AI Agent Ports 250k-Line Weather Code to GPU, 5.1x Faster

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

Researchers used an AI agent to port CReSS, a 250,000+ line legacy Fortran weather simulation, to GPUs. The validation-centric workflow extracted OpenMP regions, generated dump-based benchmarks, applied OpenACC transformations, and validated results against reference data. In a real typhoon simulation, they achieved a 5.1x speedup across 162 kernels, while detecting numerical discrepancies in five kernels due to floating-point and intrinsic-function differences. The study highlights the need for session-spanning context, runtime-state reconstruction, and recovery from static-analysis omissions in AI-assisted porting.