AI Agents Accelerate GPU Porting of Massive Legacy Weather Simulation

Researchers have developed an AI-assisted workflow to port massive legacy scientific codes to GPU-centric high-performance computing (HPC) systems. By utilizing CLI-based AI agents, the team modernized CReSS, a legacy Fortran weather simulation, while maintaining strict numerical validity.

The process targeted the CReSS codebase, which consists of more than 250,000 lines of code. The workflow functioned by extracting OpenMP regions, generating kernel benchmarks from simulation states, and applying OpenACC transformations. According to research published via arXiv, this method produced numerically validated GPU implementations for 162 target kernels. When tested using a real typhoon simulation, the AI-assisted porting achieved a 5.1x application-level speedup.

The Challenge of Legacy Science

Legacy scientific applications are critical assets, with credibility built over decades of comparison against real-world observations. However, porting these systems to modern GPU hardware is traditionally labor-intensive and risky. Because small implementation changes can compromise scientific validity, developers often face a trade-off between performance gains and the risk of introducing subtle numerical errors.

Preserving Scientific Integrity

This research shifts the role of AI in HPC from simple code generation to a "validation-centric" workflow. This approach proved essential for identifying errors that human developers might overlook. Specifically, the AI agent detected numerical discrepancies in five kernels. These errors were caused by floating-point and intrinsic-function differences, including threshold-sensitive branch divergence and cancellation effects.

Implications for HPC Modernization

This case study demonstrates that AI agents can significantly reduce the development costs associated with modernizing massive legacy codes. By automating the porting process while maintaining a rigorous validation loop, the workflow ensures that the scientific integrity of the simulation remains intact. It proves that AI can be used not just to write code faster, but to ensure that the resulting software is as accurate as the original legacy system.

Future Outlook

As HPC systems move toward GPU-centric architectures, the ability to migrate legacy Fortran and C++ codes without losing numerical precision will be vital. Future efforts will likely focus on expanding this validation-centric model to other complex scientific domains where numerical drift is unacceptable. This transition is critical for maintaining the utility of decades of scientific investment while leveraging the raw power of modern hardware.