Astra (not related to the OpenAI model of the same name), a 0.25° resolution set model driven on only eight GPU H200. It is based on an original mechanism of attention, in which each region of the globe chooses both where it reads and where it writes the information, and on a probabilistic score calculated also on its internal representations on several scales.

Figures

On the WeatherBench 2 test bench, Astra reaches GenCast, the Google DeepMind model, with five altitude variables and ten days of maturity, with less than 10% of its training budget with comparable hardware. Compared to ERA5 reanalyses, it is better than the ECMWF European Centre's operational set of 600 combinations evaluated, and ahead of 73.5 % of GenCast.

The gain is especially spectacular in use: a 15-day forecast requires 39 times less operations than GenCast, and a set of 50 scenarios is calculated in less than 40 minutes on a single consumer graphics card, in 21 GB of memory. To make the overall forecast accessible to weather services with limited means.