Publication Details

Running Large-Scale Ultrasound Simulations on Piz Daint with 512 Pascal GPUs

VAVERKA Filip, TREEBY Bradley E. and JAROŠ Jiří. Running Large-Scale Ultrasound Simulations on Piz Daint with 512 Pascal GPUs. Denver, 2017.
Type
presentation,poster
Language
english
Authors
Vaverka Filip, Ing. (DCSY FIT BUT)
Treeby Bradley E. (UCL)
Jaroš Jiří, doc. Ing., Ph.D. (DCSY FIT BUT)
Keywords

Local domain decomposition, k-Wave toolbox, GPU

Abstract

Ultrasound simulation is a critical component of model-based treatment planning for ultrasound therapy. However, the domains are typically thousands of wavelengths in size, leading to large-scale numerical models with 10s of billions of unknowns. This paper presents a novel local Fourier basis domain decomposition for full-wave ultrasound propagation simulations with a custom bell function which ensures that the numerical error stays below 0.1% while enabling almost ideal strong scaling. Realistic benchmarks with 512 Nvidia P100 GPUs in the best EU supercomputer Piz Daint achieved efficiency between 90 and 100% with a speed-up over 100 and computational cost reduction by a factor of 12 compared to 1024 Haswell cores.

Published
2017
Pages
1
Conference
SuperComputing 2017, Colorado Convention Center, 14th Street, Denver, Colorado, Spojené státy americké, US
Place
Denver, US
Files
Back to top