A method was developed to solve the combined system of the current field and the circuit. The ＂super-node＂ was used to transform the matrix for conventional nodal analyses of a circuit system from non-positive defin...A method was developed to solve the combined system of the current field and the circuit. The ＂super-node＂ was used to transform the matrix for conventional nodal analyses of a circuit system from non-positive definite to positive definite. Then, a positive definite matrix for the overall system was obtained by combining the matrix from the circuit nodal analysis method and the matrix resulted from finite element method （FEM） formulation to solve the FEM fields. This approach has been successfully applied to simulate the electrical potential and current distributions on each metal layer of printed circuit boards （PCBs） and integrated circuit （IC） packages for a given power supply. The simulation results can then be used to analyze the properties of the PCBs and IC packages such as the port resistances and IR drops. The results can also be used to optimize PCB and IC package designs, such as by adjusting the power/ground distribution networks.展开更多
忆阻器阵列(Memristor based Crossbar)在加速神经网络计算上有很好的效果。然而,忆阻器阵列会受到IR-Drop的影响,导致忆阻器阵列的计算精度下降。为此,提出一种方案来提高计算精度,该方案是基于对权值矩阵稀疏化以及对权值矩阵的行向...忆阻器阵列(Memristor based Crossbar)在加速神经网络计算上有很好的效果。然而,忆阻器阵列会受到IR-Drop的影响,导致忆阻器阵列的计算精度下降。为此,提出一种方案来提高计算精度,该方案是基于对权值矩阵稀疏化以及对权值矩阵的行向量进行聚类实现的。该方案首先通过分析IR-Drop对忆阻器阵列的影响,根据忆阻器阵列和权值矩阵的映射关系,对权值矩阵进行稀疏化训练,将受到较大IR-Drop影响的权值置零。然后对权值矩阵的行向量进行聚类,找到近似全零行向量将其权值置零,在保证零权值不变的前提下重新训练权值矩阵,接着删除全零行向量和全零列向量降低矩阵规模。最后在IR-Drop影响下计算权值矩阵行向量的权值损失,根据损失大小降序排列行向量得到新的权值矩阵,并映射到忆阻器阵列上。实验表明,经过此方案处理后,忆阻器阵列受到的IR-Drop显著降低,有效地提高了计算精度并且降低了硬件规模。展开更多
介绍了游标时数转换器(TDC)和循环时数转换器这两种用于实时多点检测电源网络IR压降的方案。游标TDC方案采用精调粗调结合的思想,在继承传统时数转换器方案高精度、单周期响应优点的基础上,对面积进行了优化。循环TDC方案采用倍频的思想...介绍了游标时数转换器(TDC)和循环时数转换器这两种用于实时多点检测电源网络IR压降的方案。游标TDC方案采用精调粗调结合的思想,在继承传统时数转换器方案高精度、单周期响应优点的基础上,对面积进行了优化。循环TDC方案采用倍频的思想,在继承单周期响应的环振(RO)方案小面积优点的基础上,对精度进行了优化。电路分别基于SMIC 28 nm高介电常数金属栅(HKMG)工艺进行仿真。仿真结果表明,与TDC方案相比,游标TDC方案在维持13.5 m V/bit高精度和单周期响应优点的同时,面积仅为163.75μm^2,减小了29%;而与同为单周期响应的环振方案相比,循环TDC方案维持92.25μm^2小面积优点的同时,精度达到40.5 m V/bit,提高了54%。展开更多
文摘A method was developed to solve the combined system of the current field and the circuit. The ＂super-node＂ was used to transform the matrix for conventional nodal analyses of a circuit system from non-positive definite to positive definite. Then, a positive definite matrix for the overall system was obtained by combining the matrix from the circuit nodal analysis method and the matrix resulted from finite element method （FEM） formulation to solve the FEM fields. This approach has been successfully applied to simulate the electrical potential and current distributions on each metal layer of printed circuit boards （PCBs） and integrated circuit （IC） packages for a given power supply. The simulation results can then be used to analyze the properties of the PCBs and IC packages such as the port resistances and IR drops. The results can also be used to optimize PCB and IC package designs, such as by adjusting the power/ground distribution networks.
文摘忆阻器阵列(Memristor based Crossbar)在加速神经网络计算上有很好的效果。然而,忆阻器阵列会受到IR-Drop的影响,导致忆阻器阵列的计算精度下降。为此,提出一种方案来提高计算精度,该方案是基于对权值矩阵稀疏化以及对权值矩阵的行向量进行聚类实现的。该方案首先通过分析IR-Drop对忆阻器阵列的影响,根据忆阻器阵列和权值矩阵的映射关系,对权值矩阵进行稀疏化训练,将受到较大IR-Drop影响的权值置零。然后对权值矩阵的行向量进行聚类,找到近似全零行向量将其权值置零,在保证零权值不变的前提下重新训练权值矩阵,接着删除全零行向量和全零列向量降低矩阵规模。最后在IR-Drop影响下计算权值矩阵行向量的权值损失,根据损失大小降序排列行向量得到新的权值矩阵,并映射到忆阻器阵列上。实验表明,经过此方案处理后,忆阻器阵列受到的IR-Drop显著降低,有效地提高了计算精度并且降低了硬件规模。
文摘介绍了游标时数转换器(TDC)和循环时数转换器这两种用于实时多点检测电源网络IR压降的方案。游标TDC方案采用精调粗调结合的思想,在继承传统时数转换器方案高精度、单周期响应优点的基础上,对面积进行了优化。循环TDC方案采用倍频的思想,在继承单周期响应的环振(RO)方案小面积优点的基础上,对精度进行了优化。电路分别基于SMIC 28 nm高介电常数金属栅(HKMG)工艺进行仿真。仿真结果表明,与TDC方案相比,游标TDC方案在维持13.5 m V/bit高精度和单周期响应优点的同时,面积仅为163.75μm^2,减小了29%;而与同为单周期响应的环振方案相比,循环TDC方案维持92.25μm^2小面积优点的同时,精度达到40.5 m V/bit,提高了54%。
