Multi-GPU FDTD study finds direct peer exchange yields 2.76x speedup

A new paper on arXiv by Victory Obieke presents a comprehensive communication-strategy study for multi-GPU three-dimensional finite-difference time-domain (FDTD) computations incorporating convolutional perfectly matched layer (CPML) boundary conditions. The work, written in the context of CUDA-accelerated clusters, evaluates runtime, throughput (millions of output points per second), strong-scaling efficiency, CPML overhead, and the performance of host-staged versus direct GPU-to-GPU peer exchange. On a single NVIDIA Quadro RTX 6000, the CPML implementation sustains an impressive 2,889–3,290 million output points per second with boundary-layer overhead below 1%. This provides a solid single-GPU baseline for the multi-GPU experiments.

The core finding is that direct GPU-to-GPU peer exchange is the dominant optimization, delivering a 2.46–2.76× speedup over host-staged exchange. Enlarged ghost regions offered only modest benefits, as reduced communication frequency was partly offset by redundant computation and additional memory traffic. On NVIDIA Quadro RTX 8000 GPUs, the implementation achieved up to a 1.51× speedup when scaling from one to two GPUs in strong-scaling tests. Four-GPU configurations enabled larger grids that approached or exceeded single-GPU memory capacity, demonstrating the potential for tackling bigger simulation domains. The study provides practical guidance for researchers and engineers optimizing electromagnetic simulations on multi-GPU systems.