Professional Documents
Culture Documents
05467625
05467625
05467625
I. INTRODUCTION
PIN torque transfer (STT) MRAM has been intensively
studied as a candidate for future memory applications due
to its nonvolatility, fast write-time, and scalability with CMOS
technology [1], [2]. The critical element of the STT-MRAM
is the magnetic tunneling junction (MTJ) with an MgO barrier
which can provide high tunneling magnetoresistance
up to 1056% at room temperature [3]. Utilizing the STT induced switching as the writing mechanism also has a scaling
advantage: upon junction size reduction, the current needed to
switch the free layer orientation is decreased. However, magnetic thermal fluctuation can be significant for room temperature operation of magnetic memory cells below 100 nm, as it
affects STT switchings speed and threshold distribution. For
reliable write, the write current has to be high enough to ensure all devices are switched to desired states, while still remain low enough to avoid junction barrier breakdown. Currently, MTJs with a 1.2 nm or less Mg-oxide layer are necessary
due to the requirement of a low resistance-area product
for a reasonably low write voltage. The dielectric breakdown
of such thin MgO barriers under voltage/current stress therefore
becomes critical to the success of STT-MRAM technology. The
reliability of the thin oxide against breakdown determines the
upper limit of the programming voltage. To commercialize the
bit error rate has to
STT-MRAM with 64 Mb density, a
be achieved in both read and write operations. The key challenge in realizing this is the write margin between the minimum
programming voltage and the dielectric breakdown of MgO for
Manuscript received October 31, 2009; revised February 01, 2010; accepted
February 01, 2010. Current version published May 19, 2010. Corresponding
author: T. Min (e-mail: tai.min@headway.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TMAG.2010.2043069
MIN et al.: STUDY OF WRITE MARGIN OF SPIN TORQUE TRANSFER MAGNETIC RANDOM ACCESS MEMORY TECHNOLOGY
2323
Fig. 1. (a) Critical switching voltage as a function of pulse width and (b)
(in Oe) as a function of measurement dwell time (b).
dwell time then fitting this data to the following formula [14],
[15]:
(3)
obtained by this
As shown in the Fig. 1(b), the value of
method can be an additional factor of 2 higher, similar to the
results reported by [16]. This experiment indicates that it is not
by simply fitting to (2), especially when
accurate to obtain
the pulse width is on the nano-second scale. The discrepancy beobtained with and without spin current
tween the values of
could be due to the fact that the spin current injected into MTJ
can induce nonuniform magnetization distributions [17], [18]
which invalidate the macrospin assumption used to derive (2).
obtained without spin current is more
For data retention, the
relevant. To determine what value and distribution of
(and
hence ) is required for adequate data retention, we adopted a
direct approach of determining the thermally induced switching
probability (TISP) function for MTJs with various volumes,
s and baking times by measuring the switched devices after
being stored at elevated temperature for various time periods
without external magnetic fields. 4-k test chips were used in this
experiment. For a given MTJ design, the percentage of switched
devices depends on baking temperature, duration, and junction
. The percentage of switched device vs Hc gives TISP as
a function of Hc. These TISP data can be fitted very well by
the cumulative distribution function (CDF) of a Gaussian distribution with two independent parameters. Once the dependence
of these two parameters of CDF on MTJ volume, baking temperature and time were determined experimentally, the data retention can be estimated for any given MTJ design. For MTJs
nm , the minimum requirement of
with sizes of
2324
Fig. 2. Weibull plots for various stressing voltage of the MTJ population with
the high breakdown distribution.
and Ave is 190 Oe and 12% for adequate data retention idle
storage at 125 C for 1 year and 1-bit error correction. Details
of this method will be published elsewhere.
B. Junction Breakdown Distributions
For desired junction
, the MgO tunnel barrier is less
than 12 thick. A dielectric at such thin thickness poses new
challenges as the dielectric breakdown was known to have
two distinct distributions [4]. The majority MTJs have high
breakdown value (HBD) while a small percentage of MTJs
exhibit low values of breakdown (LBD). The peak-to-peak
to the HBD distributions can be
separation between the
but only
as high as
for LBD distribution. As described in [4], besides the low
breakdown induced defects, some MTJs with very different
LBD characteristics, behave identically otherwise to the HBD
,
and IV curves of this type
MTJs. The values of
of LBD MTJs were identical to those of MTJs with the HBD
distribution. It is suspected therefore that the LBD is caused
by a spatially nonuniform MgO barriereither in thicknesses,
stoichiometry or nonuniform Boron distribution at the MgO
interface. These result in spatial variation of the tunneling current which can concentrate on spots with low barrier heights,
developing into a breakdown spot. MTJ structure and process
play a dominate role in determining the percentage of MTJs in
the LBD population and have to be optimized to eliminate the
LBD distribution. At present the percentage of LBD MTJ can
be lowered to 0.1% but efforts are still needed to eliminate
LBD MTJs completely.
separation between
In order to determine whether
peak-to-peak value of HBD and
distribution is enough for
a reliable product, time dependent device breakdown (TDDB)
measurements were performed on the HBD MTJs under various stress conditions, as shown by Fig. 2. The measured
TDDB follows the standard Weibull distribution. Following
the approach of [19] and applying a relatively loose criteria
that allows 20% chip failure after ten years of random access
operation with a 30% duty cycle and 2 bit ECC, the maximum
is around 350
allowed voltage across the MTJ
400 mV at an
value
m for current MTJ structure
and processes, as shown in Fig. 3. Also included in Fig. 3 are
of two different
two curves of write voltage at
Fig. 4. BER curves versus applied voltage measured three times for a single
MTJ. An existence of abnormal bifurcated switching process at lower probability is demonstrated.
MIN et al.: STUDY OF WRITE MARGIN OF SPIN TORQUE TRANSFER MAGNETIC RANDOM ACCESS MEMORY TECHNOLOGY
Fig. 5. Increasing the pulse width of applied switching voltage can reduce the
low-probability-bifurcated-switching significantly.
2325
Fig. 7. Reducing the short axis of ellipse can reduce the percentage of MTJ
with low-probability-bifurcated-switching.
Fig. 8. FMR images of two MTJs with same aspect ratio but different short
axis of ellipse. More FMR peaks for MTJ with larger short axis.
2326
H =0
H
84 2 178
H
1042 345
long axis of ellipse with respect to the reference layer magnetization direction. For the magnetic field applied along the hard
axis of ellipse, 20 40 Oe field along either hard axis direction
can reduce, enhance, create or have no effect on the LPBS. No
visible change of the percentage of LPBS MTJs was observed
for the wafers with tilted long axis MTJs. These results indicate
that changing the initial magnetization direction in our case has
no effect on LPBS and the time scale for LPBS ranges from 10
1000 ns range, suggesting the underlying mechanism of LPBS
is different from the bifurcation switching studied in the precessional region which is usually based on the macrospin model
which assumes homogenous magnetization [21][24].
The LPBS observed here is more of a thermal fluctuation
dominated process with a spatially nonuniform free layer magnetization. Since the LPBS is a thermal activation dominated
process, lowering the thermal barrier may help reduce the
appearance of LPBS MTJs. As shown by Fig. 9, applying an
easy axis field that assists switching can reduce the percentage
of LPBS MTJs significantly for various MTJ dimensions.
Increasing the easy axis field strength, the percentage of LPBS
MTJs can be reduced further. It is interesting to note that
the ratio of the percentage of LPBS MTJs normalized to the
decreases exponentially versus
percentage at
normalized to the product of
and MTJ area. The reason for
this relationship is not known at the present time.
Another interesting new phenomena that was observed when
is present is the appearance of a new
the assisting easy axis
type of LPBS (type-G) which can be fit by the probability distribution function (PDF) of a Gaussian distribution instead of CDF,
Oe. As the pulse width
as shown by Fig. 10 in which
increases from 10 to 200 ns, the peak value of type-G LPBS remains unchanged at 450 mV while its amplitude and sigma is
reduced monotonically, similar to the pulse width dependence
of the amplitude and sigma of type-C LPBS. The type-G LPBS
can be completely isolated from the main BER curve when the
pulse width is greater than 50 ns. Fig. 11 shows the effect of
strength on the type-G LPBS at fixed pulse width, 50 ns in this
case. By increasing the
from 10 Oe to 60 Oe, the peak position of type-G LPBS moves towards lower voltage and the peak
value increases rapidly.
The exact mechanism for the appearance of the two types
of LPBS is not known at this point. The speculation that it
H = 60
Fig. 11. At 50-ns pulse width, increasing the assisting easy axis field will lower
the peak position and raising the amplitude of type-G low-probability-bifurcated-switching.
MIN et al.: STUDY OF WRITE MARGIN OF SPIN TORQUE TRANSFER MAGNETIC RANDOM ACCESS MEMORY TECHNOLOGY
that for the main switching trajectory which should be the path
with the lowest energy barrier. Thermal agitation via either
longer pulses or higher substrate temperatures will allow the
spin system of MTJ to find the path with lowest energy barrier
and will also enhance the chance of escaping the local stable
node or stable limit cycle. Clearly, rigorous work on studying
the nonlinear dynamics of LLG equation of highly spatial
nonuniform magnetization system is needed to understand the
underlying physics of the LPBS so that solutions can be found
to eliminate it. The existence of LPBS increases the value of
and could raise the error floor to an unacceptable level for a
STT-MRAM product.
In summary, we have demonstrated that it is feasible to develop an MTJ design and fabrication process to meet the requirements of data retention, long term dielectric reliability with
write and read disturb error rate for a 64 Mb STT-MRAM
product. Nevertheless, continuing effort to reduce the critical
switching voltage/current is still needed and understanding and
eliminating both LPBS and LBD distribution are essential for
the STT-MRAM to be a successful product.
IV. CONCLUSION
Key design parameters of a 64-Mb STT-MRAM product
development were discussed. In the effort to reducing the
switching voltage/current, the values of free layer magnetization, thickness, anisotropic field and resistance-area product
have to be considered concurrently. It was found that there is
big discrepancy between the thermal barrier values obtained
with and without the spin polarized current. An experimental
method was developed to determine the required coercivity and
sigma for a given MTJ dimensions to have thermal barriers adequate for data retention and avoiding read disturbs. Two distinct
breakdown distributions with high and low breakdown value
were found in todays MTJ structure with less than 12-A thick
MgO spacer. For the MTJ population with high breakdown
value, time-dependent-device-breakdown study was performed
which resulted in a maximum write voltage in the range of 350
400 mV. Bit error rates were routinely collected to study
the write voltage and write margin of various MTJ designs.
Backhopping during the write operation was discussed and
eliminating its impact on write errors was shown to require
higher energy barriers. The existence of two types of bifurcated
switching modes was discovered, which are detrimental to the
write margin. These types of bifurcated switching have lower
probability than the main switching path and exist in the both
near-precessional and thermal regions. Their dependence on
pulse width of applied voltage, substrate temperature, MTJ
dimensions and constant external field were investigated. It
was found that an external field along the easy axis to assist
switching can significantly reduce the percentage of MTJ
showing multiple switching modes. Much work is needed to
understand and eliminate the bifurcated switching modes and
the low value breakdown distribution for STT-MRAM to be a
successful commercial product.
2327
REFERENCES
[1] M. Hosomi et al., in Proc. IEDM Tech. Dig., 2005, pp. 459462.
[2] H. Yoda et al., Spin torque transfer switching of perpendicular magnetoresistive elements for high density MRAMs, presented at the Dig.
Intermag, 2008, FA-04.
[3] L. Jiang et al., Large tunnel magnetoresistance of 1056% at room temperature in MgO based double barrier magnetic tunnel junction, Appl.
Phys. Exp., vol. 2, p. 083002, 2009.
[4] Q. Chen et al., Study of dielectric breakdown distributions in magnetic tunneling junction with MgO barrie, J. Appl. Phys., vol. 105, p.
07C931, 2009.
[5] R. Beach et al., A statistical study of magnetic tunnel junctions for
high-density Spin Torque Transfer-MRAM (STT-MRAM), presented
at the IEDM Tech. Dig., 2008.
[6] J. Z. Sun, Spin current interaction with a monodomain magnetic body:
A model study, Phys. Rev. B, vol. 62, p. 570, 2000.
[7] Z. Li et al., Thermally assisted magnetization reversal in the presence
of a spin-transfer torque, Phys. Rev. B, vol. 69, p. 134416, 2004.
[8] R. H. Koch et al., Time resolved reversal of spin transfer switching in
a nanomagnet, Phys. Rev. Lett., vol. 92, p. 088302, 2004.
[9] Y. Huai et al., Spin-transfer switching current distribution and reduction in magnetic tunneling junction-based structures, IEEE Trans.
Magn., vol. 41, no. 10, p. 2621, Oct. 2005.
[10] Z. Li et al., Perpendicular spin torques in magnetic tunnel junctions,
Phys. Rev. Lett., vol. 100, p. 246602, 2008.
[11] M. P. Sharrock, IEEE Trans. Magn., vol. 26, p. 193, 1990.
[12] S. Oh et al., Switching characteristics and bias voltage dependence
of spin torque switching in MgO based magnetic tunnel junctions.,
presented at the Intermag, 2008, GD-01.
[13] X. Zhu et al., Spin transfer excited regular and chaotic spin waves in
current perpendicular to plane spin valves, J. Appl. Phys., vol. 95, p.
6630, 2004.
[14] K. Lee et al., Micromagnetic investigation of dynamics of magnetization switching induced by a spin polarized current, Appl. Phys. Lett.,
vol. 88, p. 132506, 2006.
[15] J. Das et al., Area scaling and voltage dependence of time-to-breakdown in magnetic tunnel junctions, J. Appl. Phys., vol. 91, p. 7712,
2002.
[16] T. Min et al., Back-hopping after spin torque transfer induced magnetization switching in magnetic tunneling junction cells, J. Appl. Phys.,
vol. 105, p. 07D126, 2009.
[17] J. Sun et al., High-bias backhopping in nanosecond time-domain spintorque switches of MgO-based magnetic tunnel junctions, J. Appl.
Phys., vol. 105, p. 07D109, 2009.
[18] J. Sun et al., Spin-transfer-driven reversal in magnetic tunnel junctions: Torques, dynamics, and thermal activation effects, presented at
the Dig. Intermag, 2009, ED-02.
[19] H. Ohmori et al., Back hopped switching error in spin transfer torque
MRAM, presented at the Intermag, 2009, CT-05.
[20] T. Devolder et al., Electric time-domain observation of magnetization switching induced by spin transfer in magnetic nanostructures, J.
Appl. Phys., vol. 103, p. 07A723, 2008.
[21] Y. Acremann et al., Bifurcation in precessional switching, Appl.
Phys. Lett., vol. 79, p. 2228, 2001.
[22] G. Bertotti et al., Magnetization switching and microwave oscillations
in nanomagnets driven by spin-polaried currents, Phys. Rev. Lett., vol.
94, p. 127206, 2005.
[23] C. Serpico et al., Thermal stability in spin-torque-driven magnetization dynamics, J. Appl. Phys., vol. 99, p. 08G505, 2006.
[24] P. Gorley et al., Application of self-organization methods to currentinduced magnetization dynamics of a single-domain ferromagnet, J.
Appl. Phys., vol. 101, p. 034504, 2007.
[25] I. N. Krivorotov et al., Time-domain studies of very-large-angle magnetization dynamics exicted by spin transfer torques, Phys. Rev. B, vol.
77, p. 054440, 2008.
[26] J. P. Strachan et al., Direct observation of spin-torque driven magnetization reversal through nonuniform modes, Phys. Rev. Lett., vol. 100,
p. 247201, 2008.
[27] S.-C. Oh et al., Nature Phys., vol. 5, p. 898, 2009.