Yield-Aware Leakage Power Reduction of On-Chip SRAMs

• Main Author: Nourivand, Afshin
• Format: Others
• Published: 2010
• Online Access: http://spectrum.library.concordia.ca/7103/1/Nourivand_PhD_S2011.pdf; Nourivand, Afshin <http://spectrum.library.concordia.ca/view/creators/Nourivand=3AAfshin=3A=3A.html> (2010) Yield-Aware Leakage Power Reduction of On-Chip SRAMs. PhD thesis, Concordia University.

Leakage power dissipation of on-chip static random access memories (SRAMs) constitutes a significant fraction of the total chip power consumption in state-of-the-art microprocessors and system-on-chips (SoCs). Scaling the supply voltage of SRAMs during idle periods is a simple yet effective technique to reduce their leakage power consumption. However, supply voltage scaling also results in the degradation of the cells’ robustness, and thus reduces their capability to retain data reliably. This is particularly resulting in the failure of an increasing number of cells that are already weakened by excessive process parameters variations and/or manufacturing imperfections in nano-meter technologies. Thus, with technology scaling, it is becoming increasingly challenging to maintain the yield while attempting to reduce the leakage power of SRAMs. This research focuses on characterizing the yield-leakage tradeoffs and developing novel techniques for a yield-aware leakage power reduction of SRAMs. We first demonstrate that new fault behaviors emerge with the introduction of a low-leakage standby mode to SRAMs. In particular, it is shown that there are some types of defects in SRAM cells that start to cause failures only when the drowsy mode is activated. These defects are not sensitized in the active operating mode, and thus escape the traditional March tests. Fault models for these newly observed fault behaviors are developed and described in this thesis. Then, a new low-complexity test algorithm, called March RAD, is proposed that is capable of detecting all the drowsy faults as well as the simple traditional faults. Extreme process parameters variations can also result in SRAM cells with very weak data-retention capability. The probability of such cells may be very rare in small memory arrays, however, in large arrays, their probability is magnified by the huge number of bit-cells integrated on a single chip. Hence, it is critical also to account for such extremal events while attempting to scale the supply voltage of SRAMs. To estimate the statistics of such rare events within a reasonable computational time, we have employed concepts from extreme value theory (EVT). This has enabled us to accurately model the tail of the cell failure probability distribution versus the supply voltage. Analytical models are then developed to characterize the yield-leakage tradeoffs in large modern SRAMs. It is shown that even a moderate scaling of the supply voltage of large SRAMs can potentially result in significant yield losses, especially in processes with highly fluctuating parameters. Thus, we have investigated the application of fault-tolerance techniques for a more efficient leakage reduction of SRAMs. These techniques allow for a more aggressive voltage scaling by providing tolerance to the failures that might occur during the sleep mode. The results show that in a 45-nm technology, assuming 10% variation in transistors threshold voltage, repairing a 64KB memory using only 8 redundant rows or incorporating single error correcting codes (ECCs) allows for ~90% leakage reduction while incurring only ~1% yield loss. The combination of redundancy and ECC, however, allows to reach the practical limits of leakage reduction in the analyzed benchmark, i.e., ~95%. Applying an identical standby voltage to all dies, regardless of their specific process parameters variations, can result in too many cell failures in some dies with heavily skewed process parameters, so that they may no longer be salvageable by the employed fault-tolerance techniques. To compensate for the inter-die variations, we have proposed to tune the standby voltage of each individual die to its corresponding minimum level, after manufacturing. A test algorithm is presented that can be used to identify the minimum applicable standby voltage to each individual memory die. A possible implementation of the proposed tuning technique is also demonstrated. Simulation results in a 45-nm predictive technology show that tuning standby voltage of SRAMs can enhance data-retention yield by an additional 10%−50%, depending on the severity of the variations.