Online Testing of a Row-Stationary Convolution Accelerator
Authors:
- 24 May 2021
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-2
Abstract
Convolution accelerators used in safety-critical systems require robustness and resilience to runtime faults. In this paper, we design a processing element for the convolution accelerators that support a row-stationary dataflow. The proposed processing element is equipped with an embedded built-in mechanism for online testing of its computational parts without degrading performance by taking advantage of CNN data sparsity. Furthermore, the proposed architecture uses a low-cost error-detecting code to verify PE interconnections and local memory. The result obtained using post-synthesis simulations reveal, on average, 34% and 30% overheads on the power and area, respectively. This is in exchange for equipping the architecture with components that detect and locate faulty processing elements at runtime.Keywords
This publication has 6 references indexed in Scilit:
- A Perspective on Test Methodologies for Supervised Machine Learning AcceleratorsIEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2019
- On the Resilience of RTL NN Accelerators: Fault Characterization and Mitigation2018
- Using Dataflow to Optimize Energy Efficiency of Deep Neural Network AcceleratorsIEEE Micro, 2017
- EyerissACM SIGARCH Computer Architecture News, 2016
- ImageNet Large Scale Visual Recognition ChallengeInternational Journal of Computer Vision, 2015
- Hardware Redundancy2013