Algorithm Level Timing Speculation for Convolutional Neural Network Accelerators
Spéculation temporelle algorithmique pour accélérateurs de réseaux de neuro
Résumé
In this paper, we propose a technique for improving the efficiency of hardware
accelerators based on timing speculation (overclocking) and fault tolerance. We augment the
accelerator with a lightweight error detection mechanism to protect against timing errors, enabling
aggressive timing speculation. We demonstrate the validity of our approach for the convolution
layers in Convolutional Neural Networks (CNN). We present an implementation of a fault-tolerant
CNN accelerator combined with the lightweight error detection for convolution layers. The error
detection mechanism we have developed works at the algorithm level, based on algebraic properties
of the computation, allowing the full implementation to be realized using High-Level Synthesis
tools. We use a set of Zybo boards to experimentally demonstrate that overclocking boosts the
frequency by 17-36% with low chances of error, and that the infrequent errors can be detected with
a negligible overhead (only 1000 LUTs).
Origine : Fichiers produits par l'(les) auteur(s)
Loading...