Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2504.20634

On Stochastic Rounding with Few Random Bits

Andrew Fitzgibbon·Stephen Felix
View original source ↗Read source document ↗Export metadata ↓

Abstract

Large-scale numerical computations make increasing use of low-precision (LP) floating point formats and mixed precision arithmetic, which can be enhanced by the technique of stochastic rounding (SR), that is, rounding an intermediate high-precision value up or down randomly as a function of the value's distance to the two rounding candidates. Stochastic rounding requires, in addition to the high-precision input value, a source of random bits. As the provision of high-quality random bits is an additional computational cost, it is of interest to require as few bits as possible while maintaining the desirable properties of SR in a given computation, or computational domain. This paper examines a number of possible implementations of few-bit stochastic rounding (FBSR), and shows how several natural implementations can introduce sometimes significant bias into the rounding process, which are not present in the case of infinite-bit, infinite-precision examinations of these implementations. The paper explores the impact of these biases in machine learning examples, and hence opens another class of configuration parameters of which practitioners should be aware when developing or adopting low-precision floating point. Code is available at http://github.com/graphcore-research/arith25-stochastic-rounding.

Explore related subjects

math.NAcs.AIcs.LGcs.MScs.NA

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Andrew Fitzgibbon, Stephen Felix. 2025-04-29. On Stochastic Rounding with Few Random Bits. https://arxiv.org/abs/2504.20634

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Fully spectral scheme for the linear BGK equation on the whole space

In this article, we design a fully spectral method in both space and velocity for a linear inhomogeneous kinetic equation with mass, momentum and energy conservation. We focus on the linear BGK equation with a confinement potential $Φ$, even if the method could be applied to different collision operators. It is based upon the projection on Hermite polynomials in velocity and orthonormal polynomials with respect to the weight $e^{-$Φ$}$ in space. The potential $Φ$ is assumed to be a polynomial. It is, to the author's knowledge, the first scheme which preserves hypocoercive behavior in addition to the conservation laws. These different properties are illustrated numerically on both quadratic and double well potential.

math.NA↗
arXiv2026

Inverse inequalities for kernel-based approximation on bounded domains and Riemannian manifolds

This paper establishes inverse inequalities for kernel-based approximation spaces defined on bounded Lipschitz domains in $\mathbb{R}^d$ and compact Riemannian manifolds. While inverse inequalities are well-studied for polynomial spaces, their extension to kernel-based trial spaces poses significant challenges. For bounded Lipschitz domains, we extend prior Bernstein inequalities, which only apply to a limited range of Sobolev orders, to the full range of lower and upper orders, and derive Nikolskii inequalities that bound $L_\infty$ norms by $L_2$ norms. For compact Riemannian manifolds, we focus on restricted kernels, which are defined as the restriction of positive definite kernels from the ambient Euclidean space to the manifold, and prove their counterparts.

math.NA↗
arXiv2026

Error Estimates for Hyperbolic Scaling Limits of Linear Kinetic Models on Networks

This paper studies linear discrete kinetic models on networks and their asymptotic behavior in the small Knudsen number limit. For coupling conditions at an n-edge junction under a symmetric formulation, we introduce a change of variables that reformulates the system into n independent initial-boundary value problems. The asymptotic expansions are then constructed and rigorously justified by deriving an error estimate based on the energy method.

math.NA↗