.svg){kind=spacer}
Details
- NVIDIA has introduced FourCastNet3 (FCN3), an advanced AI system designed to dramatically accelerate and improve the accuracy of medium- and subseasonal weather forecasts.
- FCN3 can generate 15-day global weather forecasts at 0.25° resolution and 6-hour intervals in just one minute on an NVIDIA H100 GPU, making it 8 times faster than GenCast and 60 times faster than the widely used IFS-ENS model.
- The system utilizes a neural operator architecture featuring spherical convolutions and Morlet wavelets, and incorporates hidden-Markov models to create realistic, stochastic ensemble forecasts.
- FCN3 maintains spectral fidelity across atmospheric scales, even at 60-day forecasts, and eliminates the high-frequency artifacts often found in other machine learning models.
- It is trained on 1,024 GPUs with distributed computing, allowing for scalable, cost-effective deployment in climate analysis and preparations for extreme weather events.
Impact
FourCastNet3 is a significant step forward in medium- and long-range weather prediction, allowing industries to respond faster to extreme conditions and climate-linked risks. By outperforming traditional simulation models in speed and maintaining high accuracy, FCN3 strengthens NVIDIA’s position in the increasingly competitive AI weather forecasting market. Its integration with Earth-2 and adoption by partners mark a pivotal shift toward AI-driven environmental modeling and disaster readiness.
