一种基于rram的基于原位可转置计算和混合精度的边缘学习CIM设计

IF 2 Q3 COMPUTER SCIENCE, HARDWARE & ARCHITECTURE

IEEE Solid-State Circuits Letters Pub Date : 2025-07-16 DOI:10.1109/LSSC.2025.3589580

Qiumeng Wei;Peng Yao;Dong Wu;Qi Qin;Bin Gao;Qingtian Zhang;Sining Pan;Jianshi Tang;He Qian;Lu Jie;Huaqiang Wu

{"title":"一种基于rram的基于原位可转置计算和混合精度的边缘学习CIM设计","authors":"Qiumeng Wei;Peng Yao;Dong Wu;Qi Qin;Bin Gao;Qingtian Zhang;Sining Pan;Jianshi Tang;He Qian;Lu Jie;Huaqiang Wu","doi":"10.1109/LSSC.2025.3589580","DOIUrl":null,"url":null,"abstract":"Edge computing devices demand efficient transposable computing architectures to enable on-chip learning, necessitating novel hardware designs that balance performance and flexibility. We present an RRAM-based compute-in-memory macro capable of supporting both in-situ forward and backward propagation operations. The design incorporates an orthogonal-WL array structure with weight-level parallelism adjustment and a precision-driven input mechanism to enable flexible transposable computing. Additionally, optimized ADCs provide high throughput while maintaining area efficiency. The work exhibits a SOTA normalized area efficiency of 126.7 TOPS/mm2/bit, an energy efficiency of 2348.96 TOPS/W/bit, and a storage density of 4.84 Mb/mm2.","PeriodicalId":13032,"journal":{"name":"IEEE Solid-State Circuits Letters","volume":"8 ","pages":"205-208"},"PeriodicalIF":2.0000,"publicationDate":"2025-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"A RRAM-Based CIM Design With in-Situ Transposable Computing and Hybrid-Precision Scheme for Edge Learning\",\"authors\":\"Qiumeng Wei;Peng Yao;Dong Wu;Qi Qin;Bin Gao;Qingtian Zhang;Sining Pan;Jianshi Tang;He Qian;Lu Jie;Huaqiang Wu\",\"doi\":\"10.1109/LSSC.2025.3589580\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Edge computing devices demand efficient transposable computing architectures to enable on-chip learning, necessitating novel hardware designs that balance performance and flexibility. We present an RRAM-based compute-in-memory macro capable of supporting both in-situ forward and backward propagation operations. The design incorporates an orthogonal-WL array structure with weight-level parallelism adjustment and a precision-driven input mechanism to enable flexible transposable computing. Additionally, optimized ADCs provide high throughput while maintaining area efficiency. The work exhibits a SOTA normalized area efficiency of 126.7 TOPS/mm2/bit, an energy efficiency of 2348.96 TOPS/W/bit, and a storage density of 4.84 Mb/mm2.\",\"PeriodicalId\":13032,\"journal\":{\"name\":\"IEEE Solid-State Circuits Letters\",\"volume\":\"8 \",\"pages\":\"205-208\"},\"PeriodicalIF\":2.0000,\"publicationDate\":\"2025-07-16\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"IEEE Solid-State Circuits Letters\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://ieeexplore.ieee.org/document/11082282/\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q3\",\"JCRName\":\"COMPUTER SCIENCE, HARDWARE & ARCHITECTURE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Solid-State Circuits Letters","FirstCategoryId":"1085","ListUrlMain":"https://ieeexplore.ieee.org/document/11082282/","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"COMPUTER SCIENCE, HARDWARE & ARCHITECTURE","Score":null,"Total":0}

引用次数: 0

摘要

边缘计算设备需要高效的可转置计算架构来实现片上学习，这就需要新颖的硬件设计来平衡性能和灵活性。我们提出了一个基于rram的内存中计算宏，能够支持原位正向和反向传播操作。该设计结合了具有重量级并行调整的正交wl阵列结构和精度驱动的输入机制，以实现灵活的转座计算。此外，优化的adc在保持区域效率的同时提供高吞吐量。SOTA归一化面积效率为126.7 TOPS/mm2/bit，能量效率为2348.96 TOPS/W/bit，存储密度为4.84 Mb/mm2。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

查看原文本刊更多论文

A RRAM-Based CIM Design With in-Situ Transposable Computing and Hybrid-Precision Scheme for Edge Learning

Edge computing devices demand efficient transposable computing architectures to enable on-chip learning, necessitating novel hardware designs that balance performance and flexibility. We present an RRAM-based compute-in-memory macro capable of supporting both in-situ forward and backward propagation operations. The design incorporates an orthogonal-WL array structure with weight-level parallelism adjustment and a precision-driven input mechanism to enable flexible transposable computing. Additionally, optimized ADCs provide high throughput while maintaining area efficiency. The work exhibits a SOTA normalized area efficiency of 126.7 TOPS/mm2/bit, an energy efficiency of 2348.96 TOPS/W/bit, and a storage density of 4.84 Mb/mm2.

求助全文

通过发布文献求助，成功后即可免费获取论文全文。去求助

来源期刊

IEEE Solid-State Circuits Letters Engineering-Electrical and Electronic Engineering

CiteScore

4.30

自引率

3.70%

发文量