… Deep convolutional neural networks (CNNs) have recently shown very high accuracy in a
… Given the high computational demands of CNNs, custom hardware accelerators are vital for …

… , executing them on accelerators designed for 2D CNNs … In this paper, we present a
survey of hardware accelerators … convolution and not those that perform only 2D convolution …

… convolutional neural network accelerators, and proposes future research trends. We summarize
the mainstream CNN accelerators, … In the future, accelerators for training will become the …

… layers of neural networks. Convolutional neural networks (CNNs) are deep neural networks …
consist of four different types of layers: convolutional, classifier, pooling, and local response/…

… total execution cycles of each optimized convolutional layer. With this analysis, CNN
accelerator with unified unroll factors across convolutional layers are selected in our experiments. …

… This paper presents the Sparse CNN (SCNN) accelerator architecture for inference in
convolutional neural networks. SCNN exploits sparsity in both weights and activations using the …

… Just as DNN accelerators for specific classes of neural networks are emerging [22] [34],
so too do we expect to see further specialization of CNN accelerators for actual deployment. …

… accelerator utilization with more external memory bandwidth is bad for energy. Although
on-chip memories can help to increase accelerator … a memory-centric accelerator to improve …

… accelerators are reported [10]–[12]. Most of reported CNN accelerators only focus on accelerating
the convolution … function, which is a common layer in the CNN network. In [10], a CNN …

… In this work, we present a HW accelerator optimized for BinaryConnect CNNs that achieves
… Our accelerator outperforms state-of-the-art performance in terms of ASIC energy efficiency …