LMARSpy: A GPU‐Ready Nonhydrostatic Dynamical Core With a Sharpness‐Preserving Monotonicity Limiter and a Conservative Vertical Implicit Solver

Abstract Global numerical modeling is advancing into the era of kilometer‐scale, non‐hydrostatic simulations, while heterogeneous computing emerges as a pivotal trend in high‐performance computing (HPC). As a strong candidate for next‐generation global kilometer‐scale general circulation models, the A‐grid dynamical core based on the Low Mach number Approximate Riemann Solver must address key challenges: suppresses unphysical oscillations while preserving sharp gradients, mitigating Courant‐Friedrichs‐Lewy (CFL) constraints caused by vertically propagating sound waves, and making full use of heterogeneous computing power. This work presents LMARSpy, a GPU‐optimized non‐hydrostatic dry dynamical core with a sharpness‐preserving monotonicity limiter and a conservative vertical implicit solver. Designed for GPU efficiency, LMARSpy leverages a Python‐based HPC framework to ensure robust compatibility across heterogeneous computing environments. Benchmark tests validate the model's innovations: the limiter effectively suppresses unphysical oscillations while preserving sharpness, incurring only a 10.4% increase in GPU computational cost, as it is applied exclusively in the final sub‐step of the RK3 time‐stepping scheme; the Horizontally Explicit and Vertically Implicit scheme (employing a mass and potential temperature conservative vertical implicit solver) effectively mitigates vertical CFL limitations, delivering at least an order‐of‐magnitude efficiency improvement over explicit schemes when horizontal grid spacing substantially exceeds vertical spacing; and the Python‐based HPC framework enables seamless operation on both CPU and GPU architectures. When the computational scale is sufficient to fully utilize GPU memory, a single Nvidia RTX 3060 Ti GPU delivers 7.6 times the computational speed of a fully‐loaded 16‐thread Intel Core i5‐13490F CPU.