AI Agents Ported 250,000 Lines of Legacy Fortran Weather Code to GPU for a 5.1x Speedup — and Silently Broke 5 of 162 Kernels
Submitted 2026-08-13 (arXiv 2608.13122) and picked up on HN: a validation-centric workflow using CLI-based coding agents ported CReSS, a 250,000+ line legacy Fortran weather simulation, producing numerically validated GPU implementations for 162 target kernels and a 5.1x application-level speedup. Five kernels showed numerical discrepancies traced to floating-point differences and branch divergence — caught only because validation was built into the loop rather than bolted on afterward. The authors' honest list of what makes this hard is the transferable part: session-spanning context management, runtime-state reconstruction, and expensive recovery from small static-analysis omissions. For anyone pointing agents at a large legacy migration, the 3% silent-correctness-failure rate is the number to design your harness around.
↳ Follow the thread