BTI variability fundamental understandings and impact on digital logic by the use of extensive dataset

Bias temperature instability variation on SiON/Poly, HK/MG and trigate architectures

“…In essence, this was an indication that the over-dispersion was insufficient to capture the true spread present in the raw data, and points clearly to the use of the EP framework as the appropriate distribution model for NBTI variation at high sigma values. A very recent report by Angot et al [16] for data-sets of over 100,000 transistors also came to the same conclusion. Consequently, all subsequent data in this work have been fit by the EP framework.…”

“…Even for the larger data-set in Figure 11b, any potential deviation of post-aging VT from a normal distribution is at the limit of observability, beyond +4 σ in the extreme tails. In their recent report on a data-set of 130,000 transistors on a 28nm FDSOI technology, Angot et al [16] observed negligible deviations from normality at +4.3 σ, but the device sizes in that work were relatively larger than those presented in Figure 11b. Larger sample sizes may be needed to fully explore and understand the extreme tail of post-aging VT distribution behavior on ultra-scaled devices.…”

The impact of high V<inf>th</inf> drifts tail and real workloads on SRAM reliability

“…shows V th -shift distributions from 28LP high density bitcell due to NBTI for a symetric bitcell stress (50/50 duty cycle). The non-normality behavior is clearly observed and was previously explained to be related to Defect-centric (DC) modeling [5][6][7][8]. Several duty cycles were used to stress the bitcells by modifying the ratio between the duration where the cell stores '0' state and the one where the cell stores '1' state (from 50% '1' to 95% '1').…”

“…These two components can be added as no clear correlation was found in [7] between time zero parameters and BTI induced degradations.…”

“…Previous papers on SRAM reliability [1,2] have considered the approximation of a normal distribution for V th drifts. In our study, we have considered both the normal distribution as well as the Defect-centric one [5][6][7][8] which is a mixture of a poisson distribution (random number of defects) with gamma distribution (exponential distribution of the defects' impact on V th ).…”