AI-Assisted GPU Porting

AI-assisted GPU porting is the practice of using large-language-model coding agents to translate existing CPU-bound software — typically large legacy codebases in scientific computing, weather simulation, computational fluid dynamics, or similar domains — into GPU-accelerated implementations. The defining pattern is iterative: the agent proposes a port of a code section, the result is validated against reference CPU outputs (numerical equivalence, physics conservation, performance benchmarks), and the differences are fed back to the agent for correction. Modern AI-assisted GPU porting pipelines typically combine a foundation model with retrieval over the existing codebase, a test harness that runs the CPU reference and GPU candidate side-by-side, and an agent loop that proposes diffs and re-runs validation. The technique differs from one-shot code translation by emphasizing verification: the goal is not to produce GPU code that compiles, but GPU code whose outputs match the CPU reference within tolerance on representative workloads. AI-assisted GPU porting has become relevant as scientific software modernization budgets have grown and as foundation-model coding capabilities have crossed the threshold needed to handle non-trivial parallelization patterns (CUDA, HIP, SYCL, OpenMP target offload). Notable production examples include porting the UK Met Office's LFRic atmospheric model, portions of GROMACS, and various NWP (numerical weather prediction) codes.