← 返回 JSSC 论文列表JSSC 2022第2期Memory28nmSRAMCIM
Two-Way Transpose Multibit 6T SRAM Computing-in-Memory Macro for Inference-Train
提出一种基于6T SRAM的双向转置存内计算宏,支持AI边缘设备的高效能乘加运算
28nm CMOS, 3.8–21 ns延迟, 7–61.1 TOPS/W能效
存内计算SRAM乘加运算边缘AI双向转置
▸创新点1:双向转置乘法单元设计(电路创新):提出了一种双向转置乘法单元,支持卷积计算中的转置权重矩阵操作,解决了现有SRAM-CIM芯片仅支持推理的问题,扩展了应用场景。
▸创新点2:输入感知区域预测读取方案(方法创新):引入输入感知区域预测技术,通过预测最大部分MAC值来增强信号裕度,提高了读取的鲁棒性,确保了在复杂条件下的稳定操作。
▸创新点3:抗工艺变化的高鲁棒性(电路创新):设计了高抗工艺变化的电路结构,确保在不同工艺条件下仍能保持高性能,提升了芯片的可靠性和一致性。
▸创新点4:高性能指标(系统创新):在28nm工艺下实现的64kb TWT CIM宏,MAC操作时间范围为3.8-21ns,能效达到7-61.1 TOPS/W,显著提升了计算效率和能效比。
Abstract
Computing-in-memory (CIM) based on SRAM is
a promising approach to achieving energy-efficient multiply-
and-accumulate (MAC) operations in artificial intelligence (AI)
edge devices; however, existing SRAM-CIM chips support only
DNN inference. The flow of training data requires that CIM
arrays perform convolutional computation using transposed
weight matrices. This article presents a two-way transpose (TWT)
multiply cell with high resistance to process variation and a
novel read scheme that uses inp