In PET imaging, one main obstacle in obtaining a fully quantitative list-mode reconstruction in a run time compatible for clinical environment is the computation burden. Among parallelization and optimization architectures, graphics processing units (GPUs) represent today a powerful accelerator, especially for medical image processing. However a reconstruction on a single GPU is insufficient to handle all corrections needed (patient and scanner) with a compatible time for a clinical use. Multi-GPU is now becoming the best solution to go further for a higher performance computing. Only one method has been proposed recently on Multi-GPU context for list-mode PET image reconstruction. This method is not optimized in a high-resolution reconstruction context. In this work we propose a new Multi-GPU acceleration method, that optimizes the communication cost between GPUs. The results have shown that we obtained a linear performance computing scalability, that increases on higher resolutions, with a speedup factor of x4.4 for 4 GPUs with a communication cost decreasing of 97% against the previously proposed approach with a factor of x3.7.