AMD and Oak Ridge National Laboratory have released ORBIT-2, a global climate downscaling model designed for high-resolution weather forecasting. The model produces a best estimate of weather-related variables at high resolution, conditioned on lower-resolution inputs. ORBIT-2 uses super-resolution techniques from computer vision, treating each weather variable as a layer. It is capable of scaling to 10 billion parameters across 65,536 GPUs and achieving up to 4.1 exaFLOPS sustained throughput. AMD said the model can be run on its GPUs for inference, with a focus on global precipitation downscaling. The model's performance is evaluated against observational data at 7 km resolution, achieving R² scores between 0.98 and 0.99. AMD also provided a step-by-step guide for running ORBIT-2 inference, including data preparation and setup instructions. The model is supported by data in .zarr format, compatible with ERA5 archives. AMD said the model can be used for severe weather forecasting, with applications in disaster mitigation. The model is available for testing with a subset of the ERA5-IMERG dataset. AMD said the model's design allows for efficient computation and scalability, with a novel Reslim architecture and TILES algorithm for self-attention scaling. AMD said the model's performance is benchmarked against observational data, with high accuracy reported. AMD said the model can be used for global precipitation downscaling, with a focus on high-resolution forecasting. AMD said the model's architecture enables efficient computation and scalability, with a novel Reslim architecture and TILES algorithm for self-attention scaling. AMD said the model's performance is benchmarked against observational data, with high accuracy reported.
AMD said the model's design enables better scaling and compute efficiency, with technical innovations that can be applied beyond the currently available checkpoints. The model's current checkpoints support downscaling of global precipitation and regional US temperature variables. AMD said the model's inference process involves using low-resolution input images to generate high-resolution outputs. The model is trained on well-curated data such as the ERA5 reanalysis archives, with a focus on global weather prediction models at resolutions of 10–30 km. AMD said the model's performance is evaluated against observational data at 7 km resolution, with R² scores in the range of 0.98–0.99. AMD said the model's design allows for efficient computation and scalability, with a novel Reslim architecture and TILES algorithm for self-attention scaling. AMD said the model's performance is benchmarked against observational data, with high accuracy reported.
AMD said the model's inference process involves using low-resolution input images to generate high-resolution outputs. The model is trained on well-curated data such as the ERA5 reanalysis archives, with a focus on global weather prediction models at resolutions of 10–30 km. AMD said the model's performance is evaluated against observational data at 7 km resolution, with R² scores in the range of 0.98–0.99. AMD said the model's design allows for efficient computation and scalability, with a novel Reslim architecture and TILES algorithm for self-attention scaling. AMD said the model's performance is benchmarked against observational data, with high accuracy reported.
Source: amd
