基于16K电流的8T SRAM内存中计算宏,具有去耦读/写和1-5位列ADC

Chengshuo Yu, Taegeun Yoo, T. T. Kim, K. Chai, Bongjin Kim
{"title":"基于16K电流的8T SRAM内存中计算宏,具有去耦读/写和1-5位列ADC","authors":"Chengshuo Yu, Taegeun Yoo, T. T. Kim, K. Chai, Bongjin Kim","doi":"10.1109/CICC48029.2020.9075883","DOIUrl":null,"url":null,"abstract":"A novel 8T SRAM -based bitcell is proposed for current-based compute-in-memory dot-product operations. The proposed bitcell with two extra NMOS transistors (vs. standard 6T SRAM) decouples SRAM read and write operation. A 128×128 8T SRAM bitcell array is built for processing a vector-matrix multiplication (or parallel dot-products) with 64x binary (0 or 1) inputs, 64×128 binary (-1 or +1) weights, and 128x 1-5bit outputs. Each column (i.e. neuron) of the proposed SRAM compute-in-memory macro consists of 64x bitcells for dot-product, 32x bitcells for ADC, and 32x bitcells for calibration. The column-based neuron minimizes the ADC overhead by reusing a sense amplifier for SRAM read. The column-wise ADC converts the analog dot-product results to N-bit output codes (N=1 to 5) by sweeping reference levels using replica bitcells for 2N-1 cycles for each conversion. Monte-Carlo simulations and test-chip measurement results have verified both linearity and process variation. The largest variation (σ=2.48%) results in the MNIST classification accuracy of 96.2% (i.e. 0.4% lower than a baseline with no variation). A test-chip is fabricated using 65nm, and the 16K SRAM bitcell array occupies 0.055mm2. The energy efficiency of the 1bit operation is 490-to-15.8TOPS/W at 1-5bit ADC mode using 0.45/0.8V core supply and 200MHz.","PeriodicalId":409525,"journal":{"name":"2020 IEEE Custom Integrated Circuits Conference (CICC)","volume":"82 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2020-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"41","resultStr":"{\"title\":\"A 16K Current-Based 8T SRAM Compute-In-Memory Macro with Decoupled Read/Write and 1-5bit Column ADC\",\"authors\":\"Chengshuo Yu, Taegeun Yoo, T. T. Kim, K. Chai, Bongjin Kim\",\"doi\":\"10.1109/CICC48029.2020.9075883\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"A novel 8T SRAM -based bitcell is proposed for current-based compute-in-memory dot-product operations. The proposed bitcell with two extra NMOS transistors (vs. standard 6T SRAM) decouples SRAM read and write operation. A 128×128 8T SRAM bitcell array is built for processing a vector-matrix multiplication (or parallel dot-products) with 64x binary (0 or 1) inputs, 64×128 binary (-1 or +1) weights, and 128x 1-5bit outputs. Each column (i.e. neuron) of the proposed SRAM compute-in-memory macro consists of 64x bitcells for dot-product, 32x bitcells for ADC, and 32x bitcells for calibration. The column-based neuron minimizes the ADC overhead by reusing a sense amplifier for SRAM read. The column-wise ADC converts the analog dot-product results to N-bit output codes (N=1 to 5) by sweeping reference levels using replica bitcells for 2N-1 cycles for each conversion. Monte-Carlo simulations and test-chip measurement results have verified both linearity and process variation. The largest variation (σ=2.48%) results in the MNIST classification accuracy of 96.2% (i.e. 0.4% lower than a baseline with no variation). A test-chip is fabricated using 65nm, and the 16K SRAM bitcell array occupies 0.055mm2. The energy efficiency of the 1bit operation is 490-to-15.8TOPS/W at 1-5bit ADC mode using 0.45/0.8V core supply and 200MHz.\",\"PeriodicalId\":409525,\"journal\":{\"name\":\"2020 IEEE Custom Integrated Circuits Conference (CICC)\",\"volume\":\"82 1\",\"pages\":\"0\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2020-03-01\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"41\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"2020 IEEE Custom Integrated Circuits Conference (CICC)\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1109/CICC48029.2020.9075883\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"2020 IEEE Custom Integrated Circuits Conference (CICC)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/CICC48029.2020.9075883","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 41

摘要

提出了一种新的基于8T SRAM的位单元,用于基于当前的内存中计算点积运算。所提出的位单元具有两个额外的NMOS晶体管(与标准6T SRAM相比),可以解耦SRAM的读写操作。构建了一个128×128 8T SRAM位单元阵列,用于处理向量矩阵乘法(或并行点积),具有64x二进制(0或1)输入,64×128二进制(-1或+1)权重和128x 1-5位输出。所提出的SRAM内存计算宏的每列(即神经元)由64x位单元格组成,用于点积,32x位单元格用于ADC, 32x位单元格用于校准。基于列的神经元通过重用SRAM读取的感测放大器来最小化ADC开销。列式ADC将模拟点积结果转换为N位输出代码(N=1至5),每次转换使用复制位单元为2N-1个周期扫描参考电平。蒙特卡罗模拟和测试芯片测量结果验证了线性和工艺变化。最大变异(σ=2.48%)导致MNIST分类准确率为96.2%(即比无变异的基线低0.4%)。测试芯片采用65nm制程,16K SRAM位元阵列占地0.055mm2。在1-5bit ADC模式下,使用0.45/0.8V核心电源和200MHz, 1bit操作的能量效率为490至15.8 tops /W。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
A 16K Current-Based 8T SRAM Compute-In-Memory Macro with Decoupled Read/Write and 1-5bit Column ADC
A novel 8T SRAM -based bitcell is proposed for current-based compute-in-memory dot-product operations. The proposed bitcell with two extra NMOS transistors (vs. standard 6T SRAM) decouples SRAM read and write operation. A 128×128 8T SRAM bitcell array is built for processing a vector-matrix multiplication (or parallel dot-products) with 64x binary (0 or 1) inputs, 64×128 binary (-1 or +1) weights, and 128x 1-5bit outputs. Each column (i.e. neuron) of the proposed SRAM compute-in-memory macro consists of 64x bitcells for dot-product, 32x bitcells for ADC, and 32x bitcells for calibration. The column-based neuron minimizes the ADC overhead by reusing a sense amplifier for SRAM read. The column-wise ADC converts the analog dot-product results to N-bit output codes (N=1 to 5) by sweeping reference levels using replica bitcells for 2N-1 cycles for each conversion. Monte-Carlo simulations and test-chip measurement results have verified both linearity and process variation. The largest variation (σ=2.48%) results in the MNIST classification accuracy of 96.2% (i.e. 0.4% lower than a baseline with no variation). A test-chip is fabricated using 65nm, and the 16K SRAM bitcell array occupies 0.055mm2. The energy efficiency of the 1bit operation is 490-to-15.8TOPS/W at 1-5bit ADC mode using 0.45/0.8V core supply and 200MHz.
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
自引率
0.00%
发文量
0
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术官方微信