PreprintCC BY 4.0arXiv
Mixed-Precision Computing for Scientific Discovery: Formats, Co-Design, and Responsible Approximation
Abstract
Reduced and mixed precision have moved from a niche optimization to a central design axis in scientific computing and engineering, driven by energy constraints, heterogeneous accelerators, and the convergence of simulation and machine learning. This paper organizes the landscape around seven coupled themes---number formats, floating-point emulation, emerging architectures, hardware/software co-design, relation to other approximations, software design, and precision as a multilevel resource ---and, for each theme, synthesizes the state of the art, future directions, and open questions. We emphasize \emph{energy per trusted solution} as the core objective, and we frame \say{recklessly responsible} computing as a pragmatic doctrine: exploit low precision aggressively, but with systematic detection, escalation, and certification pathways.
Related
- Randomized algorithms in numerical linear algebra for scientific computing
- Beyond χ2: Ensemble null tests for reliable scientific discovery
- Open-ended Scientific Discovery with Possibilistic Reasoning
- KnottedGraph: Scalable knotted-graph topology for scientific and mathematical discovery
- Computing Lewis Weights to High Precision by Fixed-Point Iteration
BibTeX
@misc{agullo2026mixedprecision,
title = {{Mixed-Precision Computing for Scientific Discovery: Formats, Co-Design, and Responsible Approximation}},
author = {Emmanuel Agullo and Hartwig Anzt and Daniel Bauer and David Bindel and Alfredo Buttari and Alexandru Calotoiu and Erin Claire Carson and Pasqua D'Ambra and Ieva Daužickaitė and James W. Demmel and Jack Dongarra and Iain Duff and Massimiliano Fasi and Dominik Göddeke and Stef Graillat and Laslo Hunhold and Roman Iakymchuk and Fabienne Jézéquel and Nils Kohl and Harald Köstler and Jakub Kružík and Julien Langou and Xiaoye Sherry Li and Hatem Ltaief and Piotr Luszczek and Yuxin Ma and Theo Mary and Mantas Mikaitis and Hiroyuki Ootomo and Daniel Osei-Kuffuor and Enrique S. Quintana-Ortí and Ulrich Rüde and Jennifer Scott and John Shalf and Linda Stals and Rasmus Tamstorf and Stefan Turek and Petr Vacek and Bastien Vieublé and Rio Yokota},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.37137},
eprint = {2609.37137},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.37137},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Mixed-Precision Computing for Scientific Discovery: Formats, Co-Design, and Responsible Approximation
AU - Emmanuel Agullo
AU - Hartwig Anzt
AU - Daniel Bauer
AU - David Bindel
AU - Alfredo Buttari
AU - Alexandru Calotoiu
AU - Erin Claire Carson
AU - Pasqua D'Ambra
AU - Ieva Daužickaitė
AU - James W. Demmel
AU - Jack Dongarra
AU - Iain Duff
AU - Massimiliano Fasi
AU - Dominik Göddeke
AU - Stef Graillat
AU - Laslo Hunhold
AU - Roman Iakymchuk
AU - Fabienne Jézéquel
AU - Nils Kohl
AU - Harald Köstler
AU - Jakub Kružík
AU - Julien Langou
AU - Xiaoye Sherry Li
AU - Hatem Ltaief
AU - Piotr Luszczek
AU - Yuxin Ma
AU - Theo Mary
AU - Mantas Mikaitis
AU - Hiroyuki Ootomo
AU - Daniel Osei-Kuffuor
AU - Enrique S. Quintana-Ortí
AU - Ulrich Rüde
AU - Jennifer Scott
AU - John Shalf
AU - Linda Stals
AU - Rasmus Tamstorf
AU - Stefan Turek
AU - Petr Vacek
AU - Bastien Vieublé
AU - Rio Yokota
PY - 2026
DA - 2026/09/29
PB - arXiv
DO - 10.48550/arxiv.2609.37137
UR - https://arxiv.org/abs/2609.37137
AB - Reduced and mixed precision have moved from a niche optimization to a central design axis in scientific computing and engineering, driven by energy constraints, heterogeneous accelerators, and the convergence of simulation and machine learning. This paper organizes the landscape around seven coupled themes---number formats, floating-point emulation, emerging architectures, hardware/software co-design, relation to other approximations, software design, and precision as a multilevel resource ---and, for each theme, synthesizes the state of the art, future directions, and open questions. We emphasize \emph{energy per trusted solution} as the core objective, and we frame \say{recklessly responsible} computing as a pragmatic doctrine: exploit low precision aggressively, but with systematic detection, escalation, and certification pathways.
ER -