AMD has introduced CorrDiff, a diffusion-based model for atmospheric downscaling, enabling the reconstruction of missing high-resolution data from low-resolution atmospheric observations. The company demonstrated how to run CorrDiff inference on its Instinct GPUs, providing step-by-step instructions for users to test the model themselves. The model estimates high-resolution features based on low-resolution data, adhering to physical constraints to ensure consistency. According to AMD, CorrDiff combines existing large-scale information with physical and observational constraints to infer plausible high-resolution data. The approach is part of a broader shift toward machine learning-based methods in downscaling, which are gaining traction due to their performance and flexibility. AMD emphasized that the model can be used for both spatial and temporal data reconstruction, addressing gaps caused by sensor limitations or resolution constraints. The company also highlighted the computational efficiency of running CorrDiff on Instinct GPUs, which are optimized for AI and scientific computing tasks. The demonstration includes Python-based tools for visualizing results, making the model accessible to researchers and practitioners in climate science. *Source: [amd](https://rocm.blogs.amd.com/artificial-intelligence/corrdiff-inference/README.html)*