In this work, a high-speed and energy-efficient comparator-based Near-Sensor Local Binary Pattern accelerator architecture (NS-LBP) is proposed to execute a novel local binary pattern deep neural network. First, inspired by recent LBP networks, we design an approximate, hardware-oriented, and multiply-accumulate (MAC)-free network named Ap-LBP for efficient feature extraction, further reducing the computation complexity. Then, we develop NS-LBP as a processing-in-SRAM unit and a parallel in-memory LBP algorithm to process images near the sensor in a cache, remarkably reducing the power consumption of data transmission to an off-chip processor. Our circuit-to-application co-simulation results on MNIST and SVHN data-sets demonstrate minor accuracy degradation compared to baseline CNN and LBP-network models, while NS-LBP achieves 1.25 GHz and energy-efficiency of 37.4 TOPS/W. NS-LBP reduces energy consumption by 2.2x and execution time by a factor of 4x compared to the best recent LBP-based networks.
翻译:在这项工作中,拟议建立一个高速和节能的参照国近传感器地方二元模式加速器(NS-LBP)结构,以实施一个新的本地二元模式深神经网络。首先,在最新的LBP网络的启发下,我们设计了一个叫Ap-LBP的近似、硬件导向和乘积(MAC)无网络,以高效地提取地物,进一步降低计算复杂性。然后,我们开发了NS-LBP,作为SRAM的处理单位和平行的模擬LBP算法,在传感器附近在一个缓存处处理图像,显著减少数据传输到离芯处理器的动力消耗量。我们关于MNISST和SVHN数据集的电路到应用联合模拟结果显示,与基线CNN和LBP网络模型相比,其精确度略有下降,而NS-LBP达到1.25千兆赫,能源效率为37.4 TPS/W。NS-LBP将能源消耗量减少2.2倍,执行时间减少系数,比最近最佳的LBPP网络减少4倍。