Related Experiment Video
Updated: Jan 1, 2026
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
968
SalvageDNN: salvaging deep neural network accelerators with permanent faults through saliency-driven fault-aware
Muhammad Abdullah Hanif1, Muhammad Shafique1
1Technische Universität Wien (TU Wien), Vienna, Austria.
Summary
SalvageDNN enhances deep neural network (DNN) accelerator reliability by mapping DNN parameters onto faulty hardware. This methodology improves manufacturing yield and reduces costs for DNN accelerators.
Area of Science:
- Computer Engineering
- Artificial Intelligence
- Hardware Acceleration
Background:
- Deep neural networks (DNNs) are widely used in data processing and predictive analysis.
- Manufacturing defects in DNN accelerators necessitate improved yield and cost reduction.
Purpose of the Study:
- To present SalvageDNN, a methodology for reliable DNN execution on hardware accelerators with permanent faults.
- To improve the manufacturing yield of DNN accelerators.
Main Methods:
- Fault-aware mapping of DNN components onto hardware accelerators.
- Leveraging DNN parameter saliency and hardware fault maps.
- Modifying systolic array designs for enhanced yield and reliability.
Main Results:
- SalvageDNN enables reliable DNN execution on fault-tolerant hardware.
- The methodology improves accelerator manufacturing yield.
- Negligible overheads in area, power/energy, and performance were observed.
Conclusions:
- SalvageDNN offers a viable solution for producing cost-effective and reliable DNN hardware.
- The approach addresses the challenge of permanent faults in DNN accelerators.
- This work contributes to harmonizing energy-autonomous computing and intelligence.
