Speaker
Description
k-Wave is widely used for time-domain ultrasound simulation, but large three-dimensional problems often exceed the memory and performance limits of a single GPU. Multi-GPU implementations are therefore required to enable larger computational domains while reducing simulation time.This work compares two multi-GPU implementations of the Fourier collocation scheme employed in k-Wave. The first preserves the original pseudospectral formulation using distributed Global FFTs over the entire computational domain, while the second replaces them with Local FFTs evaluated on overlapping subdomains connected by halo exchange. Both implementations were evaluated on a single node equipped with eight NVIDIA A100 GPUs interconnected by NVSwitch.The results show that both approaches enable substantially larger simulations than a single GPU while providing good strong scaling. By replacing global all-to-all communication with nearest-neighbor halo exchange, the Local FFT implementation reduces inter-GPU communication by approximately one order of magnitude and achieves lower runtime for most tested domain sizes. The resulting approximation maintains relative errors on the order of 10^-3 for appropriately chosen halo widths, providing a practical trade-off between numerical accuracy and computational performance.