VVMex v1.0: a modular GPU-capable refactoring of the Vector Vorticity cloud-resolving Model (VVM) with LES-based hierarchical validation and performance analysis

Abstract. This study presents VVMex v1.0, a GPU-capable large-eddy simulation model with a software architecture designed for portability and extensibility on heterogeneous, exascale-oriented computing systems. Developed using the Kokkos performance-portability programming model, VVMex is a clean-start, object-oriented C++ redesign of the Fortran Vector Vorticity cloud-resolving Model (VVM). It adopts a modular architecture with a GPU-resident time-stepping workflow that minimizes host–device data transfers. The model integrates GPU-capable physical schemes, including the Noah land surface model, Rapid Radiative Transfer Model for General Circulation Model Applications–Parallel radiation scheme from the E3SM Atmosphere Model in C++ (EAMxx), and Predicted Particle Properties (P3) microphysics from EAMxx. The EAMxx P3 implementation was adapted for LES-based VVMex simulations by restoring alignment with the original Fortran P3 formulation and was evaluated against the Fortran P3 reference implementation. To establish the fidelity of the refactored model, we introduce a hierarchical verification and validation framework. The verification phase encompasses tolerance-based tests for isolated components and idealized dry dynamical-core experiments, such as mountain flows. For physics-based validation, we assess whether coupled dynamics and parameterizations preserve the intended regimes across configurations ranging from strongly to weakly forced: heterogeneous ocean–land–mountain diurnal circulations forced by land–sea contrast and terrain, a homogeneous diurnal boundary layer driven by surface processes, and radiative–convective equilibrium over a horizontally uniform ocean forced by prescribed sea-surface temperature and incoming solar radiation. In performance benchmarks using a high-resolution, full-physics configuration over complex terrain, VVMex achieves a 4- to 24-fold wall-clock speedup on NVIDIA H200 GPUs relative to the Fortran VVM baseline running on 1024 Intel Xeon Platinum 8480+ CPU cores, while demonstrating strong scaling from 8 to 64 GPUs.