IP2_2 Interactive Presentations
Date: Tuesday, 02 February 2021
Time: 17:00 - 17:30 CET
Virtual Conference Room: https://virtual21.date-conference.com/meetings/virtual/4taqRWQpMc3eRmd6X
Interactive Presentations run simultaneously during a 30-minute slot. Additionally, each IP paper is briefly introduced in a one-minute presentation in a corresponding regular session
|IP2_2.1||RECEPTIVE-FIELD AND SWITCH-MATRICES BASED RERAM ACCELERATOR WITH LOW DIGITAL-ANALOG CONVERSION FOR CNNS
Xun Liu, North China University of Technology, CN
Yingxun Fu1, Xun Liu2, Jiwu Shu3, Zhirong Shen4, Shiye Zhang1, Jun Wu1 and Li Ma1
1North China University of Technology, CN; 2North China Univercity of Technology, CN; 3Tsinghua University, CN; 4Xiamen University, CN
Process-in-Memory (PIM) based accelerator becomes one of the best solutions for the execution of convolution neural networks (CNN). Resistive random access memory (ReRAM) is a classic type of non-volatile random-access memory, which is very suitable for implementing PIM architectures. However, existing ReRAM-based accelerators mainly consider to improve the calculation efficiency, but ignore the fact that the digital-analog signal conversion process spends a lot of energy and executing time. In this paper, we propose a novel ReRAM-based accelerator named Receptive-Field and Switch-Matrices based CNN Accelerator (RFSM). In RFSM, we first propose a receptive-field based convolution strategy to analyze the data relationships, and then gives a dynamic and configurable crossbar combination method to reduce the digital-analog conversion operations. The evaluation result shows that, compared to existing works, RFSM gains up to 6.7x higher speedup and 7.1x lower energy consumption.
|IP2_2.2||AN ON-CHIP LAYER-WISE TRAINING METHOD FOR RRAM BASED COMPUTING-IN-MEMORY CHIPS
Yiwen Geng, Tsinghua University, CN
Yiwen Geng, Bin Gao, Qingtian Zhang, Wenqiang Zhang, Peng Yao, Yue Xi, Yudeng Lin, Junren Chen, Jianshi Tang, Huaqiang Wu and He Qian, Institute of Microelectronics, Tsinghua University, CN
RRAM based computing-in-memory (CIM) chips have shown great potentials to accelerate deep neural networks on edge devices by reducing data transfer between the memory and the computing unit. However, due to the non-ideal characteristics of RRAM, the accuracy of the neural network on the RRAM chip is usually lower than the software. Here we propose an on-chip layer-wise training (LWT) method to alleviate the adverse effect of RRAM imperfections and improve the accuracy of the chip. Using a locally validated dataset, LWT can reduce the communication between the edge and the cloud, which benefits for the personalized data privacy. The simulation results on the CIFAR-10 dataset show that the LWT method can improve the accuracy of VGG-16 and ResNet-18 more than 5% and 10%, respectively, with only 30% operations and 35% buffer compared with the back-propagation method. Moreover, the pipe-LWT method is presented to improve the throughput by three times further.