A review on hardware accelerators for convolutional neural network-based inference engines: Strategies for performance and energy-efficiency enhancement

In time-critical & safety-critical image classification applications, Convolutional Neural Networks (CNNs) based Inference Engines (IEs) are preferred and required to be fast, accurate, and cost-effective to meet the market demands. The self-feature extraction capabilities use millions of parameters and neurons in the stack of layers with restricted processing time. This paper reviews strategies applied in Hardware-based image classification CNN inference engines. The acceleration strategies are (1) Arithmetic Logic Unit (ALU)-based, (2) Data flow-based, and (3) Sparsity-based are considered here. Considering benchmark accuracy, the 16-bit mixed fixed/floating point could provide 99 % and 3.75 times more performance than Half-precision floating point in an application-specific CNN model. Feeding 2-dimensional or 3-dimensional data frames to the CNN layers would reuse the data. It optimizes the volume of memory usage and improves the efficiency of the processor array. The pruning of zero/near-zero valued Input Feature Maps (IFMs) and weights leads to sparsity in the data fed to the different layers. Therefore, data compression strategies and skipping the trivial computation (zero skipping approach) would reduce the complexity of the controller. There is a benchmark performance improvement of 1.17 times and 6.2 times in power efficiency compared to dense architecture. Minimizing the complexity of indexing and load balancing controller would improve the performance further.