Efficient Dual Modular Redundancy Units of Convolutional Neural Networks Based on Quantization

Edge devices powered by convolutional neural networks (CNNs) are being rapidly developed and utilized in image recognition, classification, and generation. As these devices become more widespread in mission-critical systems, fault tolerance is critical to ensure reliable behavior. Dual Modular Redundancy (DMR) is a commonly used system that compares two outputs by arraying the same modules. Although the DMR offers high reliability, it doubles both the circuit area and power consumption. This makes the DMR unsuited for resource-constrained edge devices, where CNNs already consume significant hardware resources and energy. To address the problem, this paper proposes an approximated DMR composed of an original CNN module and a low-bit quantized CNN module generated by linear quantization. This method detects faults by calculating the mean absolute difference (MAD) between the outputs of the original and low-bit quantized modules. In the experiments, the approximated DMR is evaluated in terms of fault detection accuracy using software simulation. An 8-bit ResNet-20 trained with CIFAR-10 is used, and the accuracy is measured by the Area Under the Curve (AUC). The circuits of the proposed methods are also implemented on a Field Programmable Gate Array (FPGA). Their resource and power reduction rates are evaluated compared to those of conventional DMR. We demonstrate that the approximated DMR consisting of 8-bit and 3-bit quantized modules achieves 35% Look-Up Tables (LUTs) reduction, 24% flip-flop reduction, 30% BRAM reduction, and 33% power reduction compared with the conventional DMR while achieving high fault detection performance.