Escolar Documentos
Profissional Documentos
Cultura Documentos
IEEE JOURNAL
OF SOLID-STATE
CIRCUITS,
VOL.
1987
Static-Noise
EVERT SEEVINCK,
of MOS
LOHSTROH, MEMBER, IEEE
,4bsfrad The stability of both resistor-load (R-load) and full-(2MOS SRAM cells is investigated analytically as well as by simulation. Explicit analytic expressions for the static-noise margin (SNM) as a function of
device useful tion the parameters in predicting and supply voltage are derived. The expressions SNM It are tbe effect of parameter changes on the stability as well simulawith
reductions
of 30-50
percent have
been obtained,
is presented, predicted
supply voltage reduction to avoid hot-carrier degradation. The reason is that the SNM for R-load cells becomes much lower cient larger. order than noise for 6T cells at low supply margin the R-load voltage. then For suffiIn cells cell must the SNM be made
is further
concluded
that full-CMOS
This means that 6T cells have greater potential. to explain these statements of SRAM in this paper.
I.
INTRODUCTION
In Section II the cell stability cell area and the stability of the cell. The cell area determines about two-thirds of the total chip area. The cell stability determines the soft-error rate and the sensitivity of the memory to process tolerances and operating conditions. The two aspects are interdependent cell for improved area. There has been considerable years to understand yet cells. The basic cross-coupled appearance, attempts effort over the past several of flip-flop simple in the cell Fig. R-load a read power situation l(a) and (b) shows the II.
SRAM-CELL
is discussed with the aid of Analytical expresand 6T cells are derived easy-to-use in results thus conis developed
for SRAM
representation
of the SNM.
A and B. An
investigation
since designing
V analytic
and simulation
stability
invariably
requires
a larger cell
the validity
in Section VI.
and model
the stability
STABILITY
to analytically
stability have achieved only limited success [1]. Much of the published work has been concerned with the statistical and dynamic metastable properties [1][3]. of flip-flop The stability synchronizers in the region as expressed by the
of
the
cell and the 6T full-CMOS access and with supply voltage. because the bit This is in
cell, respectively,
static-noise margin (SNM) [4] has also been investigated f& both resistor-load [5], [6] and full-CMOS [7] SRAM cells. puter This both view, The cells However, these studies have been limited to comsimulations; analytic work has not yet been reported. paper is concerned with the SNM of SRAM cells from an analytic as well as a simulation point of in the context of submicrometer MOS technology. results are useful in optimizing the design of SRAM as well as in predicting the effect of parameter (R-load) compared cells are widely SRAMS used in NMOS(6T) full-
the resistor
or p-channel
load elements
are now shunted by the n-channel access transistors, reduces the gain of the cell inverters. Both prised cell types can be represented of two inverters as shown by a flip-flop
sources V. are static-noise sources. Static noise is dc disturbance such as offsets and mismatches due to processing and variations in operating conditions. The SNM of the flip-flop is defined as the maximum value of V. that can be tolerated by the flip-flop before changing states [4]. In this paper, only static-noise sources are taken into account. A SRAM cell should be designed such that under all conditions some SNM is reserved to cope with dynamic disturbances caused by a particles, crosstalk, voltage supply ripple, and thermal noise. A basic understanding of the SNM is obtained by drawing and mirroring the inverter characteristics and finding the maximum possible square between them. This 01987 IEEE
changes on the SNM. Resistor-load and CMOScell area when (mixed-MOS) owing to their smaller
to the six-transistor
Manuscript received April 2, 1987: revised June 4, 1987, The authors are with Philips Research Laboratories, 5600 .JA Eindhoven, The Netherlands. IEEE Log Number 8716261.
0018-9200/87/1000-0748$01.00
SEEVINCK et d.:
749
l~~
a noticeable in our
voltage drop across the resistor (about the output voltage is further with transistor is in the R-load resistor and role. that
case). Subsequently
by the access transistor and drops slope for VI. >1 V when the driver on. Note that the cell noise margin in the area where do not play the maximum the load any
access
turned
current
significant it is clear
squares,
in spite of the larger value of r for the will likewise does. In decrease cell will the
addition.
(:)
Fig. 1. SRAM cells durink
R-load cell can only be used with a l~~ somewhat larger than 2V~ (say 3 V in the case of VT= 1 V) during access to avoid write-time problems.
and (b) 6T full
+m
Vn
ANALYTICAL
lIERIVATION
OF SNM
equations
and applying
11
vn -
Fig.
2.
A flip-flop
Static-noise
voltage
cally equivalent noise margin criteria [4]. For the cells of Fig. 1, we assume the right sides to be at level ZERO and the left sides at level ONE. This means that the cell circuit diagrams components can be reduced shown dotted to those shown only negligible in Fig. 4. The currents. In, Fig. are assumed to be nonconduct-
,, ,\
$~
4(a), we assume Ql, Q3, and Qq operate in the saturation region and Q ~ in the linear region. In Fig. 4(b), QI and Qd are assumed saturated while Q ~ and Q5 are in the linear region. These assumptions were verified by simulation as well as by back substitution. Explicit expressions for the SNM of the R-load cell dnd the 6T cell were obtained by using the basic MOS model
I
3 2
q, ..! II >.!J , . . \ . .
1 h
----------
-.
with
constant
threshold
voltages saturation.
(equal
for n-
Vou+ 11
[v]
and p-channel)
and neglecting
second-order
Fig. 3. GraphicaJ representation of SNh4. Curves II have been mirrored with respect to a line passing through the origin at 450 from the horizontal
reduction
and velocity
A and B, of which
is
a graphical
technique
of estimating
the SNM
in Fig. 3, using the most basic MOS model threshold voltage and a simple exponential current model. factor For for the purpose a square of illustraall are = 2.3 (/30 indicates transistor), dimensions
SNMR=~VT+r+l-r~:
r+i(VDD-VT)
(1)
2r+l VDD r+l
1[ ~+ r
VT
SNM6,
= VT
( k+l
of ~driv., to B.Cc.,, is an
important cell parameter called the cell ratio r. It determines the cell size as well as the cell stability. In this case, r is equal to 2 for the 6T cell and 3.5 for the R-load cell. From Fig. 3 we see that the R-load inverter it characteris-
k(r+l)
750
IEEE JOURNAL
OF SOLID-STATE
CIRCUITS,
VOL.
1987
DO I 1 .,= R R ;, I 1
Qw)c.zr,a;
o x
-?1 I
.
al
Fig
1
5.
SNM
estimation
squares
in a 450 rotated
(b) Fig. 4. (a), (b) Circuit diagrams of SRAM cells when static-noise sources V. inserted. accessed, with
Third,
for particular of
will
be
independent will
cients of V~~ in (2) having opposite where r = ratio= ~d/~a voltage then result of behavior through dence stability obtained lustrated q = BP/B. VT= threshold SNM6T
a positive or a negative depenrequired V~~. Thus, a particular respect choice to varying of r and V~~ can q. This be is ilr =1,
v,=
~=~ J-()
r+l derivation where right-hand
VDD
v;
q = 3/8 and r =2, hand, for the R-load with decreasing V~~, Finally, VT. of (1) was exact; no simplifying apone increasing increasing In the case of (2), only i.e., assuming Q2 /Q4 region. around both
3/8, respectively. On the other cell the SNM will always decrease as can be seen in Fig. 7. SNMCT will will increase with that the SNM decrease with temperais
SNM ~ and
The
ture, and we have already concluded that the SNM independent of the absolute value of the ~ s.
curve of inverter
Qz is in the linear
SIMULATION
based on
METHOD
the here. graphical To technique
described
is presented
estimate
values, we can draw some the SNM for both diagonals method conclusions.
a procedure which
is needed
that
finds
values
R-load and 6T cells depends only on threshold voltage, V~~, and /3 ratios, and not on the absolute value of the @s. Therefore, the increased /3 values associated with subrnicrometer processes will not by themselves lead to improved cell stability. Second, both SNM~ and SNM6T increase with r. SNMG~ remains larger than zero for all values of r > O; on the other hand, SNM R already becomes zero for r =1. To design the cells for maximum SNM, r = B~/~a must be maximized and also (in the case of 6T cells) q/r= fp/D~ by appropriate course, and proper choice of W/L ratios. This choice 1s, of constrained cell-write by the requirements operation. of small cell area
use together with a standard dc circuit simulator [7]. Fig. 5 shows a stylized version of Fig. 3 in two coordinate systems which are rotated 450 relative to each other. In the (u, u) system, subtraction of the u values of normal and mirrored inverter characteristics at given u yields curve A, which is a measure of the diagonals length. The maximum maximum Assume istics l?~(x), To find must and minimum squares. that the normal by of curve xt represent and mirrored the required charactery =
inverter
are defined FI
SEEVINCK
et U[.:
STATIC-NOISE
MARGIN
ANALYSIS
F]
.=
: 300 m .: 200
F, Fig.
= -1
?
500
- extended model z W o simp~e model
751
Loo
X theory
@q.(1) x
xx
r=
3,5
% = 100
DD
v]
7.
SNM
of R-load
> =
~ 600
1-1= (b)
Fig.
6.
Circuit
implementations
: c .: z 300 --
model
required
transformation
is
supply voltage
VDD
[1
(3a)
Fig. 8. SNM of full-CMOS cell versus supply voltage.
(3b)
Substitution
of (3) in y = Fl(.x)
gives
=u+fiF(++a
(u, u) system. now stituting The required coordinate transformation the same as (3) but with in y = Fz (x ) gives x and y exchanged;
(4)
For F;, first Fz is mirrored in the (x, y) system with respect to the u axis, and then it is transformed to the is SubFig. 9.
ok;
;
ratio
;
r
SNM
of R-load
atio
v=u+fiFz
1
U~ V .
(5) V.
ANALYTIC AND SIMULATION RESUI,TS
0;
Equations (4) and (5) represent the inverters comprising the SRAM flip-flop cell. They give u as an implicit function of u. Solutions can be found with a standard dc circuit simulator by translating the equations into circuits, using voltage-dependent as shown represented The in Fig. by U1 and voltage sources in a feedback loop solutions
Fk. b(a)
and 6T cells, as
predicted by (1) and (2), are plotted as a function of V~~ and r for VT = 0.9 V and q = 3/8. The plots extend down to V~~ = 3 V which is the approximate limit of validity of the analytic models. In the figures the analytically predicted SNM is compared with simulations which were performed according to the method outlined in Section IV. For the simulations, both the most basic MOS model and
6. The
V2 in
of (4)
and (b),
and
(5)
are
respectively.
difference
between
the
two
solutions,
UI V2, is
calculated by the simulator and is represented by curve A in Fig. 5. The absolute values of the maximum and minimum are the values of the diagonals of the maximum squares. Multiplying the smaller of the two by l/fi yields the SNM of the flip-flop.
a fully extended model with submicrometer transistor parameters (which includes subthreshold conduction, body effect, and mobility reduction) were used. The curves obtained for the 6T cell (Figs. 8 and 10) show a good correspondence between (2) and the simulations for both transistor models. Note that the SNM is
752
IEEE JOURNAL
OF SOLID-STATE
CIRCUITS,
VOL.
SC-22,
,=
T = =
VI
x 800
700 600 500 Loo 300
12345
ratio r
of
SRAM
cells.
In and are
an has
SNM been
is quick results
.5 .
velocity
causes that
full-CMOS cells
better
than
R-load memory to
voltages.
exfended model
future be
processes, 3 V or less
simple
model eq (2)
reduced conventional
x theory:
degradation, a significant
R-load to SNM by
In
at cells
SNM
of full-CMOS
voltage, or equal
that
approximately
constant
with
r =1.7.
For
APPENDIX DERIVATION OF SNM
the
smaller ratios the SNM increases with decreasing V~~ in contrast to the behavior with larger ratios. As discussed before, this behavior is predicted by (2). The curves obtained for the R-load cell (Figs. 7 and 9) show a complete fit between (1) and simulations with the simple model. with from This is expected the extended since the derivation is observed Further saturation model. of (1) for the which saturawas exact. However, simulation has shown was omitted a slope difference We
models
A CELL
4(a). The MOS
FOR R-LOAD
of Fig.
wish we
to will
circuit
ID=
;P(VGSVT)2
1 v~~ VT ~ (
DS )
(Al)
This velocity
tion effect reduces the effective ~ of the access transistor Q4 for large V~~ (see Fig. 4(a)). This reduces the influence of the high bit-line level on the low level in the cell, apparently increasing the SNM. In the case of the 6T cell of Fig. 4(b), this effect is compensated for by the mobility reduction of Q5 and the drain feedback effects of QI and Q4. As a general observation, we see that for decreasing v ~~, R-load supply cells need a significantly noise margins. the area advantage bigger ratio than 6T Hence, for reduced of R-load cells over
and linear
regions, respectively.
we must know the operating conditions of the transistors (whether saturated or linear). Clearly, Ql, Q3, and Qd are saturated. Suppose Qz is also saturated.
follows that
The voltage
for
is then J, for
equal to r. It
gain of
take
is insufficient
operation condition
than unity
be in a metastable unbalance
must linear drain the region. currents appropriate
V. and will
the
region.
When those of
that r
the Q4,
So far the noise-margin comparison has been done for the read-access situation. When the SRAM is in the retention-mode (switched-off access transistors and V~~ = 2 V) the differences by simulation parameters 600 better owing and between SNM ~ and SNMe~ are observed to be much less. For example, for the and SNMG~ are about These values are much respectively.
of
Q1
and
Q3 we
and find
using
models,
vGS3 vT=lF(vGsl-vT)
(V.SA - v,)= Now we write 2rV~s ( the Kirchhoff voltage equations: J&z -VT -~ v~sz 1
(A3) . (A4)
than the values obtained in the read-access case to the low-impedance access transistor loads being
V&l = Vn + VDS2
GS3 = VDD VGS2 V,
switched off. However, the impedances in the K--load case are much higher compared to the full-CMOS case. This makes the R-load cell much more sensitive to ac disturbances and a particles. Substituting VI. Analytic CMOS expressions
CONCLUSIONS
v G~4= VDD VDS2. these into (A3) and (A4) yields +vDs2vT)
(AS)
SRAM
expressions
are useful in predicting the effect of parameters and operating conditions on the SNM as well as in optimizing the
.s2 -T
- ~ .s72 .
(A9)
SEEVINCK
753
Eliminating
V&2
from
(A8)
and
(A9)
and
simplifying
results in a quadratic
equation (A1O)
av&2+ bvDs2 C= o +
with a=l+r+2r3/2 b=2{~(r+l) ~=vz
s
,(L)::::%, L----LJ
o v~
DD GS2
Fig, 11.
+r(fi-l)VT
-r(fi+l)V.
} 1
(All)
Linearizing
of the Q2 /Q4 P.
inverter
substitution where V, = VD~ VT. We now find marginal stability several easiest equivalent the SNM by applying a condition for to (A1O) and (All). We can choose from stability criteria [4]. For this case it is roots [4]. For requires of coinciding root, which (Al)
Equating
the drain
using
currents
the models
of
(Bl)
b2 = 4ac or b=2G since b <0. the SNM: Substituting (All) and solving for (A12) V. yields
&2 of
-V, the
-: p-
(Bz)
voltages
n-channel
q = &/Ba,
r = Bd/ba.
voltage equations
(B3)
VGS2
(B4) (B5)
v ~-5 = VDD Vn VDS2 and Next analysis inversion, this means VDS2 > we determine is valid. i.e., we QI the range and of V~~ for wfich
with
Qg have to operate
VGS1 > VT. Together
in strong
(B6)
require
VT Vn .
(A14) the
When solving for VDsz from (A1O) and (All) under condition (A12) and substituting (A13), we find
.(K
-~,
Vn 2vDs,
+ VGS2)
(B7) (B8)
(~ (A15) with When next combining find the minimum valid: 2r4~
=
- VDS2)2 = zrV~s2
VGS2- VT-
~vDS2
V. = V~~ VT, as before. V&z or V~s2 from approximation these two equations to be to a lower a fourth-degree A simplifying equation which is too complex leading
we is
vDD min
VT.
) (A16)
{ 1+
(r+l)J~r3/2l
For example, when VT= 0.9 V t~s expression 3.2 V for r = 3.5 and 2.7 V for r =1. It follows is valid for VDD down to about 3 V.
APPENDIX DERIVATION OF SNM -B
the transfer characteristic of the inverter which is ON has a fairly constant slope around its operating point. In Fig. 11 this part of the characteristic is shown, together with a straight-line approximation through point P at V~s2 = V, which is the approximate operating point when marginal noise is applied. The linear approximation is defined by the value of VDs2 and its slope at point is derived from (B8) by sulmtitUt@ (denoted by k) is determined p. Dsz at Point
&2 = V,. The
SIOPe
by first is then
differentiating at
GS2 = V,.
FOR FuLL-CMOS
CELL
(B8) with
For the circuit of Fig. 4(b) we assume QI and Q4 to be saturated and Q2 and Q5 to operate in the linear region. These assumptions were checked by simulation and back
expressed
as
v~S2 =
V(O kV&2
(B9)
754
IEEE JOURNAL
OF SOLID-STATE
CIRCUITS,
VOL.
SC-22,
NO. 5, OCTOBER
1987
with
[8]
~=~ -J()
r+l
and
VT
(B1O)
[9]
T. Ohzone, M. Fukumoto, G. Fuse, A. Shinohara, S. Odanaka, and thin polycrystalline-silicon high-vafue M. Sasago, Ion-implanted resistors for high-density poly-load static RAM applications, IEEE Trans. Electron Devices, vol. ED-32, no. 9, pp. 1749-1756, Sept. 1985, C. F. Hill, Noise margin and noise immunity in logic circuits, Microelectron., vol. 1, pp. 16-21, Apr. 1968.
V(O kv, + =
l+r y.
(B12)
Next
we
eliminate we obtain
(B7)
and
(B9).
After
simplifying,
Evert Seevinck (M75SM85) was born in Doetinchem, The Netherlands, on April 15, 1945, He was educated in South Africa, receiving the B. SC. degree in mathematics and physics in 1966, the B. SC. degree in electrical engineering in 1970, the B. SC. Hens. degree in electronic engineering (cum laude) in 1975, and the D. SC. degree in electronic engineering in 1981, all from the University of Pretoria, Pretoria, South Africa. His dissertation dealt with the anafysis and synthesis
of
X2
l+2k+Lk2 7
+2X
1kz4+A+V~-V, q
1
V,
tran.liqear
. . . . . . . . .
integrated
circuits.
+~A2=0 q (B13)
From 1970 to 1972 he was Nijmegen and Eindhoven, The design and application of analog to South Africa, where he joined
with Philips Gloeilampenfabrieken in Netherlands, where he worked on the integrated circuits. In 1973 he returned Philips in Johannesburg, continuing IC
where,
for simplicity,
we have defined
x= VDD Vn vG~2
A= As in Appendix criterion solve for V. to VO+(k+l)VnkV~~
(B14)
stability
application work. From 1975 to 1981 he was employed at the Council for Scientific and Industrial Research (CSIR) in Pretoria, where he performed research and development on novel circuit techniques and custom ICS, In 1981 he remigrated to The Netherhmds, returning to Philips and working on analog IC design. In August 1983 he became Professor of Electrical Engineering at the University of Twente, Enschede, The Netherlands, In October 1985 he returned to Philips Research Laboratories, Eindhoven, The Netherlands, where he is now performing circuit research. He maintains a uart-time professorship at the University of Twente,
to (B13).
obtain
we substitute
SNM:
(B14),
and finally
Frans J. List was born in Hong Kong in 1958. He received the Ingenieur degree from Twente University, The Netherlands, in 1984. Since then he has been with the Philips Research Laboratones, Eindhoven, The Netherlands, where he worked on development and design of memories, Currently he is working on a l-Mbit SRAM design.
REFEREFJCES
[1] R.
[2]
[3] [4]
C. Jaeger and R. M. Fox, Phase plane analysis of the upset characteristics of CMOS RAM cells, in Proc, Uniu ,/Govt./Industry Microeiectron. Syrnp. (Auburn, AL), June 1985, pp. 183-187, H. J. M. Veendrick, The behavior of flip-flops used as synchronizers and prediction of their failure rate, IEEE J. So[id-State Circuits, vol. SC-15, no. 2? pp. 169-176, Apr. 1980. and A. Alblcki, Analysls of metastable operation in T. Kacprzak RS CMOS flip- ffops, IEEE J. Solid-State Circuits, vol. SC-22, no. 1, pp. 5764, Feb. 1987. J. Lohstroh, E, Seevinck,, and J, de Groot, Worst-case static noise margin criteria for logic clrcuils and their mathematical equivalence, IEEE J. Solid-State Circuits, vol. SC-18, no. 6, pp. 803807, Dec. 1983, K. Anami, M. Yoshimoto, H. Shinohara, Y. Hirata, and T. Nakano, Design considerations of a static memory cell, IEEE J, Solid-State Circuirs, vol. SC-18, no. 4, pp 414-418, Aug. 1983. B, Chappell, S. E. Schuster, and G, S. Sai-Hafasz, Stability and SER analysis of static RAM cells, IEEE J, Solid-State Circuits, vol. SC-20, no. 1, pp. 383-390, Feb. 1985, F. J. List, The static noise margin of SRAM cells, in Dig. Tech. Papers, ESSCIRC (Delft, The Netherlands), Sept. 1986, pp. 1618.
(M79) was born in Den Haag, The Netherlands, on July 11, 1946. He received the M. SC. degree in applied physics from the Technical University of Delft, Delft, The Netherlands, in 1970. He received the Ph.D. degree in electronic engineering from the Technicaf University of Eindhoven, Eindhoven, The Netherlands, in 1981, with a dissertation on integrated Schottky logic (ISL), In 1970 he joined the Philips Research Laboratories, fxndhoven, lhe Netherlands. Initially he worked on integrated magnetic memories and silicon imaging devices for optical memories. Then he was involved with bipolar logic circuitry and memories. In this area he worked on device concepts, modeling, and application of punch through devices, 12 L and ISL logic, ECL memories, and noise-margin analysis of digital circuits. He has coauthored over 30 papers and holds severaf patents in the area of microelectronics. In 1983 he became Department Head of the research group for Digital Circuitry and Memories in the Philips Research Laboratories. From 1985 to 1987 he was Department Head of the Philips Advanced Memory Design Centre. Since 1987 he has been Head of the Centraf Application Laboratory of the Philips Component Division Elcoma in Eindhoven, The Netherlands. His current interest is in consumer, industrial, and telecommunication applications of VLSI, Dr. Lohstroh became a member of the European Program Committee of the ISSCC in 1983. Since 1985 he has been chairman of this committee.
Jan Lohstroh