Low-Field Resistance Drift in Partial-SET States in
Phase Change Memories
S. Braga, A. Cabrini, and G. Torelli
University of Pavia, Department of Electronics, Pavia, Italy
Email: [email protected]
Abstract—Multi-level programming in Phase Change
Memories (PCMs) requires adequate understanding of the
phenomena which affect the stability of the programmed
resistance levels. Although the GST (Ge2 Sb2 Te5 alloy)
crystallization process has been extensively studied, further
analysis is needed to characterize the drift of the lowfield amorphous-GST resistance in intermediate states.
In this paper, we carry out a statistical analysis on
an array of PCM cells so as to investigate the drift
dynamics of intermediate GST resistance states obtained
with partial-SET programming. Our experimental results
reveal a significant dependence of the drift dynamics on
the programmed resistance value, as in the case of partialRESET programming, but a slightly lower drift coefficient.
I. I NTRODUCTION
Phase Change Memories (PCMs) are gaining increasing interest among innovative non-volatile memory technologies due to fast read access, short programming time,
bit-level programming granularity, high endurance, good
compatibility with standard CMOS fabrication processes,
and potential of scalability beyond Flash technology
[1]. A PCM cell is based on a chalcogenide alloy,
typically Ge2 Sb2 T e5 (GST), which can be reversibly
switched between two structural phases (amorphous and
polycrystalline) by means of electrical pulses. The two
phases have significantly different electrical resistivity,
that is high in the amorphous and low in the polycrystalline state, thus enabling non-volatile data storage. The
programmed resistance of a cell changes with time due
to two physical phenomena: the crystallization of the
amorphous GST [2], [3] and the drift of the amorphous
GST resistivity [4], [5], which decreases and increases,
respectively, the cell resistance. In multi-level storage,
where the cell may be programmed to any among n > 2
different resistance levels, both the above phenomena
affect data retention since they may cause problems in
distinguishing two different resistance levels. Although
the crystallization process has been extensively studied,
the understanding of resistance drift in intermediate
levels is still incomplete.
The drift phenomenon is ascribed to the amorphous
GST material, and the drift dynamics of the low-field cell
resistance is influenced by the phase distribution inside
the active region (i.e., inside the maximum GST volume
that can undergo phase transition). For the considered
µTrench architecture, the partial-RESET programming
approach gives rise to a series-type phase configuration,
where the two phases (amorphous and crystalline) are
placed in series with respect to the current flow, the
amorphous phase forming a cap above the heater. In
this case, the GST resistance basically depends on the
thickness of the amorphous cap [6]. On the other hand,
partial-SET programming essentially gives rise to a
conductive path inside the amorphous cap [7], which,
as will be shown in the following, affects the measured
drift of the cell resistance.
In this work, we analyze the drift dynamics on an
array of PCM cells programmed to different intermediate
resistance states (i.e., to states between the full-SET
state, corresponding to a fully crystalline GST, and the
RESET state, corresponding to a largely amorphized
GST) by means of a partial-SET programming procedure
[8]. With this approach, the cell is first brought into
the RESET state and is then programmed by means of
partial-SET pulses. In addition, we studied the stability
of the full-SET state obtained by means of a conventional
staircase-down programming procedure [9].
II. E XPERIMENTAL S ETUP AND R ESULTS
The experimental characterization of the drift dynamics was performed on a 4-Mb MOSFET-selected PCM
device with µTrench memory cells fabricated in 180-nm
CMOS technology [10]. The µTrench cell consists of
a TEC (Top Electrode Contact, titanium), a thin GST
stripe, a heater (titanium nitride), and a BEC (Bottom
Electrode Contact, tungsten). The cell is selected by
means of an MOS transistor and it is connected to a
bitline biasing transistor (Fig 1a).
Our analysis was carried out on a statistical population
of cells so as to attenuate the effects of the intrinsic
V
resistance noise and fabrication process spreads.
VRST
We considered a sub-array consisting of 10k PCM
cells and carried out the operating sequence shown in
VSET,i
Fig. 1b. All measurements were performed at room temV
Time
perature. We first programmed each cell of the sub-array
to its minimum-resistance state through a full-SET operStaircase-down
Drift characterization
SET sequence
ation (VSET,in was chosen so as to avoid history effects
a)
b)
on the GST alloy). Then, we programmed each cell of
the sub-array to the full-RESET state by means of a Figure 1. Schematic of a MOSFET-selected PCM cell in the used
2
100-ns RESET pulse with an amplitude VRST = 4.8 V experimental chip (a) and operating sequence used to characterize the
GST
resistance
drift
of
partial-SET
states
(b).
(applied to the gate of Y0 ) and measured the RESET-state
resistance of each cell. Finally, we programmed the cells
0.16
Partial−RESET (from [6])
to a partial-SET state by applying a single 200-ns partial0.14
Partial−SET
SET pulse having an amplitude VSET,i . To obtain a
0.12
sufficiently wide resistance window for our analysis, we
0.1
repeated the above sequence 4 times for any cell, setting
0.08
VSET,i = 1.75 V +∆V ·i, ∆V = 0.25 V , i = 1÷4. The
0.06
programming pulse amplitudes were chosen to be higher
0.04
than the threshold voltage of the cells in the RESET
0.02
state. Then, the power delivered to the cell depends on
0
the resistance of the cell in switching conditions, which
10
10
10
igure 1: Schematic of a MOSFET-selected PCM cell in the
R
(Ω)
is almost the same for every cell, regardless of its initial
GST
onsidered experimental chip (a) and operating sequence used
low-field resistance value. After the partial-SET pulse,
o characterize the GST resistance drift of partial-SET states
∆RGST
the resistance of each cell belonging to the sub-arrayFigure
variation
of GST
the GST
resistance
)
GST (
Figure2:2. Relative
Relative variation
of the
resistance
( ∆R
) after
RGST
b).
RGST
was read by biasing the gate of Y0 with a stable readafter
30 minutes
from the
first measurement
(which was carried
out 130
s car30 minutes
from
the first measurement
(which
was
after
the130
programming
pulse):
partial-RESET pulse):
states [6] partial-RESET
(square) and
voltage (Vread =700 mV) and sensing the current flowingried
out
s after the
programming
partial-SET
states (circles).
states
[11]
(square)
and
partial-SET
states
(circles)
through the cell. The read operation was repeated at
ET state by
applying a single 200-ns partial-SET pulse
predetermined time steps for a time interval δt of about
aving an amplitude
V
. To obtain a wider resistance
30 minutes. prg,i
window, we chose VG,i = 1.75 V + ∆V · i, ∆V = 0.25 V ,
From our experimental data, we calculated the relative
In order to investigate the stability of the full-SETfrom [11].
GST
= 1 ÷ 4. The programming pulse amplitudes are chosen
increase of RGST during
a given
time interval ∆t =
that ∆R
RGST increases with RGST in
state, we applied a conventional staircase-down program- It can be noticed
o be higher than the threshold voltage of the cells in the
withcases,
respect
tothe
thedrift
initially
measured
values. Fig. differ2
both
but
dynamics
is significantly
ming algorithm [9] to a sub-array of cells, and measured
RESET state. Then, the power delivered to the cell deshows
the
comparison
of
the
relative
increase
of
the
GST
the read current under the same conditions used forent. We ascribe this disagreement to the different distriends on the resistance of the cell in switching conditions,
resistance,
functioninside
of thethe
programmed
GST , as a phase
bution
of the∆R
amorphous
active area in
partial-SET states drift measurements.
R
in
the
partial-SET
states
obtained
with
the above
which is almost the same for every cell, regardless of the
GST
In the following, we will consider the low-field activethe partial-SET states and the partial-RESET states.
experimental
procedure
and
in
the
partial-RESET
states
nitial low-field resistance value. After the partial-SET
GST resistance, RGST , which is obtained by subtracting As pointed out in the literature [7], in partial-SET proobtained
in
our
previous
work
[6].
ulse, the resistance of each cell belonging to the subis dagged inside the amorthe resistance of the heater and the crystalline GST abovegramming a conductive path
∆RGST
rray was read by biasing the gate of Y0 with a stable
with RGST thus
It
can
be
noticed
that
the increases
RESET operation,
RGST
the active region from the measured cell resistance (thephous volume obtained after
ead voltage (VG,read =700 mV) and sensing the current
in
both
cases,
but
the
drift
dynamics
is
significantly
voltage applied to the word-line was chosen so as todetermining a parallel-type phase configuration inside
We ascribe
thisthe
difference
to theresistance
different can
owing through the PCM cell. The read operation was
thedifferent.
active region.
Then,
active GST
make the ON resistance of the select transistor M
of
the
amorphous
phase
inside
the
active
area reepeated at predetermined time steps for a time interval SEL bedistribution
expressed as the parallel of the amorphous-phase
negligible as compared to the GST resistance).
in
partial-SET
and
partial-RESET
states.
As
pointed
outC :
f about 30 minutes.
RA and the crystalline-phase resistance R
In order to study the dependence of the drift dynamicssistance
in the literature [7], a conductive path is dagged inside
In order to investigate the stability of the full-SET
on the programmed RGST (i.e., on the value obtained the amorphous volume obtained after R
A RC
tate, we applied
to
the
array
a
conventional
staircaseRGST = RC kRA = the RESET
. opera- (1)
with the first resistance measurement, which was carried tion, thus determining
R
+
R
a
parallel-type
phase
configuration
A
C
own programming
algorithm [9], and we measured the
out 130 s after the programming operation, due to our inside the active region. Then, the active GST resistance
ead current
under
thesetup),
same we
conditions
of resistance
the partialRA and RC depend on the thickness xa of the
experimental
divided the
window Both
can be expressed
as the parallel of the amorphous-phase
ET drift measurements.
amorphous
cap obtained after the RESET operation and
into 12 bins and grouped together the cells having their resistance R
and
the crystalline-phase resistance RC :
A
In the following,
we willbelonging
considertothe
initial resistance
thelow-field
same bin.active
Then, wethe cross section AC of the conductive filament obtained
GST resistance,
RGSTthe
, which
obtained
by of
subtracting
can
considered
mean is
drift
dynamics
each cluster ofafter the partial-SET pulse. We R
A RCwrite RGST as a
R
=
R
kR
=
.
(1)
GST
C
A
he heater resistance
and
the
resistance
of
the
crystalline
function
of
these
parameters
as
follows
PCM cells.
RA + RC
GST above the active region from the measured cell restance (the voltage applied to the word-line was chosen
o as to make the ON resistance of the select transistor
MSEL negligible as compared to the GST resistance).
In order to study the dependence of the drift dynam-
VA
VSET
VRST
Vread
G
prog
YO
BL
Memory
cell
WL
read
∆ RGST / RGST
MSEL
3
4
RGST =
ρC ρA xa
,
AρC + AC (ρA − ρC )
5
(2)
where A is the GST-heater contact area and ρA (ρC )
are the resistivity of amorphous (crystalline) GST. The
3
When considering
RC , by assuming that the equivalent
Partial−SET
(α = 1.17,
β = 1)
resistivityModel
of the
conductive
filament to increase from its
Model (α = 1.17, β = 1.02)
0.1
initial value ρC to βρC , β > 1, during ∆t, we obtain the
following expression:
3.589
10
0.08
∆ RGST / RGST
Full−SET GST resistance (Ω)
0.12
3.588
10
0.06
α(RC + RA )
∆RGST
= α
− 1.
RGST
β RA + RC
0.04
(3)
3.587
10
2
GST
The higher is ∆R
RGST , the faster is the drift dynamics.
The value of α was set according to the drift dynamics
0
of the
RESET state obtained
with the considered10 VRST .
10
10
From data in [6], α turnsRout(Ω )to be about 1.17. According
to our fitting, the equivalent resistivity of the percolation
GST
Figure
5. increases
Relative variation
of the β
GST
( ∆R
) in to
phase
by a factor
' resistance
1.02 from
first
Rthe
GST
partial-SET
programming
after
30
minutes
from
the
first
measurethe last resistance measurement. This value is higher
ment:
comparison
between in
datathe
andfull-SET
models. case, showing that in
than
that obtained
the considered partial-SET states the fraction of residual
amorphous phase inside the conductive path is higher.
theThe
mean
GSTand
resistance
time 4.of Athevery
model
data arebehavior
comparedover
in Fig.
considered
sub-arrayis of
memorywhen
cellsthe
programmed
into
good agreement
observed
drift phenomenon
the percolation
taken into
account.
As shown
theoffull-SET
state. Wepath
canisobserve
a slight
increase
of
Fig. 3, it
can be finally
noticed
that,corresponds
since α/β <toα, if
theinfull-SET
resistance
over time,
which
we neglect
the contribution
RC to0.005
the GST
drift
a relative
increase
of RGST ofofabout
during
∆t(i.e.,
.
we set β = 1), Eq. (3) underestimates the experimental
This results is in agreement with the values presented
data.
in the literature for full-SET states [12]. Our results
The experimental analysis of the low-field resistance
suggest that the active volume of GST after the fulldrift of partial-SET intermediate states has been preSET
pulse and
is a mixture
of crystalline
grains and
a residual
sented
compared
to the resistance
drift
of partialamorphous
phase,
which
gives
rise
to
a
percolation
RESET states. The observed difference betweenstate.
the two
Similarly,
assumetothat
percolation
phase
is obtained incases iswe
ascribed
thea different
phase
configuration
inside
path
achieved
aftera our
partial-SET
sidethe
theconductive
active GST
region.
In fact,
conductive
path is
dagged inside
the amorphous phase during partial-SET
programming
pulse.
programming.
Similarly to the
case both
of theRfull-SET
state,
Then,
the drift phenomenon
affects
A and RC .
the
conductive
path
can
be
viewed
as
a
mixture
of
crysThe drift of RA can be modeled as the increase of the
talline grains and residual amorphous phase which conelectrical resistivity of amorphous GST from its initial
tributes to the drift of the GST resistance over time.
value ρA to αρA , α > 1, during ∆t.
This work has been supported by Italian MIUR in the
When considering RC , by assuming that the equivaframe of its National FIRB Project RBAP06L4S5.
0.02
3
10
10
Time (s)
3
Figure 3: Log-log plot of the experimental measurements of
Figure 3. Log-log plot of the experimental results of the full-SET
the full-SET resistance drift over time.
resistance drift over time.
0.12
0.12
0.1
∆R
/R
GST
∆ GST
RGST / R
GST
0.1
Experimental data
Model: α =1.1
Partial−SET
data
Model:(α=1.17,
α =1.17β=1)
Model
Model:(α=1.17,
α =1.2 β=1.02)
Model
0.08
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0
3
10
0
3
10
4
10
RGST (Ω )
4
10
R GST ( Ω)
Increase of
α
5
10
5
10
GST
resistance( ∆R
RGST
)
Figure 4: Relative variation of the GST
in partial-SET programming after 30 minutes from the first
∆RGST
measurement:
comparison
dataresistance(
and models.
Figure
4.
Relative
variation between
of the GST
) in
RGST
partial-SET programming after 30 minutes from the first measurement: comparison between data and model for different values of
α.giving rise to a percolation state. Similarly, a percolation
phase is obtained inside the conductive path obtained after our partial-SET programming pulse. Then the drift
phenomenon
affects
RA on
andthe
RCthickness,
. The driftxaof
RA
Both RA and
RC both
depend
, of
can
be
modeled
as
the
increase
of
the
electrical
resistivthe amorphous cap obtained after the RESET operation
ity the
of amorphous
GSTAfrom
its initial value ρ filament
to αρA ,
and
cross section,
C , of the conductive A
α
>
1,
during
∆t.
obtained after the partial-SET pulse. We can write R
GST
as a function of these parameters:
ρC ρA xa
,
RGST =
AρC + AC (ρA − ρC )
(2)
[1] L. Geppert, IEEE Spectrum 40, 48 (Mar 2003).
[2] S.A
Senkader
C. D.Wright,
Journal
where
is the and
GST-heater
contact
areaof Applied
and ρA Physics
(ρC )
is the(2004).
resistivity of amorphous (crystalline) GST. The
[3] A. Redaelli, D. Ielmini, A. L. Lacaita, F. Pellizzer, A.
value Pirovano,
of xa obtained
the considered
RESET
pulse
and R. with
Bez, IEEE
International
Electron
Deamplitude
be estimated to be on the order of 30
RST can
vices V
Meeting
(2005).
I. V.thus
Karpov,
M. us
Mitra,
D. Kau,
Y. A.
nm[4][11],
enabling
to estimate
ACG.forSpadini,
every value
Kryukov,
and V.toG.
Karpov, Journal
of Appliedof
Physics
of RGST
and, then,
determine
the contribution
RC
102, 124503 (pages 6) (2007).
and
RA Pirovano,
to the GST
resistance. In order to investigate
[5] A.
A. Lacaita, F. Pellizzer, S. Kostylev, A.
the nature
of
the
conductive
filament,Devices,
we measured
the
Benvenuti, and R. Bez, Electron
IEEE Transdrift dynamics
of
the
full-SET
state,
that
can
be
seen
actions on 51, 714 (2004).
Braga, A. state
Cabrini,
and A
G.
Physics
as [6]
a S.
partial-SET
where
= A. Applied
Fig. 3 shows
C Torelli,
Letters 94, 092112 (pages 3) (2009).
[7] D. Ielmini, D. Mantegazza, A. Lacaita, A. Pirovano,
4
5
GST
lent resistivity of the conductive filament increases from
its initial value ρC to βρC , β > 1, during ∆t, we obtain
the following expression:
and F.∆R
Pellizzer,
Device
α(RC +
RA ) Letters, IEEE 26, 799
GST Electron
= α
− 1.
(3)
(2005).
RGST
RC
β RA +C.
[8] F. Bedeschi,
R. Fackenthal,
Resta, E. Donzetti, M. Jagasivamani, F. Pellizzer, D. Chow, D. Mills,
A. Cabrini,
GST
The faster
theetdrift
the ISSCC08
higher ∆R
. The
R
G. Calvi,
al., dynamics,
Proc. of IEEE
(Feb.
GST 2008).
value[9]ofF. αBedeschi,
was set C.
according
the drift C.
dynamics
of G.
Bono, E.toBonizzoni,
Resta, and
Torelli,state
Microelectronics
Journal
1064 (2007).
the RESET
obtained with
the 38,
considered
VRST .
[10]data
F. Bedeschi,
R. Bez, C.
Bono,
E. Bonizzoni,
E. about
C. Buda,
From
in the literature
[11],
α turns
out to be
G. Casagrande, L. Costa, M. Ferraro, R. Gastaldi, O.
1.17. ItKhour,
is worthet noticing
from Fig. 4, that the observed
al., IEEE Journal of Solid State Circuits
dependence
of
drift
on
the
programmed resistance cannot
(2005).
be [11]
explained
by
the
only
effect
of Torelli,
the amorphous
phase.SciS. Braga, A. Cabrini, and G.
Semiconductor
ence
and
Technology
24,
115008
(6pp)
(2009).
According to our fitting, the equivalent resistivity of
the percolation phase increases by a factor β ' 1.02
from the first to the last measurements. This value is
higher than that obtained in the full-SET case, suggesting
that, in the considered partial-SET states, the fraction of
residual amorphous phase inside the conductive path is
higher. The proposed model and experimental data are
compared in Fig. 5. A good agreement is observed when
the drift phenomenon of the percolation path is taken into
account. As shown in Fig. 5, it can be finally noticed that,
since α/β < α, if we neglect the contribution of RC to
the GST drift (i.e., we set β = 1), Eq. (3) underestimates
the experimental data.
III. C ONCLUSIONS
The experimental analysis of the low-field resistance
drift of partial-SET intermediate states has been presented and compared to the resistance drift of partialRESET states. The observed difference between the two
cases is ascribed to the different phase configuration
inside the active GST region. In fact, a conductive path
is dagged inside the amorphous phase during partialSET programming. Similarly to the case of the full-SET
state, the conductive path can be viewed as a mixture of
crystalline grains and a residual amorphous phase, which
contributes to the drift of the GST resistance over time.
ACKNOWLEDGEMENTS
This work has been supported by Italian MIUR in the frame
of its National FIRB Project RBAP06L4S5.
R EFERENCES
[1] L. Geppert, IEEE Spectrum 40, 48 (Mar 2003).
[2] S. Senkader and C. D. Wright, Journal of Applied Physics
(2004).
[3] A. Redaelli, D. Ielmini, A. L. Lacaita, F. Pellizzer,
A. Pirovano, and R. Bez, IEEE International Electron Devices
Meeting (2005).
[4] I. V. Karpov, M. Mitra, D. Kau, G. Spadini, Y. A. Kryukov,
and V. G. Karpov, Journal of Applied Physics 102, 124503
(pages 6) (2007).
[5] A. Pirovano, A. Lacaita, F. Pellizzer, S. Kostylev, A. Benvenuti, and R. Bez, Electron Devices, IEEE Transactions on
51, 714 (2004).
[6] S. Braga, A. Cabrini, and G. Torelli, Semiconductor Science
and Technology 24, 115008 (6pp) (2009).
[7] D. Ielmini, D. Mantegazza, A. Lacaita, A. Pirovano, and
F. Pellizzer, Electron Device Letters, 26, 799 (2005).
[8] F. Bedeschi, R. Fackenthal, C. Resta, E. Donzetti, M. Jagasivamani, F. Pellizzer, D. Chow, D. Mills, A. Cabrini, G. Calvi,
et al., Proc. of ISSCC (Feb. 2008).
[9] F. Bedeschi, C. Boffino, E. Bonizzoni, C. Resta, and
G. Torelli, Microelectronics Journal 38, 1064 (2007).
[10] F. Bedeschi, R. Bez, C. Boffino, E. Bonizzoni, E. C. Buda,
G. Casagrande, L. Costa, M. Ferraro, R. Gastaldi, O. Khour,
et al., IEEE Journal of Solid State Circuits (2005).
[11] S. Braga, A. Cabrini, and G. Torelli, Applied Physics Letters
94, 092112 (pages 3) (2009).
[12] D. Ielmini, D. Sharma, S. Lavizzari and A. L. Lacaita, Proc.
of IRPS, pp 597 - 603 (2008).