Evaluating the accuracy of SRAM margin simulation through large scale Monte-Carlo simulations with accurate compact models

Accurate statistical compact model extraction and circuit simulation are key issues in contemporary SRAM design. The high statistical variability of the small SRAM cell transistors in combination with high cell density leads to yield problems determined by 5-6σ deviations from the mean. The compact modeling approach presented in this paper utilizes a firm understanding of the physical phenomenon underlying device variability. Its illustration is based on comprehensive `atomistic´ 3D device simulations. Extracted statistical models are then utilized in SRAM dynamic write simulation, and benchmarked against Gaussian V_T based simulation, an approach which has been favoured due to its ease of implementation and relative ease of technology characterisation. In addition, the Gaussian V_T approach also simplifies statistical analysis and enables margining approaches like Most Probable Vector (MPV), which can drastically reduce simulation time. It is well known that in order to do this, Gaussian V_T captures only the 1st order effects of variability, and over estimates the correlation between threshold voltage and other device figures of merit. It is of considerable industrial interest to examine quantitatively the effects of this approximation.