Evaluating GPU Code Translation Approaches to Optimize CFD Efficiency

Computational Fluid Dynamics (CFD) enables engineers to simulate and analyze fluid flow behavior in complex systems, playing a crucial role in the design, research, and development of aerospace vehicles. However, high-fidelity simulations often require extensive computational resources, with runtimes spanning several days. Graphics Processing Unit (GPU) computing has demonstrated significant speed improvements over traditional Central Processing Unit (CPU)-based methods, but transitioning CFD codes to GPU architectures requires complex adaptation and optimization for parallel execution. This study evaluates methods for converting CPU-based CFD codes to GPU execution by analyzing result accuracy, runtime performance, and power consumption. Various code transition methodologies—including code regeneration translators, compiler-based acceleration, and AI-assisted code conversion—were applied to a simple, stationary Fortran code. The various adapted versions were executed on a GPU-accelerated system and compared against the original Fortran implementation running on a CPU. Power consumption, runtime, and solution accuracy were assessed using energy metrics, pointwise error analysis, and global error norms. Preliminary results indicate that compiler-based acceleration achieves the highest code efficiency but requires the longest translation time. AI-assisted conversion speeds up the process but demands human oversight to ensure accuracy and performance. Code regeneration translators offer the fastest conversion but often produce suboptimal GPU code. Ongoing analysis will further refine these comparisons to identify the most effective approach. Final results will provide practical guidance and valuable insights for CFD engineers, supporting the efficient and reliable adaptation of legacy CPU-based code to GPU computing for faster, more cost-effective simulations.