DescriptionRealistic ultrasound simulations have found a broad area of applications in preoperative photoacoustic screening and non-invasive ultrasound treatment planing. However, the domains are typically thousands of wavelengths in size, leading to large-scale numerical models with billions of unknowns. The current trend in accelerated computing is towards the use of fat nodes with multiple GPUs per node. The multi-GPU version of our k-Wave acoustic toolbox is based on the local Fourier basis domain decomposition where 3D simulation domain is partitioned into rectangular cuboid blocks assigned to particular GPUs. This paper investigates the benefits of using the CUDA-Aware MPI and CUDA peer-to-peer transfers on an 8-GPU server equipped with Nvidia P40 GPUs. The server has a total GPU memory of 192 GB and a single-precision performance of 96 Tflops. These techniques reduces the overall simulation time a factor of 2-3.6.