Takuto Nishimura, Yuya Ichikawa, Akira Goda, Naoko Misawa, C. Matsui, Ken Takeuchi
{"title":"基于随机计算的dnn内存计算(SC CiM)体系结构及非易失性存储器容错和缺陷容错的分层评估","authors":"Takuto Nishimura, Yuya Ichikawa, Akira Goda, Naoko Misawa, C. Matsui, Ken Takeuchi","doi":"10.1109/IMW56887.2023.10145982","DOIUrl":null,"url":null,"abstract":"A Stochastic Computing-based Computation-inMemory architecture (SC CiM) with non-volatile memory (NVM) stochastic weights has been proposed. Both of the input and weight values are converted into stochastic bit streams. The multiply-and-accumulate (MAC) operation of the SC multiplication is performed in memory. The computational accuracy is characterized for Resnet-56 and CIFAR-10, with the bit error rate (BER) range from 1$0^{-4}$ to 1$0^{-2}$ representing the actual NVM characteristics. The effects of NVM BER are analyzed hierarchically at all computational layers (bit streams, MAC calculation and DNN inference) and compared both qualitatively and quantitatively with the conventional CiM. The results show the excellent robustness of the proposed SC CiM architecture in the wide range of BER. The tolerance to the manufacturing defects is even better than that of the conventional CiM. Furthermore, the desired weight distributions are discussed by exploiting the unique behaviors against NVM BER in the SC CiM.","PeriodicalId":153429,"journal":{"name":"2023 IEEE International Memory Workshop (IMW)","volume":"25 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2023-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Stochastic Computing-based Computation-in-Memory (SC CiM) Architecture for DNNs and Hierarchical Evaluations of Non-volatile Memory Error and Defect Tolerance\",\"authors\":\"Takuto Nishimura, Yuya Ichikawa, Akira Goda, Naoko Misawa, C. Matsui, Ken Takeuchi\",\"doi\":\"10.1109/IMW56887.2023.10145982\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"A Stochastic Computing-based Computation-inMemory architecture (SC CiM) with non-volatile memory (NVM) stochastic weights has been proposed. Both of the input and weight values are converted into stochastic bit streams. The multiply-and-accumulate (MAC) operation of the SC multiplication is performed in memory. The computational accuracy is characterized for Resnet-56 and CIFAR-10, with the bit error rate (BER) range from 1$0^{-4}$ to 1$0^{-2}$ representing the actual NVM characteristics. The effects of NVM BER are analyzed hierarchically at all computational layers (bit streams, MAC calculation and DNN inference) and compared both qualitatively and quantitatively with the conventional CiM. The results show the excellent robustness of the proposed SC CiM architecture in the wide range of BER. The tolerance to the manufacturing defects is even better than that of the conventional CiM. Furthermore, the desired weight distributions are discussed by exploiting the unique behaviors against NVM BER in the SC CiM.\",\"PeriodicalId\":153429,\"journal\":{\"name\":\"2023 IEEE International Memory Workshop (IMW)\",\"volume\":\"25 1\",\"pages\":\"0\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2023-05-01\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"2023 IEEE International Memory Workshop (IMW)\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1109/IMW56887.2023.10145982\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"2023 IEEE International Memory Workshop (IMW)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/IMW56887.2023.10145982","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Stochastic Computing-based Computation-in-Memory (SC CiM) Architecture for DNNs and Hierarchical Evaluations of Non-volatile Memory Error and Defect Tolerance
A Stochastic Computing-based Computation-inMemory architecture (SC CiM) with non-volatile memory (NVM) stochastic weights has been proposed. Both of the input and weight values are converted into stochastic bit streams. The multiply-and-accumulate (MAC) operation of the SC multiplication is performed in memory. The computational accuracy is characterized for Resnet-56 and CIFAR-10, with the bit error rate (BER) range from 1$0^{-4}$ to 1$0^{-2}$ representing the actual NVM characteristics. The effects of NVM BER are analyzed hierarchically at all computational layers (bit streams, MAC calculation and DNN inference) and compared both qualitatively and quantitatively with the conventional CiM. The results show the excellent robustness of the proposed SC CiM architecture in the wide range of BER. The tolerance to the manufacturing defects is even better than that of the conventional CiM. Furthermore, the desired weight distributions are discussed by exploiting the unique behaviors against NVM BER in the SC CiM.