AI agents speed legacy weather code GPU porting by 5.1×

The paper documents a practical case study of GPU porting for CReSS, a large legacy scientific application written in Fortran with more than 250,000 lines. Scientific codes like this one are not mere legacy software—they are assets whose credibility rests on decades of development, validation against observational data, and use in published research. Any GPU porting effort must therefore preserve this scientific validity, not just speed up execution. The researchers' approach was validation-centric from the start: they used an AI agent to extract OpenMP regions, generate kernel benchmarks from physically meaningful simulation states, apply OpenACC transformations, and validate results through element-wise comparison with dumped reference data and application-level validation. This design proved critical because the workflow uncovered five kernels with numerical discrepancies—subtle errors in floating-point arithmetic and intrinsic-function behavior that pure performance optimization might have overlooked. The 5.1× speedup achieved within practical wall-clock development cost suggests that AI-assisted porting can scale to real production codebases, but only when the workflow is designed to catch and report validation failures back to domain experts.