Learning to live with errors: a fresh look at floating-point computation

Josh Milthorpe, Alistair Rendell

Research output: Contribution to conferencePaperpeer-review

Abstract

Interval analysis is an alternative to conventional floating-point computation that offers guaranteed error bounds. Despite this advantage, interval methods have rarely been applied in high performance scientific computing. In part, this is because of the additional cost associated with performing interval operations
over the corresponding floating-point operations. The aim of this work is to take a fresh look at interval arithmetic and the viability of using intervals in large scale scientific computations. In this paper we will report on the performance of interval arithmetic on the UltraSPARC architecture, with a focus on the Sun Studio implementation. Different methods of calculating interval results will be discussed, including one novel approach to interval multiplication. Based on the benchmark results for different interval implementations, changes to existing interval algorithms are suggested. The hardware modification of floating point units to provide.
additional architectural support for intervals is also considered.
Finally, brief consideration will be given to the application of interval techniques to a real-world problem
in computational chemistry, namely the evaluation of pairwise interactions under the Coulomb potential.
Original languageEnglish
Number of pages8
Publication statusPublished - 17 Nov 2005
Externally publishedYes
Event3rd Australian Undergraduate Students’ Computing Conference 2005 -
Duration: 17 Nov 2005 → …

Conference

Conference3rd Australian Undergraduate Students’ Computing Conference 2005
Period17/11/05 → …

Fingerprint

Dive into the research topics of 'Learning to live with errors: a fresh look at floating-point computation'. Together they form a unique fingerprint.

Cite this