We aim to develop AI-based surrogate models to predict full-vehicle aerodynamic performance with high accuracy and efficiency. The training data is generated from high-fidelity large-scale CFD simulations, where each case involves high-resolution meshes and flow-field data represented as point clouds of tens of millions to over one hundred million points. Managing and processing such large datasets requires both high-memory GPUs and substantial storage resources. To accomplish this, we request access to FAT-type GPU nodes NVIDIA A100 80GB GPUs, which provide the necessary memory and compute capability for efficient training of multi-fidelity surrogate models.