一种面向存算一体的具有片上自校准的三级流水线读出电路OA
A three-stage pipeline readout circuit with on-chip self-calibration for computing-in-memory
提出了一种具有片上自校准功能的三级流水线存算一体读出电路,旨在解决模拟计算中的精度误差挑战,并提升计算吞吐率.同时,引入了一种电流对减误差校准技术,通过采用负反馈和电容存储方案,有效补偿了电流镜失配,从而提高了对减电流的精度.此外,利用ADC辅助失调检测实现了电阻变化校准,以抑制电流-电压转换电路的线性误差.为进一步提升性能,设计了一种三级流水线方案,通过并行处理提高计算吞吐率.基于55 nm工艺的仿真结果表明,与未校准情况相比,所提出的电流对减误差校准技术将相邻列电流相减的精度提高了58.3%;读出电路的计算吞吐量达到1 908 GOPS;在8位权重输入和8位输出量化的条件下,应用所提出的校准技术在MNIST、CIFAR-10和ImageNet数据集上的推理精度分别提升了7.6%、6.6%和8.9%.
This paper presents a high-precision three-stage pipeline readout circuit for computing-in-memory(CIM)architectures with on-chip self-calibration.The proposed readout circuit addresses the challenges of precision errors in analog computing and enhances computing throughput.A current subtraction error calibration is proposed to effectively compensate for current mirror mismatches,ensuring accurate bypass current by employing negative feedback and capacitor storage.In addition,a resistance variation calibration is proposed to suppress the non-linearity of current-to-voltage conversion circuits with ADC-assisted offset detection.Furthermore,a three-stage pipeline scheme is proposed to enhance computing throughput by parallel processing.Simulated in a 55 nm process,the proposed current subtraction error calibration improves the precision of adjacent column current subtraction by 58.3%compared to the case without calibration.Moreover,the computing throughput of the proposed readout circuit reaches 1 908 giga operations per second(GOPS).Under 8-bit weight input and 8-bit output quantization,the proposed readout circuit demonstrates significant improvements in inference accuracy on MNIST,CIFAR-10,and ImageNet datasets by 7.6%,6.6%,and 8.9%,respectively.
陈建涛;乔洋;虞致国;顾晓峰
江苏省MEMS传感器与专用集成电路制造技术重点实验室,江南大学 集成电路学院,江苏 无锡 214401江苏省MEMS传感器与专用集成电路制造技术重点实验室,江南大学 集成电路学院,江苏 无锡 214401江苏省MEMS传感器与专用集成电路制造技术重点实验室,江南大学 集成电路学院,江苏 无锡 214401江苏省MEMS传感器与专用集成电路制造技术重点实验室,江南大学 集成电路学院,江苏 无锡 214401
存内计算读出电路片上自校准流水线架构电流失配校准
computing-in-memory(CIM)readout circuiton-chip self-calibrationpipeline architecturecurrent error calibration
《测试科学与仪器》 2026 (3)
446-458,13
This work was supported by Key Research Project of Jiangsu Province(No.BE2023019-3).
评论