Impact of Memory Voltage Scaling on Accuracy and Resilience of Deep Learning Based Edge Devices

Editor’s note: As more and more artificial intelligence capabilities are deployed onto resource-constrained devices, designers explore several techniques in an effort to boost energy efficiency. Two techniques are quantization and voltage scaling. Quantization aims to reduce the memory footprint, as well as the memory accesses. Therefore, this article explores the resilience of convolutional neural networks to SRAM-based errors and analyzes the relative energy impact of quantization and voltage scaling, when used separately and jointly. —Theocharis Theocharides, University of Cyprus —Muhammad Shafique, Technische Universität Wien