Convolutional neural networks (CNNs) have proven to be a disruptive technology in most vision, speech and image processing tasks. Given their ubiquitous acceptance, the research community is investing a lot of time and resources on deep neural networks. Custom hardware such as ASICs are proving to be extremely worthy platforms for running such programs. However, the ever-increasing complexity of these algorithms poses challenges in achieving real-time performance. Specifically, CNNs have prohibitive costs in terms of computation time, throughput, latency, storage space, memory bandwidth, and power consumption.

Hence, in the last 5 years, a lot of work has been done by the scientific community to mitigate these costs. Researchers have primarily focused on reducing the computation time, the number of computations, the memory access time, and the size of the memory footprint. In this survey paper, we propose a novel taxonomy to classify prior work, and describe some of the key contributions in these areas in detail.

在大多数视觉，语音和图像处理任务中，卷积神经网络（CNN）已被证明是一种破坏性技术。鉴于它们的普遍接受，研究界正在深度神经网络上投入大量时间和资源。事实证明，像ASIC这样的自定义硬件是运行这些程序的非常有价值的平台。然而，这些算法的不断增加的复杂性对实现实时性能提出了挑战。具体而言，CNN在计算时间，吞吐量，等待时间，存储空间，内存带宽和功耗方面具有高昂的成本。

因此，在过去的5年中，科学界为减轻这些成本已经做了很多工作。研究人员主要集中在减少计算时间，计算次数，内存访问时间和内存占用空间方面。在这份调查报告中，我们提出了一种新颖的分类法来对先前的工作进行分类，并详细描述这些领域中的一些关键贡献。