M
A
S
T
E
R
S
E
R
I
E
S
Feature
I. Introduction
I
n the last decade, we have witnessed an increasing
interest in MOS CurrentMode Logic (also named
SourceCoupled LogicSCL) circuits, which repre
sents an alternative to traditional CMOS logic styles in
several applications. Despite of their recent adoption,
MCML circuits actually have quite old ancestors in their
family tree, as they directly descend from the bipolar
CurrentModel Logic (CML) which has the same topology,
despite of the different adopted technology [1].
The fundamental structure of an ninput MCML gate is
depicted in Figure 1, where an NMOS network (consisting
of properly stacked sourcecoupled pairs) steers the bias
current I
SS
to one of the two output branches, according
to the value of the differential inputs v
i1=
v
i1,1
v
i1,2
, . . .
v
in=
v
in,1
v
in,2
. The steered current is then converted
into a differential output voltage v
o=
v
o,1
v
o,2
by the two
resistances R
D
(in red line) which can be often imple
mented by physical resistors, or alternatively by PMOS
transistors (working in the triode region) active load. As
opposite to previous works dealing with the powerdelay
tradeoff management in MCML
gates [2][5], in the following a
physical resistor will be assumed.
The current source I
SS
in Figure 1
is usually implemented by a simple
current mirror, which is not shown
for the sake of simplicity. The load
capacitance C
L
represents the
external capacitance due to the
input capacitance of the following
gates and the wiring capacitance.
The general topology in Figure 1
allows the implementation of both
combinational and sequential gates
whose logic function only depends
on the connection of the source
coupled pairs. The implemented
function can also be modified by
negating the inputs and the output,
i.e., by simply swapping the corre
sponding pairs of differential sig
nals. As an example of the simplest
logic gate, the topology of an MCML
inverter is depicted in Figure 2,
where the NMOS network consists
of only one sourcecoupled pair. As
other examples, the NMOS network
topology of a 2input Multiplexer
41
Massimo Alioto is with the DII (Dipartimento di Ingegneria dellInformazione), UNIVERSIT DI SIENA, v. Roma, 56, I53100 SIENA ITALY,
Email: malioto@dii.unisi.it. Gaetano Palumbo is with the DIEES (Dipartimento di Ingegneria Elettrica Elettronica e dei Sistemi), UNIVERSIT
DI CATANIA, Viale Andrea Doria 6, I95125 CATANIA ITALY, Email: gpalumbo@diees.unict.it
FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
V
DD
v
o,1
v
o,2
C
L
R
D
R
D
I
SS
NMOS Source
Coupled Pairs
Network
v
i1,1
...
v
i1,2
...
C
L
v
i2,1 v
i2,2
v
in,1
v
in,2
Figure 1. Topology of
a generic MCML gate.
NMOS Source
Coupled Pairs
Network
M1 M2
v
i,1 v
i,2
V
DD
v
o,1
v
o,2
C
L
R
D
R
D
I
SS
C
L
M1 M2
v
i,1
v
i,2
Figure 2. Topology of
a MCML inverter gate.
(MUX), XOR and Dlatch gate are shown in Figures 35,
respectively, (for an overview on the design techniques to
derive the topology of arbitrary MCML gates, the reader
is directed to [2]). Their static operation is easily under
stood by considering the simplest case of the inverter
gate in Figure 2: when the input voltage v
i=
v
i,1
v
i,2
is
high (low), the sourcecoupled pair M1M2 completely
steers the current I
SS
to the drain of M1 (M2), thus the
output voltage is equal to the low (high) value
V
OL=
(V
DD
R
D
I
SS
) V
DD
= R
D
I
SS
(V
OH =
R
D
I
SS
).
The obtained logic swing is
V
SWING
= V
OH
V
OL
= 2R
D
I
SS
(1)
which is rather small, typically in the order of a few hun
dreds of millivolts. Due to the symmetry of the IV trans
fer characteristics of the sourcecoupled pair and of the
circuit, the logic threshold V
LT
is equal to zero (i.e.,
v
o
= 0 when v
i
= 0).
Unfortunately, the power dissipated by the MCML gate
is dominated by the static power consumption V
DD
I
SS
due to the bias current source since the dynamic contri
bution (associated with the capacitance charge during
the gate switching) is rather small due to the reduced
logic swing. For this reason, various techniques have
been adopted to dynamically reduce the static power
consumption [1]. The static power consumption is the
fundamental weak point of MCML gates, thus in their
design it must be kept as small as possible for a given
required performance by consciously managing the
powerdelay tradeoff, both to efficiently design MCML
cells and develop an automated design flow. In the fol
lowing sections, poweraware design strategies will be
derived to address this problem.
Compared to traditional CMOS logic, MCML gates
exhibit various interesting features that make them suit
able for an increasingly wide range of applications:
1) MCML gates are faster. The higher speed allows for
implementing circuits for fast communication sys
tems (e.g., multiplexing/demultiplexing ICs in the
range of 10 Gb/s for SONET/SDH opticfiber links
and highspeed crosspoint switches) and RF cir
cuits (e.g., PLL, prescalers, circuits for clock
recovery and VCOs), as well as highspeed cur
rentmode buffers [6][13]. This speed improve
ment is due to the tremendous CMOS technology
scaling, and allows for replacing previous bipolar
CML logic [14]. However, in Section V it will be
shown that the high speed performance is not due
to the small logic swing, as opposite to the com
mon belief.
2) MCML gates have a better power efficiency at high
frequencies. This enforces the suitability of MCML
gates for highfrequency applications, since from
the last decade a low power consumption is also
required in highspeed circuits for reasons related
to the heat removal, as well as to the battery life
time in portable devices [15], [16]. This has extend
ed the range of applications of MCML gates to the
implementation of highspeed lowpower arith
metic and signal processing cores [15].
3) MCML gates generate a much lower switching
noise during switching. Indeed, the power supply
must provide a static power and thus a constant
current to each gate. This avoids the typical cur
rent spikes of CMOS logic that determine large
voltage variations on the supply voltage V
DD
[17][22] which in turn couple with the eventual
analog circuits sharing the same substrate (as
occurs in current SystemsonChip) and degrade
their resolution. In particular, the almost con
stant supply current leads to an almost zero volt
age drop in the bonding
wires/supply rails inductance due
to current variations di/dt [6],
which will be increasingly impor
tant in next technology nodes, in
which both this inductance and
the supply current variations are
expected to dramatically grow
because of the increased clock
frequency [23], [24].
The switching noise generated
by MCML circuits is typically
reduced by two orders of magni
tude, thus this logic style is cur
rently adopted in most highspeed
highresolution mixedsignal ICs for
digital audio and video signal pro
42 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
V
DD
v
o,1
v
o,2
C
L
R
D
I
SS
C
L
M1 M2 SEL
1
M3 M4 A
2
M5 M6
A
1
B
2
B
1
SEL
2
0
1
A
B
SEL
OUT
2
:
1
M
U
X
R
D
Figure 3. Topology of a 2:1 Multiplexer.
cessing (such as sigmadelta A/D and D/A convert
ers) [25][32].
4) MCML gates have a better signal integrity and a lower
delay noise. This is due to the much lower supply
voltage noise (discussed in point 3) and the differ
ential operation of MCML gates, which are insensi
tive to commonmode signals, including the supply
noise. This greatly simplifies the design of the sup
ply distribution network and reduces the size (and
area) of decoupling capacitors needed to ensure
low V
DD
variations.
Interestingly, MCML circuits can also be made
insensitive to the noise arising from the (capaci
tive) coupling with other switching circuits. Indeed,
this coupling noise becomes a commonmode sig
nal if the cells and the interconnects are carefully
designed with a symmetric layout. This also avoids
the delay variations due to the capacitive coupling
with other switching gates
(often named delay noise
[33]), which is a major source
of delay uncertainty in cur
rent CMOS logic circuits.
5) MCML gates potentially have a
lower sensitivity to process,
supply and environmental
variations. Simple techniques
to significantly lower the
effect of process tolerances
have been developed for
MCML gates circuits [1], [16],
even though this aspect has
not completely understood
and is currently under inves
tigation [34]. As an example,
the variation in the logic
threshold V
LT
due to process
tolerances determines an
uncertainty on the input
switching time (in which
v
i
= V
LT
) and thus on the
delay. However, the V
LT
varia
tions in MCML gates are
mainly due to the mismatch
of sourcecoupled NMOS
transistors (or load resistanc
es), whereas the CMOS varia
tions are due to the poorer
mismatch between a PMOS
and an NMOS transistor [35].
Thus, a lower uncertainty in
V
LT
and in the delay is
expected in MCML circuits.
Due to the essentially static power consumption,
the chip temperature and the supply voltage also
tends to be constant (according to point 3), there
by minimizing the delay variations associated with
supply and environmental variations.
The potentially lower delay uncertainty is an
appealing property since it is becoming an
increasing fraction of the clock period [36], and
thus a major limit to the speed improvement in
current DeepSubMicron (DSM) technologies [33].
6) MCML gates suffer from a lower degradation of the
electrical transistor properties due to DSM effects.
This is due to the lower logic swing, which reduces
the voltages across the transistors terminals, and
thus the electric field under the transistor channel.
This reduces DSM effects, such as the carrier mobil
ity degradation and velocity saturation, when com
pared to standard CMOS logic.
43 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
V
DD
A
B
OUT
v
o,1 v
o,2
C
L
R
D
R
D
I
SS
C
L
M1 M2
M3 M4
A
2 M5 M6 A
1
A
2
B
2
B
1
Figure 4. Topology of a XOR gate.
V
DD
D
CLK
OUT
v
o,1 v
o,2
C
L
R
D
I
SS
C
L
M1 M2
M3 M4 A
2
M5 M6 A
1
D
Q
Q Clk
CLK
CLK
R
D
Figure 5. Topology of a D Latch.
Moreover, the worse speed performance of PMOS
transistors does not impose a speed limit in MCML
gates, since the switching sourcecoupled pairs
are made up of only NMOS transistors [7].
According to points 13, the range of applications in
which MCML gates exhibit significant advantages has
continuously broadened. This trend is expected to con
tinue according to points 36, since MCML gates are less
sensitive than CMOS logic to limitations arising in sub
100 nm technologies.
II. Static Analysis Through the AlphaPower Law
The noise margin NMis the fundamental requirement on
the static behavior of any logic style, and for a single
input gate (i.e., the inverter in Figure 2) it is defined as
V
OH,min
V
I H,min
(or equivalently as V
I L,max
V
OL,max
,
due to the symmetry of the transfer characteristics) from
the critical points of the DC transfer voltage characteris
tics (V
I L,max
, V
OH,min
) and (V
I H,min
, V
OL,max
) in Figure 6.
In the more general case with multiple inputs, the
noise margin is evaluated from the DC characteristics
associated with a given input v
i
driving a sourcecoupled
pair M1M2 with the other inputs being preliminarily
assigned [6]. The transistors driven by the latter con
stant inputs do not affect the static behavior of the gate,
since they are switched off or can be assumed as short
circuited, thus the DC behavior of multiple gates is equal
to that of a simple inverter made up of the sourcecou
pled pair M1M2, according to Figure 7.
In general, the noise margin NM might depend on the
considered input v
i
, but in practical MCML gates all
sourcecoupled pairs are made identical to have the same
noise margin for all inputs, hence the noise margin
expression of the inverter is immediately extended to
arbitrary logic gates.
IIA. Evaluation of the Noise Margin:
A Simple Approach.
In this subsection, a novel simplified approach is
adopted to evaluate the noise margin of nanometer
MCML circuits by assuming a IV relationship of MOS
transistors given by the wellknown AlphaPower law
44 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
R
D
I
SS
v
o
(v
i
)
v
i
V
ILmax
V
LT
= 0
0
V
IHmin
V
OHmin
R
D
I
SS
V
OLmax
Figure 6. Typical DC transfer characteristics of a MCML gate.
V
DD
v
o,1
R
D
M1 M2
v
i,1
Upper Source
Coupled Pairs
Lower Source
Coupled Pairs
I
SS
V
DD
v
o,2
I
SS
M1 M2
R
D
v
i,2
v
i,1
v
i,2
R
D
R
D
v
o,2
v
o,1
I
SS
Figure 7. Noise margin: equivalence of multipleinput gates to an inverter gate.
[37]. The latter expresses the NMOS drain current i
D
as
a function of the gatesource voltage v
GS
in the satura
tion region
i
D
= K W (v
GS
V
TH
)
(2)
W being the effective channel width, V
TH
the transistor
threshold voltage, K and technologydependent coeffi
cients (channel length modulation effect was neglected, as
usual). In particular, old longchannel technologies have a
square IV law, thus parameter is equal to 2 and K is
equal to
n
C
OX
/2L [6], where
n
is the electron mobility,
C
OX
is the gate oxide capacitance per unit area and L is
the effective channel length. In this case, the MCML noise
margin has been previously found to be given by, [2], [38],
NM =
V
SWING
2
1
A
V
(3)
where is a constant coefficient equal to
2, and A
V
is
the magnitude of the smallsignal voltage gain around the
logic threshold given by, [2],
A
V
= g
m
R
D
(4)
g
m
being the transistor transconductance around the
logic threshold.
In the limit case of a very shortchannel device with
a completely saturated carrier velocity, the IV relation
ship is linear, = 1, and K = v
sat
C
OX
, [6], [37]. In this
case, by performing the simple calculations reported in
Appendix I, the noise margin turns out to be still given
by (3) but with a different value of , which in this case
is equal to 1. In actual nanometer devices, as shown for
example by the data in Table 1 referring to a 90nm
technology, is somewhat intermediate between 2 and
1, and thus is expected to range from
2 1.4 and 1.
As a reasonable approximation, can be set to the
intermediate value 1.2, which leads to
NM =
V
SWING
2
1
1.2
A
V
(5)
Extensive simulations were performed by varying the
logic swing from 240 mV to 800 mV, with A
V
ranging from
1.6 to 2.5, adopting a 90 nm technology whose main
parameters are reported in Table 1. The error of the ana
lytical model (5) was found to be always lower than 14%
and typically in the order of a few percent. Typical values
of NM are in the order of 100 mV in current nanometer
technologies.
IIB. Considerations on the Technology Scaling
and Circuit Design.
From (5), the noise margin is proportional to the logic
swing, and roughly equal to half of it, if A
V
is suffi
ciently high. Next, the comparison of (5) with (3)
shows that the noise margin achieved with nanometer
devices is greater than that with old longchannel tran
sistors, for assigned values of the logic swing and the
voltage gain. This is a good news, since it means that
DSM effects are beneficial in terms of the noise margin
in MCML gates, and that the longchannel model in (3)
is pessimistic for current technologies. However, the
maximum logic swing which ensures the transistor
operation in the saturation region is equal to 2V
TH
[2],
which slightly decreases when scaling the technology,
thus the maximum noise margin tends to decrease
slowly.
Now, let us derive simple design equations to size R
D
and transistors in order to obtain assigned values of
V
SWING
and A
V
satisfying the noise margin requirement
(more detailed design guidelines to preliminarily assign
these two parameters will be discussed in Section V).
Solving (1), a given logic swing is achieved by properly
setting the resistance R
D
to V
SWING
/2 I
SS
, whereas from
(4) an assigned value of A
V
is achieved by setting the
NMOS transconductance g
m
in (6) to A
V
/R
D
g
m
=
di
D
dv
GS
i
D
=
I
SS
2
= (K W)
1
I
SS
2
1
1
(6)
where V
GS
under the drain current I
SS
/2 was evaluated
from (2). By substituting (6) into (4) and solving for W,
the transistor channel width needed to achieve a given
A
V
is
W =
2
21
K
A
V
V
SWING
I
SS
. (7)
From (7), the channel width of NMOS transistors must
be set to a value which is proportional to the bias current,
and increases proportionally to the ratio A
V
/V
SWING
. Of
45 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
1.45
K 0.83 E3 A/(mV
2
)
V
TH
0.35 V
W
min
120 nm
L
min
90 nm
effective L
min
65 nm
C
OX
18 fF/m
2
maximum V
DD
1 V
resistance per unit length r 1.23 kW/m
(unsilicided p+POLY)
capacitance per unit length c 0.07 fF/m
(unsilicided p+POLY)
Table 1.
AlphaPower law coefficients and main process parameters
in a 90nm technology.
course, it is set to the minimum value allowed by the
technology in the cases where (7) is lower than it.
III. Gate Delay Modeling Methodology
The delay in MCML gates can be evaluated by resorting to
the general approach in [2], [3], [37], where the circuit is
first properly linearized around the logic threshold and
eventually simplified by resorting to the halfcircuit con
cept by exploiting the symmetry. The linearized (half) cir
cuit is then approximated to a firstorder circuit with a
pole time constant and a zero time constant
z
(respec
tively equal to the negative of the reciprocal of the pole
and the zero). The propagation delay
PD
of this first
order approximation is equal to, [39],
PD
= 0.69 (
z
) (8)
where parameters and
z
can easily be evaluated by
applying the wellknown opencircuit timeconstant
method, [40], [41].
When linearizing the circuit, NMOS transistors cannot
be rigorously modeled with the wellknown smallsignal
MOS model in Figure 8 due to the strong nonlinearity
involved in logic gates. However, it is well known [2] that
46 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
I
n the limit case of a very shortchannel device with a completely saturated carrier velocity, i.e. with = 1 [38] and K = v
sat
C
OX
, [6],
the DC transfer characteristics of an MCML gate can easily be evaluated by solving the usual set of two equations encountered in the well
known analysis of atraditional sourcecoupled pair, [46],
v
i
= v
GS1
v
GS2
(KVL at input loop)
i
D1
+i
D2
= I
SS
(KCL at the source node)
(A1.1)
By expressing v
GS
as a function of i
D
from (2) and substituting itinto the first equation in (A1.1), the solution of the set of two equations eas
ily gives the expression of the transistor currents as a function of the input voltage v
i
i
D1
(v
i
) =
0 if v
i
<
I
SS
KW
I
SS
2
+K W
v
i
2
if v
i

I
SS
KW
I
SS
if v
i
>
I
SS
KW
(A1.2a)
i
D2
(v
i
) = I
SS
i
D1
(v
i
) (A1.2b)
from which, considering that v
o1
= V
DD
R
D
i
D1
and v
o2
= V
DD
R
D
i
D2
, as well as substituting the voltage gain expression
A
V
= K W R
D
(achieved from (4), with g
m
equal to di
D
/dv
GS
= K W from (2) with = 1) and V
SWING
by solving (1), the differ
ential output voltage is equal to
v
o
(v
i
) =
V
SWING
2
if v
i
<
V
SWING
2A
V
A
V
v
i
if v
i

V
SWING
2A
V
V
SWING
2
if v
i
>
V
SWING
2A
V
(A1.3)
which according to Figure 20 is a piecewise linear curve, as expected due to the linear IV relationship. From this figure, the critical points
that define the noise margin are
(V
IL,max
, V
OH,min
) =
V
SWING
2A
V
,
V
SWING
2
(V
IH,min
, V
OL,max
) =
V
SWING
2A
V
,
V
SWING
2
(A1.4)
Thus the noise margin is equal to
NM = V
OH,min
V
IH,min
=
V
SWING
2
1
1
A
V
. (A1.5)
APPENDIX I
the same topology in Figure 8 can
be used to model NMOS transistors
since they work in the saturation
region most of the time, even
though the linearized parameters
(i.e., the transistor transconduc
tance and capacitances) must be
evaluated in a largesignal condi
tion to account for the wide voltage
variations during a switching tran
sient. The largesignal transconduc
tance G
M
can be evaluated as the
ratio of the drain current variation
i
D
and the gatesource voltage variation v
GS
during
the gate switching (in place of the smallsignal transcon
ductance di
D
/dv
GS
). In a complete switching the transis
tor, the current changes from 0 to I
SS
or vice versa (i.e.,
i
D
= I
SS
), and this change is determined by a gate
source voltage from V
TH
to [V
TH
+( I
SS
/KW )
1/
] by
solving (2) (thus v
GS
= ( I
SS
/KW)
1/
), hence G
M
is
equal to
G
M
=
i
D
v
GS
=
I
SS
( I
SS
/KW)
1
=
g
m
2
1
1
g
m
0.6 +0.4
(9)
where the smallsignal transconductance g
m
around the
logic threshold (6) was substituted. In (9), G
M
is lower
than g
m
by a factor equal to /2
(11/)
, which is very well
approximated by (0.6 +0.4.) with an error smaller than
1% for ranging from 1 to 2. For scaled processes having
the values of closer to unity, the largesignal transcon
ductance is only slightly lower than the smallsignal
value.
In Figure 8, the sourcebulk and drainbulk capaci
tances C
sb
and C
db
can be linearized by multiplying their
zerobias value by a factor which depends on the junction
builtin potential, the grading coefficient and the mini
mum/maximum direct voltage across the junction, [2],
[6]. The gatedrain and the gatesource capacitances C
gd
and C
gs
in the saturation region are approximately linear,
thus no linearization must be performed. It is worth not
ing that all these NMOS parasitic capacitances are pro
portional to the channel width W.
To model the load resistance R
D
, observe that it is
actually implemented by a strip of a highlyresistive layer
(to reduce its area occupation) with length L according to
Figure 9, which also has a distributed parasitic capaci
tance to ground, with an overall value C
R,TOT
. By follow
ing the analysis in Appendix II, this RC strip can be
represented by a lumped RC circuit consisting of the
resistance R
D
with a parallel capacitance C
RD
, as shown
in Figure 9. According to Appendix II, the capacitance
C
RD
is equal to one third of the total parasitic capaci
tance, and is also proportional to 1/ I
SS
C
RD
=
C
R,TOT
3
=
C
R,unit
I
SS
(10a)
C
R,unit
=
V
SWING
6
c
r
(10b)
C
R,unit
being the load parasitic capacitance for a unit
bias current (c and r are the capacitance and resistance
per unit length of the layer implementing the resistance,
which are provided in the technology design kit). To val
idate this approximate firstorder RC circuit, several
physical resistances were simulated by extracting para
sitics from the layout and applying a step current, in
order to evaluate the equivalent time constant
eq
of the
corresponding voltage waveform. In particular, by con
sidering an unsilicided pdoped polysilicon layer with
the resistance r and capacitance c per unit length
reported in Table 1 for the 90nm adopted technology,
results showed that (10) agrees very well with simula
tions, with an error always lower than 4%.
IV. Delay Versus Bias Current
in Nanometer MCML Gates
In this section, the methodology and the circuit models
of transistors and the load resistances discussed in
Section II are applied to an inverter gate (Subsection A)
and to more complex MCML gates (Subsection B).
Compared to [2], [3] a strongly simplified procedure is
adopted to express the powerdelay tradeoff in a very
simple manner.
IVB. MCML Inverter Gate
Let us consider the inverter gate in Figure 2, in which
transistors M1M2 work in the saturation region most of
the time, and their source voltage is the same for both
input logic values (it is fixed by the NMOS transistor in
47 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
G
S
D
C
gd
G
S
C
gs
G
M
v
GS
C
db
D
C
sb
Figure 8. Equivalent linear model of NMOS transistors.
the ON state). Thus, the circuit can be linearized around
the logic threshold v
i
= 0, and the halfcircuit concept
applies due to the symmetry and the differential signal
ing. As shown in Figure 10, where the transistor model in
Figure 8 is substituted, the linearized halfcircuit is a sim
ple commonsource circuit. By applying the timeconstant
method to this circuit (i.e., by evaluating the time con
stants associated with each capacitance when the others
are opencircuited), the time constants and
z
in (8) are
easily found to be
= R
D
(C
db
+C
gd
) +C
RD
+C
L
= R
D
(C
drain
+C
RD
+C
L
) (11a)
z
=
C
gd
G
M
=
C
gd
g
m
(0.6 +0.4) (11b)
48 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
A B ...
Physical Implementation
Load Resistance
...
L
A B
R
D
n
R
1
A B
Circuit Model (RC Ladder)
Decomposition Into
n Sections
R
D
B
A
Simplified RC Circuit Model
A
B
R
D
C
RD
=
C
R,TOT
3
= =
C
2
=
n
C
n
=
C
R,TOT
2n
C
R,TOT
2n
R
2
R
n
C
R,TOT
C
R,TOT
n
C
R,TOT
2n
C
R,TOT
2n
C
1
C
0
Figure 9. Physical implementation of the load resistance: derivation of its lumped circuit model.
where it was observed that all capacitances see the
same resistance R
D
in the evaluation of , and the sum
of C
gd
and C
db
was interpreted as the transistor capaci
tive contribution C
drain
at the drain node. The (nega
tive) zero time constant in (11b) is that of the
wellknown commonsource circuit, and from (8) tends
to increase the delay more significantly in downscaled
technologies.
1
From (8) and (11a)(11b), the delay
PD
is equal to
PD
= 0.69R
D
__
C
drain
+
C
gd
(0.6 +0.4)
A
V
_
+C
RD
+C
L
_
.
(12)
Now, let us consider the explicit dependence of the
delay (12) on the bias current I
SS
, considering that in prac
tical designs R
D
= V
SWING
/2 I
SS
as discussed in Section II
B, and NMOS transistors are sized according to (7). Since
all NMOS capacitances are proportional to W, as pointed
out in Section III, from (7) the transistor capacitance
C
drain
(C
gd
) turns out to be proportional to I
SS
by a con
stant C
drain,N
(C
gd,N
) which represents its value per unit
current (i.e., C
drain
= C
drain,N
. I
SS
and C
gd
= C
gd,N
. I
SS
).
By substituting (10), the MCML inverter delay in (12) is
equal to
PD
= 0.35 V
SWING
_
C
MOSnet,N
+
C
R,unit
I
2
SS
+
C
L
I
SS
_
(13)
where the NMOS network capacitive contributions per
unit current were lumped into a single contribution
C
MOSnet,N
C
MOSnet,N
= C
drain,N
+
C
gd,N
(0.6 +0.4)
A
V
. (14)
49 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
1
This is because the (overlap) gatedrain capacitance scales more slowly than the other parasitic capacitances, since the direct overlap size cannot lin
early scale as reducing the minimum feature size. As another important aspect, the recent adoption of high dielectrics tends to further increase this
capacitance [35].
2
When v
i
is applied to the upper transistors, the capacitances of lower transistors (that have already switched) do not contribute to the overall delay.
T
o model the effect of the distributed resistance and capacitance associated with the load resistance physical layer, we develop an equiv
alent lumped RC circuit which has approximately the same dynamic behavior. To this aim, divide the strip in Figure 9 into a high number
n of small sections, each of which represented by a lumped resistance R
D
/n and a capacitance C
R,TOT
/n (split into two symmetric con
tributions C
R,TOT
/2n, according to Figure 9). Thus the distributed RC strip can be described by the ladder network in Figure 9 with
C
0
= C
n
= C
R,TOT
/2n, C
1
= C
2
= . . . = C
n1
= C
R,TOT
/n, with C
n
being shortcircuited to ground.
The equivalent impedance Z
D
of the RC ladder circuit in Figure 9 can be approximated to a firstorder RC circuit with an equivalent time
constant
eq
, [47], [48],
Z
D
(s) = R
D
1 +b
1
s +b
2
s
2
. . .
1 +a
1
s +a
2
s
2
. . .
R
D
1
1 +s
eq
(A2.1)
which apparently consists of a resistance R
D
with a parallel equivalent capacitance C
R
such
eq
= R
D
C
R
. In (A2.1), the equivalent time
constant
eq
is equal to a
1
b
1
[39], which in turn is easily evaluated through the timeconstant method, [40], [41]. After simple but tedious
calculations, a
1
and b
1
for n we obtain
a
1
=
R
D
C
R,TOT
2
(A2.2)
b
1
= lim
n
R
D
C
R,TOT
n
2
n1
i=1
_
i
i
2
2
_
= lim
n
_
R
D
C
R
n
2
_
n(n 1)
2
1
n
_
(n 1)
3
3
+
(n 1)
2
2
+
(n 1)
6
___
=
R
D
C
R,TOT
6
(A2.3)
therefore the equivalent time constant
eq
is equal to R
D
C
R,TOT
/3 (thereby yielding C
RD
= C
R,TOT
/3).
The equivalent capacitance C
R
can be expressed as an explicit function of the bias current by observing that the resistance R
D
is equal
to r L, r being the resistance per unit length of the considered physical layer and L the strip length. The same observation holds for C
R,TOT
equal to c L, c being the capacitance per unit length of the considered layer. Accordingly, by expressing the strip length L as R
D
/r and
substituting the expression of R
D
= V
SWING
/2I
SS
we get the relationships (10).
APPENDIX II
It is worth noting that relationship (13) analytically
expresses the PowerDelay tradeoff, since the delay is an
explicit function of I
SS
(which defines the static power
consumption P = V
DD
. I
SS
).
IVB. Complex MCML Gates and Input Capacitance
In [2], it was shown that the powerdelay interdependence
(13) actually holds for arbitrary MCML gates, as will be
shown in the following for various MCML gates. First, let
us consider the MCML MUX in Figure 3, whose worstcase
delay
2
PD,MUX
is obtained by applying the switching
input vi to transistors M1M2 and keeping inputs A and B
constant. Without loss of generality, A and B can respec
tively be assumed to be at the low and high level, thus M3
and M6 are in the saturation region, while M4 and M5 are
in cutoff. Observe that the XOR gate has the same delay
as the MUX, since its topology is obtained from the latter
by setting B =
A, hence in the following only the MUX
gate will be considered. By applying the adopted modeling
methodology, the MUX/XOR linearized halfcircuit is
depicted in Figure 11, whose delay (8) is easily found to be
PD,MUX
=0.69
R
D
C
drain,3
+C
drain,5
+
C
gd
(0.6 +0.4)
A
V
+ C
RD
+C
L
+
1
G
M
(C
drain,1
+C
source,3
+C
source,4
)
0.69 R
D
2C
drain
+
C
drain
+2C
source
A
V
(0.6 +0.4)
+ C
RD
+C
L
(15)
where the sum of C
gs
and C
sb
is interpreted as the transistor
capacitive contribution C
source
at the source node. It has been
observed that all transistors have the same C
drain
(C
source
),
and the zero time constant
z
(given by (11b)) is negligible
when compared to the sum of the other capacitances, since
the latter is much greater than in the case of the inverter. By
following the same approach as the inverter, and remembering
that NMOS parasitic capacitances are proportional to I
SS
(i.e.,
C
drain
= C
drain,N
I
SS
, C
source
= C
source,N
I
SS
), the
delay is still given by (13) with an overall NMOS capacitance
per unit current equal to
C
MOSnet,N
= 2C
drain,N
+
C
drain,N
+2C
source,N
A
V
(0.6 +0.4). (16)
In regard to the Dlatch, whose worstcase delay is the
clocktooutput delay occurring when input CLK switch
es, this gate differs from the MUX/XOR gate only for the
sourcecoupled pair M5M6 storing the previous output
value for C LK = 0 due to their positivefeedback connec
tion. Thus, the capacitive contributions of the D latch are
the same as the MUX/XOR gate, except for the additional
capacitance C
input
in (15) seen from the gate of M5 (M6).
As a consequence, the D latch clocktooutput delay is still
given by (13a),with an overall NMOS capacitance equal to
C
MOSnet,N
= 2C
drain,N
+
C
drain,N
+2C
source,N
A
V
(0.6 +0.4) +C
input,N
. (17)
where C
input,N
is obtained from the gatesource capaci
tance expression and (7)
C
input,N
=
C
input
I
SS
=
2
3
W
I
SS
L C
OX
=
2
3
2
21
K
A
V
V
SWING
L C
OX
(18)
Observe that the generalization of (13) to arbitrary
gates is easily justified by considering that in arbitrary
MCML gates the parasitic capacitance C
RD
of the load
resistance is always responsible for the delay term
inversely proportional to I
2
SS
, and the external load capac
itance C
L
determines the term inversely proportional to
I
SS
. Analogously, the NMOS transistor capacitances have
the same dependence on I
SS
and are responsible for the
delay term independent of I
SS
in (13), which is given by
the sum of capacitances at the output node and the other
capacitances multiplied by (0.6 +0.4)/A
V
.
IVC. Simulation Results and Numerical Examples
The delay model was compared to Cadence Spectre simula
tions with I
SS
widely ranging from 1 A to 100 A and load
ing each gate with a number of equal gates (i.e., the fanout
FO) ranging from 0 to 4, using the 90nm CMOS technology
previously described. The delay obtained for FO equal to 0
and 4 is plotted in Figure 12 versus I
SS
in logarithmic scale
50 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
C
drain,N
1.38 E11 F/A
C
input,N
1.67 E11 F/A
C
R,unit
1.88 E20 FA
C
source,N
= C
drain,N
+C
input,N
3.05 E11 F/A
Table 2.
Delay coefficients for MCML gates in a 90nm technology
(with V
SWING
= 700mV, A
V
= 2.2).
(due to the wide considered range of
the bias current), assuming V
SWING
equal to 700 mV and A
V
equal to 2.2.
In the same figure, the predicted
delay (13) with C
L
= C
input
F O
(with C
input
given by (18) is plotted
versus I
SS
, where the numerical
data reported in Table 2 were used.
The error, which is plotted versus
I
SS
in Figure 13, is always within 10%
and is typically in the order of a few
percent, with an average value of
4.7%. It is worth noting that the max
imum error almost doubles (19%)
when the zero effect
C
gd,unit
(0.6 +0.4)/A
V
in (14) is
neglected, there by confirming that it
is an increasingly important contri
bution in nanometer technologies.
In Figure 12, as expected the
delay does not depend on the fan
out for very low values of the bias
current, since the dominant capac
itive contributionis due to the par
asitic capacitance associated with
the load resistance.This confirms
that the widely adopted assump
tion of an ideal load resistor is far
from being realistic, since its para
sitic capacitance in (10) must be
accounted for. Similar curves are
obtained for the other considered
MCML gates which are omitted for
the sake of compactness, and the
obtained numerical value of
C
NMOS,unit
in (13) is reported in
Table 3.
V. PowerDelay TradeOffs and
Design Guidelines
From the general relationship (13),
different powerdelay tradeoffs
and several interesting properties
of MCML gates can be derived, by eventually measuring
the efficiency in the powerdelay tradeoff with the Power
Delay Product PDP (i.e., the product of P = V
DD
I
SS
and
(13)), [6]. In any MCML gate with assigned values of
V
SWING
and A
V
, three different regions can be identified
when varying the power consumption; see Figure 14
which plots the trend of (13) versus I
SS
:
1) LOW POWER REGION: for low values of I
SS
such
that the term C
R,unit
/ I
2
SS
dominates over the other
two in (13), the parasitic capacitance associated
with the load resistance dominates over the others,
thus
PD
is inversely proportional to I
2
SS
. Accord
ingly, PDP is inversely proportional to I
SS
, i.e., it
greatly increases when reducing the power con
sumption. Thus, in lowpower designs, a power sav
ing is achieved at the cost of a much greater speed
penalty. Moreover, the delay (13) does not depend
on the NMOS network, thus it is the same for all
MCML gates, regardless of the implemented logic
function.
51 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
C
gd
G
M
v
i1,2
C
db
+ C
RD
+
V
i1,2
v
o1,2
R
D
C
L
Figure 10. Equivalent circuit of a MCML inverter.
C
gd1
G
M
v
i1
C
db1,2
+ (C
gs,3
+ C
sb,3
) + (C
gs,4
+ C
sb,4
)
+
v
i1
M1
G
M
v
gs3
v
o1
R
D
C
RD
+ C
L
+
v
gs3
(C
db3
+ C
gd,3
) + (C
db,5
+ C
gd,5
)
M3
Figure 11. Equivalent circuit of an MCML MUX gate.
1
10
100
1,000
10,000
1 2 4 5 6 7 8 9 10 20 30 40 50 60 70 80 90 100
I
SS
(A)
P
O
(
p
s
)
FO = 0 (Simulated) FO = 0 (Predicted)
FO = 4 (Simulated) FO = 4 (Predicted)
3
Figure 12. Inverter delay versus I
SS
with a fanout of 0 and 4.
2) POWEREFFICIENT REGION: for moderate values of
I
SS
such that the term C
L
/ I
SS
dominates over the
other two,
PD
is inversely proportional to I
SS
,
hence PDP is roughly constant. A power saving is
achieved at the cost of an equal speed penalty. In
this case, the delay mainly depends on the load.
3) INEFFICIENT DESIGN REGION: for high values of
I
SS
, the MCML delay can no longer be lowered
despite of a power increase, since it asymptotically
tends to a minimum value (achieved from (13) with
I
SS
) set by the NMOS capacitances: the gate
tends to be selfloaded due to the large transistor
size (7) which determines large NMOS capaci
tances. In this case, MCML gates are very inefficient
in terms of the powerdelay trade
off, and the delay is mainly deter
mined by the considered gate
through its NMOS network.
It is worth noting that all MCML
gates with the same V
SWING
and A
V
have the same powerdelay inter
dependence, with the only differ
ence being the value of the
constant term C
MOSnet,N
in (13).
Hence more complex gates have a
greater C
MOSnet,N
and thus a
greater asymptotic minimum delay
PD,min
. Therefore the delay curves
versus I
SS
which analytically
describe the powerdelay tradeoff
of two different MCML gates only
differs for a different up/down shift
by the difference of the two differ
ent minimum delay values, as
graphically reported in Figure 15.
According to the previous con
siderations on the powerdelay
tradeoff, MCML gates will usually
be designed in the powerefficient
region where power and speed per
formance are reasonably balanced,
whereas the lowpower region will
be used only for noncritical paths.
In the following subsections, sim
ple design criteria will be derived
from (13) in the three typical cases
(powerefficient, highspeed and
lowpower design), and design con
siderations on the power supply
voltage will be made.
VA. PowerEfficient Design
To achieve an optimum power
delay balance, it is necessary to
minimize the powerdelay product
PDP = V
DD
I
SS
PD
(with
PD
given by (13), which is obtained by
setting its derivative to zero and
solving for I
SS
. The obtained bias
current which minimizes PDP is
52 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
12
10
8
6
4
2
0
2
4
6
8
1 2 3 9 10 20 30 40 50 60 70 80 90 100
I
SS
(A)
E
r
r
o
r
(
%
)
FO = 0 FO = 4
4 5 6 7 8
Figure 13. Error of the delay model versus I
SS
for an inverter gate with a fanout
of 0 and 4.
Low
Power
Power
Efficient
I
SS
1
I
SS
1
2
Inefficient
Design
Delay
I
SS
Bias Current
(Power)
Gate
Independent
Delay
LoadDependent
Delay
GateDependent
Delay
Constant
2
PD,min
PD,min
PD
Figure 14. General delay dependence on the bias current (or equivalently the power
consumption) in MCML gates.
I
SS,opt PDP
=
C
R,unit
C
MOSnet,N
. (19)
First, (19) yields C
MOSnet,N
I
SS,opt PDP
= C
R,unit
/
I
SS,opt PDP
, which means that a powerefficient design
leads to equal capacitive contributions of the NMOS net
work and the load resistance, as reported in Figure 16.
Moreover, the optimum bias current (19) is independent
of the load, and the minimum powerdelay product
(obtained by substituting (19) into PDP) turns out to be
PDP
opt,MC ML
0.35 V
DD
V
SWING
C
MOSnet,N
C
R,unit
+C
L
(20)
from which a PDP increase (i.e., a worse power effi
ciency) is observed when increasing the load capaci
tance C
L
, as well as C
R,unit
and C
MOSnet,N
. Observe that
C
R,unit
is proportional to V
SWING
(according to Appen
dix II) and does not depend on A
V
, whereas the NMOS
contribution C
MOSnet,N
is proportional to W, which in
turn is proportional to ( A
V
/V
SWING
)
according to (7),
hence PDP in (20) is proportional to V
SWING
(3)/2
and
A
V
/2
. As a general result, in MCML gates designed for
powerefficiency the logic swing and the voltage gain
should be kept as low as possible within the range
allowed by the noise margin requirement. These con
siderations are summarized in Figure 16, where it is
considered that for I
SS
= I
SS,opt PDP
the terms propor
tional to 1/ I
SS
2
and the constant one are equal, thus it
lies at the boundary of the lowpower and the power
efficient region.
VB. HighSpeed Design
When a high speed performance is the principal goal, two
situations may occur. In the first one, a delay constraint
PD
derived from considerations at the gate level has to
be met by properly setting I
SS
to
I
SS
= 0.17 V
SWING
C
L
PD
PD,min
1 +
1 +11.4
C
R,N
C
2
L
PD
PD,min
V
SWING
(21)
that was obtained by solving (13) for I
SS
and substituting
its asymptotic minimum expression
PD,min
= lim
I
SS
PD
= 0.35 V
SWING
C
MOSnet,N
(22)
In the second case the speed potential must be
exploited as much as possible, thus
PD
has to be close
to (22) while keeping I
SS
within reasonable values, i.e.,
I
SS
should only be increased as long as a significant
speed improvement is achieved. To this aim, observe
that for sufficiently high values of I
SS
such that
C
MOSnet,N
> (C
R,unit
/ I
2
SS
+C
L
/ I
SS
), the constant term in
(13) dominates over the other two (i.e. the gate is self
loaded), thus a high speed is achieved, but a further
increase in the bias current does not lead to a significant
speed advantage. In contrast, for lower values of I
SS
such that C
MOSnet,N
< (C
R,unit
/ I
2
SS
+C
L
/ I
SS
), the terms
depending on I
SS
dominate over the constant one, thus
a worse speed performance is achieved, but
PD
is highly
sensitive to a bias current increase. As a compromise, a
reasonable choice of I
SS
is achieved in the intermediate
case C
MOSnet,N
= (C
R,unit
/ I
2
SS
+C
L
/ I
SS
), which appar
ently makes
PD
only twice the minimum achievable
(i.e.
PD
= 2
PD,min
), as reported in Figure 16. Thus, the
I
SS,opt delay
needed for such highspeed criterion is at
the boundary of the powerefficient and the inefficient
region in Figure 16. Moreover, under this current
C
MOSnet,N
I
SS,opt delay
is equal to (C
R,unit
/ I
SS,opt delay
+
C
L
), thus under this design criterion the NMOS capaci
tance contribution equals the sum of C
L
and that of the
load resistance.
The bias current I
SS,opt delay
is easily found from (21)
by substituting
PD
=2
PD,min
I
SS,opt delay
= 0.17 V
SWING
C
L
PD,min
1 +
1 +11.4
C
R,unit
C
2
L
PD,min
V
SWING
(23)
which is easily found to be always greater than
I
SS,opt PDP
in (19) (or equal to, in the limit case C
L
= 0).
This means that a high speed is achieved at the cost of a
worse power efficiency, when compared to the case dis
cussed in the previous subsection.
By reiterating the reasoning in Subsection A, C
MOSnet,N
is proportional to ( A
V
/V
SWING
)
PD
= 2
PD,min
is proportional to A
V
/V
SWING
1
from
(22), therefore in highspeed designs the voltage gain
should be kept low, whereas the logic swing should be set
as high as possible, cf. Figure 16. Surprisingly, this is in
contrast with the usual belief that the highspeed feature
of MCML gates is due to the small logic swing, [16], that
probably is due to a superficial extension of wellknown
properties of CML bipolar gates [2]. This consideration
can be intuitively justified by observing that an increase
in the logic swing reduces the transistor size (5) needed
to achieve a given A
V
, thereby reducing the NMOS capac
itances which are the dominant contribution in the high
speed region.
53 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
VC. LowPower Design
In lowpower design, e.g., the design of noncritical paths,
the power consumption per gate allowed is usually an
assigned parameter that is derived from the requirements
at the system level. Therefore, the only design parameter
is the logic swing, whereas I
SS
is set to a very low value
chosen from the system considerations, thus the gate
works in the lowpower region where the dominant term
is C
R,unit
/ I
2
SS
, and the delay is approxi
mately
PD
= 0.35 V
SWING
C
R,unit
I
2
SS
(24)
which shows that in lowpower design the
logic swing has to be set as low as possi
ble, as in the case of powerefficient
design, whereas the voltage gain does not
affect the speed performance.
VD. Remarks on the Power Supply
Voltage Sizing
Since the NMOS network in MCML gates
consists of stacked sourcecoupled
pairs associated with different levels,
according to Figure 17, only the transis
tors at the first (upper) level can be
directly driven by the output of an
MCML gate, whereas the input voltages
of transistor pairs at lower levels are
progressively reduced through level
shifter stages to ensure operation in
the saturation region (for the reader
interested in the design of level shifter
stages, the subject is thoroughly
addressed in [2]). Each level
shifter stage is implemented
with a commondrain stage as in
Figure 17, [2], [18]. The mini
mum V
DD
is found by consider
ing the input v
i,n
at the nth
lowest level in Figure 17, which
is set by the output voltage of
the preceding gate and the gate
source voltage drop
(n 1)V
GS,shift
of (n 1) level
shifters, is equal to
V
DD
(n 1)V
GS,shift
in the case
of a high input. According to Fig
ure 17, this voltage must accom
modate the gatesource voltage
drop of the lowest transistor
driven by v
in
and the minimum
voltage drop across the bias
current source V
I SS,min
(equal to
a small V
DS,sat
100 mV in the
case of a simple current mirror
implementation), thus
54 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
I
SS
Low
Power
Power
Efficient
2
1
I
SS
1
Inefficient
Design
Delay
PD,min
I
SS
I
SS,opt_PDP
I
SS,opt_delay
PowerEfficient
Design
(Low V
SWING
, Low A
V
)
C
MOSnet
= C
RD
HighSpeed
Design
(HighV
SWING
, Low A
V
)
C
MOSnet
= C
RD
+ C
L
PD
Bias Current
(Power)
Constant
2
PD,min
LowPower
Design
(Low V
SWING
, any A
V
)
C
RD
>>C
MOSnet
+ C
L
Figure 16. Summary of design criteria of MCML gates.
Gate 1 (Simpler)
Delay
PD,min2
Bias Current
(Power)
Gate 2 (More Complex)
PD,min1
PD,min2
PD,min1
Figure 15. Delay curves versus I
SS
for two different MCML gates with the
same logic swing and load.
V
DD,min
= V
GS
+(n 1) V
GS,shift
+V
I SS,min
(25)
Equivalently eq. (26) sets the maximum number of
levels n for a given V
DD
(typically 23 [2], [18]). The
level shifter voltage drop V
GS,shift
is usually kept very
close to the transistor threshold voltage V
TH
by set
ting the bias current to a rather low value. In regard to
V
GS
, it is the gatesource voltage of a transistor in the
ON state, i.e., with a current I
SS
, thus it is obtained by
solving (2)
V
GS
= V
TH
+
I
SS
K W
1
= V
TH
+
1
2
2
1
V
SWING
A
V
(26)
where (7) was substituted. From (26), in order to reduce
the supply voltage, the voltage swing should be kept as low
as possible, and the voltage gain should not be too low.
VI. A Design Example
Let us apply the concepts presented until now to the
carry logic of a Full Adder, which evaluates the carry out
put C
out
= A B +C
in
A B
CARRY
is given by (13) with C
MOSnet,N
given by the con
tribution at the output node 3C
drain
, because the other
capacitances have already switched during the carry
input transition. In a highspeed design under the above
conditions, the obtained optimum current and delay are
respectively 18 A and 28 ps.
VIII. CONCLUSIONS
In this paper, an overview of techniques to manage the
powerdelay tradeoff in nanometer MCML circuits has
been presented. Compared to pre
vious works, a strongly simplified
and comprehensive approach was
adopted which also account for
DeepSubmicron effects. As oppo
site to the previous works of the
same authors, a physical resistance
load was assumed, whose distrib
uted parasitic capacitance was sim
ply modeled as a lumped circuits. It
was also shown that the usual
assumption made in the previous
papers of an ideal resistor (i.e.,
without parasitic capacitance) is
strongly unrealistic, especially in
lowpower designs.
To understand better the
design tradeoffs, simple models of
the noise margin and the delay
have been discussed. Further
more, a simple approach to write
the delay by inspection of the gate
56 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
Y
V
DD
R
D
R
D
c
out,1
I
SS
C
out,2
C
L
C
L
M1 M2
A
2
A
1
A
1
A
2 M5 M6
M3
M3
M4
M4
M3 M4
B
1
B
2
C
in1
C
in2
C
in1
C
in2
X
Figure 18. Topology of the carry logic in an MCML Full Adder, with worstcase current
path in dashed line.
logic gate C
MOSnet,N
inverter 1.48 E11
MUX/XOR 6.76 E11
D latch 8.44 E11
Full Adder (carry logic) 1.21 E10
Table 3.
Overall NMOS capacitance per unit current of different MCML
logic gates(with V
SWING
= 700mV, A
V
= 2.2).
topology was extrapolated by generalizing the results of
a few gates. Interesting properties on tradeoffs and
effect of scaling have been derived from these analytical
models: for example, it is shown that the DSM effects are
beneficial in terms of the noise margin. In particular,
three design targets have been discussed (i.e., low
power, powerefficient and highspeed), and simple
design criteria to size the bias current, the logic swing
and the voltage gain have been found. These results,
which are summarized in Figures 14 and 16, provide pow
erful information for decision taking in the design
process. Interestingly, it was shown that a high speed is
achieved by increasing the logic swing, as opposite to
the incorrect traditional belief that low logic swings
make MCML circuits faster. The
practical design of the carry
logic of a Full Adder has been
discussed presenting numerical
examples by considering a 90
nm CMOS process.
Several challenges must still
be faced in the understanding of
MCML circuits, which are an
approach that is less mature than
the traditional CMOS logic. First,
the understanding of the interde
pendence of design parameters
and the design criteria here
derived should be exploited to
implement automated design
flows to optimize effectively com
plex MCML circuits with a rea
sonable computational effort.
Secondly, although MCML cir
cuits were shown to be less sensi
tive to the problems related to the
technology downscaling than tra
ditional CMOS logic, further prob
lems will arise due to the
continuous reduction of the supply
voltage. Indeed, the latter will
increasingly limit the number of
logic levels within a gate (accord
ing to (25)(26)), and thus the com
plexity that can be implemented
into a single gate. This will trans
late into a greater number of bias
current sources (and thus a
greater overall power consump
tion) and interconnects (which
degrade the speed performance).
To overcome this limit, novel cir
cuit approaches will be needed,
such as the lowvoltage tripletail cell approach that was
previously adopted in bipolar integrated circuits, [42][44].
Moreover, the logic swing reduction that will be forced by
the supply voltage scaling will determine a decrease in the
available noise margin, which will have to be recovered by
increasing the voltage gain (according to (5)) by means of
novel circuit techniques such as the introduction of posi
tive feedback [45].
Third, efficient powerdown techniques will be needed
to reduce eventually the static power consumption in
MCML blocks that do not perform useful computations,
while still keeping supply current variations within
reasonable bounds, in order to maintain the advantages
due to the almost constant supply current of MCML gates.
57 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE
v
o
(v
i
)
v
i
0
0
V
SWING
2
V
SWING
2
V
SWING
2A
V
V
SWING
2A
V
Figure 20. DC transfer characteristics of a MCML gate with a completely saturated car
rier velocity.
1
10
100
1,000
10,000
4 5 6 7 8 9 10 20 1 2 3 30 40 50 60 70 80 90 100
I
SS
(A)
(
p
s
)
FO = 1 (Simulated) FO = 1 (Predicted)
Figure 19. Carry logic delay versus I
SS
with a unity fanout.
References
[1] M. Mizuno et al., A GHz MOS adaptive pipeline techniques using
MOS currentmode logic, IEEE Journal of SolidState Circuits, vol. 31,
no. 6, pp. 784791, June 1996.
[2] M. Alioto and G. Palumbo, Model and Design of Bipolar and MOS Cur
rentMode Logic (CML, ECL and SCL Digital Circuits), Springer, 2005.
[3] M. Alioto and G. Palumbo, Design strategies for source coupled logic
gates, IEEE Trans. on CAS Part I, vol. 50, no. 5, pp. 640654, May 2003.
[4] M. Alioto and G. Palumbo, Powerdelay optimization of D
Latch/MUX source coupled logic gates, International Journal of Circuit
Theory and Applications, vol. 33, no. 1, pp. 6586, Jan./Feb. 2005.
[5] M. Alioto and G. Palumbo, Oscillation frequency in CML and ESCL
ring oscillators, IEEE Trans. on CAS Part I, vol. 48, no. 2, pp. 210214,
Feb. 2001.
[6] J. Rabaey, A. Chandrakasan, and B. Nikolic, Digital Integrated Circuits
(A Design Perspective), Prentice Hall, 2003.
[7] B. Razavi, Prospect of CMOS technology for highspeed optical
communication circuits, IEEE Jour. of SolidState Circ., vol. 37, no. 9,
pp. 11351145, Sep. 2002.
[8] B. Razavi (Ed.), Monolithic PhaseLocked Loops and Clock Recovery
Circuits (Theory and Design), IEEE Press, 1996.
[9] C. Hung, B. Floyd, B. Park, and K. O, Fully integrated 5.35GHz CMOS
VCOs and prescalers, IEEE Trans. on Microwave Theory and Techniques,
vol. 49, no. 1, Jan. 2001.
[10] C. Lam and B. Razavi, A 2.6GHz/5.2GHz Frequency Synthesizer in
0.4m CMOS Technology, IEEE Jour. of SolidState Circ., vol. 35, no. 5,
pp. 788794, May 2000.
[11] H. Nosaka, K. Isshii, T. Enoki, and T. Shibata, A 10Gb/s datapattern
independent clock and data recovery with a twomode phase compara
tor, IEEE Jour. of SolidState Circuits, vol. 38, no. 2, pp. 192197, Feb. 2003.
[12] S.T. Yan and H. Luong, A 3V 1.3to1.8GHz CMOS voltagecon
trolled oscillator with 0.3ps Jitter, IEEE Trans. on Circuits and Systems
Part II, vol. 45, no. 7, pp. 876880, July 1998.
[13] B. Razavi, Design of Integrated Circuits for Optical Communications,
McGrawHill, 2003.
[14] T.H. Lee, The Design of CMOS Radio Frequency Integrated Circuits,
Cambridge University Press, 2nd edition, 2003.
[15] J. Musicer and J. Rabaey, MOS current mode logic for low power,
low noise CORDIC computation in mixedsignal environments, Proc. of
ISLPED 2000, pp. 102107, 2000.
[16] A. Tanabe, M. Umetani, I. Fujiwara, T. Ogura, K. Kataoka, M. Okiara,
H. Sakuraba, T. Endoh, and F. Masuoka, 0.18m CMOS 10Gb/s
Multiplexer/Demultiplexer ICs Using Current Mode Logic with Tolerance
to Threshold Voltage Fluctuation, IEEE J. of SolidState Circuits, vol. 36,
no. 6, June 2001.
[17] R. Senthinatan and J. Prince, Application specific CMOS output
driver circuit design techniques to reduce simultaneous switching
noise, IEEE Jour. of SolidState Circuits, vol. 28, no. 12, pp. 13831388,
Dec. 1993.
[18] S. Maskai, S. Kiaei, and D. Allstot, Synthesis techniques for CMOS
folded sourcecoupled logic circuits, IEEE J. Of Solid State Circuits, vol.
27, no. 8, pp. 11571167, Aug. 1992.
[19] D. Allstot, S. Chee, S. Kiaei, and M. Shristawa, Folded source
coupled logic vs. CMOS static logic for lownoise mixedsignal ICs, IEEE
Trans. on CASPart I, vol. 40, no. 9, pp. 553563, Sep. 1993.
[20] S. Kiaei, S. Chee, and D. Allstot, CMOS sourcecoupled logic for
mixedmode VLSI, Proc. Int. Symp. Circuits Systems, pp. 16081611, 1990.
[21] H. Ng and D. Allstot, CMOS current steering logic for lowvoltage
mixedsignal integrated circuits, IEEE Trans. on VLSI Systems, vol. 5,
no. 3, pp. 301308, Sep. 1997.
[22] B. Stanistic, N. Verghese, R. Rutenbar, L. Carley, and D. Allstot,
Addressing substrate coupling in mixedmode ICs: simulation and
power distribution synthesis, IEEE Jour. of SolidState Circuits, vol. 29,
pp. 226238, Mar. 1994.
[23] International Technology Roadmap for Semiconductors, Available:
http://public.itrs.net.
[24] R. Singh (Ed.), Signal Integrity Effects in Custom IC and ASIC Design,
IEEE Press, 2002.
[25] B. Del Signore, D. Kerth, N. Sooch, and E. Swanson, A monolithic 20b
deltasigma A/D converter, IEEE J. SolidState Circuits, vol. 25, pp. 13111317,
Dec. 1990.
[26] H. Leopold, G. Winkler, P. OLeary, K. Ilzer, and J. Jernej, A mono
lithic CMOS 20b analogtodigital converter, IEEE J. SolidState Circuits,
vol. 26, pp. 910916, July 1991.
[27] I. Fujimori et al., A 5V single chip deltasigma audio A/D converter
with 111 dB dynamic range, IEEE J. Of Solid State Circuits, vol. 32, pp.
329336, Mar. 1997.
[28] S. Jantzi and K. Martin, A. Sedra, Quadrature bandpass modu
lator for digital radio, IEEE J. Solid State Circuits, vol. 32, pp. 19351949,
1997.
[29] B. Kup, E. Dijkmans, P. Naus, and J. Sneep, A bitstream digitalto
analog converter with 18b resolution, IEEE J. SolidState Circuits, vol. 26,
pp. 17571763, Dec. 1991.
[30] J. Kundan and S. Hasan, Enhanced folded sourcecoupled logic
technique for lowvoltage mixedsignal integrated circuits, IEEE Trans.
on CASPart II, vol. 47, no. 8, pp. 810817, Aug. 2000.
[31] H. Lee, D. Hodges, and P. Gray, A selfcalibrating 15bit CMOS A/D
converter, IEEE Jour. of SolidState Circuits, vol. 19, pp. 813819, Dec.
1984.
[32] D. Su, M. Loinaz, S. Masui, and B. Wooley, Experimental results and
modeling techniques for substrate noise in mixedsignal integrated cir
cuits, IEEE Jour. of SolidState Circuits, vol. 28, pp. 420430, Apr. 1993.
[33] K. Bernstein et al., High Speed CMOS Design Styles, Kluwer Academic
Publishers, 1999.
[34] S. Bruma, Impact of onchip process variations on MCML perform
ance, Proc. IEEE International SystemsonChip Conference (SOCC03), pp.
135140, 2003.
[35] B.P. Wong, A. Mittal, U. Cao, and G. Starr, NanoCMOS Circuit and
Physical Design, John Wiley & Sons, 2005.
[36] D. Chinnery and K. Keutzer, Closing the Gap between ASIC & Custom,
Kluwer Academic Publishers, 2002.
[37] T. Sakurai and A.R. Newton, AlphaPower law MOSFET model and
its applications to CMOS inverter delay and other formulas, IEEE Jour.
on SolidState Circuits, vol. 25, no. 2, pp. 584594, Apr. 1990.
[38] M. Alioto, G. Palumbo, and S. Pennisi, Modeling of Source Coupled
Logic Gates, International Journal of Circuit Theory and Applications, vol.
30, no. 4, pp. 459477, 2002.
58 IEEE CIRCUITS AND SYSTEMS MAGAZINE FOURTH QUARTER 2006
[39] W. Elmore, The transient response of damped linear networks, J.
Appl. Phys., vol. 19, pp. 5563, Jan. 1948.
[40] B. Cochrun and A. Grabel, A Method for the Determination of the
Transfer Function of Electronic Circuits, IEEE Trans. on Circuit Theory,
vol. CT20, no. 1, pp. 1620, Jan. 1973.
[41] G. Palumbo and S. Pennisi, Feedback Amplifiers Theory and Design,
Kluwer Academic Publishers, 2002.
[42] B. Razavi, Y. Ota, and R. Swartz, Design techniques for lowvoltage
high speed digital bipolar circuits, IEEE Jour. of SolidState Circ., vol. 29,
no. 2, pp. 332339, Mar. 1994.
[43] G. Schuppener, C. Pala, and M. Mokhtari, Investigation on lowvolt
age lowpower silicon bipolar design topology for highspeed digital cir
cuits, IEEE Jour. Of SolidState Circ., vol. 35, no. 7, pp. 10511054, July
2000.
[44] M. Alioto, R. Mita, and G. Palumbo, Performance evaluation of the
lowvoltage CML DLatch topology, IntegrationThe VLSI Journal, Spe
cial Issue in Analog and MixedSignal IC Design and Design Methodolo
gies (edited by Francisco V. Fernandez), vol. 36, no. 4, pp. 191209, Nov.
2003.
[45] M. Alioto, L. Pancioni, S. Rocchi, and V. Vignoli, Modeling and Eval
uation of PositiveFeedback SourceCoupled Logic, IEEE Trans. on
CASPart I, vol. 51, no. 12, pp. 23452355, Dec. 2004.
[46] P. R. Gray and R. G. Meyer, Analysis and Design of Analog Integrated
Circuits, John Wiley & Sons, 1977.
[47] J. L. Wyatt, Jr., Signal propagation delay in RC models for intercon
nect, Circuit Analysis, Simulation and Design, Part II: VLSI Circuit Analysis
and Simulation, A. Ruehli (Ed.), vol. 3 in the series Advances in CAD for
VLSI, NorthHolland, 1987.
[48] M. Alioto G. Palumbo, and M. Poli, Evaluation of energy consump
tion in RC ladder circuits driven by a ramp input, IEEE Trans. on VLSI
Systems, vol. 12, no. 10, pp. 10941107, Oct. 2004.
Massimo Alioto (M01) was born in
Brescia, Italy, in 1972. He received the lau
rea degree in Electronics Engineering and
the Ph.D. degree in Electrical Engineering
from the University of Catania (Italy) in
1997 and 2001, respectively. In 2002, he
joined the Engineering faculty of the Uni
versity of Siena as a Research Associate and in the same
year as an Assistant Professor. In 2006, he became Associ
ate Professor in the same faculty.
Since 2001 he has been teaching undergraduate and
graduate courses on basic electronics, microelectronics
and advanced VLSI digital design. He has authored or co
authored over 80 journals and conference papers. He is
coauthor of the book Model and Design of Bipolar and
MOS CurrentMode Logic: CML, ECL and SCL Digital Circuits
(Springer, 2005). His primary research interests include:
modeling and optimized design of CMOS highperform
ance digital circuits in terms of highspeed or lowpower
dissipation, transistor and gatelevel design of arithmetic
circuits, design of circuits for cryptographic applications
(e.g., random number generators, circuits resistant to Dif
ferential Power Analysis), and design for variability. His
research was previously focused also on the modeling
and the design of bipolar CML/ECL circuits, as well as adi
abatic logic.
Gaetano Palumbo was born in Catania,
Italy, in 1964. He received the laurea
degree in Electrical Engineering in 1988
and the Ph.D. degree from the University
of Catania in 1993. Since 1993 he conducts
courses on Electronic Devices, Electron
ics for Digital Systems and basic Elec
tronics. In 1994 he joined the DEES (Dipartimento
Elettrico Elettronico e Sistemistico), now DIEES (Diparti
mento di Ingegneria Elettrica Elettronica e dei Sistemi), at
the University of Catania as a researcher, subsequently
becoming associate professor in 1998. Since 2000 he is a
full professor in the same department.
His primary research interest has been analog circuits
with particular emphasis on feedback circuits, compensa
tion techniques, currentmode approach, lowvoltage cir
cuits. Then, his research has also embraced digital circuits
with emphasis on bipolar and MOS currentmode digital
circuits, adiabatic circuits, and highperformance building
blocks focused on achieving optimum speed within the
constraint of low power operation. In all these fields he is
developing some the research activities in collaboration
with STMicroelectronics of Catania.
He was the coauthor of three books CMOS Current
Amplifiers, Feedback Amplifiers: theory and design and
Model and Design of Bipolar and MOS CurrentMode Logic
(CML, ECL and SCL Digital Circuits) all by Kluwer Academ
ic Publishers, in 1999, 2001 and 2005, respectively, and a
textbook on electronic devices in 2005. He is a contributor
to the Wiley Encyclopedia of Electrical and Electronics Engi
neering. He is the author of more than almost 300 scientific
papers on referred international journals (over 110) and in
conferences. Moreover he is coauthor of several patents.
Since June 1999 to the end of 2001 and since 2004 to
2005 he served as an Associated Editor of the IEEE Trans
actions on Circuits and Systems part I for the topic Ana
log Circuits and Filters and Digital Circuits and
Systems, respectively. Since 2006 he is serving as an
Associated Editor of the IEEE Transactions on Circuits and
Systems part II.
In 2005 he was a panelist in the scientificdisciplinaire
area 09industrial and information engineering of the
CIVR (Committee for Evaluation of Italian Research), which
has the aim to evaluate the Italian research in the above
area for the period 20012003.
In 2003 he received the Darlington Award. Prof. Palum
bo is an IEEE Senior Member.
59 FOURTH QUARTER 2006 IEEE CIRCUITS AND SYSTEMS MAGAZINE