Physics-informed neural networks (PINNs) combined with domain decomposition have shown promise for multiphase flow simulation, yet their application is often constrained by high training costs. To address this, we enhance the phase-field PINN framework with JAX for high-performance computing, automatic mixed precision for memory efficiency, and redesigned communication strategies for multi-GPU parallel training. Additionally, we evaluate data parallel, domain decomposition, and a hybrid data-domain parallel approach using a coaxial bubble rising case. Results indicate that the data parallel method offers the best trade-off between accuracy and efficiency, while the hybrid method reduces inter-partition communication and achieves lower loss than domain decomposition. These strategies collectively enable accurate, efficient direct numerical simulation of multiphase flows with acceptable training costs.