2021/07/09 by Z. Fahimi, Zahra Fahimi, M. R. Mahmoodi +15 · 1 citation
Computer Science · Engineering · #Advanced Memory and Neural Computing #CCD and CMOS Imaging Sensors #Emerging Technologies (cs.ET) #FOS: Computer and information sciences #Ferroelectric and Negative Capacitance Devices #cs.ET
paper · pdf · doi:10.48550/arxiv.2107.04236
arxiv created 2021/07/09 · openalex publication_date 2021/07/09 · arxiv updated 2021/07/12 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
The progress in neuromorphic computing is fueled by the development of novel nonvolatile memories capable of storing analog information and implementing neural computation efficiently. However, like most other analog circuits, these devices and circuits are prone to imperfections, such as temperature dependency, noise, tuning error, etc., often leading to considerable performance degradation in neural network implementations. Indeed, imperfections are major obstacles in the path of further progress and ultimate commercialization of these technologies. Hence, a practically viable approach should be developed to deal with these nonidealities and unleash the full potential of nonvolatile memories in neuromorphic systems. Here, for the first time, we report a comprehensive characterization of critical imperfections in two analog-grade memories, namely passively-integrated memristors and redesigned eFlash memories, which both feature long-term retention, high endurance, analog storage, low-power operation, and compact nano-scale footprint. Then, we propose a holistic approach that includes modifications in the training, tuning algorithm, memory state optimization, and circuit design to mitigate these imperfections. Our proposed methodology is corroborated on a hybrid software/experimental framework using two benchmarks: a moderate-size convolutional neural network and ResNet-18 trained on CIFAR-10 and ImageNet datasets, respectively. Our proposed approaches allow 2.5x to 9x improvements in the energy consumption of memory arrays during inference and sub-percent accuracy drop across 25-100 C temperature range. The defect tolerance is improved by >100x, and a sub-percent accuracy drop is demonstrated in deep neural networks built with 64x64 passive memristive crossbars featuring 25% normalized switching threshold variations.