The in-memory computation (IMC) is a potentialtechnique to improve the speed and energy efficiency of data-intensive designs. However, the scalability of IMC to largesystems is hindered by the non-linearities of analo...
详细信息
The in-memory computation (IMC) is a potentialtechnique to improve the speed and energy efficiency of data-intensive designs. However, the scalability of IMC to largesystems is hindered by the non-linearities of analog multiply-and-accumulate (MAC) operations and process variation, whichimpacts the precision of high bit-width MAC operations. In thispaper, we present an IMC architecture that is capable of perform-ing multi-bit MAC operations with improved speed, linearity, andcomputational accuracy. To improve the speed/linearity of theIMC-MAC operations, the image and weight data are applied byusing the pulse amplitude modulation (PAM) and thermometrictechniques, respectively. Although the PAM technique improvesthe speed of the IMC-MAC operations, it has linearity issuesthat need to be addressed. Based on the detailed linearity analysisof the IMC-MAC circuit, we proposed two approaches to improvethe linearity and the signal margin (SM) of the IMC *** proposed configurable current steering thermometric digital-to-analog converter (CST-DAC) array is employed to provide thePAM signals with various dynamic ranges and non-linear gapsthat are required to improve the linearity/SM. The proposedcombined PAM and thermometric IMC (PT-IMC) architectureis designed and fabricated in the TSMC 180-nm CMOS *** post-silicon calibration of the design point mitigates theprocess-variation issues and provides the maximum SM (closeto the simulation results). Furthermore, the proposed PT-IMCarchitecture performs MNIST/CIFAR-10 data set classificationwith an accuracy of 98%/88%. In addition, the PT-IMC archi-tecture achieves a peak throughput of 12.41 GOPS, a normalizedenergy efficiency of 30.64 TOPS/W, a normalized figure-of-merit(FOM) of 3039, a loss in the SM of 8.3% with respect to theideal SM, and a computational error of 0.41%
暂无评论