arXiv subjects

Laslo Hunhold

Publications and source records attributed to Laslo Hunhold.

10 recordsLinked to original sources

arXiv2025

Tekum: Balanced Ternary Tapered Precision Real Arithmetic

In light of recent hardware advances, it is striking that real arithmetic in balanced ternary logic has received almost no attention in the literature. This is particularly surprising given ternary logic's promising properties, which could open new avenues for energy-efficient computing and offer novel strategies for overcoming the memory wall. This paper revisits the concept of tapered precision arithmetic, as used in posit and takum formats, and introduces a new scheme for balanced ternary logic: tekum arithmetic. Several fundamental design challenges are addressed along the way. The proposed format is evaluated and shown to exhibit highly promising characteristics. In many respects, it outperforms both posits and takums. As ternary hardware matures, this work represents a crucial step toward unlocking the full potential of real-number computation in ternary systems, laying the groundwork for a new class of number formats designed from the ground up for a new category of next-generation hardware.

cs.ET↗

arXiv2025

Spectral Methods via FFTs in Emerging Machine Number Formats: OFP8, Bfloat16, Posit, and Takum Arithmetics

The Fast Fourier Transform (FFT) is one of the most widely used algorithms in high performance computing, with critical applications in spectral analysis for both signal processing and the numerical solution of partial differential equations (PDEs). These data-intensive workloads are primarily constrained by the memory wall, motivating the exploration of emerging number formats -- such as OFP8 (E4M3 and E5M2), bfloat16, and the tapered-precision posit and takum formats -- as potential alternatives to conventional IEEE 754 floating-point representations. This paper evaluates the accuracy and stability of FFT-based computations across a range of formats, from 8 to 64 bits. Round-trip FFT is applied to a diverse set of images, and short-time Fourier transform (STFT) to audio signals. The results confirm posit arithmetic's strong performance at low precision, with takum following closely behind. Posits show stability issues at higher precisions, while OFP8 formats are unsuitable and bfloat16 underperforms compared to float16 and takum.

math.NA↗

arXiv2025

Numerical Performance of the Implicitly Restarted Arnoldi Method in OFP8, Bfloat16, Posit, and Takum Arithmetics

The computation of select eigenvalues and eigenvectors of large, sparse matrices is fundamental to a wide range of applications. Accordingly, evaluating the numerical performance of emerging alternatives to the IEEE 754 floating-point standard -- such as OFP8 (E4M3 and E5M2), bfloat16, and the tapered-precision posit and takum formats -- is of significant interest. Among the most widely used methods for this task is the implicitly restarted Arnoldi method, as implemented in ARPACK. This paper presents a comprehensive and untailored evaluation based on two real-world datasets: the SuiteSparse Matrix Collection, which includes matrices of varying sizes and condition numbers, and the Network Repository, a large collection of graphs from practical applications. The results demonstrate that the tapered-precision posit and takum formats provide improved numerical performance, with takum arithmetic avoiding several weaknesses observed in posits. While bfloat16 performs consistently better than float16, the OFP8 types are generally unsuitable for general-purpose computations.

math.NA↗

arXiv2025

Streamlining SIMD ISA Extensions with Takum Arithmetic: A Case Study on Intel AVX10.2

Modern microprocessors extend their instruction set architecture (ISA) with Single Instruction, Multiple Data (SIMD) operations to improve performance. The Intel Advanced Vector Extensions (AVX) enhance the x86 ISA and are widely supported in Intel and AMD processors. The latest version, AVX10.2, places a strong emphasis on low-precision, non-standard floating-point formats, including bfloat16 and E4M3/E5M2 float8 (OCP 8-bit Floating Point, OFP8), primarily catering to deep learning applications rather than general-purpose arithmetic. However, as these formats remain within the IEEE 754 framework, they inherit its limitations, introducing inconsistencies and added complexity into the ISA. This paper examines the recently proposed tapered-precision takum floating-point format, which has been shown to offer significant advantages over IEEE 754 and its derivatives as a general-purpose number format. Using AVX10.2 as a case study, the paper explores the potential benefits of replacing the multitude of floating-point formats with takum as a uniform basis. The results indicate a more consistent instruction set, improving readability and flexibility while offering potential for 8- and 16-bit general-purpose SIMD arithmetic.

cs.AR↗

arXiv2024

Integer Representations in IEEE 754, Posit, and Takum Arithmetics

Although not primarily designed for this purpose, floating-point numbers are often used to represent integral values, with some applications explicitly relying on this capability. However, the integral representation properties of IEEE 754 floating-point numbers have not yet been formally investigated. Recently, the bfloat16, posit and takum machine number formats have been proposed as alternatives to IEEE 754, while OCP 8-bit floating point (OFP8) types (E4M3 and E5M2) have been introduced as 8-bit extensions of IEEE 754, albeit with slight deviations from the standard. It is therefore timely to evaluate IEEE 754 and to assess how effectively the new formats fulfil this function in comparison with the standard they aim to replace. This paper presents the first rigorous derivations and proofs of the integral representation capabilities of IEEE 754 floating-point numbers, OFP8, bfloat16, posits, and takums. We examine both the exact number of bits required to represent a given integer and the largest consecutive integer representable with a specified bit width. The results show that OFP8 yields mixed outcomes, bfloat16 generally underperforms, and posits fail to scale effectively, whereas takums consistently match or outperform the other formats, maintaining backward compatibility with IEEE 754.

cs.AR↗

arXiv2024

Evaluation of Bfloat16, Posit, and Takum Arithmetics in Sparse Linear Solvers

Solving sparse linear systems lies at the core of numerous computational applications. Consequently, understanding the performance of recently proposed alternatives to the established IEEE 754 floating-point numbers, such as bfloat16 and the tapered-precision posit and takum machine number formats, is of significant interest. This paper examines these formats in the context of widely used solvers, namely LU, QR, and GMRES, with incomplete LU preconditioning and mixed precision iterative refinement (MPIR). This contrasts with the prevailing emphasis on designing specialized algorithms tailored to new arithmetic formats. This paper presents an extensive and unprecedented evaluation based on the SuiteSparse Matrix Collection -- a dataset of real-world matrices with diverse sizes and condition numbers. A key contribution is the faithful reproduction of SuiteSparse's UMFPACK multifrontal LU factorization and SPQR multifrontal QR factorization for machine number formats beyond single and double-precision IEEE 754. Tapered-precision posit and takum formats show better accuracy in direct solvers and reduced iteration counts in indirect solvers. Takum arithmetic, in particular, exhibits exceptional stability, even at low precision.

math.NA↗

arXiv2024

Design and Implementation of a Takum Arithmetic Hardware Codec

The takum machine number format has been recently proposed as an enhancement over the posit number format, which is considered a promising alternative to the IEEE 754 floating-point standard. Takums retain the useful posit properties, but feature a novel exponent coding scheme that yields more precision for small and large magnitude numbers and a much higher and bounded dynamic range. This paper presents the design and implementation of a hardware codec for both takums (logarithmic number system, LNS) and linear takums (floating-point format). The codec design is emphasised, as it constitutes the primary distinguishing feature compared to logarithmic posits (LNS) and posits (floating-point format), which otherwise share similar internal representations. Furthermore, a novel internal representation for LNS is proposed. The presented takum codec, implemented in VHDL, demonstrates near-optimal scalability and performance on an FPGA. It achieves latency reductions of up to 38% and reduces LUT utilisation up to 50% compared to the best state-of-the-art posit codecs.

cs.AR↗

arXiv2024

Beating Posits at Their Own Game: Takum Arithmetic

Recent evaluations have highlighted the tapered posit number format as a promising alternative to the uniform precision IEEE 754 floating-point numbers, which suffer from various deficiencies. Although the posit encoding scheme offers superior coding efficiency at values close to unity, its efficiency markedly diminishes with deviation from unity. This reduction in efficiency leads to suboptimal encodings and a consequent diminution in dynamic range, thereby rendering posits suboptimal for general-purpose computer arithmetic. This paper introduces and formally proves 'takum' as a novel general-purpose logarithmic tapered-precision number format, synthesising the advantages of posits in low-bit applications with high encoding efficiency for numbers distant from unity. Takums exhibit an asymptotically constant dynamic range in terms of bit string length, which is delineated in the paper to be suitable for a general-purpose number format. It is demonstrated that takums either match or surpass existing alternatives. Moreover, takums address several issues previously identified in posits while unveiling novel and beneficial arithmetic properties.

math.NA↗

arXiv2023

Modern Methods for Signal Analysis: Empirical Mode Decomposition Theory and Hybrid Operator-Based Methods Using B-Splines

This thesis examines the empirical mode decomposition (EMD), a method for decomposing multicomponent signals, from a modern, both theoretical and practical, perspective. The motivation is to further formalize the concept and develop new methods to approach it numerically. The theoretical part introduces a new formalization of the method as an optimization problem over ordered function vector spaces. Using the theory of 'convex-like' optimization and B-splines, Slater-regularity and thus strong duality of this optimization problem is shown. This results in a theoretical justification for the modern null-space-pursuit (NSP) operator-based signal-separation (OSS) EMD-approach for signal decomposition and spectral analysis. The practical part considers the identified strengths and weaknesses in OSS and NSP and proposes a hybrid EMD method that utilizes these modern, but also classic, methods, implementing them in a toolbox called ETHOS (EMD Toolbox using Hybrid Operator-Based Methods and B-splines) and applying them to comparative examples. In the course of this part a new envelope estimation method called 'iterative slope envelope estimation' is proposed.

math.NA↗

arXiv2017

The Unum Number Format: Mathematical Foundations, Implementation and Comparison to IEEE 754 Floating-Point Numbers

This thesis examines a modern concept for machine numbers based on interval arithmetic called 'Unums' and compares it to IEEE 754 floating-point arithmetic, evaluating possible uses of this format where floating-point numbers are inadequate. In the course of this examination, this thesis builds theoretical foundations for IEEE 754 floating-point numbers, interval arithmetic based on the projectively extended real numbers and Unums.

math.NA↗