所属单位：The Hong Kong University of Science and Technology (HKUST)

发表刊物：Integration, the VLSI Journal（CCF-C类）

项目来源：This work is partially supported by Hong Kong Research Grant Council (RGC) under Grant 619813.

关键字：Convolutional Neural Network (CNN), Resistive RAM, data encoding, dynamic quantization, computation saving

摘要：Although the CMOS-based CNN accelerators have achieved impressive progress, the memory wall issue and the high power density are still the major barriers for substantial improvement in energy efficiency and throughput. As an attractive alternative, recently the Resistive RAM-based accelerators have delivered significant breakthroughs by leveraging the in-situ computation. However, there are still some challenges, including the high computation complexity and the large energy overhead at the analog/digital interfacing circuits. In this work, we take advantage of the subword-based computation in the Resistive RAM-based accelerator to achieve energy saving and speedup. First, an encoding method is proposed for the weights and activations to reduce the energy consumption of the in-situ computation and the resolution requirement of ADC. Then the resolution of ADC is further optimized based on the distribution of the subword computation results. Furthermore, a dynamic quantization scheme is proposed to skip 67%–87% of the subword computations which outperforms the conventional quantization schemes. We fully investigate the influences of the encoding scheme and the layer-wise quantization range scaling on the performance of dynamic quantization. Finally, we demonstrate the effectiveness of the proposed algorithms under different hardware configurations and network complexities. A dedicated architecture, SubMac, is proposed to implement the above schemes. Experimental results show that the energy efficiency and the throughput are improved by 2.8–5.7 and 2.5–7.9 times, respectively, when compared with the state-of-the-art Resistive RAM-based accelerators.

备注：中国计算机学会 CCF-C 类

合写作者：Jingbo Jiang,Jingyang Zhu,Chi-Ying Tsui

第一作者：Xizi Chen

是否译文：否

发表时间：2019-01-01

收录刊物：SCI