CASSIS: Characterization with Adaptive Sample-Size Inferential Statistics Applied to Inexact Circuits

Justine Bonnot (1) , Vincent Camus (2) , Karol Desnos (1) , Daniel Menard (1)

1 IETR - Institut d'Électronique et des Technologies du numéRique
2 EPFL - Ecole Polytechnique Fédérale de Lausanne

To design faster and more energy-efficient systems, numerous inexact arithmetic operators have been proposed, generally obtained by modifying the logic structure of conventional circuits. However, as the quality of service of an application has to be ensured, these operators need to be precisely characterized to be usable in commercial or real-life applications. The characterization of inexact operators is commonly achieved with exhaustive or random bit-accurate gate-level simulations. However, for high word lengths, the time and memory required for such simulations become prohibitive. Besides, when simulating a random sample, no confidence information is given on the precision of the characterization. To overcome these limitations, CASSIS, a new characterization method for inexact operators is proposed. By exploiting statistical properties of the approximation error, the number of simulations needed for precise characterization is drastically reduced. From user-defined confidence requirements, the proposed method computes the minimal number of simulations to obtain the desired accuracy on the characterization. For 32-bit adders, the CASSIS method reduces the number of simulations needed up to a few tens of thousands points.

Domaines: Traitement du signal et de l'image [eess.SP]; Architectures Matérielles [cs.AR]; Modélisation et simulation; Systèmes embarqués; Mathématique discrète [cs.DM]

Justine Bonnot, Vincent Camus, Karol Desnos, Daniel Menard. CASSIS: Characterization with Adaptive Sample-Size Inferential Statistics Applied to Inexact Circuits. EUSIPCO: EUropean SIgnal Processing Conference, Sep 2018, Rome, Italy. ⟨10.23919/eusipco.2018.8553451⟩. ⟨hal-01879628v2⟩