CFD has become a vital tool for understanding and optimizing fluid flow phenomena in engineering. The recent incorporation of AI into CFD has opened new prospects for faster and more reliable simulations. This work delves into the accuracy of integrating CFD with AI and evaluates its performance on an HPC system using multiple NVIDIA HG200 chips - one of the most powerful GPU-based accelerators for AI. This research explores the potential of mixed precision techniques with diverse data formats to accelerate the distributed data-parallel training of our DNN model proposed for CFD motorBike simulations. Among the considered formats are BF16, TF32, and FP32. Especial emphasis is given to validating and tuning the accuracy of training concerning the impact of mixed precision methods and partitioning a large training dataset into smaller batches. We aim to understand better how various number formats impact the performance-accuracy trade-off in training DNN models for CFD simulations on modern HPC platforms with multiple GPUs and nodes.