Characterization of Drain Current Variations in FeFETs for PIM-based DNN Accelerators
Characterization of Drain Current Variations in FeFETs for PIM-based DNN Accelerators
复制标题
DOI：
10.1109/aicas51828.2021.9458437
复制
发表时间：
2021-06
期刊：
影响因子：
--
通讯作者：
N. Miller;Zheng Wang;Saurabh Dash;A. Khan;S. Mukhopadhyay
中科院分区：
文献类型：
--
作者：
N. Miller;Zheng Wang;Saurabh Dash;A. Khan;S. Mukhopadhyay
我们分析了28 nm高K金属栅铁电FET器件的漏电流$（I_{DS}）$变化对基于FeFET的存储器处理（PIM）深度神经网络（DNN）加速器的影响。IDS的非正态变化是由于在不同读取频率下对具有不同沟道尺寸的FeFET器件进行重复读取操作而观察到的。使用测得的电流分布进行的器件-电路协同分析显示，当使用LeNET-5 DNN模型对Fashion-MNIST数据集进行分类时，基于FeFET的PIM平台的准确度下降了1%至3%。这种精度下降可以用变化感知训练方法完全恢复，表明在许多读取周期内的单个FeFET器件电流变化对DNN加速器的设计并不禁止。
We analyze the impact of drain current $(I_{DS})$ variation in 28 nm high-K metal-gate Ferroelectric FET devices on FeFET-based processing-in-memory (PIM) deep neural network (DNN) accelerators. Non-Normal variation in IDS is observed due to repeated read operation on FeFET devices with different channel dimensions at various read frequencies. Device-circuit co-analysis using the measured current distribution shows a 1 to 3 percent accuracy degradation of an FeFET-based PIM platform when classifying the Fashion-MNIST dataset with the LeNET-5 DNN model. This accuracy drop can be fully recovered with variation-aware training methods, showing that individual FeFET device current variation over many read cycles is not prohibitive to the design of DNN accelerators.