📖
Wordlength Optimization for Custom Floating-point Systems
Quentin Milot, Mickaël Dardaillon, Justine Bonnot, Daniel Menard

Wordlength Optimization for Custom Floating-point Systems

Science & Mathematics Published 2024 Verification Pending

Overview & Synopsis

Algorithm complexity and problem size explosion during the last few years brought new challenges for hardware implementation of data-oriented applications. In particular, they have caused a large increase in memory, execution time, and power needed. One solution to address those challenges is to reduce data sizes. Custom floating-point is a good candidate that brings a large dynamic range in a compact representation. Thanks to improvement in floating-point units, it is more and more explored as an alternative to fixed-point. This paper proposes an automatic optimization flow to optimize floating-point wordlength for a given quality metric. This flow is generic and can be used for any C or C++ algorithm. Different strategies are proposed to optimize the exponent and mantissa wordlengths. These strategies take advantages of analyzing the data dynamic range to reduce the optimization time. The obtained results show that better implementation cost can be obtained compared to 16-bit floating-point for a same quality.

Indexed from repository: HAL Open Science (hal.science)
Loading document from verified source...
Keyboard Shortcuts: ← / → for pages • + / - for zoom • 0 to fit width Powered by Mozilla PDF.js v3.11

Document Information

Document Title Wordlength Optimization for Custom Floating-point Systems
Author / Organization Quentin Milot, Mickaël Dardaillon, Justine Bonnot, Daniel Menard
Subject & Discipline Multidisciplinary Academic Open Research
Categories Science & Mathematics
Publication Year 2024
License / Rights Open Access / Creative Commons
Resource Verification Listed PDF (Verification Pending)
Original Host Website HAL Open Science (hal.science)
File Format application/pdf (Portable Document Format)
Language EN
Catalog ID DOC-244899
Original Source Link Visit Source Page on hal.science »