May_EDFA_Digital

24 / 58

ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 18 NO. 2
24
planar SRAM are higher than those in the FinFET SRAM,
as expected because Qcoll is suppressed by the geometry
of the FinFET.[14] Figure 12 shows Qcoll as a function of the
relative distance of the strike locations along the Dx direc-
tion. Qcoll as a function of distance decreases in a similar
manner as the measured MCU probability distributions
as a function of the MCU cluster length, as expected for a
charge-sharing mechanism. Moreover, it is clear that Qcoll
is a function of the technology node for any given cluster
length, implying that the differences observed in MCU
probabilities at any given distance are in fact related to
variations in Qcoll between technologies.
For a charge-sharing mechanism of upset, the abso-
lute rates of SBU and MCU are highly correlated, and so it
is assumed that the MCU probability is determined only
by the Qcoll associated with the particular cluster length.
The relative MCU probability that may be anticipated
from charge sharing at a distance Dx from the drain can
be approximated as:
and the number of MCU events is NMCU= NSBU·PMCU. Relative
Fig. 12 Qcoll as a function of Dx
Fig. 13 Modeled MCU probabilities (dashed lines) of SRAMs normalized to Qcoll of 40 nm at D x = 0 compared to experimental data
(symbols), which is the average of all test patterns, along the BL direction
(Eq 2)