Impact of Reduced accuracy and Mixed in Computing Lattice-Boltzmann Method for Multiple GPUs

• Gabriel Freytag UFRGS
• João Vicente Ferreira Lima UFSM
• Paolo Rech UFRGS
• Philippe Olivier Alexandre Navaux UFRGS

DOI: https://doi.org/10.5753/eradrs.2020.10795

Abstract

The heterogeneity of modern computing architectures allows engineers refine algorithms to maximize the level of affinity with the architecture and, consequently, the effectiveness of computing. In this study, we investigated the impact of the accuracy and reduced mixed in computing Lattice-Boltzmann method in a multi-GPU platform. Using half precision for storage and for simple arithmetic operations, we obtained a speedup of 3:44 consuming 71 \% less energy and a loss 00:02 \% accuracy.

Keywords: Algorithms Parallel and Distributed, Computer architecture, Dedicated architectures and Specific, Evaluation, Performance Measurement and Prediction, Heterogeneous computing

References

Mittal, S. (2016). A survey of techniques for approximate computing. ACM Computing Surveys, 48(4).