GPU Acceleration of the SUMMA Land Model

Process-based land models are increasingly used for large-domain hydrologic simulations and ensemble prediction. Such simulations are computationally challenging, especially on modest computing resources. This work presents a Graphics Processing Unit (GPU) implementation of the Structure for Unifying Multiple Modeling Alternatives (SUMMA) modeling framework. SUMMA represents the geographical domain as hundreds of thousands of topographically distinct spatial units, modeled as independent vertical columns that solve local water and energy balance equations without direct lateral interactions. We exploit this structure by mapping independent sub-basins to GPU threads and integrating GPU-enabled numerical solvers from the SUNDIALS library. Numerical experiments demonstrate that the GPU implementation reproduces the results of the reference SUMMA code while providing improved computational throughput and scalability. For large batches of sub-basins, the GPU implementation achieves speedups of up to 30 times relative to the corresponding serial CPU execution making large-domain and ensemble simulations more practical on modest computing resources.