Page 1

Magnetic tunnel junction design margin exploration for self-reference

sensing scheme

Z. Sun,1,a)H. Li,1,b)and X. Wang2,c)

1Electrical and Computer Engineering Department, Polytechnic Institute of New York University,

6 Metrotech Center, Brooklyn, New York 11201, USA

2Seagate Technology LLC, 7801 Computer Avenue South, Bloomington, Minnesota 55435, USA

(Presented 31 October 2011; received 6 October 2011; accepted 19 December 2011; published

online 13 March 2012)

This work investigates the magnetic tunnel junction (MTJ) design requirements for the application

of nondestructive self-reference sensing scheme, a novel sensing scheme featuring high tolerance

of process variations, fast sensing speed, and no impact on device reliability. Unlike the

conventional sensing scheme that requires a large TMR ratio and the uniform antiparallel and

parallel resistances for MTJs, the nondestructive self-reference sensing scheme is more sensitive to

the roll-off slope of MTJ’s R-I or R-V curve. Our purpose is to provide a guidance to facilitate

MTJ design used in the nondestructive self-reference scheme. In this work, we comprehensively

investigate and analyze the design matrix by considering MTJ device physical properties, such as

bias voltage dependent conductance, spin torque, etc. The manuscript suggests the approaches to

optimize MTJ design for better trade-off between device properties and circuit design. V

American Institute of Physics. [doi:10.1063/1.3679647]

C 2012

I. INTRODUCTION

Spin-transfer torque random access memory (STT-RAM)

becomes a promising memory candidate for the future com-

puting systems because of its fast access time, high density,

nonvolatility, small write current, and good scalability. How-

ever, similar to all the other nano-scale technologies, STT-

RAM suffers from significant process variations, which causes

large distributions of the antiparallel and parallel resistances

of memory cells. As the fabrication technology keeps scaling

down, the process variations demonstrate more and more

severe impacts. Consequently, the read failures of the conven-

tional sense scheme using dummy cells as references increase

dramatically and hence resultinintolerablelowyield.

Some self-reference sensing schemes were proposed to

overcome bit-to-bit process variations in magnetic tunnel

junction (MTJ) resistance. The conventional self-reference

sensing scheme1compares the original value to a reference

value written into the same MTJ. The two write steps incur

high power consumption and long access latency, as well as

raise the concern about the device reliability.

Recently, a nondestructive self reference sensing scheme

(NSRS)2–4have been proposed by leveraging the different

current/voltage dependence of the antiparallel and the parallel

resistance states of an MTJ. Compared to the conventional

self reference scheme, NSRS significantly reduces read la-

tency and power consumption by eliminating write steps.

However, it requires a larger read current or voltage which

could induce unwanted MTJ flips during read. In this work,

we investigate and analyze MTJ design matrix by comprehen-

sively considering device physical properties (such as bias

voltage dependent conductance, spin torque, etc.) and circuit

considerations in NSRS. The purpose of the work is to pro-

vide a guidance to facilitate MTJ optimization with a better

trade-off between device properties and circuit design.

II. NONDESTRUCTIVE SELF-REFERENCE SENSING

SCHEME

The nondestructive self reference scheme (NSRS)2is

based on the unique feature of MgO-based MTJ: the current

dependence of the antiparallel and the parallel resistance

states of an MTJ are quite different. Figure 1 shows the

measured R-I sweep curve of an MgO-based MTJ with

90nm?180nm cell size, under a 40ns voltage pulse.

Some points missing from the 40ns pulse measurement are

extrapolated based on DC (static) measurement. When the

applied current or voltage increases, the antiparallel-state

resistance of MTJ decreases rapidly. On the contrary, the

parallel-state resistance merely changes as the current or

voltage varies.

FIG. 1. (Color online) The measured R-I sweep curve of a typical MgO-

based MTJ (Ref. 3).

a)Electronic mail: szheny01@students.poly.edu.

b)Electronic mail: hli@poly.edu.

c)Electronic mail: xi-aobin.wang@seagate.com.

0021-8979/2012/111(7)/07C726/3/$30.00

V

C 2012 American Institute of Physics 111, 07C726-1

JOURNAL OF APPLIED PHYSICS 111, 07C726 (2012)

Page 2

NSRS can be classified as current- or voltage-driven

designs, which is determined by applying given read currents

or voltages, respectively. The read operation of NSRS

includes three steps. For example, in a current-driven design,

first we apply a small read current IR1to generate a sensing

voltage VO1¼IR1RAP1or IR1RP1. Here, RAP1and RP1repre-

sent the MTJ resistance under current IR1at antiparallel (AP)

or parallel (P) states, respectively. VO1is stored in a capaci-

tor. Next, a larger current IR2¼2IR1is applied. The corre-

spondingsensing voltage

depending on the MTJ resistance state. To cancel out the

impact of the increased read current amplitude, we use a

voltage divider to produce V0

state can be detected by checking DVO¼VO1?V0

shown in Fig. 1, DVO,AP¼IR1(RAP1?RAP2) ? 0 if the MTJ

is at antiparallel state. On the contrary, when the MTJ is at

parallel state, DVO,P¼IR1(RP1?RP2) ? 0. In NSRS design,

DVO,APis defined as the sense margin, which must be large

enough to be detected by the sense amplifier.

Since the current or voltage in read operations are

injected from the free layer to the reference layer of MTJ,

the MTJ state could be flipped from AP to P, which is called

as read disturbance. Note that the MTJ read disturbance

from P to AP never happens.

VO2¼2IR1RAP2

or2IR1RP2,

O2¼0.5VO2. Finally, the MTJ

O2. As

III. MTJ DEVICE PHYSICS FOR SELF REFERENCE

SCHEME

MTJ resistance and spin torque dependence upon bias-

ing voltage recently have been extensively investigated

through experimental and theoretical approaches.5–17This

dependence in general can be written as:

GP;AP¼ GP;AP;0ð1 þ bP;APV þ cP;APjVjcþ dP;APV2Þ;

T==¼ T==0ðb==V þ c==jVjcþ d==V2Þ;

T?¼ T?0ð1 þ b?V þ c?jVjcþ d?V2Þ:

(1)

(2)

(3)

Where G is conductance, T is spin torque with == denoting

longitudinal component and ? denoting perpendicular com-

ponent. The coefficients (a, b, c, d, c, l) are determined by

electronic and spin tunneling process across the barrier.

For elastic tunneling, the terms depending upon absolute

value of voltage can be dropped: CP;AP¼ 0, C==;?¼ 0. For

elastic tunneling through a symmetric ferromagnetic-insula-

tor-ferromagnetic FM/I/FM junction structure, above equa-

tion can be simplified as

GP;AP¼ GP;AP;0ð1 þ dP;APV2Þ;

T==¼ T==0ðb==V þ d==V2Þ;

T?¼ T?0ð1 þ d?V2Þ:

(4)

(5)

(6)

Notice only quadratic voltage dependence terms appear in

conductance and perpendicular spin torque.

For elastic tunneling through symmetric FM/I/FM struc-

ture, in the case of high tunneling barrier, coefficient dP,AP

dependence upon barrier height and thickness can be analyti-

cally obtained through WKB approximation. According to

Brinkman model:

dP;AP¼me2

4? h

d2

/;

(7)

where d is barrier thickness and / is barrier height, reducing

tunneling barrier height for fixed barrier thickness could

increase slope of resistance dependence upon voltage. How-

ever this will decrease tunneling resistance and increase spin

torque magnitude. For high barrier case, MTJ resistance and

spin torque is inversely proportional to exponential of energy

barrier, decreasing tunneling barrier height alone could signif-

icantly decrease resistance and increase spin torque. This will

cause more disturbance and error in self-reference sensing

scheme. However, if barrier thickness and barrier height

could be tuned in a way such that

dffiffiffi

/

p

increases and

ffiffiffiffi

/

p

d

remains the same, there is a chance to increase resistance

roll-off sensitivity and remain the same or even reduce the re-

sistance and spin torque magnitude for less read disturbance.

Another way to manipulate resistance roll-off curve and

spin torque roll-off is through asymmetric structure and/or

inelastic scattering. For example, in inelastic scattering, dif-

ferent physics mechanism will generate different resistance

roll-off and spin torque roll-off. This is reflected as different

exponential of absolute value of voltage as in above formula.

We will explore more general case in future publications.

IV. SELF REFERENCE DESIGN SPACE EXPLORATION

To reflect the latest research status, we selected the MTJ

device demonstrated by Zhao et al.18as the baseline design,

and then explore the MTJ design space in NSRS by varying

the key parameters. The baseline MTJ design has a dimen-

sion of 130nm?50nm. Its tunneling magnetoresistance

(TMR) ratio is 135%. The resistance area (RA) product at

parallel state is 4.3 Xlm2. Accordingly, RAPand RPat a

close to zero current are 1567X and 667X, respectively. The

thermal stability D is an important parameter affecting MTJ

critical switching current IC0. Zhao et al.18tested it in two

different ways and obtained 50 and 68, respectively. Previ-

ous research works19–21show that the roll-off slope could

range from 1KX=mA to 8KX=mA. To be consistent with

Ref. 18, we selected a moderate value 2KX=mA as the base-

line. IR1is fixed at 0.025mA and the read current pulse dura-

tion2is 7ns. IC0is about 0.152mA.

Figure 2(a) shows the sense margin of the current-driven

NSRS under three roll-off slopes when fixing IC0at 0.152mA.

The sense margin increases as IR2increases. The relationship

between sense margin and roll-off slope as well as IR2are

both in linear manner. When changing IR2from 0.1 mA to

0.15 mA without considering read disturbance, the corre-

sponding sense margin sweeps from 25mV to 190.5mV.

In reality, IR2is constrained by the read disturbance

probability (Prsw). Theoretically, Prswof an MTJ at a read

current IRcan be expressed as2

?

Prsw¼ 1 ? exp

?t

s? exp ?D 1 ?IR

IC0

??? ??

:

(8)

07C726-2Sun, Li, and WangJ. Appl. Phys. 111, 07C726 (2012)

Page 3

Here, t is the sensing current pulse width. D is the thermal stabil-

ity. s is write pulse width; IC0is critical switching current, which

is the minimum current to switch the MTJ resistance for s.

If the minimum requirement of the raw bit yield is set to

be 99.999%, the read disturbance probability then cannot

exceed 0.0001%. Figure 2(b) shows the read disturbance prob-

ability versus IR2by varying IC0. Constrained by the yield, the

sense margin of design with 2KX=mA roll-off slope and 0.102

mA critical switching current turns out to be 50mV. According

to the simulation results, the applied IR2must below 0.07 mA if

IC0¼0.102 mA. If we relax the requirement of raw bit yield to

be 99.99%, the applied IR2must below 0.078 mA.

In summary, once the roll-off slope and spin torque (i.e.

IC0) are determined, the applied read current decides both the

sense margin and read disturbance probability. The absolute

value of MTJ resistance does not affect the effectiveness of

the current-driven NSRS.

NSRS can also be implemented by applying read voltage,

called as voltage-driven NSRS. Two read voltages VR1and

VR2are added on the memory cell successively to generate

sensing currents IO1¼VR1/RAP1 and IO2¼VR2/RAP2 when

MTJ is at antiparallel state, or IO1¼VR1/RP1and IO2¼VR2/

RP2when MTJ is at parallel state. The data stored in MTJ can

be detected by comparing IO1and IO2. Unlike current-driven

NSRS, the absolute values of MTJ resistances have a large

impact on voltage-driven NSRS. Explicitly, RAPand RPdeter-

mine not only the sense margin but also the generated sensing

current and hence the read disturbance probability.

Figure 3(a) shows seven MTJ designs with different pa-

rameters including roll-off slope, critical switching current

(IC0) and resistance at antiparallel state (RAP). IC0is deter-

mined by spin torque. And both IC0and RAPare inversely

proportional to the exponential of energy barrier. Therefore,

as the roll-off slope increases from 2KX=mA to 3KX=mA by

reducing the tunneling barrier. Both IC0and RAPreduce as

shown in Fig. 3(a). Here, Design 1 is the baseline as we

presented in section III. The we change the roll-off slope

from 2KX=mA to 3:2KX=mA. Accordingly, IC0varies from

0.152mA to 0.08mA and RAPvaries from 1567X to 859X.

With different design parameters, Fig. 3(b) shows the

corresponding sense margin and read disturbance probability

of voltage-driven NSRS. The sense margin is proportional to

roll-off slope similar to that of current-driven NSRS. when

the second read voltage is applied, the decreasing of resist-

ance can result in increasing of current flow through MTJ.

So the read switching probability is more sensitive than that

of current-driven NSRS. If the requirement of yield is con-

strained at 99.99%, only Design 1 and Design 2 can meet the

requirement. Their sense margin are 0.08 mA and 0.09 mA,

respectively. If we tight the constraint of the yield require-

ment to be 99.9999%, only Design 1 is legal.

V. DISCUSSIONS

We comprehensively study the MTJ device physics to

provide guidance to facilitate nondestructive self-reference

scheme. Both current- and voltage-driven non-destructive self-

reference schemes have been studied analytically and quantita-

tively based on various MTJ device parameters with the tuning

of the MTJ tunneling barrier, which manipulates antiparallel-

state roll-off slope and spin torque. The simulation results

demonstrate that antiparallel-state roll-off slope and spin tor-

que of MTJ determine the sense margin and read disturbance

probability, respectively for both current- and voltage-driven

schemes. Only voltage-driven scheme is sensitive to the MTJ

absolute resistance in its read disturbance probability.

1G. Jeong et al., IEEE J. Solid State Circuits 11, 1906 (2003).

2Y. Chen et al., Int. Symp. Low Power Electron. Des., 1 (2010).

3Z. Sun et al., Int. Conf. Comput. Aided Des., 432 (2010).

4Y. Chen et al., US Patent 7852665, 14 December 2010.

5J. C. Sankey et al., Nat. Phys. 4, 67 (2008).

6H. Kubota et al., Nat. Phys. 4, 37 (2008).

7S. Petit et al., Phys. Rev. Lett. 98, 077203 (2007).

8I. Theodonis et al., Phys. Rev. B 97, 237205 (2006).

9A. Kalitsov et al., Phys. Rev. B 79, 174416 (2009).

10Y.-H. Tang et al., Phys. Rev. B 103, 057206 (2009).

11A. Manchon et al., J. Phys.: Condens. Matter 20, 145208 (2008).

12A. Manchon et al., Phys. Rev. B 82, 174420 (2010).

13X. Wang et al., Phys. Rev. B 79, 104408 (2009).

14J. Xiao et al., Phys. Rev. B 77, 224419 (2008).

15M. Wilczynski et al., Phys. Rev. B 77, 054434 (2008).

16C. Heiliger et al., Phys. Rev. B 100, 186805 (2008).

17J. C. Slonczewski et al., J. Magn. Magn. Mater. 310, 169 (2007).

18H. Zhao et al., J. Appl. Phys. 109, 07C720 (2011).

19N. Mojumder et al., J. Appl. Phys. 108, 104306 (2010).

20E. Evarts et al., Appl. Phys. Lett. 95, 132510 (2009).

21K. Ono et al., Int. Electron Devices Meet., 1 (2009).

FIG. 2. (Color online) (a) Sense margin and (b) read disturbance probability

vs IR2under different design parameters.

FIG. 3. (Color online) Sense margin and read disturbance probability com-

parison of different MTJ designs.

07C726-3 Sun, Li, and WangJ. Appl. Phys. 111, 07C726 (2012)