DescriptionRealistic ultrasound simulations are becoming integral part of many novel medical procedures such as photoacoustic screening and non-invasive treatment planning. The common denominator of all these applications is the need for cheap and relatively large-scale ultrasound simulations with sufficient accuracy. Typical medical applications require full-wave simulations which take frequency-dependent absorption and non-linearity into account.
This poster investigates the benefits of high-bandwidth low-latency interconnects to k-Wave acoustic toolbox in dense multi-GPU environment. The k-Wave multi-GPU code is based on a variant of the local Fourier basis domain decomposition. The poster compares the behavior of the code on a typical PCI-E 3.0 machine with 8 Nvidia Tesla P40 GPUs and a Nvidia DGX-2 server. The performance constraints of PCI-E platforms built around multiple socket servers on multi-GPU applications are deeply explored. Finally, it is shown the k-Wave toolbox can efficiently utilize NVlink 2.0 and achieve over 4x speedup compared to PCI-E systems.