Nvidia makes AI weather forecasting more accessible, no supercomputer needed
Nvidia released two open-source, transformer-based weather models, Earth-2 Medium Range and Nowcasting, aiming to democratize forecasting beyond supercomputers.Summary
Nvidia has launched two new open-source weather forecasting models, Earth-2 Medium Range and Earth-2 Nowcasting, which utilize the transformer architecture common in LLMs, moving away from specialized architectures. The Medium Range model offers high-accuracy forecasts up to 15 days and reportedly outperforms DeepMind's GenCast across many variables. The Nowcasting model provides high-resolution, country-scale forecasts, which organizations like the Israeli Meteorological Service and The Weather Company plan to adopt for localized severe-weather applications. Crucially, Nvidia emphasizes accessibility; the Medium Range model requires only 26GB of GPU memory to run, and the Nowcasting model needs just 5GB on a single GPU, significantly lowering the barrier to entry compared to traditional supercomputer-dependent methods. Furthermore, Nvidia released the Global Data Assimilation model to rapidly generate the necessary starting conditions for these forecasts, a process that previously consumed about 50% of traditional supercomputing loads.
(Source：The New Stack)