Qirui Zhang, Hyochan An, Zichen Fan, Zhehong Wang, Ziyun Li, Guanru Wang, Hun-Seok Kim, D. Blaauw, D. Sylvester
{"title":"A 22nm 3.5TOPS/W Flexible Micro-Robotic Vision SoC with 2MB eMRAM for Fully-on-Chip Intelligence","authors":"Qirui Zhang, Hyochan An, Zichen Fan, Zhehong Wang, Ziyun Li, Guanru Wang, Hun-Seok Kim, D. Blaauw, D. Sylvester","doi":"10.1109/vlsitechnologyandcir46769.2022.9830340","DOIUrl":null,"url":null,"abstract":"We present a highly flexible micro-robotic vision SoC featuring a hybrid Processing Element (PE) for efficient processing of both Convolutional Neural Network (CNN) and non-CNN vision tasks with 2MB embedded MRAM for retentive fully-on-chip weight storage. Fabricated in 22nm, the design achieves 0.22nJ/pix for Harris corner detection (a non-CNN vision task) and 3.5TOPS/W (INT16) for CNN, a 60% efficiency improvement over state-of-the-art NVM-based NN ASICs.","PeriodicalId":332454,"journal":{"name":"2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits)","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2022-06-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/vlsitechnologyandcir46769.2022.9830340","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
We present a highly flexible micro-robotic vision SoC featuring a hybrid Processing Element (PE) for efficient processing of both Convolutional Neural Network (CNN) and non-CNN vision tasks with 2MB embedded MRAM for retentive fully-on-chip weight storage. Fabricated in 22nm, the design achieves 0.22nJ/pix for Harris corner detection (a non-CNN vision task) and 3.5TOPS/W (INT16) for CNN, a 60% efficiency improvement over state-of-the-art NVM-based NN ASICs.