Numerical Precision Effects on GPU Simulation of Massive Spatial Data, Based on the Modified Planar Rotator Model

TL;DR

This paper investigates how reducing numerical precision in GPU implementations of the modified planar rotator model affects the speed and accuracy of spatial data predictions, finding significant efficiency gains with minimal accuracy loss.

Contribution

It provides an analysis of precision effects on GPU-based spatial prediction using the MPR model, highlighting optimal settings for large data sets.

Findings

Reduced precision can significantly improve GPU speed.

Minimal accuracy degradation with appropriate precision settings.

Optimal precision choices depend on data size and application needs.

Abstract

The present research builds on a recently proposed spatial prediction method for discretized two-dimensional data, based on a suitably modified planar rotator (MPR) spin model from statistical physics. This approach maps the measured data onto interacting spins and, exploiting spatial correlations between them, which are similar to those present in geostatistical data, predicts the data at unmeasured locations. Due to the short-range nature of the spin pair interactions in the MPR model, parallel implementation of the prediction algorithm on graphical processing units (GPUs) is a natural way of increasing its efficiency. In this work we study the effects of reduced computing precision as well as GPU-based hardware intrinsic functions on the speedup and accuracy of the MPR-based prediction and explore which aspects of the simulation can potentially benefit the most from the reduced precision. It is found that, particularly for massive data sets, a thoughtful precision setting of the GPU implementation can significantly increase the computational efficiency, while incurring little to no degradation in the prediction accuracy.