Unleashing the Potential of Mixed Precision in AI-Accelerated CFD Simulation on Intel CPU/GPU Architectures

CFD has emerged as an indispensable tool for comprehending and refining fluid flow phenomena within engineering domains. The recent integration of CFD with AI has unveiled novel avenues for expedited simulations and computing precision. This research paper delves into the accuracy of amalgamating CFD with AI and assesses its performance across modern server-class Intel CPU/GPU architectures such as the 4th generation of Intel Xeon Scalable CPUs (codename Sapphire Rapids) and Intel Data Center Max GPUs (or Ponte Vecchio). Our investigation focuses on exploring the potential of mixed-precision techniques with diverse number formats, namely, FP32, FP16, and BF16, to accelerate CFD computations through AI-based methods. Particular emphasis is given to validating outcomes to ensure their applicability across a CFD motorBike simulation. This research explores the performance/accuracy trade-off for both AI training and simulations, including OpenFOAM solver and interference with the trained model, across various data types available on Intel CPUs/GPUs. We aim to provide a thorough understanding of how different number formats impact the performance and accuracy of the DNN-based model in various application scenarios running on modern HPC architectures.