A Method for Swift Selection of Appropriate Approximate Multipliers for CNN Hardware Accelerators

As convolutional neural networks (CNNs) gain traction for embedded device implementation, there's a burgeoning interest in approximate computing technologies for increasing hardware efficiency. Most of the works in this field focus on proposing novel approximate hardware units and structures, but structured guidance for selecting optimal approximate calculation techniques for CNN accelerators remains scant. This paper introduces a novel error injection technique, leveraging the error rate matrix of approximate multipliers (AxMs), called Error Matrix Based Error Injected (EMEI). This facilitates the swift selection of appropriate AxMs for each PE in the CNN hardware accelerator. In addition, this approach is applied to a MobileNetV2-based CNN model on the CIFAR-10 dataset to demonstrate the performance. Experimental results show that our method adeptly optimises hardware resources by combining AxMs with different accuracy levels while ensuring accuracy. This innovation paves the way for streamlined CNN accelerator designs in embedded systems.