376
P. Huang et al.
Fig. 30 Retention characterization of TiO x /Al 2 O 3 RRAM. a Temperature-dependent LRS retention
failure at 300 µA, b formation of nonmetallic filament, c typical retention failure process starts with
a conduction drift followed by a certain jump. Reprinted from [68]
square root of baking time, and the distribution changes with baking time [69]. In
detail, Fig. 31 introduces the important results in this work. The current distribution
is tight at the beginning and becomes wider as Fig. 31a shown. Figure 31b–d shows
that the conductance distribution of each level at a specific time can be fitted as
normal distribution. A compact model based on the standard deviation of different
levels is proposed from Fig. 31e, f. When applying the compact model to the neural
network, it is found that the recognition rate decreases due to the retention (Fig. 31g).
Fig. 31 a Read current distribution of 8 levels as a function of baking time at 125 °C, b–d the
normal distribution fitting for read current in (a) ~2 µA at 30, 300, and 600 s; standard deviation
(sigma) of normal distribution fitted from (a) as a function of square root of baking time in e 3 h can
be fitted independently, f 5 LRS can be only fitted with one line, g recognition rate as a function of
retention time at 85 °C. Reprinted from [69]
P. Huang et al.
Fig. 30 Retention characterization of TiO x /Al 2 O 3 RRAM. a Temperature-dependent LRS retention
failure at 300 µA, b formation of nonmetallic filament, c typical retention failure process starts with
a conduction drift followed by a certain jump. Reprinted from [68]
square root of baking time, and the distribution changes with baking time [69]. In
detail, Fig. 31 introduces the important results in this work. The current distribution
is tight at the beginning and becomes wider as Fig. 31a shown. Figure 31b–d shows
that the conductance distribution of each level at a specific time can be fitted as
normal distribution. A compact model based on the standard deviation of different
levels is proposed from Fig. 31e, f. When applying the compact model to the neural
network, it is found that the recognition rate decreases due to the retention (Fig. 31g).
Fig. 31 a Read current distribution of 8 levels as a function of baking time at 125 °C, b–d the
normal distribution fitting for read current in (a) ~2 µA at 30, 300, and 600 s; standard deviation
(sigma) of normal distribution fitted from (a) as a function of square root of baking time in e 3 h can
be fitted independently, f 5 LRS can be only fitted with one line, g recognition rate as a function of
retention time at 85 °C. Reprinted from [69]