UC Berkeley

Robust SRAM design is one of the key challenges of process technology scaling. The steady pace of process technology scaling allows doubling memory array sizes, which requires thorough characterization of the impact of sources of process variability on SRAM operation. SRAM arrays are traditionally designed using static noise margins which are known to be optimistic in writeability and pessimistic in read stability. This work presents techniques for characterizing SRAM using dynamic stability metrics, which better represent actual SRAM operating conditions. Quantitative relationships between static and dynamic stability metrics are explored through statistical circuit simulations. Nano-scale SRAM design is traditionally complicated by sources of variability related to physical variability in the structure of the transistors, such as random dopant distribution. This work identifies temporal sources of variability in transistor intrinsic parameters, caused by random telegraph signaling (RTS) noise, which is directly correlated with fluctuation in SRAM performance. A large-scale dynamic stability characterization architecture is introduced and implemented in an early commercial low-power 45 nm CMOS process. This is used to experimentally verify the expected correlations between static and dynamic stability metrics. Outliers of up to 100X which are not correlated between static and dynamic stability metrics were observed and identified to be due to enhanced sensitivity of dynamic stability metrics to variability. Measurement techniques for characterizing temporal sources of variability caused by RTS noise, with particular emphasis on the large-signal bias change response typically encountered in SRAM operation, are used to collect large-scale statistics and to estimate the statistical impact of RTS noise on large SRAM arrays.

An importance sampling algorithm adapted for dynamic stability metrics is developed in this work. This algorithm is used to estimate improvements in SRAM robustness expected from new process technology options such as FDSOI, different bitcell designs such as the 8T-SRAM, as well as several read-assist and write-assist techniques. An optimization framework enabled by this importance sampling algorithm is used to design SRAM arrays with maximum array efficiency through joint-optimization between process technology, bitcell design, and array organization. In conclusion, this dissertation identifies important sources of variability in nano-scale SRAM and also introduces the relevant optimization tools for performing variability-aware SRAM design.