You are on page 1of 22

Mi

er

Design Report The guide of designing a Mixer for Monkey and Dog DESIGNERS

The goal of this paper is to make a little bit more light in mixer design, but this does not mean that we will really succeed in this... This paper, as everyone in the world should do, will respect the freedom of choice and in general the freedom of the designer and it is good to say, from the beginning, that there is not so much freedom to find in mixers design.., but happily enough has left for original ideas.. . Lets see: Respecting the market law we will talk about a double balanced mixer (Fig.1).. So first questions when we start designing are: 1. 2. If we know that Vdd=2.5V what else should we need, W/L reports? So how big W/L report for the transconductors and how small for switching pairs, or rehearsal? What should the relation between W/L switch and W/L tranconductor?...

3.

But if we start with this questions we will never end, that is why we its required to pay attention to the parameters of a mixer. A good mixer seems to be defined by such values for his defining parameters: NF=12dB IIP3=+5dbm Gain=10dB Port-to-Port isolation 10-20dB

It is nice to propose yourself to do better, but is that possible? Lets find out... Because we dont want people to believe in genies there is to say that for writing such a report where necessary this kind of papers: * Noise in Current-Commutating CMOS Mixers A simple

Manolis T. Terrovitis IEEE Journal of Solid State Circuits, Vol. 34,No. 6, June 1999

* Noise in RF-CMOS Mixers: Physical Model Hooman Darabi and Asad

A. Abidi IEEE Transactions on Solid State Circuits, Vol. 35, No. 1, January 2000

* Distortions in Elementary Transistors Circuits Will Sansen IEEE Transactions on Circuits and
Systems II; Analog and Digital Signal Processing, Vol. 46, No. 3, March 1999

Tranconductance Based Cmos Circuits Circuit Generation, Classification and Analiysis Eric A.M. Klumperink 1997 Why dog and monkey? Dog is the one who is tracing statistically the design and find the right approximate solution or the best one for it(if he is lucky) and monkey is trying all the extreme cases and than conclude that the mean solutions area is the best one without any real motivation... (just all the time in the middle baby thats the safest place, yes...).

NOISE FIGURE:
We are using the first paper now: Manolis T. Terrovitis. A defining formula(1) for MIXER Noise Figure is the one below, assuming all MOST working in strong inversion:
NF := c
2

(2 3 )
c gm3 Rs
2

( 4 1 G) 2 c ( gm3 ) 2 R s

( 4 rg1 G )
2

c ( g m3 ) Rs
2 2

( 2 rg3 )
c Rs
2

c ( g m3 ) 2 Rs Rl
2

This formula describe the N.F. (SSB) for double balanced mixer, but in this way it is now does not say too much about the mixer we want, just it seems we need a big gm3; Rs=50 ohms and Rl seems to be preferred as big as possible. For making the things easier we will consider =1(supposing we are switching very hard (VLO=0.8V, LO amplitude), c=2/ and considering rg1 and rg3 very small, almost zero and taking in account this two formulas, which are approximate for a real situation:

MeanG :=

( 2 I B)
Vlo of the

and the mean square transconductance of the switching pairs:


2 ( K ) 0.5 ( I ) 3 B 1 MeanSqG := 4.64

Tlo

and also

We will find for Noise Figure a formula(2), which approaches this one, where we also have to say that it is considered for short channels, for (can depend on bias and can be affected by hot electron phenomena; 2/3 for long channel MOST or higher for shorter channel length ), such values as bellow:
1 := 1 2 := 1

Vlo := 2 Tlo

This formula(2) can approximately to show us how important is the role that is played by gm in Noise Figure, but the graph bellow will exaggerate a little bit ...so be careful, do not trust in it too much, but just as an idea...(Vx=VGT of the transconductor-M3 is the same with the Vx from the papers of Manolis...)

1 1 2 3 1 NF1 ( gm ) := + + Vx 2 2+ 2 2 2 ( gm Rs ) gm Rl Vlo
2

Any way this is one of the few possibilities of seeing how can we reduce noise on a mixer, see (Fig.2):
4.935 10 12 1 .10 1 .10
13

12

Noise Figure function of gm on Log scale

1 .10 1 .10

11

10

1 .10 1 .10 1 .10

NF1 ( gm )

1 .10 1 .10 1 .10 1 .10

100 10 3.939 1 8 1 .10 1 10 8

1 .10

1 .10

1 .10

1 .10

1 .10

0.01

gm gm of the transconductor on Log scale

0.1 0.09

This graphic is considered(Federicco; I do not totally agree) that has less error that the next one, which shows also the Noise Figure, but as a function of the W of the transconductor (its clear that a bigger W means a bigger gm). The next graphic is based on such formula(3), where we have substitute gm with:

gm := k

w Vgt 2 l

gm := 2 or

Ib Vgt 2 and the final formula looks

almost like this, we hope people are going to like it(3):


N ( w) :=

1 3 1 + + + 2 2 2 2 2 Vlo w w Rs 2 k 2 k 2 w ( Vgt ) 2 Rs Rl 2k Vgt Rs l l l


2

and for clear reasons, we say that it wise to use a Vgt>sqrt(10)*100mV ,Vgt of the transconductor, but we will see later that there are not only personal reasons, but are also linearity reasons(explained latter)...O.K. Lets see the plot and discuss a little bit about it(Fig.3):

7.274 10 4

1 .10

Noise Figure as a function of W on Log

1 .10

1 .10

N ( w)

100

10

2.501

1 0.1 0.25

10

100

1 .10

w W on Log Scale

1 .10 9.999 10 3

From simulation point of view this plot is much relevant, even if is not exactly like in reality but its approximating it well... We would say that this is quite some light or oxygen... So first point is that we need really big transconductance or big transistors as drivers... There is to say that this graphic can give us an rough idea about the variation of N.F. with W, there are some errors in it of course...with respect to a real situation. Now, we are going to play a little bit tricky, considering gm or the gain of the mixer(Gain) fixed, trying to find new things. Lets see if it is possible. For this we will write the Noise Figure in a different way, a new formula(4),we choose some reasonable values for the other parameters which are kept constant and (Fig.4): Gain := 7 Vx := 0.36
2

gm := 10 8 , 2 10 7 ..

5 10 3

Vlo := 0.8

1 1 1 2 + + + Vx NFG ( gm ) := 2 2 2 2 Rs gm Vlo Gain

1 .10

Noise Figure with fixed Gain

1 .10

1 .10

1 .10

NFG(gm)
1 .10
4

1 .10

100

10 8 1 .10

1 .10

1 .10

1 .10

1 .10

1 .10

0.01

gm gm Log Scale

This plot take in consideration, that also the amplitude of the LO is fixed and there is no second order effect in which Vx=Vgt of the switches is modifying somehow, especially if the Vtsw is changing. But the problem is that Gain and gm are related from Gain=c*gm*Rl so the question still remains: How can we create a big gm in conditions of a small bias voltage? And the answer is again big W/L ratio. But this plot can show as how important is also the role played by the load Rl, which is directly contributing at the Gain. It is clearly wanted to have a bigger Rl, but the size of the Rl is limited by the condition of working in saturation for switches, which means that the size of Rl is dependent of the amplitude of LO. If we want to switch harder it is clear that for keeping switches in saturation we need smaller Rl. So here comes one of the many trades you can do in a mixer(there is very few to trade, but very much trade ..that is fanny):
Trade the value of the LO amplitude with Rl is equivalent with trading value of the flicker noise with Gain and as usual you loose more gain than noise.

But, if there are big transistors we will be able hard and win more gain, because of a bigger Rl. And now we are going to keep gm fixed and vary the more or less equivalent with varying Rl. But it is gain of a mixer can not vary too much or be too high

to switch less Gain, which is clear that the and so if...

gm3 := 5 10 6

and Gain is varying like this


2

Gain := 1 ..

200

1 1 2 3 1 NF ( Gain ) := + + Vx 2 2+ 2 2 gm3 R Gain Vlo The equation being as we can see above, there is also a graphic to show(Fig.5):
2.851 106

2.9.10

2.8.10

Noise Figure func of Gain/Log Scale

2.7.10

2.6.10

NF(Gain )
2.5.10
6

2.4.10

2.3.10

2.288 106

2.2.10

20

40

60

80

100

120

140

160

180

200
200

Gain

What if we can vary the Gain?Lin Scale

As we can see varying too much the gain does not help for noise figure of a mixer, but this just for playing with an equation... The main thing we might be able to say is that varying the load it does not help to much the noise figure, if the load does not have too small values, something like 0.1 ohms. So, for noise figure load is not so relevant that is why we can say: Noise Figure does not trade too much with Gain, from the load point of view(because also gm is contributing to the Gain and as we had seen big gm means low Noise Figure). This is may be a good thing, but we can not find the right value for the load from this formulas...

Another question might be what is the importance of Vlo and Vx, when we suppose that, some how , Gain and gm are fixed or more appropriate to say that Rl and gm are fixed, because Gain depends on how hard we are switching. So using the same equation :
Vx:= 0.365 Vlo := 0.1.. 2 1 := 1 3 := 1 gm := 0.08 Rs := 50 Rl := 10

2 1 1 1 + 2 3 + NF1( Vlo) := + Vx 2 2 ( gm Rs) 2 2 gm Rl Vlo 2

We will notice that if gm is low and Vx is also high the contribution of the Vlo to the reduction of noise is important. We can see a graphic now(Fig.6):
7.339
7.5

Noise Figure reported to Vlo

6.5

NF1 ( Vlo )

5.5

4.733

4.5

0.2

0.4

0.6

0.8

1.2

0.1

Vlo
The amplitude of the switching LO

1.1

But if gm is small, gm=0.0008 we will be able to see that Vlo is contributing more significant to the reduction of the Noise Figure and also we will show that if Vx is smaller the power of noise figure reduction of Vlo is smaller too(Fig.7):

8.123 10

8150

8100
Noise Figure reported to Vlo/Small gm

8050

NF1 ( Vlo )

8000

7950

7900

7.863 10

7850

0.2

0.4

0.6

0.8

1.2

0.1

Vlo
The amplitude of the switching LO

1.1

But if Vx=0.065V is small and gm=0.08 the Noise Figure reduction, because of Vlo is smaller(Fig.8)...
4.983

Noise Figure reported to Vlo/Smaller Vx

4.9

4.8

NF1 ( Vlo )
4.7

4.6

4.519

4.5

0.2

0.4

0.6

0.8

1.2

0.1

Vlo
The amplitude of the sw. LO/Vx=0.065V

1.1

As we can see, not only a big gm is required but also a big Vx, which is the difference between Vgt and Vt of the transconductor, this one of the reasons we recommended Vx> 10 *100mV (Fig.9).

10

14.29

15

10
Noise Figure reported to Vx/Vlo=0.8

5
NF1 ( Vx )

5.345

10

10

0
Vx

10
10

10

The amplitude of the sw. LO/Vlo=0.8V

But if higher:
9.407 10
3

gm<<0.08
4 1 .10

or

Vlo<<0.8V

the

slope

becomes

higher

and

Noise Figure reported to Vx/Vlo=0.5

9000

NF1 ( Vx )

8000

7000

6.266 10

6000

10

10

10

Vx
The amplitude of the sw. LO/gm=0.0008V

10

Fig.10 This graphics are a little bit strange, because it would show that as small Vx as small Noise Figure, this would be true if Vx would be much higher then one, or much higher than Vlo. But in normal case Vx is also contributing to the size of gm, and gm goes up much faster and is canceling more noise than the increasing of Vx is producing.

11

This show indirectly that a good bias makes a good mixer, so there is also a trade bias-Noise Figure... A good bias is giving you the possibility to switch much harder and to improve more the noise figure in conditions of a small Vx=Vgt. O.K. we had find out that we need some big stuff(MOST), really big, but how big? Who is imposing the upper limit for W/L report or for gm? For the amplitude of the LO... we know is limited by the bias conditions and also Gain and Linearity, but again what about W/L?

Luckily also some noise equation are going to help us with this question. So its important to observe that up to now we were working with Noise Figure(which is of course more convenient for a designer)but now we are going to discuss about output Noise values... The main interest here is the so called Indirect Switching Noise(flicker noise). To make more easy to understand what is about this Indirect Noise Mechanism, we can say that as the LO slope(amplitude, frequency or the LO sine wave is just approaching by a square wave )is increasing, output Flicker noise appears through another way which depends on LO frequency and circuit capacitance. But are more important the formulas that describe the phenomenon than just talking about it... So:
Ar := A gms gms + ( Cp lo)
2 2

:= atan Cp

lo

gms

io, n := 2

Cp T

Vn

where gms is the tansconductance of the switching transistor; Cp is the equivalent parasitic capacitance seen from the source of the switching transistor to up the ground; A is the amplitude of the LO; omega_lo is the LO frequency; T is the period of the local oscillator; Vn is the equivalent input noise at the gait of the switching transistor; Ar is the LO amplitude that rich at the drain of the tranconductor and is a kind the of phase noise(which, for our days, yet and designer sake is not so important)is more to show what are the consequences of o big capacitance which as we can see will reduce the port to port isolation(Lo to Rf). Because the transconductor is supposed to work all the time in saturation, the capacitance that contributes to Cp is small and linear, so the main problem is Cgs capacitances of the of the switching pairs(more correctly is to say that mainly only one Cgs is available because they are switching) and if they are all the time in strong inversion, we can be happy because Cgs is more or less linear and so will not affect to much linearity, but Gain. How is that? Well a big Cgs in saturation means a big W/L ratio with an approximate formula:

(Cp lo)2 2 2 gms + ( Cp lo)

12

so it seem that a big W/L for a switch is good for improving the Noise Figure and useful in bias settings, but for linearity(incase that the switched transistor gets out of saturation) or high frequency behavior (the speed of the device is becoming slower and the mixer is loosing very fast Gain with frequencys increasing...) this is not quite a happy thing. Situation becomes worse for linearity if the load is big and the switches get out of saturation, which means also that the equivalent Cp is no longer linear... Realizing now that the contribution to Cp is split between Cgs of the switch and the equivalent capacitance of the transconductor and the fact that big transistor means big parasitic capacitance we can see again that(with respect to the equation above): There is more than a trade Noise Figure_Gain_Linearity, but also a trade Noise Figure_Frequency, or Noise Figure_Bandwidth. It is also nice to realize whose is contributing more to Cp the transconductor or the switch? But the answer depends on bias and W/L or the transistors also. Any way the idea is that we would like that the main contribution to belong to the transconductor and not to the switch(because this means a weaker port to port Isolation or a bigger LO to RF/IF feed through). This kindof noise(we are talking about Flicker Noise) is not significant if omega_lo<<gms*Cp, but in our case, having a big gms, it may be significant. It looks a little bit strange because Noise is increasing because of current and gms, but Noise Figure is decreasing with W/L ration(gm) of the Transconductor. How can be this possible?.. This means that we need big transconductors but not big switches. O.k. but small switches means more bias consuming, and bias consuming means at least that we will not be able to switch so hard(good for linearity not to switch hard) and we will have more flicker noise near to the IF band... It look like we should study marketing first, and after that to try to design a mixer... But it can not be so difficult... Now that we know approximately how big we need the devices(because we know what kind of noise figure we want) and what are the disadvantages of big transistor(from above discussion) we can look at Fig.3 and say that a W in range of 80..150 is acceptable; of course if you are not looking for high gain you can choose bigger values, or even so you can have high gain(20dB) but bad linearity(1 up to 3dBm )...It looks like we are very poor but steel trading few... With respect to Cp we can say if it is comparable with Cgs of the transconductor that it is possible to estimate the role of this capacitance in to the SNR from the following formulas:
gm . . Cp Vin 2 . 2. Vn flo

Cgs :=

2 3

Cox W L

SNR3

13

where gm is of transconductor, flo is the frequency of the LO(local oscillator), Vin the voltage level at the gate of the trasconductor and Vn the input equivalent noise at the gait of the switching transistor1. So a big Cp does not brig a good signal to noise ratio, but it seems that we still can keep up because of the big gm we use(in this case gm and Cp are some how related because Cp is comparable with Cgs, which is linear-working in saturation- and also proportional with W/L ratio; gm/Cp in case of small switch devices could remain constant, but there are a lot of troubles here because for hard switching(Cp is bias dependent)Cp starts varying). So, at least is nice to know we have an upper limit for W/L ratio for both transconductor and switches, limit defined by Cp. It also important to see that for high frequency we need also big gm so are required fast devices for design, with low parasitic capacitances. At the final we want to show that electrical engineers know to make a 3D plot for Noise Figure:

NF0

Noise Figure 3D plot for very very small gm and Rl

14

G0 , k := ( 0.000000001 + k 5 ( 0.00000000001 Rl0 , i := 1 + i 0.025 ) )

this graphic can succeed only for small values and is more for fan. Now, that we almost finished with noise equations lets make a step forward...
Gain:

It is one of most traded parameters we can say, comparing with Noise Figure . Trading gain with Linearity, Port to Port Isolation and Frequency or Bandwidth. Gain has this expression: 2 appears because of using a double balance mixer, if we consider gm as being the tranconductance of one of the two transconductors. From this formula we can clearly understand that gain has something good from the point of gm: both Noise Figure and Gain need big gm and c(it is needed to switch harder for reducing the flicker noise). And not only from the point of view of gm but also from the point of view of Rl: a big Rl means both high gain and lower Noise Figure. So there should be a conclusion: Gain and Noise Figure for a mixer can have very good values , because there are no contradictions(as log as we can keep switches saturated) in obtaining them with the same circuit. But even if switches get out of saturation Gain is still improving and Noise Figure does not seem to change at all, but it might be possible that the LO to IF feed through do increase(anyway, nobody seems to like to study the behavior of switches out of saturation, yet ). That is way we will go to the third parameter Linearity ... But first of all we have to realize that the main trade for a mixer will be between linearity and noise figure, Gain being the consequence of their trade. We have to find the maximum of the value gm*Rl, the proper current and the proper load, because for small MOST we can have very high gain(25dB) and very good linearity but N.F. is really bad. But the trade between Gain and Linearity is described by this two equations:

Gain

2 . c . gm. Rl

VDSswitch :=

W Cox( VGTtransconductor ) 2 Rl 2 L
1

( 1)
and the other one
( 2)

gmtranconductor :=

W L

CoxVGTtransconductor

so the main idea would be that we need a low VDSswitch(for having a higher dynamic range) and a high gm*Rl product, if we use a small load, that means that for good gain is required high gm but, for linearity is required that Vgt>sqrt(10)*100mV and this means that linearity does not like too much difference between IDtrasconductor and gm. So it looks like this: While current IDtrasconductor is the result of VGS(bias) and gm is related to signal, linearity(IP3) is very small if the signal vgs is amplified by a high gm resulting a high signal current Ids which

15

becomes comparable with IDtrasconductor(bias current). This allows as another way to see linearity/non linearity in mixer: the signal current has to be much smaller than the bias current for a good linearity and that IP3 values are not sensitive to the Rl values(as long as switches are in strong inversion). Or the same thing can be sad like this: the higher the ration between vgs and VGT is the better linearity we have. Linearity: Because we have a combination between two devices a transconductor and a differential pair we consider that is proper to study both of the separately and after that, if possible to conclude about their functional behavior in a mixer, as one device. Tranconductors Linearity :

(* IV distortion in a MosFet Amplifier)because the transfer characteristic Ids-Vgs is quadratic and not exponential(like for bipolars) less distortion is expected. The drain current iDS and vGS of a MOST are related by :

( v GS VT ) 2 iDS := K ( 1 + v DS) or iDS := ( vGS VT ) 2 2 1 + ( v GS VT )


where K(or ) parameter depends on technology and W/L ratio and models to a first order the source series resistance, mobility degradation due to the vertical field and velocity saturation due to the lateral field in short-channel devices(it depends on channel length and is independent of the body effect). Also: We are going to make some IM2 and IM3 and case of both formulas. But all the calculation we are going to make for the easiest one and after that just presenting the results for the other one.
IDS + ids := 2
v GS := VGS + v gs iDS := IDS + ids

as you can see we make the difference between signal and bias, and subtracting the bias current component and reporting also to it, the result we will reach this:
ids IDS := 1 +

( VGS VT + v gs)
2

2 1 VGS VT

v gs

i :=

ids IDS

Idsp

or

IDS

:=

2 v gsp VGS VT

2 v gsp 4 VGS VT
1

ip :=

Idsp IDS

where

ip and vgsp are the peak values of the relative current swing and the gate-source input voltage, respectively. For small signals we need only the first term of the equation above, which yields:

16

Idsp :=

2 IDS VGS VT

Vgsp

Idsp := g m Vgsp

Also, the peak relative current swing is related to the input drive by: Now if we pay attention to the equation that describe an amplifier and compare the upper equation with it we can see something like this(y is the output and u the input of the amplifier, but we have practic an transconductor which act also like an amplifier, but the output is current):
y := a0 + a1 u + a2 u + a3 u + .... identifying ip := y and a1 := 1 a2 := 1 a2 u 2 a1 1 4 and a3 := 0
2 3

ip :=

(2 Vgsp)

VGT VT

u :=

2 Vgsp VGS VT

we can identify:

and calculate IM2=2*HD2 and IM3=3*HD3:


1 a3 2 u 4 a2
Vgsp

HD2 :=

HD3 :=
1

and we find out through substitution that:


IM2 := ip 4

IM2 := 2 HD2 and

IM2 :=

2 VGS VT IP3 := u 1

IM3 := 3 HD3

IM3 := 0

so it seems that IP3 is infinite so no third order distortion occurs and the only one we should pay attention is the second order harmonic. And we can say that is recommended to have less current but more VGT and these means it requires a small w/l ratio(but we know that for low noise figure we need a bigger w/l ratio - that might be a trade). As soon as the complete expression is taken of MOST, including the terms with 3/2 exponents, then the third order distortion appears in limited amounts, but we will realize that IP3 will depend on Vds, and the Vds of the transconductor is not so constant anymore. But even that if IP3 is varying periodically this is simulators problem.
As a general conclusion for the transconductor we can say that a bigger Vgs is more convenient for linearity, actually if the ratio between the input signal and Vgt remain the same the value of IM2 or IM3 also roughly remain constant.

IM3

But lucky us that we have small input signals fact that does not require too high Vgt(because in 2.5V there is not so much room for both hard switching and big Vgt+W/L for transconductor which means big current)for an acceptable value of IP3(+5dBm or higher). If the reader will try an approximate calculation for the other formula(for iDS) he will see that IM3 is no quite zero and IP3 is high sensitive when the device is going in deep triode region(this is a subject that we will talk later about it).

17

Switches Linearity (Distortion in a MOST differential Amplifier). We will try to find the contribution of the switches to the linearity through the point of view of the differential pair. The transfer characteristic of a differential pair with MOST is symmetrical around the origin so no second order distortion can thus occur. We can say that this combination transconductorswitches attenuates all the IM2 produced by the transconductor but, as we are going to see, switches will produce their own IM3. The differential output current can be calculated with such a formula (1):
ID1 ID2 := 1 2 n Cox W L ( Vin1 Vin2) 4 Iss

C n ox L

( Vin1 Vin2)

( 1) IB := ISS

n Cox

W L

:=

v Id := Vin1 Vin2

iOd := ID1 ID2

( 2)

and using (2) we can say that the relative current swing i is given by (3). Notice that the differential output current IOd is twice the ac current in each transistor.
i := iOd IB i :=

where vId
i := iOd IB i :=

vId ( 3) VGS VT 4 ( VGS VT ) is the differential output and we ca always use (4).
v Id 1 1 v Id VGS VT 1

v Id 4 ( VGS VT )
1

( 3)

VGS VT := I B

( 4)

for small values of vId<VGS-VT, the square root can be approximated by a power of series:
1 x := 1 x 2 x
2

16

....

( 5)

and applying this to (3) and switching to peak values ip, iOdp, VIdp we can find something like in (6):
ip := IOdp IB IOdp IB

for a pure small signal analysis, the power series can be limited to the first term only and we might find something like this:
IOdp :=

vId := .... (VGS VT ) 8 ( VGS VT )


v Idp 1

( 6)

( VGS VT )

VIdp IB

or

IOdp := ( VGS VT ) VIdp

or

IOdp := g m1 VIdp

( 7)

where gm1 is the gm of one of the switches and VGS the gate-source voltage. From (6) we can find (8):

18

ip :=

( VGS VT )

VIdp

or

ip :=

VIdp IB

( 8)

and doing in the same way like to the transconductor ip=y and u=VIdp/(VGS-VT), the coefficients are a0=0, a1=1, a2=0, a3=-1/8. Now we can see that IM2=0 so no second order distortion occurs. Coefficient a3 is negative which shows that compression distortion happens, so we finally can say that:
3 VIdp 2 ( 9.1) or IM3 := ( ip) from IM3 := 3 HD3 ( 9) 32 ( VGS VT ) 32 so we can see from (8) how big is the value of the Vgt that does not produce too much distortion for a differential pair but, any how because we use them like switches is clear that we have IM3 and if we want to make it smaller there are two solutions: IM3 := 3
2

IP3 := 4

2 3

( VGS VT )

( 10)

because we a referring to a mixer we are interested by the mean value of Vgt when the transistor is on and... swiching a little bit harder: THIS IS A SURPRIZE!!!(9.1); a smaller W/L ratio; But both of them produce inconvenience that means trade: switching less herder will brig more noise and a smaller W/L ratio for switches will mean more voltage consuming, which mean that the transconductor will require a smaller Vgs that meaning more IM2 and IM3 distortion in tranconductor stage which is more sensitive to Vgs that to Vds. So from what we have seen up to know we can say:
The distortion introduced by the transconductor is more sensitive to Vgs and a bigger Vgs for the transconductor means more linearity while the distortion introduced by the switches is more sensitive to values of Vds(it is wanted a smaller Vds which is imposed by the transconductor through the current that pass through the load, that means a smaller load and less gain...) but if the ratio Vds to the sqr. of the ratio IB/, remains the same we will have the same distortion. So for the transconductor we need big W/L and Vgs which means big current. In conclusion if we want bigger gain but the same linearity we have to switch harder(while keeping switches in SATURATION; but it is not such a recommended option voltage headroom consuming) and a bigger Vgs for transconductor. We see also that a MOST differential stage does generate third-order distortion, because of the limiting action of its transfer characteristic. It even generates a somewhat more thirdorder distortion than a bipolar stage. The input voltage allowed is much larger and depends also on Vgs-Vt. So for better results in linearity and not only, is recommended to use bipolar switches and MOS transconductor for a mixer.

19

What it can be nice is that because we have 4 MOST and mainly 2 all the time on, they act like a differential pair with Vgs equal to the mean of the LO amplitude plus the bias Vgs of the switches(that is why we are quite close by a differential stage case). So looking carefully to (8) we understand that using big or small transistors(but keeping the ratio between W/L(switch) and W/L(transconductor) constant we can have almost the same linearity. There is also to see that in (8) is depends on the W/L(switch) and IB is the one imposed by the transconductor. And a smaller (switches) it will mean(at least from paper) that well have more linearity but, again we trade bias voltage and this finally mean gain. But we can not exactly say what should be the ratio between W/L(switch) and W/L(transconductor) just that should be between 0.5 and 3, higher values will bring less linearity(and also less noise figure) and lower values will be more voltage headroom consuming. Because the equation we use is not considering second order effect we can not clearly reveal what is the consequence of such a high variation of Vds upon linearity for a mixer. That is why we will try to see another way to analyze linearity and consider the mixers transconductors as VCCS...
Distortion in 2VCCS Circuits: *Taylor Series and Intercept Point Definitions(just trying to use some extra high quality information from Erics book). As an equation for V to I conversion we are going to use the such formulas:

where g1,g2,g3 are the Taylor coefficient of the DC I(V) transfer characteristic of the VCCS, calculated in bias point (V0,I0):
g 1 :=

it3 := g 1 v + g 2 v + g 3 v

( 1)

where g1=g the transconductance of the MOST or VCCS. Discussing about a single VCCS we can write the inverse Taylor series:
v t3 := r1 i + r2 i + r3 i
2 3

d I dV

g 2 :=

2! dV2

d2

g 3 :=

3! dV3

d3

( 2a , b , c)

( 3)

and substituting (3) in (1) we can collect equal order terms up to third order and we find for I this kind of expression:
( 4) By equating the i-terms and equating the coefficients of the i2and i3-terms to zero, we find such relation between ri and gi : r1 := 1 g1 ( 5.1) r2 := g 2 i := g 1 r1 i + g 1 r2 + g 2 ( r1) i + g 1 r3 + ( 2 g ) 2 ( r1 r2) + g 3 ( r1) i
2 2 2 3

from this we can find the values for the intercept current and voltages for second order distortion:

(g1)

( 5.2)

r3 := 2 ( g 2) g 3 g 1
2

( 5.3)

20

VIP2 := 2

g1 g2 g1 g2

and

IIP2 :=

r1 r2 r1 r3

( 6)

and for third order distortion:


( 7)

and conform to 6.3.2 Distortion in Single MOST VCCSs table 6-8 for SVCCS we can identify Taylor coefficient g1,g2,g3 and r1,r2,r3(using the approximated values) and calculate using (6) and (7) VIP2,VIP3,IIP2,IIP3 . Now lets do it:
g 1 := 2 k ( V0 VT ) r1 := 2 k ( V0 VT ) 1 g 2 := k r2 := 1 8 k ( V0 VT )
2

VIP3 := 2

and

IIP3 := 2

g 3 := k r3 :=

( 8) 1 16 k ( V0 VT )
3 3

( V0 VT )

+ ( 9)

substituting in (6) and (7) we will find out


VIP2 := 2 ( V0 VT ) ( 10) IIP2 := 4 k ( V0 VT ) VIP3 := 2 2
2

and something like this:

( 11) ( 12) 2 1 + ( V0 VT )

(V0 VT )

IIP3 := 4 k ( V0 VT )

or

IIP3 := 8 I0

2 1 + ( V0 VT )

( 13)

Because we are interested on transconductor we will pay attention on IIP2,IIP3(because the load of the mixer is quite linear) and observe that we need(as already predicted above) a big Vgs for transconductor and is also possible to see why: so as not to lose one order when we square it and from simulation this seems to be quite sensitive. And also we might quote some rough useful conclusions from Erics book: *A MOST in weak inversion has a significant smaller intercept point than in strong inversion(that is why we have to keep both in strong inversion); *The saturated MOST has a third order intercept voltage proportional with square root of VGT/. Its value is an order of magnitude higher than for a weak inversion MOST, however somewhat smaller than for a triode MOST. However, the advantage of a triode MOST over a saturated MOST is lost for high values of . Now if we want to discuss about the mixer as a 2VCCS circuit we can benefit of some conclusions from {V,V} Class and {V,I} Class from Erics book: For {V,V} class allows linearity improvement in at least two ways: a) even order distortion cancellation for equal VCCSs(sfrom square model of the transconductor in our case); the odd terms are doubled(balancing) in this case, so that the odd order distortion remain the same. b) reduce the coefficients g1,g2 and g3 to the difference of g1,g2 and g3 of the 2VCCSs involved; if g2 and/or g3
VGT > 10 100mV

21

are relatively bias independent, whereas g1 strongly varies with bias, this can be advantageous; a general disadvantage of this technique is its sensitivity for mismatch and noise , because the subtraction of signals. For {V,I} class compared to a single VCCS, a reduction in distortion is often possible with a 2VCCS circuit, if the input signal is divided over two VCCSs(our two transconductors); if each VCCS experience half of the input current(we are in the case even so...the two transconductors share equal currents and voltages), this leads to a double 2nd and 3rd order intercept point or current. That is almost all with our story, because about port to port isolation is only one thing to say: smaller parasitic capacitance and take care that the Cgs capacitance of the transconductors to be the dominating one, but if you walk on the above predicted ways, you will be still safe. So you will trade this parameter with the size of the transistors you will use, and somehow also with bias. The discussion about Cp is useful also for this case, because more LO-IF feed through means at the output more noise(flicker noise and not only).
O.K. now is quite the moment to question our self is this paper is good of something... if we did not spread some toner on a paper for testing a printer or showing to people how fast and accurate we can write an internal report... Oh, by the way... When you start designing pay attention that linearity can be affected by how many transistors you put fold for realizing a certain W.

You might also like