From ‘Uncomputable’ to ‘Computable’ How Multi-GPU Breaks the VRAM Bottleneck in Large-Scale 3D3C-PIV Flow Field Reconstruction

This video presents a quantitative experimental validation of multi-GPU parallel computing for large-scale 3D3C-PIV flow field reconstruction. Through a spatially partitioned scheduling strategy, the system distributes interrogation window subsets across multiple GPUs while maintaining numerical consistency with single-GPU results.

Three datasets (966, 35,568, and 49,608 windows) demonstrate that multi-GPU architecture not only accelerates computation (up to 56.5% time reduction) but also enables reconstructions that are infeasible on single-GPU systems due to VRAM limitations — achieving full 49,608-window reconstruction in approximately 106 seconds.