FourCastNet3 (FCN3) is NVIDIA Earth-2’s latest AI global weather forecasting system, offering unprecedented accuracy and efficiency. It matches top machine learning models in forecasting and outperforms traditional numerical weather prediction systems. A single 60-day FCN3 rollout is computed in under four minutes on a single NVIDIA H100 GPU.

FCN3 employs a unique neural operator architecture with convolutional and spectral convolutions based on Morlet wavelets. An innovative hidden-Markov formulation introduces stochasticity, enabling efficient ensemble generation. FCN3 is trained on up to 1,024 GPUs using model-parallelism for scalability and computational efficiency.

FCN3 outperforms physics-based models like IFS-ENS and matches Gencast in predictive skill. It produces accurate, well-calibrated forecasts with stable spectral signatures even at extended lead times of 60 days. FCN3’s predictions of atmospheric and surface variables remain sharp and physically consistent, showcasing its reliability in weather forecasting.

To run FCN3 inference, Earth2Studio offers an easy method with sample code provided. Running a 4-member ensemble inference yields accurate predictions, showcasing FCN3’s performance. Optimizing FCN3’s performance involves installing torch-harmonics with custom CUDA extensions and using automatic mixed precision in bf16 format during inference. Custom inference or training code is available on GitHub for further exploration. A study on ensemble member 2 of FCN3 shows the total column water vapor and 10-meter zonal wind velocity fields. The standard deviation of these fields across all four ensemble members is also displayed in the research.

Authors of the study include Boris Bonev, Thorsten Kurth, Ankur Mahesh, Mauro Bisson, Karthik Kashinath, Anima Anandkumar, William D. Collins, Mike Pritchard, and Alex Keller. The research provides valuable insights into weather forecasting and climate modeling.

Read more at Nvidia
Nvidia has announced a new data center chip called Grace that is designed for artificial intelligence workloads. The chip is set to be released in 2023 and is expected to significantly improve performance for AI applications. The chip is named after Grace Hopper, a computer programming pioneer.: FourCastNet 3 Enables Fast and Accurate Large Ensemble Weather Forecasting with Scalable Geometric ML