Skip to Main content Skip to Navigation

Analyse d'erreurs pour les systèmes utilisant des calculs approximés

Abstract : Approximate Computing is an energy- aware computing technique that relies on the exploitation of the tolerance to imprecision of an application. Developed to face the end of Moore’s law, it answers the growing demand in computing capacity. Approximation techniques have been proposed at different abstraction levels, from circuit to system level. This thesis focuses on the development of methods and tools to quickly evaluate the impact of different Approximate Computing techniques on the application quality metric. The study of the induced errors is critical to use approximations in the industry. Approximate Computing techniques have been considered at two different levels, the hardware level with the study of inexact arithmetic operators and the data level with the study of fixed-point arithmetic. First, efficient simulation-based characterization methods have been proposed to derive statistics on the errors induced by the considered approximation. Inferential statistics have been proposed to reduce the time for error characterization. The proposed characterization methods are based on adaptive simulations and statistically characterizes the approximation error according to user-defined confidence requirements. Then, the obtained error metrics are linked with the application quality metric. For inexact operators, a simulator has been proposed for the approximation design space exploration process to select the best approximation for the considered application. For fixed-point arithmetic, the proposed error model has been implemented in a fixed-point refinement algorithm to determine the optimized word-lengths of the internal variables in an application. The results of this thesis are proposing concrete methods to ease the implementation of Approximate Computing in industrial applications, speeding up state-of-the-art methods from one to three orders of magnitude.
Complete list of metadata

Cited literature [373 references]  Display  Hide  Download
Contributor : Abes Star :  Contact Connect in order to contact the contributor
Submitted on : Tuesday, April 6, 2021 - 4:50:10 PM
Last modification on : Wednesday, October 27, 2021 - 4:18:34 AM
Long-term archiving on: : Monday, February 15, 2021 - 6:03:12 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03005325, version 1


Justine Bonnot. Analyse d'erreurs pour les systèmes utilisant des calculs approximés. Traitement du signal et de l'image [eess.SP]. INSA de Rennes, 2019. Français. ⟨NNT : 2019ISAR0008⟩. ⟨tel-03005325⟩



Record views


Files downloads