AI-Assisted GPU Porting of a 250k Line Legacy Weather Simulation Code

Researchers successfully used AI agents to port a 250,000+ line legacy Fortran weather simulation code (CReSS) to GPUs while maintaining scientific validity through a validation-centric approach. The workflow achieved a 5.1x application-level speedup and detected numerical discrepancies in five kernels caused by floating-point differences, demonstrating that effective AI-assisted GPU porting requires both code generation and rigorous validation design rather than code conversion alone.