NBTI degrades the reliability and lifetime of digital devices. In case of SRAM, NBTI decreases both SNM stability and the cell performance depending on the average time that the PMOS gates are biased with zero input. The cache soft-error is a function of the average SNM. In 6T-SRAM cell in Figure 1a, the Q and QB nodes save complementary values. When the Q node is zero, the gate of M3 and M4 transistors are biased with “0” and “1” respectively. In other words, when one of these transistors is in stress mode, the other one is in recovery mode. We used the NBTI model in paper to evaluate the change in the threshold voltage of PMOS transistors. In 5T-SRAM cell, NBTI has huge impact on the stability because the cell just has one pull-up PMOS transistor. The average SNM stability improvement of 5T-SRAM after three years NBTI effect with 1000 Monte Carlo simulation and 10% variations are listed on Table 4. We used statistical evaluation to extract the data activity of SRAM part of the proposed cache in different THR running SPEC CPU 2006 benchmarks. Data in proposed cache increases the percentage of the ones in SRAM part. Although this increment does not have any effect on 6T-SRAM lifetime due to its symmetry, this improves the SNM degradation of the 5T-cell by 35.1% after three years NBTI effect with THR=65%. To further reduce the write energy on STT-RAM cache, we use early write termination techniques. Before write operation, this technique compares the bitline values with input data bit by bit and neglects writing the same bits on the cache. In addition, we are writing mostly zero data in the cache line with majority zero. So, EWT significantly reduces
the number of writes on STT-RAM cache depending on the value of THR. As Table 4 shows, the proposed cache decreases the number of writes in STT-RAM by 28.9% using THR=75%.