IEEE TRANSACTIONS ON SOLID STATE CIRCUITS, VOL. 35, NO.

1, JANUARY 2000 15

Noise in RF-CMOS Mixers: A Simple

Physical Model
Hooman Darabi and Asad A. Abidi, Fellow, IEEE

Abstract—Flicker noise in the mixer of a zero- or low-interme-

diate frequency (IF) wireless receiver can compromise overall re-
ceiver sensitivity. A qualitative physical model has been developed
to explain the mechanisms responsible for flicker noise in mixers.
The model simply explains how frequency translations take place
within a mixer. Although developed to explain flicker noise, the
model predicts white noise as well. Simple equations are derived
to estimate the flicker and white noise at the output of a switching
active mixer. Measurements and simulations validate the accuracy
of the predictions, and the dependence of mixer noise on local os-
cillator (LO) amplitude and other circuit parameters.
Index Terms—Active mixers, CMOS integrated circuits, com-
munication systems, integrated circuits, mixers, mixer noise, noise, Fig. 1. A typical switching active mixer with noise source shown at the
nonlinear circuits, receivers. transconductor input.

I. INTRODUCTION believe it should be as simple to estimate and optimize noise

figure in mixers as it is in amplifiers.

I N THE course of design of a CMOS mixer, it was found

that flicker ( ) noise appears at its output even though
the resistor loads are free of flicker noise. This is unexpected, II. LOW-FREQUENCY NOISE IN A MIXER
because at first glance one expects flicker noise in the input An active mixer comprises an input transconductance,
stage transistors and mixer switches to translate in frequency. switches, and an output load. Noise is present in all the transis-
There are important repercussions in a direct-conversion tors making up these functions. First the noise contribution of
receiver [1], where the signal downconverts to baseband after the loads and transconductance FET’s is described, followed by
only minimal amplification at radio-frequency (RF). Mixer the noise contribution of the switches. Frequency translation of
noise degrades signal-to-noise ratio (SNR), and as a conse- noise is dealt with approximately, using a simple analysis and
quence the overall noise figure of the receiver suffers. By using qualitative reasoning. More accurate analytic [3] and numerical
a modulation such as wideband FSK, it is sometimes possible [4] methods have been presented elsewhere.
to position the downconverted signal beyond the flicker noise
corner [2]. However, there is no room to do so in a narrowband A. Load Noise
communication system such as FLEX used for wireless paging,
where adjacent channels are only 25 kHz apart. In a zero-IF receiver, flicker noise in the loads of the down-
This paper describes the processes whereby flicker noise ap- conversion mixer competes with the signal [5]. This noise may
pears at the output of a CMOS downconversion mixer. Some- be lowered in one of many ways. PMOSFET’s show lower
times the noise is so large that it disqualifies the mixer for use flicker noise [6], [7], ascribed to buried channel behavior,
in a direct-conversion receiver. In investigating these processes, compared to NMOSFET’s of the same dimensions. Therefore,
a simple, qualitative model of noise in mixers has been devel- PMOS loads may be used. Otherwise at the expense of some
oped whose predictions agree very well with sophisticated sim- voltage headroom the mixer may be loaded with polysilicon
ulations of mixer noise, and with measurements. The model is resistors, which are free of flicker noise.
extended to white noise. Straightforward equations capture the
noise (both flicker and white) originating in different parts of the B. Transconductance Noise
mixer. The main contribution of this work is to demystify mixer A prototype CMOS transconductance mixer is shown in
noise, which is so far dealt with in the literature using either ex- Fig. 1. Noise in the lower transconductance FET’s accompanies
haustive analyzes or specially developed simulation tools. We the RF input signal, and is translated in frequency just like the
signal is. Therefore, flicker noise in these FET’s is upconverted
to and to its odd harmonics, while white noise at (and
Manuscript received January 27, 1999; revised September 8, 1999. its odd harmonics) is translated to DC. If the output of interest
The authors are with the Department of Electrical Engineering, University of
California, Los Angeles, CA 90095-1594 USA. lies at zero IF, then the transconductance FET’s only contribute
Publisher Item Identifier S 0018-9200(00)00122-0. white noise after frequency translation, since the flicker corner
0018–9200/00$10.00 © 2000 IEEE
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
16 IEEE TRANSACTIONS ON SOLID STATE CIRCUITS, VOL. 35, NO. 1, JANUARY 2000

(a)
Fig. 2. Single-balanced mixer with switch noise modeled at gate.

Fig. 3. Assumed switch I –V characteristic.

frequency of these devices is usually much lower than the LO

frequency. (b)
Due to mismatches in the switching transistor quad, a small
Fig. 4. (a) Switch input voltage and (b) mixer output current decomposed into
amount of the flicker noise in the transconductance FET’s can noiseless response and pulse of noise.
appear at the loads. This is dealt with at the end Section II-C.
half period, the current switches to the right branch. Thus, the
C. Direct Switch Noise output is a square-wave at frequency of with zero DC value.
Without loss of generality, consider the single-balanced mixer LO feedthrough is a characteristic of single-balanced mixers.
in Fig. 2. The bias current in the switch FET’s M1 and M2 is Next, include noise. The slowly varying modulates the
periodic at a frequency . Flicker noise arises from traps with time at which the pair M1, M2 switches [Fig. 4(a)]. At every
much longer time constants than the typical period of oscillation switching event the skew in switching instant modulates the dif-
at RF, and it may be assumed that the time-average inversion ferential current waveform at the mixer output. Although the
layer charge in the channel determines the root mean square height of the square-wave signal at the output remains con-
(rms) flicker fluctuations. These charge fluctuations are referred stant, noise advances or retards the time of zero-crossing by
as a voltage to the gate of one FET in the differential pair with , where is the slope of the LO voltage at
a constant rms value and a spectral density proportional to the switching time [Fig. 4(a)]. Now the waveform at the mixer
( in Fig. 2). This equivalent voltage is like a slowly varying output consists of a square-wave of frequency and am-
offset voltage associated with differential pair. It should be noted plitude , representing the LO feedthrough, superposed with a
that based on the carrier-density fluctuation model, the input pulse train of random widths and amplitude of at a fre-
referred flicker noise of MOSFET’s is independent of . This quency of , representing noise [Fig. 4(b)]. Over one period
bias independence is experimentally verified in NMOSFET’s the average value of the output current is
[6], [7].
To further simplify analysis, it is assumed that the circuit (1)
switches sharply, that is, a small differential voltage excursion
( ) causes the current ( ) to completely switch from one side where is the period of LO, equal to . This means
of the differential pair to the other side (Fig. 3). that low-frequency noise at the gate of switch, , appears at
Now consider the direct effect of the switch noise at the mixer the output without frequency translation, and corrupts a signal
output. The transconductance RF input stage is replaced by a downconverted to zero IF. The zero-crossing modulation, ,
current source, , at the tail. In the absence of noise, for positive depends on the low-frequency noise, , and the LO-voltage
values of LO voltage M1 switches ON and M2 switches OFF, slope ( ) at zero-crossing normalized to LO frequency, .
and a current equal to appears at the left branch. In the next For a sine-wave LO, 4 π , where is the amplitude
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
DARABI AND ABIDI: NOISE IN RF-CMOS MIXERS: A SIMPLE PHYSICAL MODEL 17

fully commutating mixer from RF input to the differential cur-

rent output is , the SNR is

(5)

where the last term applies to a sinewave LO of amplitude

. This relationship shows that SNR improves by raising the
product of the slope of the LO waveform at zero-crossing and
its period (in the case of a sinewave LO, this amounts to raising
); by increasing the gate area of the switch FET’s to lower
flicker noise ; and by lowering the transconductance FET
over-drive. However, increasing switch gate area or lowering
Fig. 5. Mixer output spectrum in presence of direct noise. usually degrade mixer bandwidth.
The double-balanced mixer (Fig. 1) works in much the same
and a factor of two accounts for the fact that is compared to way. The main difference is that there is no LO feedthrough,
a differential LO signal with an amplitude of . and represents the equivalent noise of four switches in the
It is instructive to examine the complete spectrum of the mixer. induces a pulse train with the width of (as defined
mixer output noise. As the pulses are approximated earlier) and frequency of at the mixer output [Fig. 4(b)].
with ideal delta-function impulses of amplitude at The low-frequency component of this waveform’s spectrum is
twice the LO frequency [Fig. 4(b)]. With this approximation, exactly equal to what was calculated in (1). Therefore, the noise
the spectrum is found by invoking well-known sampling theory: at the output is the same as in (2).
the noise waveform is scaled by , then impulse-sampled This analysis also answers the earlier question of how flicker
at a rate of rad/s. The frequency spectrum of baseband noise originating in the transconductance devices leaks through
noise current at the output is switch FET’s unbalanced by an offset. Replace the noise voltage
with an offset voltage, . The mixer output current is now
(2) a square-wave signal of constant offset in its zero-crossings,
, whose height is modulated by the noisy current,
where the last term applies to a sinewave LO of amplitude . , where is the input-referred flicker noise of the
Sampled images of this spectrum appear at integer multiples of transconductance FET. In this case the output current has a con-
(Fig. 5). At , the LO feedthrough amplitude is equal stant DC offset equal to ( )/( ) and a low-frequency
to , which is independent of LO amplitude or switch size, noise component equal to
but it has no flicker noise sidebands.
If the mixer is used for upconversion, the switches contribute (6)
no flicker noise to the output at , although flicker noise in
the transconductance stage is upconverted to this frequency. This is similar to (2), with replaced by . The corre-
Short channel MOSFET’s obey the following – character- sponding SNR is
istic, which takes into account velocity saturation [8]
(7)
(3)
If it is assumed in comparing (5) and (7) that the noise volt-
where is the drain current, is the transistor gate over-drive ages are of a similar order of magnitude, then as
voltage ( ), and are the transistor channel width , switch-induced noise is much larger than transconductance
and length, respectively, is the gate oxide capacitance, stage flicker noise leaking to the output due to DC offsets. In
is the saturated velocity, and is the saturated electric field. general, in (7) can represent any unwanted interference at
Hence, the ratio of the transconductance ( ) to its DC bias the mixer input, such as low-frequency noise on the ground line.
current ( ) is
D. Switch Noise: Indirect
The analysis so far suggests that flicker noise at the mixer
(4) output may be eliminated if the LO waveform is a perfect
square-wave with infinite slope at zero crossing ((1)). How-
In long channel transistors ( ), approaches ever, as the LO slope rises, output flicker noise appears via
as predicted by the classic square-law. However, for short another mechanism that depends on LO frequency and circuit
channel devices at large over-drive voltages, approaches capacitance. This is called the “indirect” mechanism.
. The short-channel formulas are used from here on. Consider the mixer shown in Fig. 6, with a perfect
These expressions are used to predict the SNR at the mixer square-wave applied to the LO port, alternating between
output, which is also an indication of its noise figure. The gain voltages and . The voltage at the tail of the switching
from the RF input, , depends on because the input FET pair is . In one half cycle of the LO, switch M1 is ON and
transconductance is ). As the conversion gain of a M2 OFF. Therefore at the gate of transistor M1 a voltage
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
18 IEEE TRANSACTIONS ON SOLID STATE CIRCUITS, VOL. 35, NO. 1, JANUARY 2000

Fig. 6. Single-balanced mixer with a square-wave LO.

Fig. 9. Waveforms resulting from a square-wave LO: (a) voltage at tail, (b)
(a) (b) capacitance current, and (c) output current.
Fig. 7. Mixer with square-wave LO at each half cycle: (a) first half cycle and
(b) second half cycle.
the mixer output. The amount of noise is the average of the
output current

(8)

Therefore the output noise current is

(9)
Fig. 8. Mixer in Fig. 6, replaced with a source follower.
The conversion gain to flicker noise in due to the indirect
is present in series with the noise source, , as shown in Fig. process is ( ) . This gain grows with LO frequency, but is
7(a). In the next half cycle, switch M2 is ON and now only the usually smaller than the gain due to the direct mechanism ((1)).
DC voltage, , is at the gate [Fig. 7(b)]. Associating all the If the total capacitance at the tail is comparable to of the
noise with one FET’s gate correctly captures the differential transconductance stage, then the SNR for indirect noise is given
noise in the switching pair. by
To find the tail voltage, , the hard-switched mixer is thought
of as a single source follower continuously connected to the tail
because of an assumed instantaneous crossover from M1 to M2, (10)
whose gate voltage alternates between zero and around a bias
(Fig. 8). where it is assumed that the switched mixer conversion gain is
The voltage at the tail, , is estimated using a linear model equal to . Therefore, the effect of flicker noise at the mixer
because is much smaller than the LO voltage, , and be- output can be reduced by applying a square-wave LO with sharp
cause the source follower is relatively linear for large signals. transitions and reducing the tail parasitic capacitance, or equiv-
Assuming that the transconductance is , the time constant at alently, increasing the unity current gain frequency of the tran-
the source in Fig. 8 is , which is normally much smaller sistors ( ). When the junction capacitance of the transconduc-
than the LO period, . As a result, the tail voltage waveform tance stage dominates the total capacitance at the tail, increasing
charges exponentially to at one half cycle, and discharges to the size of switches improves the SNR since goes down.
zero in the other half cycle, as in Fig. 9(a). However, when the capacitance of the switches dominates, then
This voltage produces the current waveform shown in Fig. making the switches larger degrades the SNR since is low-
9(b) in the capacitance at the tail, . The capacitive current, ered as the square root, while reduces linearly with channel
has a frequency equal to the LO frequency, with zero DC length. Unlike the direct mechanism, now if the over-drive of
value. At the mixer output in Fig. 6, when M1 is ON the differen- the transconductance stage is reduced decreases and SNR
tial current is , while in the next half cycle, when M2 is ON, degrades.
the output current commutates to , [Fig. 9(c)]. The output When a sine-wave LO is applied to the mixer, the source fol-
current alternates at twice the LO frequency with non-zero dc lower model of Fig. 8 remains valid with the difference that its
value, which indicates that baseband flicker noise is present at gate voltage consists of the switched noise super-imposed on a
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
DARABI AND ABIDI: NOISE IN RF-CMOS MIXERS: A SIMPLE PHYSICAL MODEL 19

The current produced in the tail capacitance by the recti-

fied sine-wave, after downconversion by the noisy LO (train of
pulses), has an average value of

(13)
Fig. 10. Indirect mechanism for a sine-wave LO.

where is equal to , as calculated before, and is the

rectified sine-wave at tail. This also contributes to the low-fre-
quency noise at the mixer output. Notice the similarity of the
above equation to the one derived for direct noise in Section
II-C. If the capacitive current [the term in (13)] is
replaced by the constant tail current, , the result is exactly the
same obtained in (1) for direct noise. After some simplifications
in (13), the low-frequency noise spectrum at the mixer output is

(14)

This has the opposite sign to the term arising from the
square-wave noise, given in (9). The total noise at the output is
the sum of the two components, given in (9) and (14)

(15)
Fig. 11. Waveform at switching pair tail due to a sine-wave LO.

When , the noise due to indirect mechanism is

full-wave rectified sine-wave with a peak equal to the LO am- negligible for a sine-wave LO. However, it rapidly goes up as
plitude (Fig. 10). The rectified sine-wave captures the action of the tail capacitance increases. Downconversion of the switched
the alternating source followers; the tail voltage follows the gate noise at the tail by the noisy LO results in a second-order output,
with the larger voltage. Again, the tail voltage waveform ( ) is and is neglected.
estimated by superposing two signals, consisting of the expo- In most practical situations, then, flicker noise due to a
nentially charging/discharging [Fig. 9(a)] and a large recti- sine-wave LO is attributable to the direct mechanism, which is
fied sine-wave ( ) whose amplitude is frequency independent. However, even a LO waveform with
infinitely fast risetime and falltime does not eliminate flicker
noise but pushes it down to a level determined by the tail
(11) capacitance. In general, LO waveforms with a large
product, that is, a low-frequency LO with sharp transitions will
lower flicker noise.
and whose phase difference with respect to the LO is
III. HIGH FREQUENCY NOISE IN A MIXER
A. White Noise in Mixer Switches
(12) The model developed to understand the origins of flicker
noise at the mixer output is now extended to white noise in the
where is the LO amplitude and is the switch transcon- switches. Starting with the direct mechanism, the noise current
ductance. Fig. 11 shows the resultant voltage waveform at the at the output [Fig. 12(a)] consists of train of pulses, with a rate
source. of twice the LO frequency, with a height equal to and a
After downconversion, the exponential waveform leads to a width randomly modulated by noise. To simplify analysis, this
noise component at the mixer output as calculated earlier and is approximated by a train of perfect rectangular pulses with
given in (9). The rectified sine-wave is downconverted by a some constant width of and a height of ), sampling
noisy LO, consisting of a perfectly periodic square-wave and the broadband noise which fluctuates at rates comparable
a train of pulses at twice the LO frequency, with a height equal to, or greater than the LO frequency [Fig. 12(b)].
to 2 but a width modulated by noise [Fig. 4(b)]. The height of The width of each pulse in Fig. 12(b), , is calculated from
the noisy pulses in Fig. 4(b) is , since these pulses downcon- the – characteristic of the switching pair. Fig. 13(a) shows
vert the tail bias current ( ). a simplified piecewise linear – curve which approximately
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
20 IEEE TRANSACTIONS ON SOLID STATE CIRCUITS, VOL. 35, NO. 1, JANUARY 2000

(a)

Fig. 14. Frequency translations of white noise originating in mixer switches.

which is white and cyclostationary can be expressed as the

product of , a periodic and deterministic sampling
(b)
function, and the white and stationary switch input-referred
Fig. 12. (a) Noise current at the output and (b) output noise approximated by noise [10]. The autocorrelation of the output noise is calculated
a train of pulses sampling the input noise.
as

(18)

The autocorrelation of white noise, , is a delta function.

However, the output noise autocorrelation is a function of both
and , which indicates that the output noise is not stationary
but periodic, proving that the output noise is white and cyclo-
stationary, as expected. The power spectral density is calculated
by averaging over one period and transforming the result,
now only a function of , into the frequency domain [10], [11].
The power spectral density of the output current noise is

(19)

Fig. 13. (a) Switching pair I –V curve, (b) the transconductance of the
The input noise is white and stationary and its power spectral
switching pair in voltage domain, and (c) transconductance in time domain. density is

applies to any differential pair, independent of the detailed tran- (20)

sistor characteristics. In a well-designed switching mixer the ex-
cursion is usually much smaller than the LO amplitude, where is the channel noise factor, traditionally 2/3 for long-
. The derivative of this characteristic [Fig. 13(b)] gives the channel MOSFET’s, higher in practice due to hot carrier effects
bias dependent transconductance of the switching pair. For a [9] (also it lumps other sources of noise in a transistor, such
sinewave LO ( ), which produces a time-pe- as back-gate noise), and is the switch transconductance at
riodic transconductance, is then nonzero over the zero-crossing
time window [Fig. 13(c)].
Switches contribute noise to the mixer output over the time (21)
when they are both ON. If one switch is OFF, it obviously con-
tributes no noise, and neither does the other switch that is ON Combining (19)–(21), the power spectral density of the output
because it acts as a cascode transistor whose tail current is fixed current noise due to one switch is
to by the RF input transconductance stage.
The sampling window, , in Fig. 12(b), is the time when both (22)
switches are ON, that is . This means that the switch
When the LO waveform is a sine-wave, and then
noise, , is transferred to the output only at each zero-crossing.
The sampling function, , can be expressed as
(23)

(16) which says that the output noise density of switches only de-
pends on LO amplitude and the bias current, and not on tran-
where is periodic at twice the LO frequency, since there are sistor size!
two zero-crossings over every cycle of the LO. There is a physical explanation of this surprising result.
The mixer output noise The discrete sampling action at the zero-crossings aliases the
broadband white noise of the switches (Fig. 14). The finite
(17) bandwidth of the sampling pulses, determined by the inverse
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
DARABI AND ABIDI: NOISE IN RF-CMOS MIXERS: A SIMPLE PHYSICAL MODEL 21

where the first term is the white noise at downcon-

verted by the fundamental of the commutating waveform, the
second term is noise at downconverted by the third
harmonic of the LO, whose amplitude is one third of the main
harmonic, and so on. The noise is uncorrelated at each sideband
and frequency, and the various contributions add as the mean
square.
Including noise due to switches and the load, the total white
noise at the mixer output is
Fig. 15. Frequency translations of white noise originating in RF input
transconductor. (26)

of the time when both switches are on ( in Fig. 12), limits which simplifies to
the number of aliases. In fact, the train of finite pulses in Fig.
12(b) may be replaced by a train of impulses (an ideal sampling (27)
function) passed through a filter with a sinc-shape frequency
response, whose bandwidth is proportional to (the zeros
where the first term is due to the two load resistors , the
of the sinc will lie at frequencies of , where 1, 2,
second term is the output noise due to the two switches, and
3, ). Thus, with all else the same, as the switches get larger
the third term shows the noise of the transconductance stage
they turn ON for a shorter time and decreases, widening the
transferred to the mixer output, assuming a conversion gain of
sampling bandwidth. However, as the switch size gets larger its
.
input referred noise density is also lowered. The integrated rms
This equation clearly shows how mixer output noise varies
output noise, therefore, remains constant, as does the equivalent
with different circuit parameters, such as LO amplitude ( ) or
white noise spectral density. This is similar to the well known
mixer DC bias current ( ). Most importantly, it allows the circuit
integrated voltage noise on a switched capacitor, which
designer to straightforwardly design the mixer to meet a target
is independent of the size of the switch.
noise figure. In the double-balanced mixer there are twice as
Also, when some other external white and stationary noise
many FET’s in the transconductance stage and the switches, so
source is present at the mixer LO port such as in LO buffers,
the output noise is
its effect at the mixer output can be found by adjusting the noise
factor, , in (20). Similarly, (17) can be applied to find the influ-
ence of any interferer in the LO port at the mixer output, where (28)
will represent the interferer signal (the interferer should be,
however, small relative to the LO magnitude so that the assump- where is the bias current in each side of the mixer. Now com-
tions remain valid). An interferer at a frequency of produces pare a scaled double-balanced mixer with the same total current
harmonics at the mixer output at , , , as a single-balanced mixer (that is, the former is biased at half
since is periodic at twice the LO frequency. the current per branch but the same as the latter). The
equations show that the output noise is the same for both mixers.
B. Transconductor Noise and Total Mixer Noise However, since the gain of the double-balanced mixer from the
differential input is half, the input referred noise voltage is twice
With the noise due to the switches accounted for, what
as large. Referred to a differential 100 source, its noise figure
remains is the contribution of the transconductor stage to mixer
is 3 dB larger than that of a single-balanced mixer referred to a
output noise. As far as the mixer is concerned, white noise
single-ended 50 source resistance. The main advantage of the
originating in the transconductor is indistinguishable from the
double-balanced mixer is that it suppresses LO feedthrough, as
RF input signal. Therefore, as mixer commutatation is assumed
well as noise or interferers superimposed on the LO waveform
square wave-like, the LO frequency and its odd harmonics will
applied to the mixer. It cannot suppress the uncorrelated noise
downconvert the respective components of white noise to the
in the switches.
IF (Fig. 15). After including the mixer conversion gain of ,
the noise at IF is C. Mixer Noise Optimization
The expression for total noise at the mixer output is expressed
(24)
in terms of bias quantities by replacing for a short channel
MOSFET by , as discussed earlier
The factor in the last term represents accumulated noise after
aliasing. Any periodic LO waveform, sine-wave or otherwise,
(29)
which switches the mixer results in square-wave commutation
of the transconductance stage output current. Therefore, using
the well-known harmonic amplitudes of the square-wave This shows that the relative noise contribution of the switches
to the transconductance FET is . As the gate
over-drive bias on the transconductance FET approaches the
(25)
sine-wave LO amplitude, the switches, and transconductance
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
22 IEEE TRANSACTIONS ON SOLID STATE CIRCUITS, VOL. 35, NO. 1, JANUARY 2000

Fig. 16. Single-balanced mixer used in simulations.

(a)

stage contribute comparable noise at the mixer output. This is

the fundamental tradeoff between noise and linearity in active
mixers.
“Linear” mixers may bias the transconductance FET at
a large over-drive to enhance the linearity, accompanied by
modest LO swing to keep the switch transistors operating in
saturation. These conditions boost the relative noise contribu-
tion of the switches. Also, as the DC voltage drop across the
load resistor approaches the gate over-drive of the transcon-
ductance FET, the noise contribution of the load becomes more
important. (b)
Simple estimates of mixer noise sometimes neglect the contri-
Fig. 17. Gain from noise port to output at: (a) 1 GHz and (b) 250 MHz.
bution of the switches and load resistors, which underestimates
noise figure by 2–4 dB. Also ignoring the aliasing of white noise
in the transconductance underestimates noise figure by up to 1
dB [ compared to 2 in (25)].
from the noise port to the mixer output is comparable to the gain
due to the direct mechanism [Fig. 17(a)].
IV. VALIDATING THEORY BY SIMULATION The white noise of the switches is also simulated using
To validate this theory a single-balanced mixer, shown in Fig. SPECTRE-RF periodic steady-state analysis, which can
16 is simulated using SPECTRE-RF at different LO frequencies analyze noise of nonlinear circuits such as mixers. Fig. 19
and amplitudes. The FET’s are described by Philips MOS model compares the simulated output white noise of one switch with
9 parameters for the ST Microelectronics HCMOS7 process. A predictions based on (23), at different bias currents and LO
sine-wave voltage with an amplitude of 10 mV and frequency amplitudes, for the mixer circuit of Fig. 16. The simulated noise
of 2 MHz models the low-frequency flicker noise at the gate of is very close to predictions. Specifically, (23) predicts that noise
one of the switches ( ). The amplitude of the output signal at 2 is independent of switch size. To confirm this, mixer noise is
MHz is taken from the FFT of the output signal. The gain from simulated versus switch width and the results are plotted in
the noise port to the output is then deduced. Fig. 20. As the width varies from 4.5–18 m, corresponding to
The direct noise mechanism is simulated by applying a about 100% variation in switch , the white noise contributed
sine-wave LO with different amplitudes and frequencies, by the switch to the mixer output changes only by 15%. This
shown in Fig. 17. Very close agreement is seen between the suggests that switches should be sized no further than large
simulations and theory. As predicted by (15), the gain at 1 GHz enough to turn on and off quickly for good mixer linearity,
and 250 MHz is almost the same. The gain decreases as the otherwise they lower mixer bandwidth and raise the load on the
LO amplitude goes up. At moderate LO swings, the gain from LO buffer without benefit of lower noise.
noise port to the output is almost unity, which means that the
input-referred flicker noise of the switches can severely raise V. MEASUREMENT RESULTS
the mixer’s noise floor at low frequencies.
Indirect noise is simulated by applying a perfect square-wave The mixer shown in Fig. 21 was fabricated in the ST Micro-
at the LO port. As seen in Fig. 18, simulations again agree electronics 0.25 m CMOS process, and its output noise was
closely with theory. In Fig. 18(a) when the tail capacitance in- measured. Measurements are now compared with theory and
creases beyond a certain value, the assumption that simulations. The flicker noise measurements are at an output
becomes invalid, and the theoretical curve deviates from the frequency of 2.5 kHz, where the spectrum is clearly . White
simulated gain. Fig. 18(b) shows that at 1 GHz, the voltage gain noise is measured at about 1 MHz, where the spectrum is flat.
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
DARABI AND ABIDI: NOISE IN RF-CMOS MIXERS: A SIMPLE PHYSICAL MODEL 23

(a)

Fig. 20. Switch white noise versus its width.

(b)
Fig. 18. Indirect noise at a constant: (a) frequency and (b) tail capacitance.

Fig. 21. Single-balanced mixer fabricated in ST HCMOS7 process.

The factor of two represents the uncorrelated noise of two

switches. The value of was initially predicted by scaling
the results of actual flicker noise measurements done at UCLA
[12]. Its exact value in this CMOS process was later extracted
from actual noise measurements. Equation (30) leads to an
input-referred flicker noise power spectral density for two
switches of −117.1 dBm/Hz with respect to 50 , which was
then transferred to the output by the noise gain given in the
plots of Fig. 17.
The output noise was also predicted using the periodic steady-
state (PSS) noise simulation in SPECTRE-RF. The results of the
output noise (with respect to 50 ) is shown in Fig. 22, for two
Fig. 19. Switch white noise at the mixer output for different LO amplitudes
and bias currents.
frequencies of LO: 1000 and 250 MHz.
The simulations agree very well with theory. The flicker noise
behavior of a mixer can be very simply calculated based on
The input referred flicker noise voltage of the switch FET’s the equations presented here, and from knowledge of the de-
is vice characteristics. In all cases, the noise level goes down as
the LO amplitude increases. Measurements also match theory
very well, with the largest discrepancy of 2.5 dB at the highest
(30)
amplitude of the LO power. One possible source of this discrep-
ancy is that both theory and Philips MOS 9 models assume that
where input-referred flicker noise is independent of bias. Any bias de-
= 1.2 10 ; pendence in reality will be most pronounced in the measure-
= 7.8 µm; ments at large LO amplitudes.
= 0.18 µm; Most of the flicker noise at the mixer output originates in the
= 6.9 fF/µm ; switches, since the load is free of flicker noise. Also very little
= 2.5 kHz. flicker noise in the transconductance FET is expected to leak
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
24 IEEE TRANSACTIONS ON SOLID STATE CIRCUITS, VOL. 35, NO. 1, JANUARY 2000

(a)
(a)

(b)
Fig. 22. Flicker noise at the mixer output at different LO amplitudes and (b)
frequencies: (a) LO = 1000 MHz and (b) LO = 250 MHz.
Fig. 23. White noise at the mixer output at different bias currents and LO
amplitudes: (a) LO = 0.6 V and (b) LO = 1 V.

to the output in the measurements as the offset is much smaller

than the bias . independent of transistor type, and may be applied to bipolar ac-
The mixer white noise is measured at different bias currents tive mixers.
and LO amplitudes, and compared to the theoretical white noise Switches in an active mixer contribute flicker noise to the
at the mixer output predicted by (27), and PSS simulations in mixer output in two different ways. One way, labeled direct, is
SPECTRE-RF (Fig. 23). Again the measurement results are by random modulation of the time instants of mixer switching.
close to theory. A maximum error of about 1 dB between theory This contributes flicker noise at or near zero IF. Flicker noise is
and measurements is seen, while the theory and SPECTRE-RF lower with sharper LO transitions relative to LO frequency. In
simulations agree to within a fraction of 1 dB. the extreme case of a square-wave LO waveform with infinite
slope, flicker noise at the mixer output does not disappear but
may remain significant at GHz LO frequencies due to a newly
VI. CONCLUSIONS AND DISCUSSION discovered indirect mechanism. This noise grows smaller with
lower LO frequency, or with higher device and diminishing
This paper presents a simple but accurate model to predict parasitic capacitances.
flicker and white noise in mixers. By pinpointing the origins Although mixer flicker noise is only practically important
of the noise and the mechanisms of frequency translation, it in zero-IF or low-IF receivers, it is useful in probing all noise
allows the flicker noise to be optimized in a downconversion mechanisms in a time-varying circuit, because unlike white
mixer intended for use at or near zero IF. The model uses the noise it does not accumulate after frequency translation, and
simplifying assumption that the switch FET’s commutate in- therefore can be uniquely traced to the original sources at
stantly, and that the flicker noise of switches can be represented baseband.
by a low-frequency bias-independent input voltage source. In The flicker noise model has been extended to evaluate white
spite of these apparently gross simplifications, good agreement noise in the switches. Simple expressions are derived to capture
is obtained with measurements. More complicated analytical the noise contribution of different sources in a switching mixer,
methods using time-varying gain and stochastic concepts [3], or requiring only basic circuit parameters, such as bias current, LO
exhaustive analyzes [13] also deviate from measurements by as amplitude, and load resistance. White noise in switches sampled
much as 2 dB. This analysis culminates in simple analytical ex- by pulses of nonzero width results in a -like expression for
pressions to capture mixer noise, analogous to well-known ex- rms noise, which is independent of switch size. No complicated
pressions for noise in an amplifier. The physical model is largely mathematical expression or MATLAB simulation is necessary.
Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.
DARABI AND ABIDI: NOISE IN RF-CMOS MIXERS: A SIMPLE PHYSICAL MODEL 25

Contrasted to previously reported methods [3], [13], the model [11] J. G. Proakis, Digital Communications: New York: McGraw Hill, 1995.
presented here is specific to hard-switched commutating mixers, [12] J. J. Min, “Characterization of CMOS FET’s flicker noise in VLSI tech-
nologies,” masters dissertation, Dept. Elect. Eng., Univ. of California,
but it allows the noise to be predicted by a very simple equation. Los Angeles, CA, 1988.
Obviously this is simple and flexible, and offers the circuit de- [13] M. T. Terrovitis and R. G. Meyer, “Noise in current-commutating CMOS
signer direct insights on mixer noise optimization and receiver mixers,” IEEE J. Solid-State Circuits, vol. 34, pp. 772–783, June 1999.
frequency planning.
We believe that the method described here not only sheds
light on mixers, but also on closely related circuits such as os-
Hooman Darabi was born in Tehran, Iran, in 1972.
cillators, which still lack a physical model to comprehensively He received the B.S. and M.S. degrees in electrical
explain their noise characteristics. engineering from Sharif University of Technology,
Tehran, Iran, in 1994 and 1996, respectively. He re-
ceived the Ph.D. degree from the University of Cali-
REFERENCES fornia, Los Angeles, in 1999.
[1] A. A. Abidi, “Direct-conversion radio transceivers for digital commu- He is currently with Microlink Corporation, CA.
nications,” IEEE J. Solid-State Circuits, vol. 30, pp. 1399–1410, Dec. His interests include analog and RF IC design for
1995. wireless communications.
[2] A. Rofougaran, G. Chang, J. J. Rael, J. Y.-C. Chang, M. Rofougaran, P.
J. Chang, M. Djafari, J. Min, E. W. Roth, A. A. Abidi, and H. Samueli,
“A single-chip 900-MHz spread-spectrum wireless transceiver in 1-µm
CMOS—Part II: Receiver design,” IEEE J. Solid-State Circuits, vol. 33,
pp. 535–547, Apr. 1998.
[3] C. D. Hull and R. G. Meyer, “A systematic approach to the analysis of Asad A. Abidi (S’80–M’81–SM’95–F’96) received
noise in mixers,” IEEE Trans. Circuits and Systems—I: Fundamental the B.Sc.(Hon.) degree from Imperial College,
Theory and Applications, vol. 40, pp. 909–919, Dec. 1993. London, in 1976 and the M.S. and Ph.D. degrees in
[4] R. Telichevesky, K. Kundert, and J. White, “Receiver characterization electrical engineering from the University of Cali-
using periodic small-signal analysis,” in Proc. Custom Integrated Cir- fornia, Berkeley, CA, in 1978 and 1981, respectively.
cuits Conf., 1996, pp. 449–452. He was with Bell Laboratories, Murray Hill, NJ,
[5] A. Rofougaran, J. Y. C. Chang, M. Rofougaran, and A. A. Abidi, “A 1 from 1981 to 1984 as a Member of the Technical
GHz CMOS RF front-end IC for a direct-conversion wireless receiver,” Staff in the Advanced LSI Development Laboratory.
IEEE J. Solid-State Circuits, vol. 31, pp. 880–889, July 1996. Since 1985, he has been with the Electrical Engi-
[6] J. Chang, A. A. Abidi, and C. R. Viswanathan, “Flicker noise in CMOS neering Department of the University of California,
transistors from subthreshold to strong inversion at various tempera- Los Angeles, where he is a Professor. He was a
tures,” IEEE Trans. Electron Devices, vol. 41, pp. 1965–1971, Nov. Visiting Faculty Researcher at Hewlett Packard Laboratories during 1989. His
1994. research interests are in CMOS RF design, high-speed analog integrated circuit
[7] D. M. Binkley, J. M. Rochelle, B. K. Swann, L. G. Clonts, and R. N. design, data conversion, and other techniques of analog signal processing.
Goble, “A micropower CMOS direct-conversion, VLF receiver chip for Dr. Abidi served as the Program Secretary for the International Solid-State
magnetic-field wireless applications,” IEEE J. Solid-State Circuits, vol. Circuits Conference from 1984 to 1990, and as General Chairman of the Sym-
33, pp. 344–358, Mar. 1998. posium on VLSI Circuits in 1992. He was Secretary of the IEEE Solid-State
[8] N. G. Einspruch and G. Gildenblat, “VLSI electronics microstructure Circuits Council from 1990 to 1991, and from 1992 to 1995 he was an Editor of
science,” in Advanced MOS Device Physics: Ney York: Academic, 1989, the IEEE JOURNAL OF SOLID-STATE CIRCUITS. He has received the 1988 TRW
vol. 18. Award for Innovative Teaching and the 1997 IEEE Donald G. Fink Award, and
[9] A. A. Abidi, “High frequency noise measurements on FET’s with was the corecipient of the Best Paper Award at the 1995 European Solid-State
small dimensions,” IEEE Trans. Electron. Devices, vol. ED-33, pp. Circuits Conference, the Jack Kilby Best Student Paper Award at the 1996 In-
1801–1805, Nov. 1986. ternational Solid-State Circuits Conference (ISSCC), the Jack Raper Award for
[10] W. A. Gardner, Cyclostationarity in Communications and Signal Pro- Outstanding Technology Directions Paper at the 1997 ISSCC, and the Design
cessing: Piscataway, NJ: IEEE Press, 1993. Contest Award at the 1998 Design Automation Conference.

Authorized licensed use limited to: gnana sai raghu yerrapothu. Downloaded on September 18,2023 at 04:20:41 UTC from IEEE Xplore. Restrictions apply.