Energy-Efficient Transprecision Techniques for Iterative Refinement

Research output: Contribution to conferencePoster

54 Downloads (Pure)

Abstract

This paper presents transprecision techniques for iterative refinement, which utilize various precision arithmetic dynamically according to numeric properties of the algorithm and computational latencies depending on precisions. The transprecision techniques were plugged into a mixed precision iterative refinement on an Intel Xeon E5-2650 2GHz core with MKL 2017 and XBLAS 1.0. The transprecision techniques brought further 2.0-3.4X speedups and 3.0-4.1X energy reductions to a mixed precision iterative refinement when double precision solution accuracy was required for forward error and a matrix size was ranged from 4K to 32K.
Original languageEnglish
Publication statusAccepted - 01 Sep 2017
EventSupercomputing'17 (SC17): International Conference on High Performance Computing, Networking, Storage and Analysis - Denver, United States
Duration: 11 Nov 201717 Nov 2017

Conference

ConferenceSupercomputing'17 (SC17): International Conference on High Performance Computing, Networking, Storage and Analysis
Abbreviated titleSC17
CountryUnited States
CityDenver
Period11/11/201717/11/2017

Keywords

  • transprecision, mixed precision, iterative refinement

Cite this

Lee, J., Vandierendonck, H., & Nikolopoulos, D. (Accepted/In press). Energy-Efficient Transprecision Techniques for Iterative Refinement. Poster session presented at Supercomputing'17 (SC17): International Conference on High Performance Computing, Networking, Storage and Analysis, Denver, United States.