University of Padova
Department of Information Engineering
Amplifier Design for a Pipeline ADC in 90nm Technology
SUPERVISOR: Prof. Andrea Gerosa CANDIDATE: Alessandro Michielin
Padova, 5th October 2010
I would like to thank my family for the constant support, prof. Gerosa for the guidance throughout the whole Master Degree, and the nSilition staff; in particular Thierry and Laurent for their continuous technical and logistic support on my internship period.
1.2 Thesis organization
2.0 Ideal A/D Converter
2.1 Pipeline ADC
2.3 MDAC operation
2.4 Pipelined ADC Performance Characteristics
2.5 Double sampling tecnique
2.6 Derivation of the OperationalAmplifier Parameters
2.6.1 Open loop DC-Gain
2.6.2 Gain Bandwidth
2.6.3 Slew Rate
2.6.4 Noise considerations
3.1 Current losses
3.3 Surface scattering
3.4 Velocity saturation
3.5 Impact ionization
3.6 Hot electrons
3.7 The modification of the threshold voltage due to SC Effects
- 1 -
3.8 Output conductance reduction
3.8.1 Lightly doped drain implant
3.8.2 Halo Implant
3.8.3 Anti Punchthrough Implant
4.1.1 Fifth stage amplifier: differential
4.1.2 Folded cascode architecture
4.1.3 Cascode structure
4.1.4 Double input pair
4.1.5 Biasing strategy
4.1.6 Power down switches
4.2.0 Stage 3 amplifier
4.2.1 Telescopic cascode
4.2.2 Output stage
4.2.3 Two stages amplifier compensation
5.1 Stage 5 common mode regulator topologies
5.1.1 Inverter based comparator
5.1.2 Current based comparison
5.1.3 Voltage buffers comparison
5.2 Stage 5 common mode rejection
5.3 Stage 3 commmon mode feedback regulator
6.2.1 Stage 5 op-amp AC behaviour
6.2.2 Stage 5 CM behaviour
6.2.3 Max dynamic configuration
- 2 -
6.2.6 Start-up and switch down
6.2.6 INL simulation
6.3.0 Stage 3 characterization
6.3.1 AC behaviour
6.3.2 CM behaviour
6.3.3 Max dynamic configuration
7.1 Main sources of variations
7.2 Interdigit structure
7.3 Antenna effect and antenna diodes
7.4 Dummy transistors
7.6 General consideration for layouting
8.1 Future Work
- 3 -
- 4 -
Abstract This paper explains the choices taken for the design of two full differential operational amplifiers. These op amp have been designed for the third and the fifth stage of a pipelined A/D Converter. It shows also the solutions found to reach high gain, wide bandwidth and short settling time, without degrading too much the output swing. First the operational amplifier specification are extracted starting from the ADC architecture, then the issues related to the sub-micrometrical design are analysed; the different structures tested are then presented and the motivation of the final topology choice are shown. It presents then the op amp schematic implementation, the simulation results and the layout with the 90nm TSMC design kit.
- 5 -
- 6 -
1.0 Introduction Operational Amplifiers are one of the most widely used building blocks for analog and mixed-signal
technology has become dominant over bipolar technology for analog circuit design in a mixed-signal system due to the industry trend of applying standard process technologies to implement both analog circuits and digital circuits on the same chip. While many digital circuits can be adapted to a smaller device level with a smaller power supply, most existing analog circuitry requires considerable change or even a redesign to accomplish the same feat. With transistor length being scaled down to tens of nanometers, analog circuits are becoming increasingly more difficult to improve upon, in fact, if small geometries can improve speed decreasing the parasitic capacitance, the gain can be heavily affected. So, gain enhancement techniques are required, but these methods often require more complicated circuit structures and higher power supply voltage, and may produce a limited output voltage swing or introduce a significant noise contribution. This thesis summarizes the work produced during a six months stage by nSilition sprl, a fabless company specialized in the design of high performances, low power converters. The design object was a 14bit, 200MS/s ADC. The converter is - 7 -
implemented with a six stages pipeline architecture; the design is based on switchcapacitor circuitry. Each stage consists of an OTA and a subADC, and stage 5 and stage 3 OTA are the main objects of this dissertation. The devices are implemented through TSMC 90nmRF process technology. Analog circuit design requires a good understanding of how the system and circuit work. Unlike digital circuitry which works with two distinct states, many parameters are under consideration for analog circuits which work with continuous values. Due to the multi-dimensional variables of an analog circuit, any slight change in the analog configuration like current, voltage, a transistor parameter, a device model, a manufacturing process, or a modified layout may cause significantly different performance. For analog design engineers, a good design methodology including intuition, mathematical methods, and specialized tools are assets. The
consists on Virtuoso Front to Back Design Environment for the schematics and layout, Matlab and Excel for the specifications extraction. All the specifications required have been met.
1.2 Thesis organization The thesis is organized into eight chapters.
Chapter 1 introduces the problem.
Chapter 2 reviews the basic theory of A/D converters and the principle of the pipelining; the main characteristics of the converter of the project are described as well as the methodology used for the extraction of the amplifier specifications in each stage.
Chapter 3 describes the main side effects related to the use of short channel devices and how they will affect the schematic modelization and the layout. - 8 -
In Chapter 4 the design of two differential amplifiers is discussed. The different trade-off between gain, bandwidth and stability are presented, and the chosen solutions are explained.
In Chapter 5 the designs of the common mode regulators are discussed. Several different architectures are presented, the one chosen is described as well as the modifications applied to the main amplifier to reach the stability specifications.
Chapter 6 deals on the simulation sets: the testbenches are presented as well as the simulation results.
In Chapter 7 the layout work is shown. The main source of issues are presented as well as the solution chosen. The whole amplifier layout is shown.
Chapter 8 analyses the power consumption and describes the possible future works
- 9 -
- 10 -
Chapter 2: pipeline converters
2.0 Ideal A/D Converter An analog-to-digital converter performs the quantization of analog signals into a number of amplitude-discrete levels at discrete time points. A basic block diagram of an A/D converter is shown in Fig. 2.1.
Fig 2.1: basic A/D converter A sample-and-hold amplifier is added to the input to sample the analog input and to hold the signal information at the sampled value during the time needed for the conversion into a digital number. The analog input value V IN is converted into an N-bit digital value using the equation N −1 V in =Dout eq = ∑ B m 2 me q V ref m =0
In the equation, Rref represents a reference value, which may be a reference - 11 -
voltage, current or charge. BN−1 is the most significant bit and B 0 is the least significant bit of the converter. The quantization error e q represents the difference between the analog input signal Vin divided by Rref and the quantized digital signal Dout when a finite number of quantization levels is used. Eq. 2.1 can be partly rewritten as N −1
Dout = ∑ Bm 2m
The sampling operation of analog signals introduces a repetition of input signal spectra at the sampling frequency and multiples of the sampling frequency. To avoid aliasing of the spectra, the input bandwidth must be limited to not more than half the sampling frequency (Nyquist criterion). 2.1 Pipeline ADC The pipelined is a popular architecture for modern applications of analog-to-digital converters due to its high sustained sampling rate, low power consumption, and linear scaling of complexity. Figure 2.2 shows a block diagram of a pipelined ADC. The term “pipelined” refers to the stage-by-stage processing of an input sample VIN.
Fig 2.2: basic pipeline ADC architecture In the above diagram, the analog input voltage VIN enters the ADC. Each subsequent pipeline stage of the ADC resolves a certain n number of bits to be contributed to the final conversion output. The number of bits that each stage is responsible for quantizing is usually on the order of 1–5 bits. Simultaneously, after each stage has finished quantizing its input sample to n bits, it outputs an analog residue voltage that serves as the input to the next stage. After s stages of conversion, an m-bit ADC resolves the lower bits of the overall ADC digital output. Each stage’s digital decision is then passed to a digital block that properly timealigns the output bits and corrects for any errors in each stage. The final digital decision - 12 -
Chapter 2: pipeline converters is then produced. 2.3 MDAC operation Each stage displayed in the block diagram shown above can be explored further. A typical pipeline stage is displayed in Fig. 2.3.
Fig 2.3: MDAC architecture The input voltage is sampled and held in the sample-and-hold circuit embedded in each stage. Subsequently, an n-bit flash ADC quantizes the analog voltage and produces a digital decision of n bits. The digital decision is then fed through an n-bit flash DAC to be re-converted into an analog signal. The summation node presented in the above diagram takes the input voltage from the sample-and-hold circuit and subtracts the DAC voltage from it. This difference voltage is then fed through a gain stage with gain G to produce the residue voltage, the output voltage of this stage. In a typical pipelined ADC implementation, like the one under design, the sample-and-hold circuit and flash DAC are implemented in a single switched-capacitor circuit called a multiplying DAC, or MDAC. The amplification of the residue usually occurs with a closed-loop operational amplifier. In equation form, the output of each pipeline stage can be described as:
Vres = G(V in – DVres)
The residue voltage, VRES, becomes the input voltage to the next stage. The digital decisions versus input voltage and the residues versus input voltage of a typical pipelined ADC are displayed in Fig 2.4.
- 13 -
Fig 2.4: output of a MDAC In Fig 2.4 the input voltage is swept through the whole operative range. The residue represents the amplified remainder from the subtraction of the DAC output voltage from the stage input voltage. The pipelined ADC theory of operation is that each stage is responsible for quantizing a certain set of bits that will eventually become integrated into the final conversion output. For the generalized pipeline ADC described previously, each stage is responsible for quantizing n bits of the input sample. The final ADC output consists of a weighted sum of each stage’s digital decision. The weightings are determined by the interstage gains, or the gains of the residue amplifiers within each stage. The final output is weighted according to: x=
D N s−1 D0 D1 ... G0 G0G1 G 0 G 1 ... G N −1
where D(i) and Gi represent the digital decision and the residue amplifier gain of each pipeline stage. The above equation suggests that later stages have a smaller weight in the final ADC output. This is indeed the case, as later stages resolve the lower bits of the overall conversion. In the above example, D(Ns-1) represents the digital decision made by the final flash ADC, responsible for resolving the least significant bits of the output. As mentioned before, each stage in a generalized pipelined ADC is responsible for resolving n bits of the ADC output, while the final flash ADC is responsible for - 14 -
Chapter 2: pipeline converters quantizing the m least significant bits of the ADC output. It is evident from the serialized operation of the pipelined ADC that some sort of time-alignment and errorcorrection circuitry is required for aligning each stage’s digital decision to produce the final output. 2.4 Pipelined ADC Performance Characteristics In general, the pipeline architecture enables the implementation of relatively highresolution ADCs without sacrificing processing speed or power draw. Additionally, the linear complexity scaling inherent to the pipeline architecture makes the implementation of higher-resolution pipeline ADCs more manageable than with another ADC architecture. The architecture of the pipelined ADC enables it to have a high throughput rate. This is evident in that pipelined ADCs can have sampling rates of a few MSps up to 200Msps, like the device discussed here. The reasoning for this is that the sample-andhold circuit can begin processing the next analog input voltage sample as soon as the DAC, summation node, and gain amplifier have finished processing the previous sample. This pipelining action allows a high sustained sampling rate. Additionally, since each stage is only responsible for quantizing a low number of bits relative to the overall resolution of the pipeline ADC, each stage processes each sample relatively quickly. The architecture of the pipelined ADC also allows it to scale linearly as complexity increases. In the generalized pipeline ADC discussed earlier, each stage has a small flash ADC that performs the quantization of the input sample. These flash ADCs are comprised of many comparators that are responsible for quantizing the sample. For an n-bit flash ADC, 2n comparators are needed to perform the conversion. In a pipeline ADC, higher overall resolution is obtained effectively by adding additional small flash ADCs in the form of having more stages. A 14-bit pipeline ADC with 6 stages, 2.8 bits per stage, is implemented using only 42 comparators. This is in stark contrast to a 14-bit pure flash ADC, which would require 214 = 16384 comparators in order to quantize the sample. The complexity in a pipeline ADC scales linearly and not exponentially, as is the case in a flash ADC. It also follows that fewer required comparators translates to much less power dissipation and power draw, another advantage of the pipeline architecture.
- 15 -
Although the pipeline ADC allows for high speed, lower power dissipation, and low complexity, there are still tradeoffs. For instance, the serialized nature of the conversion process means that there is a significant time delay between the sample that enters the first sample and hold of the first stage and when the digital alignment circuitry produces the correct output code. Each stage in a pipeline ADC delays the data output by approximately one additional clock cycle. This data latency has to be accounted for when implementing a pipelined ADC. Even in spite of these tradeoffs, the pipelined ADC architecture enables an ADC to have relatively high resolution, high speed, and low power dissipation, all with very few tradeoffs. 2.5 Double sampling tecnique The property of the successive ADC stages working in opposite clock phases can be exploited by sharing the operational amplifier, the comparators and the all the logic part between two parallel component ADCs. This approach uses the double-sampling concept of switched capacitors circuits. By using this technique, the equivalent sampling rate is doubled, but still the power dissipation remains almost the same as for an ADC having traditional single sampled pipeline stages with a half sample rate. The area can be reduced up to 40%. In contrast, the complexity of the pipeline stage is increased and more clock signals with different phases are needed. Scheme of the double sampling multiplying D/A converter is shown in fig 2.5.
- 16 -
Chapter 2: pipeline converters
Fig 2.5: double-sampling MDAC architecture The capacitors of two parallel channels working on opposite clock phases share the same amplifier. While the pipeline1 samples the Vin1 signal onto the Cs and Cf capacitors independently of the amplifier, the pipeline2 switches to the amplification phase. Two important side effects are caused by the amplifier sharing. First, the amplifier load capacitance is increased and affects its bandwidth. Second, the amplifier input offset is never reset; this can be tolerated by an adequate amplifier open loop DCgain. The second one is not so critical in this design because of the differential architecture used, and thereby a symmetric compensation is possible.
2.6 Derivation of the OperationalAmplifier Parameters To calculate the DC-gain of the amplifier in a multiplying D/A converter it is necessary to deal with the resolution; instead, for the slew rate and GBW specifications, the sampling speed of the A/D converter is the key parameter.
- 17 -
Fig 2.6: hold mode MDAC The topology presented in fig 2.6 is assumed to be in the hold mode as single-ended for simplicity. However, all the calculations are performed for a fully differential topology. In this configuration, the input signal is sampled to the sampling capacitors n−1
C s =∑ C s , j
and feedback capacitor Cf . 2.6.1 Open loop DC-Gain The settling error at the output of the operational amplifier in a multiplying D/A converter, resulted from the finite open loop DC gain A0 = gmro is approximately given by
where f is the feedback factor f=
C f ∑ C s , j C par
which can be approximated in case of C s , C f ≫C i n to equal f≈
Assuming that the errors εA0,i, caused by the finite DC-gain in all the m = k-1 stages with a resolution of Bi + r bits, are the only error sources, the total input error of a N-bit A/D converter is
- 18 -
Chapter 2: pipeline converters m
tot =∑ i=1
Applying some substitutions, the total error at the input can be rewritten m
tot =∑ i=1
A 0, i⋅∏ 2 B
where A0,i is the open loop DC-gain of the amplifier in the ith stage. At the same time, the total error at the ADC input must be less than LSB/2, which corresponds to εtot < 1/2N for an N-bit ADC. The inequality for the dimensioning of the amplifier open loop DCgains becomes in general case m
A0, i⋅∏ 2 B
2.6.2 Gain Bandwidth The successive pipeline stages operate in opposite clock phases, which gives a settling time of a half of the clock cycle (T/2). The settling time is determined first by the slew rate (SR) and finally by the gain bandwidth of the amplifier, as indicated in Fig. 2.7. Again, the MDAC topology of Fig. 2.6(a) is considered as fully differential.
Fig 2.7: settling of the output The most commonly used OTAs can be modeled with a single-pole small-signal model of Fig. 2.6(b). The GBW frequency of an OTA is related to the transconductance gm by - 19 -
equation GBW =
gm 2πC L, tot
where the total load capacitance CL;tot = CL + Cout includes the parasitic output capacitance Cout. Using the symbols of Fig. 2.6(b), the corner frequency for the settling in the hold mode is ω−3dB=
gm gm ⋅f = n−1 C L, H C L ,tot ∑ C s , jC par f j=0
It was arbitrarily chosen to reserve one third of the settling time for the SR limited part and two thirds for the GBW limited exponential settling. The error ε t caused by the incomplete exponential settling during T/3 = 1/(3fS) is given by −ω−3dB⋅1 3f s
ε τ =e
gm n−1 C L ,tot ∑ C s , j C par 3f s f j=0
In order to fulfill the resolution requirement, the settling error must be less than LSB/2, this case reduced to the input of the stage i, which results in a condition ε τ ,i
where Ni is the resolution of the remaining back-end pipeline including the ith stage. By combining Eq 2.8, 2.14, and 2.15, and solving the amplifier transconductance gm yields Bi
gm3ln2⋅2 ⋅N i⋅f S⋅kC L;tot
where the constant k>1 is the ratio between the effective load capacitance in the feedback configuration CL;H and in open loop CL;tot, resulting in N−1
C L, H =1 C L ,tot
C f ∑ C s , jC par j =0
C LCout C f ∑ C s , jC par j=0
On the other hand, the transconductance is related to the width W, length L, and drain current ID of the transistor by
gm= 2μC ox
W I L D
where μ is the mobility and Cox the gate oxide capacitance. By substituting Eq. 2.18 into
- 20 -
Chapter 2: pipeline converters Eq. 2.16, a condition for the minimum drain current of one transistor of the amplifier input differential pair ID can be derived, and this can be expressed in terms of the minimal gain bandwidth
∑ C s , j C par 3ln2 B GBW ⋅N i f s⋅ 2 j=0 2π C L, tot i
An interesting special case occurs when all the pipeline stages are identical, having Bi=B with equal correlated settling errors εt;i = εt being the only error sources. The total error reduced to the input of an N bit pipeline ADC is given by 1 1− mB 2 ε tot =ε τ⋅ B 2 −1
Again, for an N bit ADC it must hold that εtot < 1/2N . By combining this to Eqs. 2.8, 2.14 , and 2.20 , for the transconductance gm holds B
gm3 Nln2−ln2 B−1⋅2 ⋅f S⋅kC L; tot
2.6.3 Slew Rate The slew rate of a single-stage OTA, like a folded cascode amplifier, is linearly dependent on the maximal current Imax charging and discharging the load capacitance. To assure symmetrical slewing of the output, the currents of the output stages have to be equal to the current of the input stage, which indicates Imax = 2ID. In a pipeline stage, the load capacitance during the slewing depends on the capacitor charging in the previous operation phase. In the worst case, the total load capacitance is CL;tot+Cf. Using the symbols of Fig. 2.6(b), the slew rate is given in this case by SR=
I max 2I D = C L Cout C f C L, tot C f
For a worst-case slewing of the differential full-scale voltage Vpp;diff, the SR limited part being one third of the settling time, holds the condition T ⋅SRV FS ,diff 6
Substituting Eq. 2.22 into the inequality of Eq. 2.23, the minimum drain current set by the slew rate is given by ID3 f S⋅V FS ; diff⋅C L;tot Cf - 21 -
2.6.4 Noise considerations The errors in each stage come from different sources: •
the finite gain of the amplifier;
the incomplete settling;
the mismatch in the capacitances or in the transistors.
It can be demonstrated that the last source is not critical for a stage number lower than 6; moreover the mismatch in the capacitor can be corrected by calibration. Furthermore, considering an equal contribution from all the sources, the allowed error increase from stage to stage by a factor equal to the interstage gain; this implies that the largest contribution will come from the first stages. So the capacitors can be scaled down from stage to stage with a factor equal to the square of the interstage gain, a relation coming from the area dependency model: σ2
A 2 A2 C ΔC = C = C ox , C WL C
down to a minimum dictated by other constraints.
Fig 2.7: noise through the pipeline Concerning the input referred noise V
2 n ,inref
V 2n 0 G20
V 2n 1 G20 G21
V 2n N s−1 G20 G21 ... G 2N −1
where Ns is the number of stages and Gx is the residual gain of stage x. The thermal noise comes mainly from the switches and the amplifiers. It can be modelized in a MDAC as following:
- 22 -
Chapter 2: pipeline converters
Fig 2.8: noise source modelization Considering: G=
C MDAC C0
1 αg m
with α a design variable, Req ≈ B eq≈
π gm 2 2πCeq
C eq=C MDAC C L
C MDAC C L C0
it is possible to express the noise at the output as: V 2n , outref =
kT C MDAC
4kTR on2 4kTR eq ⋅Beq⋅G 2
where the first addend represents the sampled and held contribution, the first addend inside brackets is the broadband contribution from the switches during the amplification phase, and the second addend inside brackets is the contribution coming from the amplifier.
- 23 -
- 24 -
Chapter 3: Design issues in short channel devices
Design issues in short channel devices
3.0 Introduction As the technology scales beyond 100-nm sizes, the traditional design approach needs to be modified to take into account the increased process variation, interconnect processing difficulties, and other physical effects. It can be experienced a significant increase in gate tunnelling current, due to the thin oxide. Subthreshold leakage and gate tunnelling are no longer second-order effects. If these effects are not taken care of, the result will be a dysfunctional devices, especially for digital circuits, but also the analog environment will be heavily changed. Typically, processor designers tape out their design when the verification confidence level is high enough. Debug continues on silicon, which is usually several orders of magnitude faster and would result in getting a product to market sooner. Now, due to the increased mask cost and longer fabrication turnaround time, the trade-off to arrive at the most cost-effective product and shortest time to market will certainly be different . The transistor figure of merit is now deviating from the reciprocal of the gate length. Furthermore, global wiring is not scaling, whereas wire resistance below 0.1 μm is increasing exponentially. This is due primarily to surface scattering and grain-size - 25 -
limitations in a narrow trench, resulting in carrier scattering and mobility degradation. The gate dielectric thickness is approaching atomic dimensions and at 1.2 nm in the 90nm node is about five atomic layers of oxide. Source–drain extension resistance is getting to be a larger proportion of the transistor “on” resistance. Source–drain extension doping has been increased significantly, and the ability to reduce this resistance has to be traded off with other short-channel effects, such as hot-carrier injections and leakage current due to band-to-band tunnelling. Source–drain diffusions are getting so thin that implants are at the saturation level and resistance can no longer be reduced unless additional dopants can be activated. The main effects related to the reduced dimensions of the devices are the following: -the current losses -the mobility degradation -the threshold shift -the gate capacitance shift -the gds degradation
3.1 Current losses The device under design has no consumption particularly stringent specification; of course power must be minimized, but, since it is not the main goal, the leakage in devices like the amplifier can be neglected (it is not the case instead for the switches, the bootstrap or other other circuital elements, but, since it is not object of this dissertation, the problem will be ignored). The mail losses are related on: -gate tunnelling through the oxide -junction losses -hot carrier current -gate induced drain leakage 3.1.1. Tunnelling
- 26 -
Chapter 3: Design issues in short channel devices The electron wave function, for oxides below 8-10nm, can spread up to the anode; in particular, depending on the electric field, several cases are possible: a) Eox > φB / tox : the voltage drop is higher than the barrier. In this case, since the tunnelling is an energy conservative process, the electron sees a triangular potential barrier, having an effective depth of teff = φB / Eox, as depicted in fig 3.1.
Fig 3.1: Fowler-Nordheim tunnelling This process, named Fowler-Nordheim Tunnel, implies that the electron reaches the conduction band of the oxide and crosses it up to the anode, following the relation: J FN = A⋅E 2OX⋅exp
where B a constant related on the barrier height and on the type of oxide. b) Eox < φB / tox : the voltage drop is lower than the barrier height and the electron sees a trapezoidal barrier; it jumps directly from the conduction band of the cathode to the one of the anode, so this process is named Direct Tunnel.
Fig 3.2: Fowler-Nordheim tunnelling The expression is similar to the Fowler-Nordheim case: 3 /2
q⋅EOX⋅t OX −B J D≃ A⋅E ⋅exp ⋅1−1− EOX φB 2 OX
c) the Hole Valence Band tunnelling, where the process is similar to the previous cases - 27 -
but involving holes instead of electrons; although it is less probable, since the barrier is intrinsically higher (5eV). d) the Electron Valence Band tunnelling, where the hypothesis is that the electron has an energy at the same level of free energy level in the anode. The gate voltage required is around 1.5V, but, when an electron crosses the oxide, it sees a barrier of about 4.2eV, so this happens only for an oxide thickness lower than 2-3nm.
Fig 3.3: different tunnelling e) the Band-to-band tunnelling; when the doping is high, the charge space region drops: if also the potential between source and drain is high, then the band bending is strong. This means that the electrons from the cathode see a triangular voltage barrier Eg high; so the electron can jump in the conduction band by tunnelling. The expression is similar to the one for the Fowler-Nordheim; the difference is in the carrier concentration and in the state function density: A⋅E⋅V d −β⋅E 3/g 2 J b−b = ⋅exp E Eg
3.1.2. GIDL It happens typically when the gate is grounded and the drain is at Vdd. The MOS is off, there is no channel and the substrate is in accumulation. All the silicon surface is in accumulation, so it behaves like a high doped p semiconductor, where the Fermi level is close to the valence band. Then, at the Si/SiO 2 interface, it forms a p+/n/n+ junction which can create a band-to-band tunnel leakage.
- 28 -
Chapter 3: Design issues in short channel devices
Fig 3.4: gate induced drain lowering 3.1.3 DIBL The current flow in the channel depends on creating and sustaining an inversion layer on the surface. If the gate bias voltage is not sufficient to invert the surface (VGS
Fig 3.5: Drain-induced barrier lowering 3.2 Punchtrough The expressions for the drain and source junction widths are:
- 29 -
2εSi V DSφ Si V SB qN a
and x dS=
2ε Si φ Si V DB qN a
where VSB and VDB are source-to-body and drain-to-body voltages. When the depletion regions surrounding the drain extends to the source, so that the two depletion layers merge (i.e., when xdS + xdD = L), punchtrough occurs. Punchthrough can be minimized with thinner oxides, larger substrate doping, shallower junctions, and obviously with longer channels. It can be a destructive effect, so it must be strictly avoided.
Fig 3.6: punchtrough 3.3 Surface scattering When the channel length becomes smaller due to the lateral extension of the depletion layer into the channel region, the longitudinal electric field component ey increases, and the surface mobility becomes field-dependent. Since the carrier transport in a MOSFET is confined within the narrow inversion layer, and the surface scattering (that is the collisions suffered by the electrons that are accelerated toward the interface by ex) causes reduction of the mobility, the electrons move with great difficulty parallel to the interface, so that the average surface mobility, even for small values of ey, is about half as much as that of the bulk mobility.
- 30 -
Chapter 3: Design issues in short channel devices
Fig 3.7: MOS cross section showing the field contributions
3.4 Velocity saturation The performance short-channel devices is also affected by velocity saturation, which reduces the transconductance in the saturation mode. At low ey, the electron drift velocity vde in the channel varies linearly with the electric field intensity. It can be noted that the drain current is limited by this effect instead of pinchoff. This occurs in shortchannel devices when the dimensions are scaled without lowering the bias voltages. Using vde(sat), the maximum gain possible for a MOSFET can be defined as gm = WcoxVde(sat)
3.5 Impact ionization Another undesirable short-channel effect, especially in NMOS, occurs due to the high velocity of electrons in presence of high longitudinal fields that can generate electronhole pairs by impact ionization, that is, by impacting on silicon atoms and ionizing them. It happens as follow: normally, most of the electrons are attracted by the drain, while the holes enter the substrate to form part of the parasitic substrate current. Moreover, the region between the source and the drain can act like the base of an npn transistor, with the source playing the role of the emitter and the drain that of the - 31 -
collector. If the holes are collected by the source, and the corresponding hole current creates a voltage drop in the substrate material of the order of 0.6V, the normally reversed-biased substrate-source pn junction will conduct appreciably. Then electrons can be injected from the source to the substrate, similar to the injection of electrons from the emitter to the base. They can gain enough energy as they travel toward the drain to create new e-h pairs. The situation can worsen if some electrons generated due to high fields escape the drain field to travel into the substrate, thereby affecting other devices on a chip.
3.6 Hot electrons Another problem, related to high electric fields, is caused by so-called hot electrons. This high energy electrons can enter the oxide, where they can be trapped, giving rise to oxide charging that can accumulate with time and degrade the device performance by increasing VT and affect adversely the gate’s control on the drain current.
Fig 3.8: hot-electron damages 3.7 The modification of the threshold voltage due to Short-Channel Effects (SCE) The equation giving the threshold voltage at zero-bias V T0 =V FB 2φ F
qD I 1 2q⋅ε Si⋅N A 2φF Cox C ox
is accurate in describing large MOS transistors, but it collapses when applied to smallgeometry MOSFET. In fact that equation assumes that the bulk depletion charge is only due to the electric field created by the gate voltage, while the depletion charge near n+
- 32 -
Chapter 3: Design issues in short channel devices source and drain region is actually induced by pn junction band bending. Therefore, the amount of bulk charge the gate voltage supports is overestimated, leading to a larger VT than the actual value. The electric flux lines generated by the charge on the MOS capacitor gate electrode terminate on the induced mobile carriers in the depletion region just under the gate. For short-channel MOSFET, on the other hand, some of the field lines originating from the source and the drain electrodes terminate on charges in the channel region. Thus, less gate voltage is required to cause inversion. This implies that the fraction of the bulk depletion charge originating from the pn junction depletion and hence requiring no gate voltage, must be subtracted from the VT expression.
Fig 3.9: gate-induced bulk depletion region The figure shows the simplified geometry of the gate-induced bulk depletion region and the p-n junction depletion regions in a short channel MOS transistor. It can be noted that the bulk depletion region is assumed to have and asymmetric trapezoidal shape, instead of a rectangular shape, to represent accurately the gate-induced charge. The drain depletion region is expected to be larger than the source depletion region because the positive drain-to-source voltage reversed-biases the drain-substrate junction. We recognize that a significant portion of the total depletion region charge under the gate is actually due to the source and drain junction depletion, rather than the bulk depletion induced by the gate voltage. Since the bulk depletion charge in the short channel device is smaller than expected, the threshold voltage expression must be modified to account
- 33 -
for this reduction: VT0short ch = Vt0 – ΔVt0 where VT0 is the zero-bias threshold voltage calculated using the conventional longchannel formula and ΔVT0 is the threshold voltage shift (reduction) due to the shortchannel effect. The reduction term actually represents the amount of charge differential between a rectangular depletion region and a trapezoidal depletion region. Let ΔLS and ΔLD represent the lateral extent of the depletion regions associated with the source junction and the drain junction, respectively. Then, the bulk depletion region charge contained within the trapezoidal region is:
Q B0=− 1−
LS L D 2L
4q Si N A f
To calculate ΔLS and ΔLD, the simplified geometry shown in the figure can be useful.
Fig 3.10: geometry of the depletion region Here, xdS and xdD represent the depth of the pn-junction depletion regions associated with the source and the drain, respectively. The edges of the source and drain diffusion regions are represented by quarter-circular arcs, each with a radius equal to the junction depth, xj. The vertical extent of the bulk depletion region into the substrate is represented by xdm. The junction depletion region depths can be approximated by x dD =
2 Si V DS 0 qN A
- 34 -
Chapter 3: Design issues in short channel devices x dS =
2 Si 0 qN A
with the junction built-in voltage 0 =
NDNA kT ln q n2i
From figure, L D≈ x j
2xdD −1 xj
Similarly, the length ΔLS can also be found as follows: L S≈ x j
2x dS −1 xj
Now, the amount of the threshold voltage reduction ΔVT0 due to short-channel effects can be found as: V T0 =
x 1 ⋅ 4q Si N A F⋅ j ⋅ C ox 2L
2x dD −1 xj
2xdS −1 xj
The threshold voltage shift term is proportional to xj/L. As a result, this term becomes more prominent for MOS transistors with shorter channel lengths, and it approaches zero for long channel MOSFET where L >> xj.
3.8 Output conductance reduction High performance logic devices are optimized for good drive current, low leakage and SCE control which incorporate super halo and double-pocket structures. These structures however, often result in low output resistance, device gain, transconductanceto-drive current ratio and matching properties . Low output resistance is the result of increase ID with VD in saturation regime. Three components are associated with this increase, namely channel length modulation (CLM), drain-induced-barrier-lowering (DIBL) and substrate current body effect (SCBE). It can be found that gds is most sensitive to LDD dose, halo dose, halo tilt and APT dose and energy.
- 35 -
3.8.1 Lightly doped drain implant
Fig 3.11: Junction profile for different LDD dose Fig 3.12 shows the sensitivity of VA to LDD dose for a PMOS transistor; VA is defined as ID/gds – VD, the Early voltage. If the LDD dose is increased, VA decreases because the effective channel becomes shorter, as shown in Figure 3.11. Shorter channel length results in larger residual DIBL thus causing output resistance to decrease.
Fig 3.12: VA dependence on PLDD dose 3.8.2 Halo Implant - 36 -
Chapter 3: Design issues in short channel devices The halo is a p+ implant applied in the proximity of the source and drain junctions; its purpose is to reduce the effects of charge sharing, DIBL and punchthrough. Figure 3.13 shows the drawback of this technique: increasing the halo dose increases VA.
Fig 3.13: VA dependence on halo dose Increasing the halo tilt angle also increases V A as shown in the inset: a larger tilt angle places the halo implant almost at the centre of the channel. The additional arsenic in the channel lessens the effect of DIBL as shown in Figure 3.14. However, increasing halo dose also causes Idsat to decrease. Experimental studies have also shown that pocket implant has tradeoff effects on VA and Idsat.
Fig 3.14: Potential barrier shifts at different halo tilt angle
- 37 -
3.8.3 Anti Punchthrough Implant The natural thresholds of the NMOS is about 0V and of the PMOS is about 1.2V. An p implant is used to make the NMOS harder to invert and the PMOS easier resulting in threshold voltages balanced around zero volts.
Fig 3.15: Anti punchthrough implants Also an implant can be applied to create a higher-doped region beneath the channels to prevent punch-through from the drain depletion region extending to source depletion region. This technique is typically named anti punchthrough (APT) implant.
Fig 3.16: Effects of APT implant on VA and Idsat Fig 3.16 shows the plot of Idsat versus VA at different APT implant conditions. As APT energy implant is increased, VA shows a contradict trend depending on the implant dose used. At high dose and high energy, the plot is shifted to the up-left when energy is - 38 -
Chapter 3: Design issues in short channel devices increased. However, the plot shifted to the up-right as energy is increased at low APT implant dose and energy. At low dose and energy, increasing the APT implant energy forms super steep retrograde channel, which has positive effects for analog applications.
- 39 -
- 40 -
Chapter 4: Amplifier design
Differential Amplifier design
4.0 Introduction In general, operational amplifiers are amplifiers with an open loop gain high enough to ensure that the closed loop transfer characteristic with negative feedback is approximately independent of the op amp gain. To ensure wide swing and noise immunity, a fully differential architecture is used; so a differential and a common mode behaviour will be investigated
4.1.1 Fifth stage amplifier: differential The design effort is directed towards the power consumption minimization. Different circuital topologies have been taken in exam: the goal is to reach the following specifications:
- 41 -
Current Capability 0.28mA Table 4.1: OpAmp stage 5 specifications The dynamic range specification implies that at least 0.55 over the 1.2V available must be dedicated to the output swing; this limits the number of devices than can be stacked in the output branches and, consequently, the resistance. At the same time, previous considerations demonstrate that the gmr0 in this technology is low, so it seems to be mandatory to cascode the output. At the state of the art, three different topologies seem to meet the specifications required: -the folded-cascode; -the active-cascode; -the two-stage architecture. Working in an purely intuitive way, if the first met the specifications, it would automatically be better than the second, because the power consumption is mainly related on bandwidth, so the boosters would only increase the current consumption. Finally, assuming that the folded-cascode topology has enough gain, what to do is to choose between a single stage or a two stages architecture. In an industrial environment, where the human effort is a parameter to take into account in a design, the more efficient methodology is probably to design the single stage device, to study the technology limits, then to design a two stage amplifier for a more sophisticated MDAC (for instance the stage 3 MDAC) and finally to choose which one is more indicated for the project purpose.
- 42 -
Chapter 4: Amplifier design 4.1.2 Folded cascode architecture
Fig 4.1: folded cascode architecture The most important advantage of the folded structure lies in the voltage output swing because it does not “stack” the cascode transistor on top of the input device. The lower swing of the output is given by Vmin = Vds,sat3+Vds,sat5,
Vmax = Vdd-(Vds,sat7+Vds,sat9).
and the upper end by Thus the peak to peak swing on each side is therefore: Vswing = Vdd-4*Vds,sat.
Using the half circuit depicted in fig. 4.2(a), and writing that |Av|=GmRout,
it is possible to calculate the equivalent Gm and Rout. As shown in Fig. 4.2(b), the output of the circuit current is approximatively equal to the drain current of M1, as the - 43 -
impedance seen looking into the source of M3 is much lower than Ron1||Ron5.
Fig 4.2: half circuit representation The use of a cascode structure allows to reach a very high impedance seen from the output node. In fact, as previously discussed, in 90nm technology the g ds is relatively high, due to physical (the channel length) and technological (the halo structure) aspects.
4.1.3 Cascode structure To increase the gain of the CMOS stage, the transconductance of the stage can be improved or the output resistance can be enhanced. The output resistance increases in proportion to a decrease in bias current as shown in Eq. 4.5
r ds =
1 λI DP
where IDP is the drain pinchoff current and λ is the channel length modulation factor. Instead the transconductance increases as the square root of the increase in bias current in a relation that can be simplified by the following: g m=
∂i D = 2μC ox W /L ID ∂ v gs
It is power efficient to increase the output resistance by lowering the bias current. - 44 -
Chapter 4: Amplifier design Fig. 4.3 shows a single stage amplifier using a conventional cascode connection where the common-gate stage device M2, biased by a voltage supply VG2, is added to the input common-source stage M1. VG1, VG2, and Ibias are chosen to make M1 and M2 to operate in their active regions.
Fig. 4.3: cascode structure Assuming the current source Ibias is ideal, the output resistance is rout = rds1 + rds2 + (gm2 + gmb2)rds1rds2.
The midband voltage gain for the circuit of Fig. 4.3 is A0 = −[gm1rds1 + gm1rds1(gm2 + gmb2)rds2],
where gm1 and gm2 are the transconductance of M1 and M2 individually, rds1 and rds2 denote the drain to source resistance of M1 and M2 at the bias point used, and gmb2 represents the transconductance that models the body effect of M2. As indicated by Eq. 4.8, it is clear that the cascode structure can achieve significantly higher voltage gain than a simple MOS stage by providing a higher output resistance. However, this configuration requires that the bias voltage VG2 for M2 be VT + 2Veff . The drain of M2 is set higher than VG2 in order to allow for the voltage swing. The operation of this cascode connection has limitations for low voltage, low power applications due to the bias voltage requirement and limited output swing. To achieve an even higher gain, more cascode devices can be added in the cascode stack connection to form a “triple cascode”. But this further reduces the output swing, so in 1.2V Vdd technology cannot be implemented. - 45 -
Another aspect to take into account is the rds variation related to the V ds reduction: in fact, as depicted in fig4.4, the gds can drop when cascode V ds comes close to the overdrive value. As previously described, this is a consequence of the Halo implant that reduces the effective channel length.
Fig 4.4: effect of the drain-source voltage on the output resistance So the gain of the whole device will be G mRout, and the bandwidth, or better the dominant pole location will be:
ω-3dB = 1/ RoutCout
which allows to calculate the gain bandwidth product: GBW =
g m1 2π⋅C out
4.1.4 Double input pair Before deciding which type of transistors to use as input-pair, several aspects must be taken into account: 1) The electrons mobility is considerably higher than the mobility of the holes. g m, and thereby the gain and unity gain frequency will be higher when using NMOS instead of PMOS-transistors, for the same input capacitance (that is, the W/L ratio). 2) With an NMOS input pair the impedance at the folding points will be lower, due to the intrinsic lower impedance of the PMOS transistors. Both the gain as well as the - 46 -
Chapter 4: Amplifier design phase margin will be lower than if using PMOS at the input, for the same sizes of the transistors. 3) NMOS transistors have lower thermal noise (a parameter related to the gm). 4) PMOS transistors have lower 1/f noise. Since high gain is needed, NMOS transistors at the input has been preferred over PMOS. Also, since correlated double sampling will be used, the 1/f noise is reduced, which also make NMOS a better choice from the point of view of the noise. Another degree of freedom in this circuital topology is the amount of current flowing into the input and the output branches. Typically, once the current capability specification is known, this value is applied at the output, and at the same time at the input to reach the same slewing behaviour when the current is flowing in the two directions relatively the output node. Intuitively, a small amount of current at the output branch will implies good gain and lower bandwidth, since the MOS effective resistance is inversely proportional to this parameter; on the other side, more current in the input pair will consists in increased bandwidth and gain, since gm is proportional to the square root of the current:
g m = 2k
W I L D
Here the bottleneck is the gain, so the choice has been to use the minimum amount of current at the output meeting the current capability specifications and increase the one at the input until the gain and bandwidth requirements were met. This will clearly implies a different behaviour at the output when the signal rises and drops, due to the different current capability available. Simulations demonstrate that this can be not a problem in switching capacitor circuit until the minimum current available is enough to met the slewing specifications. A drawback can be found in the gds of the transistors where both the current for the output and the one for the input pair flow. In fact huge transistors will be necessary to have enough Vds, and since the current is high, also the metal width of the connections will be increased. This implies a high parasitic capacitance which, in - 47 -
parallel with the input pair cascode and the output cascode, determines the secondary pole of the amplifier. This imposes the limit to the maximum over-bias for the input pair. Another aspect to take into account is that, in the nodes of fig 4.5, besides the sum of
Fig 4.5: the bottleneck the parasitic capacitance, there is also the parallel of the resistances of the mirror transistors with the input pair transistors. Since the node between the input pair and the tail current generator is a virtual ground, and since the input pair devices have a very small channel length to minimize the input capacitance maximizing the g m, the relative gds,input_pair will be also high. So several dB of gain can be lost; the solution chosen has been to stack a cascode transistors between the input pair and the node highlighted in fig 4.5. These devices were chosen identical to the input pair devices, mainly to obtain an easier symmetry when layouting, and biased at Vdd so no additional biasing circuitry was needed. An other peculiarity of the topology lies in the presence of two input pairs. Each of them is driven by a switch connected to a clock signal so that when a pair is turned - 48 -
Chapter 4: Amplifier design on, the other is off. In fact verifications by simulations show the presence of an unexpected ΔQ at the input of the amplifier, and further investigations demonstrate that it was due to the charges pumped by the bootstrap circuit and by Saap, Sabp, Saan and Sabn switches.
Fig 4.6: MDAC in double sampling configuration In particular, at the moment these switches close, there is a ΔQ pumped from the ground through the parasitic capacitor, and this, of course, generates an error on the value sampled on the capacitor, and an error on the bottom plate voltage of the capacitors. This will imply an extra error later at the end of the amplification phase. In a single sampling configuration this problem does not exist, because there is no switch in the feedback path of the op-amp.
- 49 -
Fig 4.7: MDAC in single sampling configuration From this consideration comes the solution chosen, which is to remove the critical switches by creating a multiple path. This means that the amplifier will have two input pairs, but the extra stored charge is avoided. This clearly implies the need of extra area and a clock signal, but the power consumption remains nearly unaltered.
- 50 -
Chapter 4: Amplifier design
Fig 4.8: MDAC in double sampling configuration, proposed topology
4.1.5 Biasing strategy In the project it is assumed to have a 50uA current source, generated by a bandgap circuit. This means that this current will be proportional to the supply voltage. What is needed is to bias: – the transistors used as current source; – the cascode transistors. For the first case the 50uA can be simply fed into a device in diode configuration; the voltage at the gate can then be used to bias a scaled version of the same device to reach the current desired.
- 51 -
Fig 4.9: basic current mirror The two devices have the same gate voltage, but the V ds can be different. This implies a mismatch in the current (assuming that the gates have the same dimensions), which can be represented: ΔI = I 2−I 1 ≈
V 2−V 1 ΔV DS ,2 = r0 r0
Fig 4.10: current mismatch due to Early effect To avoid this deviation, since every current mirror is part of a cascode, an easy solution is to use the cascode itself to fix the Vds, as shown in fig 4.11.
- 52 -
Chapter 4: Amplifier design
Fig 4.11: cascoded current mirror Using this strategy the topology has robust current mirrors. What still remains is to bias the cascode transistors. Investigating a particular case, for instance the NMOS cascodes in the output branch, it is clear that the goal is to have a DC voltage high enough to leave M 3 and M1 in saturation, but low enough to leave M2 also in saturation when Vout is at its minimum, or, better, when the main amplifier operates at its full swing. So the desired voltage is: Vgs7 = Vt2 + Vov2 + Vov1
Fig 4.12: cascode transistor biasing In fig 4.12 it is shown a diode connected MOS working as voltage reference; the current is generated by a cascoded current mirror.
- 53 -
Several different solutions have been taken into account; but the diode connected MOS has the advantage that it is simple, small and has a low and adjustable power consumption. Its output node has also a low impedance, which means that the rejection to the supply noise will be good.
4.1.6 Power down switches It is mandatory to introduce in the design the possibility to power down the circuit even if the supply is still connected. It simply consists in adding to some nodes switches to power down the current sources or the biasing of some transistors so that it is impossible for the current to flow in certain conditions. In fig 4.13 the power down switches are highlighted
Fig 4.13: power down switches The strategy rely mainly on: -to stop the current coming from the external current source, obtained applying a switch in series between the current source and the mirror transistor;
- 54 -
Chapter 4: Amplifier design
Fig 4.14: current stopping switch -switch off the current mirrors, by shorting the gate voltage to ground or Vdd;
Fig 4.15: turning off switches for current mirrors
- 55 -
-shorting the current stored in big capacitance node to a power rail
Fig 4.16: discharge for big capacitance nodes The problem arises on the turning on of the circuit. In fact the biasing of the different cascode devices depends on the current generated by the current mirrors, which, at the same time, contain a cascode.
Fig 4.17: start-up for the cascode bias The solution adopted is the one shown in fig 4.17: the circuit guarantees a low current - 56 -
Chapter 4: Amplifier design flowing in the diode connected MOS which bias the cascodes, allowing it to generate some voltage gap which will converge at the desired bias value once Vdd stabilizes. 4.1.7 Noise Every transistor can be considered as a noise source, so it can be modellized as a current source having a spectral current density: KFg 2m
I out =4kTγg m
For simplicity a single pole model can be considered
Fig 4.18: single pole model where Rds3 indicates a cascoded structure. The power spectral density of each device can be referred to the input: V
I 2ds1 f gm 21
I f = 2ds2 2 2 gm 1 gm1 r ds1
I 2ds3 f = gm21
Finally the total input referred noise will be the sum of each contribution: 3
V 2nOTA =∑ V 2njin j=1
- 57 -
4.2.0 Stage 3 amplifier These are the specifications to reach Gain
Some considerations: stage 5 op-amp was already at the limit for the technology given: in fact no more gain was affordable. Also the bandwidth was no freely increasable, since the bottleneck at the discussed mirror transistor would make the gain drop if the current was raised. So a simple single stage amplifier with that kind of architecture can not reach the specifications required. Also a boosted folded-cascode device would be ineffective, mainly for three reasons:
1) the current to have stable boosters can augment the power consumption up to 40%;
2) the doublet pole-zero which can degrade the slew rate; 3) the noise injected at the output node by the boosters, which have nearly the same amplification than the noise emitted by the input pair transistors. The third reason is probably the most important, because noise specification, here but especially in the first and in the second MDAC, is the key parameter, or, better, the most difficult specification to achieve. So it seems the case to use a two stages device. The behaviour concerning the voltage swing is the same as in the previous case: the second stage, which has a topology similar to the output part of stage 5 amplifier, can easily satisfy the requirements, and has the advantage that the gain requirements are less stringent. Then, supposing that the second stage has a gain between 30 up to 40dB, a value easily reachable using a cascode structure, the first stage will need something around
- 58 -
Chapter 4: Amplifier design 20mV of swing. So a folded architecture is no more mandatory for neither of the stages, and this means that the current consumption can be reduced since there is a single path between the supply and the ground for each stage. 4.2.1 Telescopic cascode The first stage is a telescopic cascode OTA, it means that a current source (cascoded, to improve the impedance, as described in the previous chapter) feeds of current an input pair which drives a cascoded load.
Fig 4.19: telescopic cascode architecture The DC output voltage of the first stage is forced by a negative feedback to a value which, fed at the input of the second stage, bias this transistor to a fixed current value. The gain characteristics can be easily extracted by the following: A0 = gmRout
where gm is the transconductance of the input pair transistors and R out is the real part of the impedance seen by the output node, that is Rout = (rds1 + rds3 + (gm3 + gmb3)rds1rds3) || (rds7 + rds5 + (gm5 + gmb5)rds5rds7)
Instead the frequency behaviour depends on the location of the dominant pole, which is - 59 -
1 2π⋅Rout C out
which allows to calculate the gain bandwidth product: GBW =
g m1 2π⋅C out
Fig 4.20: telescopic cascode amplifier schematic As previously described, the input pair is doubled and driven by a switch like in stage 5 amplifier.
4.2.2 Output stage This second stage is a couple of common-source transistors with a cascoded load. The DC gain is, as usual A0 = gmRout
and the dominant pole depends on the load and on Rout f p1=
1 . 2π⋅Rout C out
There is no common path between ground and supply, but the previous stage differential architecture guarantees that the sum of the current in the two output branches remains constant.
- 60 -
Chapter 4: Amplifier design
Fig 4.21: stage 3 amplifier: output stage 4.2.3 Two stages amplifier compensation The main difference between a single stage and a two stages amplifier lies in the stability of the architecture itself. In fact, supposing that each stage has an unique dominant pole, it would be necessary a stage having challenging bandwidth characteristics.
Fig 4.22: Bode plot of a two stages amplifier - 61 -
As depicted in fig 4.21, supposing that each stage has a dominant pole behaviour, the whole system will have a dominant pole at the same frequency of the slowest stage and the secondary one at the frequency of the other pole. So stability in feedback configuration is difficult to achieve, since the two stage have a similar bandwidth, and since each stage secondary poles are not considered. So a compensation is needed; the solution chosen is to use a “Miller” capacitor, which, intuitively, is a capacitor at the output of the first stage which impedance is amplified by the second stage OTA.
Fig 4.23: small signal schematic of a two stages amplifier If Cc was placed in parallel to C1, then the behaviour of the two stages would be the following:
Fig 4.24: parallel compensation of a two stages amplifier - 62 -
Chapter 4: Amplifier design Since stability is a parameter linked to the ratio between the secondary pole and the 0dB frequency, a phase margin improvement is clear from the graph (both axes are logarithmic). Instead, applying Cc in series between the outputs of the two stages, the small signal response becomes:
Fig 4.25: Miller compensation of a two stages amplifier which is clearly better than the previous case, because the secondary pole shifts at higher frequencies leaving the 0dB cross unaltered. Intuitively, what happens is that the effective impedance seen by the first stage is amplified by the second stage; but, at higher frequencies, when the gain of second stage starts to drop at its dominant pole, also the efficient value of the capacitor drops. This is why the first stage seems to have a zero located at the second stage dominant pole. Also the second stage has a variation respect to the parallel case, in fact its bandwidth increases due to the fact that the efficient value of the Miller capacitor is reduced due to the attenuation imposed by the first stage. This effect is called “pole splitting”, since in the small signal plot of the whole amplifier the dominant pole move towards DC and the secondary pole increases its value. An other fact to note is the presence of a zero in the Miller capacitor path. Intuitively, - 63 -
the problem can be analized as follow:
Fig 4.26: small signal representation of a two stages amplifier The two stages gain and bandwidth can be assumed to be very similar: this is mainly due to the fact that the gm is proportional to the square root of the current. This assertion allows to assume that C c is much bigger than C1 and C2, condition necessary to achieve stability. So, for a certain range of frequencies, C 1 and C2 can be seen as open circuits; instead Cc can be assumed as a short, as shown in fig 4.25:
Fig 4.27: middle band representation of a two stages amplifier So the whole second stage can be assumed as a diode connected transistor in parallel to a coscoded load, which means a total impedance of 1/g m2, making the Rds3 contribution uninfluent. From the current point of view, instead, assuming the previous hypothesis, the system can be represented as follows:
- 64 -
Chapter 4: Amplifier design
Fig 4.28: second stage equivalent schematic so the zero can be easily found: ω zero =
where all the current generated by the first stage flow through the second stage transconductance. The problem can be easily solved placing a resistor in series with the Miller capacitor.
Fig 4.29: nulling resistor So: V out =V x⋅1−gm2 Rz −
g m2 sC c
If Rz = 1/gm2, then Vout can never be null for any frequency.
- 65 -
- 66 -
Chapter 5: Common mode feedback design
Common mode feedback design
5.0 Introduction The common mode feedback circuit is a critical component in this kind of architecture: in fact what is needed is a device having
1) a wider bandwidth than the amplifier itself; 2) a low gain to achieve easily the stability; 3) an input dynamic range equal to the output range of the amplifier. For high bandwidth structures, it is typically used a switched capacitor structure: it has a negative gain and no problems concerning the input range. But it needs an operative clock that is the double of the one used for the main converter: this means the necessity of a clock generator and can be risky from the noise point of view (it would implies a tone at the double of the operative frequency). So a continuous time feedback regulator is mandatory.
5.1 Stage 5 common mode regulator topologies The input dynamic range is probably the most critical specification for this device in low voltage technology, so it has been the first taken into account. Different architectures have been investigated.
- 67 -
5.1.1 Inverter based comparator The first topology considered consists in two inverters having two degeneration resistors, as shown in fig
Fig 5.1: inverter based comparison This structure has a wide input range and a gain adjustable by changing the resistors value (which are made using transistors); the current consumption is instead related to the inverters sizes. The voltage value at the central node would be compared by an amplifier to the one produced by the same structure biased at AGND. The drawback of this topology, however, is that it has only two degrees of freedom, the W/L ratio and the value of the resistors; so, even if it can be good enough for stage 5 amplifier, it seems no possible to use it for the wider bandwidth amplifiers.
5.1.2 Current based comparison Another solution proposed is the one shown in fig 5.2. Here the current generated by the pair connected to the main amplifier output is fed, through a mirror, into a MOS biased at AGND, which is the desired voltage value for the DC output. M 1 and the couple M2+ M2- together have the same effective W/L ratio.
- 68 -
Chapter 5: Common mode feedback design
Fig 5.2: current based comparison The main disadvantage of this architecture is that the current consumption is defined only by the device geometries; in particular, since M2+ e M2- has to be in saturation in the whole swing (that means from 0.325 up to 0.875V), their Vth must be low (at maximum 0.3V): so their overdrive must be high, which means also an high current consumption. Simulations and calculations demonstrates that this architecture is not power efficient. 5.1.3 Voltage buffers comparison A different solution proposed is the use of a buffered Miller amplifier, as depicted in fig 5.3:
Fig 5.3: voltage buffers solution
- 69 -
The buffer is a simple source followerMOS in common drain configuration made using a native transistors. Its purpose is to widen the input range of the amplifier, which would not be enough instead. Native transistors are devices made using a very low doping, or leaving the silicon intrinsic. So they are big devices, but with the advantage that the V th, can be very low or eventually negative (see Appendix B). The current can be tuned by the current mirror which bias the native MOS, so the bandwidth over the power consumption can be optimized. A capacitor has also been added to introduce a zero in the common mode path: this can reduce the current flowing through the buffer to achieve the same bandwidth, so its dimensions can be minimum, reducing also the capacitive load for the main amplifier.
Fig 5.4: amplifier and actuator Finally a simple Miller, single ended amplifier is used to regulate the actuator for the common mode feedback: it compares the voltage provided by the sensor to the voltage provided by another sensor, identical to the first one, but biased at AGND. Riassuming, the differential output of the OTA is converted into a common mode voltage by the sensor, compared to the desired value by the amplifier and then fed into
- 70 -
Chapter 5: Common mode feedback design an actuator, which is a simple MOS converting the previous voltage into a current.
Fig 5.5: stage 5 common mode regulator schematic As previously said, the common mode open loop gain must be as low as possible (the lower bound can be the voltage offset for the DC behaviour and the CMR specification for the AC behaviour) to satisfy the stability requirement, so the g m of the actuator must be the lowest possible. The solution chosen is to split the current flowing in the output branches of the main amplifier in two parts: a DC current provided by a mirror, and an other current regulated by the actuator. The ratio of these two values is related on the difference of the current magnitude flowing in the output branches at the full dynamic; in few words, there must be some current flowing in the actuator when the amplifier is completely unbalanced, otherwise the common mode regulation would not be effective anymore. Another advantage coming from this current splitting is linked to the fact that the MOS actuator can have a very small size (but big enough to guarantee saturation), so the load of the Miller amplifier (the comparator block) can be minimum, a fact that permits to reach stability in a easier way.
5.2 Stage 5 common mode rejection As previously said, one of the purposes of the common mode regulator is to attenuate the common mode signals which can affect the behaviour of the fully differential amplifier. But at the same time the gain must be minimized in the loop, to reach stability
- 71 -
without affecting the current consumption. Clearly these two affermations contraddict each other: in fact the common mode rejection is directly proportional to the loop gain. So, to improve the rejection, the solution chosen has been to improve the intrinsic rejection of the main amplifier. By definition, the CMRR is: CMRR = Adm/Acm To calculate the common mode gain in a differential structure it is convenient to split the input pair:
Fig 5.6: common mode signal equivalent schematic So, Acm =
gm ⋅ R∥r 0 [12g m R tail ] 12g m Rtail
which means that the only parameter available to manage the common mode rejecton is Rtail: increasing Rtail, the rejection is improved. This is why the current source for the input pair has been cascoded:
- 72 -
Chapter 5: Common mode feedback design
Fig 5.7: cascoded tail resistance
5.3 Stage 3 commmon mode feedback regulator The architecture used for the blocks of the common mode regulator is the same used for the single stage amplifier: two buffers to sense the common mode voltage, a comparator and an actuator. The main problem rises from the second stage of the main amplifier: there, in fact, the common mode signal is amplified instead of being rejected, since there is no common path for both the positive and negative signals, so there is no intrinsic common mode rejection. In fact, in the first stage, the common mode attenuation will be proportional to the impedance of the current generator which feeds the input pair; instead in the second stage the same signal will have an amplification: - 73 -
ACM = gm2r02
There is also a second aspect to take into account: when the OTA is connected in negative feedback (or, at least, in a configuration detected by the differential signal perceives as a negative loop), there is a positive path for the common mode through the capacitors, as shown in fig 5.8:
Fig 5.8: different common mode loops In fact, through the yellow path in fig 5.8, the signal crosses three inverting devices, the two stages of the amplifier and the CM Miller amplifier. Instead, along the red path, the inverting blocks are only two. This happens because the (negative) amplification introduced by the capacitive path: feedback factor =
Cf C LC f C S
is bigger than the one of the regulator path. In fact, although the Miller amplifier can have a gain of approximately 15dB, the actuator behave like an attenuator: the MOS used is very small, and its load is a mirror current source which, since its bias is in saturation but near the liner region, has a very high g ds. The connection was made there, and not at the output node, because this actuator must act as a current source, and if the load was too high, there would be some Miller effect on the C gd, which would deteriorate the bandwidth behaviour of the common mode loop, already critical.
- 74 -
Chapter 5: Common mode feedback design So a solution could be to increase the gain of the regulator, but it is not possible because of the stability requirement. In fact, to obtain a stable loop and at the same time a CMR high enough for, at least, the whole bandwidth of the ADC (that would be something like 20dB of attenuation for 200MHz), a comparative block having much more bandwidth than the main amplifier is needed (and this means that the common mode regulator would consume approximately as much as the OTA, that is not acceptable). The first improvement is to reduce the common mode gain in the main amplifier. Since it is not possible on the second stage (leaving the architecture as previously designed), it must concern only the first stage. There, as already discussed, the rejection is proportional to the equivalent resistance of the tail transistors: Acm =
gm ⋅ R∥r 0 [12gm Rtail ] 12g m Rtail
A simple solution is to boost the cascode:
Fig 5.9: boosted tail resistance In this way the first stage rejection improves, so the gain of the second stage can be
- 75 -
partially compensated. The booster architecture is the following:
Fig 5.10: booster implementation It is a telescopic cascode single ended architecture which compares the input voltage to a reference value, which is the biasing gate voltage for the cascode transistor. The second improvement is to split the whole common mode loop in two, one for every stage of the amplifier. In this way the total gain of the loop would be the sum of the gains of the two loops, and it is convenient, since gm has a square root relation with current, and a single amplifier would cost much more power.
- 76 -
Chapter 5: Common mode feedback design
Fig 5.11: common mode regulators Both the loops have the same architecture as in stage 5. The first, in particular, instead of taking AGND as the reference parameter, uses a value extracted from a dummy architecture that replies the bias voltage needed by the input of the second stage, as shown in fig 5.12.
Fig 5.12: reference voltage for the first CM loop Of course, the current generator and the diode connected transistor are scaled version of
- 77 -
the effective one to reduce the power consumption.
Fig 5.13: complete common mode regulator schematic
- 78 -
Chapter 6: Amplifier Characterization
6.0 Introduction To verify the proper behaviour of the device, several parameters have been taken into account. The whole characterization set has been performed considering the corners setup provided by the foundry at a temperature compatible with the specifications and at the full span of the voltage supply; in fact variations in fabrication process, ambient temperature and supply voltage affect the electrical performance of the transistors. For example, a higher temperature and a lower supply voltage make the transistor operate slower. This is why the operation of the circuit has been verified by simulating the design in slow (SS) corner, typical corner (TT) and fast corner (FF), and also by simulating the design with fast NMOS and slow PMOS corner (FN), and slow NMOS and fast PMOS corner (SN).
6.1 Reusability In an industrial environment a key point is the reusability of the design and the efficiency in sharing the components between different team members. So it is important to use the same strategy in making the symbols to optimize the work during the different design steps. - 79 -
In fig 6.1 the symbol used to represent the complete amplifier is depicted:
Fig 6.1: complete amplifier symbol This symbol contains the differential amplifier and also the common mode feedback regulator, which are represented in fig 6.2; splitting differential and common mode part allows to have an easier debug during the design process.
Fig 6.2: differential and common mode symbols - 80 -
Chapter 6: Amplifier Characterization 6.2.1 Stage 5 op-amp AC behaviour The first testbench relies on the DC gain, the bandwidth and the stability of the amplifier. As depicted in fig 6.3 the main goal is to obtain the environment in which the device is supposed to operate. As already said, it is important to use of modular hierarchy in the symbols to make it simpler to interface other blocks in successive steps of the design. This justify the use of dummy blocks, in order to simulate the effective load of the device: in fact, in the converter, there will be a chain of stacked amplifier, so a dummy amplifier can be an efficient way to simulate the effective impedance seen.
Fig 6.3: AC behaviour testbench - 81 -
In table 6.1 the AC behaviour is plotted, considering all corners at the minimum and maximum of the temperature range. Corners Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
gainBwProd(Hz) 1,12E+009 1,25E+009 1,36E+009 9,43E+008 1,02E+009 1,24E+009 1,35E+009 9,34E+008 1,00E+009 1,23E+009 1,33E+009 9,30E+008 9,97E+008 1,27E+009 1,38E+009 9,56E+008 1,03E+009 1,26E+009 1,36E+009 9,51E+008 1,02E+009
gain(dB) openloop @100kHz 57,24 56,71 57,88 55,5 56,71 55,62 56,73 53,98 55,1 56,86 57,97 55,76 56,93 56,49 57,74 54,9 56,21 57,54 58,87 56,17 57,7
Table 6.1: stage 5 AC characterization It can be noted that the DC gain is not too much sensitive to the corner variation, it is around 6dB for the whole lot, and that the secondary pole is enough close to the 0dB intercept. This means that the device will be stable, in particular these are the values extracted for the phase margin: Corners PhaseMargin (°) Alltyp1 70,14 Alltyp2 68,03 Alltyp3 69,6 Alltyp4 70,68 Alltyp5 71,9 TffRtypCtyp1 75,32 TffRtypCtyp2 77 TffRtypCtyp3 77,25 TffRtypCtyp4 78,71 TfnspRtypCtyp1 68,71 TfnspRtypCtyp2 70,8
Corners PhaseMargin (°) TfnspRtypCtyp3 70,98 TfnspRtypCtyp4 72,77 TsnfpRtypCtyp1 72,38 TsnfpRtypCtyp2 71,19 TsnfpRtypCtyp3 75,58 TsnfpRtypCtyp4 73,89 TssRtypCtyp1 61,98 TssRtypCtyp2 61,54 TssRtypCtyp3 67,56 TssRtypCtyp4 65,6
Table 6.2: stage 5 amplifier stability
- 82 -
Chapter 6: Amplifier Characterization The way the phase margin is calculated is the following. At first, all the signal sources are turned off; then, in an arbitrary point of the loop of interest, a block named AC_killer is added. As depicted in fig this is an ideal block which cuts the whole AC behaviour leaving unchanged the DC value of the signal.
Fig 6.4: the “AC_killer” block The input of this device is connected at the same point of the loop to a dummy amplifier; in this way the impedance seen in the loop under exam remains approximatively unchanged. The output instead is used to bias a small-signal source and then fed in the point of the topology where the loop was broken. So it is possible to plot the AC behaviour of the loop without affecting neither the internal impedances, neither the bias point of the devices, in a certain way recreating the same condition in which the real amplifier is expected to operate. In fig 6.5 the Bode plot is shown; the device seems to be robust among the whole corner lot, since the variations affect the gain and the secondary pole location, but leaving it still inside the specifications required.
- 83 -
Fig 6.5: stage 5 Bode plot 6.2.2 Stage 5 CM behaviour The parameters that the common mode controller must guarantee are an acceptable DC output offset, a sufficient common mode error suppression and the stability of the regulator itself. Corners Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
DC offset [V] 0,0034 0,0079 0,0044 0,0056 -0,0002 0,0045 0,0004 -0,0043 -0,0118 0,0059 0,0024 0,0019 -0,0039 0,0114 0,0063 0,0122 0,0033 0,0112 0,0074 0,0153 0,0082
Effective_swing 1,3992 1,4164 1,6304 1,1708 1,3876 1,4484 1,6664 1,2272 1,4256 1,4124 1,6264 1,1820 1,3996 1,4124 1,6348 1,1508 1,3788 1,3816 1,5984 1,1168 1,3376
Table 6.3: stage 5 offset and swing - 84 -
Chapter 6: Amplifier Characterization The offset is inversely proportional to the common mode loop gain, and, since its maximum absolute value is about 15mV, its main effect is to reduce the effective output voltage swing. In fig 6.6 the maximum ripple is shown, it is obtained forcing the amplifier to represent a sinusoidal wave at full swing. Clearly the different mean values of these sine are linked to the fact that VAGND (which is the DC voltage output) depends on Vdd, in fact VAGND = Vdd/2 by definition. In table the common mode loop gain and the phase margin are shown: Corners Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
loop_gain @100kHz (dB) 39,08 39,73 40,63 37,16 37,92 37,38 37,89 34,18 34,43 40,06 40,62 37,61 38 37,35 39,67 33,99 36,47 40,39 42,35 37,31 39,55
phaseMargin (°) 70,14 68,03 69,6 70,68 71,9 75,32 77 77,25 78,71 68,71 70,8 70,98 72,77 72,38 71,19 75,58 73,89 61,98 61,54 67,56 65,6
Table 6.4: stage 5 CM AC behaviour The way these values have been extracted is depicted in fig 6.7; the goal is to break the common mode loop in the amplifier in feedback configuration, and sense the transfer function between a common mode input and a point anywhere in the loop. As in the differential case, an AC_killer block has been used, and a dummy block has been inserted to recreate the effective impedance seen where the loop was broken. The ratio between the input of the AC_killer and the output of the AC source represent the loop behaviour.
- 85 -
Fig 6.6: stage 5 DC ripple
Fig 6.7: common mode testbench - 86 -
Chapter 6: Amplifier Characterization Finally it is possible to plot the CMR by applying a common mode signal at the input and sensing the output:
Fig 6.8: stage 5 common mode rejection The rejection is good, which means that it is about -40dB up to tenth of MHz and then it increases up to -25dB. This is due to a combination of the limited band of the comparing block in the common mode loop, and to the parasitic capacitance in the tail transistors. Another parameter to test is the PSR, extracted by applying a signal source at the voltage supplies, and sensing the output. The PSR is plotted in fig 6.9. There is the same peak seen in the CMR plot, it depends on the fact that the supply is perceived as a common mode signal and so it is processed in the same way.
- 87 -
Fig 6.9: stage 5 power supply rejection
6.2.3 Max dynamic configuration As previously described, the worst case for the INL extraction is located at the extreme points of the voltage gap. This means that the differential outputs of the amplifier will be completely unbalanced, reducing the value of Rout according to the previous considerations, and forcing the gain to drop. So, to extract the effective INL behaviour, it is necessary to characterize the amplifier in this particular bias configuration. It can be easily achieved by the use of a DC feedback made by ideal devices. The role of this loop is to force the DC output to be a constant: V+,DC + V-,DC = Vmax_swing while the DC input is unbalanced by the loop itself, as can be seen in fig
- 88 -
Chapter 6: Amplifier Characterization
Fig 6.10: Max dynamic configuration testbench The values extracted are still in spec
- 89 -
Corners BW @maxSwing PhaseMarg @maxSwing Gain @maxSwing Alltyp1 3,14E+008 60,96 54,69 Alltyp2 3,44E+008 61,63 54,66 Alltyp3 3,76E+008 60,39 56,17 Alltyp4 2,57E+008 63,07 50,83 Alltyp5 2,87E+008 61,28 53,45 TffRtypCtyp1 3,40E+008 62,03 53,71 TffRtypCtyp2 3,72E+008 60,88 55,1 TffRtypCtyp3 2,52E+008 63,61 50,06 TffRtypCtyp4 2,82E+008 62,04 52,23 TfnspRtypCtyp1 3,43E+008 60,63 54,75 TfnspRtypCtyp2 3,75E+008 59,48 56,2 TfnspRtypCtyp3 2,56E+008 62,23 51,02 TfnspRtypCtyp4 2,86E+008 60,52 53,57 TsnfpRtypCtyp1 3,45E+008 62,69 54,51 TsnfpRtypCtyp2 3,77E+008 61,33 56,11 TsnfpRtypCtyp3 2,56E+008 64,3 50,24 TsnfpRtypCtyp4 2,87E+008 62,33 53,05 TssRtypCtyp1 3,46E+008 61,65 55,27 TssRtypCtyp2 3,78E+008 60,21 57,04 TssRtypCtyp3 2,57E+008 63,62 50,5 TssRtypCtyp4 2,88E+008 61,41 53,84
Table 6.5: stage 5 max_dyn behaviour There are approximatively no variations in bandwidth from the previous characteristics; instead there is a gain drop of about 4dB.
6.2.4 Noise The noise figure and the relative integral can be extracted using the tools provided by the design tools: the tools calculate the noise generated by each device and represents it at the input and at the output of the amplifier designed. The tools can estimate the noise power spectral density figure; here it is only reported the integral among the noise bandwidth.
- 90 -
Chapter 6: Amplifier Characterization
Corner Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
Input_ref (nV/sqrt(Hz), 100-20G) output_rms (mVrms) 1,86E-008 1,12E-003 2,07E-008 1,24E-003 2,07E-008 1,24E-003 1,79E-008 1,07E-003 1,80E-008 1,08E-003 2,38E-008 1,42E-003 2,39E-008 1,43E-003 1,93E-008 1,16E-003 1,96E-008 1,17E-003 1,93E-008 1,15E-003 1,93E-008 1,16E-003 1,68E-008 1,01E-003 1,70E-008 1,02E-003 2,21E-008 1,33E-003 2,21E-008 1,33E-003 1,90E-008 1,14E-003 1,91E-008 1,14E-003 1,82E-008 1,09E-003 1,81E-008 1,09E-003 1,67E-008 9,97E-004 1,67E-008 1,00E-003
Table 6.6: stage 5 noise 6.2.5 Montecarlo The Montecarlo tool simulates the deviations introduced into the silicon by the process variations. It allows to check the robustness of the device against mismatches. In fig 6.11 the testbenches used for the simulations are shown. A feedback loop is added: its goal is to eliminate the differential offset before being fed into the input by the capacitive path, otherwise it would be amplified by the OTA forcing the output to clip.
- 91 -
Fig 6.11: Montecarlo testbench
Fig 6.12: bandwidth and phase margin variations in Montecarlo simulation - 92 -
Chapter 6: Amplifier Characterization Bandwidth and phase margin are enough robust against mismatch
Fig 6.13: noise and offset variations in Montecarlo simulation 6.2.6 Start-up and switch down The specifications require that the device can be turned on at -40° degree. The simulation is made by applying a 1us ramp at the power supply from 0V to Vdd.
- 93 -
Fig 6.14: ramp-up simulation
Fig 6.15: turn-off, turn-on simulation
- 94 -
Chapter 6: Amplifier Characterization 6.2.6 INL simulation To check the effectivity of the device designed, it is important to simulate if it works properly inside the environment it was made for. The testbench realizes the complete MDAC as it will be printed on the silicon, only the subADC is substituted by a behavioural model to avoid huge simulation times (therefore the time machine needed is more than a day). The input of the system is fed by a ramp; the output obtained is then elaborated by a Matlab routine (see Appendix A).
Fig 6.16: INL simulation result
- 95 -
Fig 6.17: INL testbench
- 96 -
Chapter 6: Amplifier Characterization 6.3.0 Stage 3 characterization Only the results of the characterization set are reported, since the methodology is the same as in stage 5 amplifier. 6.3.1 AC behaviour CORNER Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
bandwidth(Hz) 0dBcloseloop gain(dB) openloop @100Hz phaseMargin(°) 3,13E+008 89,49 61,67 3,71E+008 90,96 61,43 4,09E+008 92,54 63,03 2,47E+008 85,23 61,28 2,76E+008 86,67 61,9 3,67E+008 87,51 62,92 4,06E+008 89,17 64,66 2,44E+008 80,58 63,6 2,74E+008 82,09 64,17 3,63E+008 86,28 61,46 4,00E+008 88,69 63,26 2,41E+008 80,71 61,72 2,70E+008 83,17 62,41 3,74E+008 92,92 61,47 4,13E+008 94,19 63,07 2,50E+008 85,43 61,13 2,79E+008 86,34 61,77 3,71E+008 94,3 60,25 4,07E+008 95,75 61,96 2,49E+008 89,43 59,6 2,77E+008 90,81 60,31
Table 6.6: stage 5 noise There are 12dB of variation in gain among corners, the double as in the previous case as expected, since the stages are two and the deviations are uncorrelated. 6.3.2 CM behaviour As previously described there are two different common mode loops. They are simulated in two steps
- 97 -
Fig 6.18: first testbench for double CM loop simulation The first loop, the one regulating the common mode behaviour of the first stage has the following characteristics: CM Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
Cm loop1_gain @10kHz (dB) phaseMargin (°) 45,95 69,25 47,99 67,23 48,78 67,63 40,3 71,34 40,64 71,71 44,54 75,06 45,2 76,08 34,39 77,53 34,65 78,55 49,57 67,37 50,23 68,23 43,16 71,26 43,54 71,81 45,7 69,24 47,03 68,26 35,74 73,16 36,26 73,09 50,19 59,3 51,74 57,61 44,5 63,16 45,42 62,62
Table 6.7: stage 7 CM behaviour of the first loop - 98 -
CM DC offset stage1 2,45E-003 3,11E-003 1,33E-003 3,86E-003 1,35E-003 7,98E-004 -1,63E-003 -3,27E-003 -6,53E-003 2,60E-003 7,95E-004 3,25E-003 6,30E-004 3,71E-003 1,74E-003 4,23E-003 1,60E-003 4,53E-003 2,82E-003 6,96E-003 4,61E-003
Chapter 6: Amplifier Characterization The second one, which regulates the output of the whole amplifier:
Fig 6.19: second testbench for double CM loop simulation CM Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
Cm loop2_gain @10kHz (dB) phaseMargin (°) 46,15 70,89 46,7 69,22 47,87 71,15 44,22 70,98 45,25 73,29 44,3 79,05 45,39 82,07 41,92 82,83 42,74 85,47 46,37 69,5 47,44 72,11 43,83 70,42 44,79 73,56 46,53 70,02 48,08 70,64 44,41 71,79 45,56 73,27 48,81 60,81 50,18 61,5 46,32 60,59 47,47 62,09
Table 6.8: stage 7 CM behaviour of the second loop - 99 -
The offset at the output: Vout_cm-V(AGND) (V) CM Vout_DC Effective_swing CM Alltyp1 2,88E-003 0,6029 1,4140 Alltyp2 2,78E-003 0,5728 1,4664 Alltyp3 2,77E-003 0,6328 1,6688 Alltyp4 2,64E-003 0,5726 1,1468 Alltyp5 3,39E-003 0,6334 1,3436 TffRtypCtyp1 2,55E-003 0,5726 1,4800 TffRtypCtyp2 2,44E-003 0,6324 1,6864 TffRtypCtyp3 6,85E-003 0,5769 1,1676 TffRtypCtyp4 7,63E-003 0,6376 1,3608 TfnspRtypCtyp1 2,20E-003 0,5722 1,4656 TfnspRtypCtyp2 1,88E-003 0,6319 1,6724 TfnspRtypCtyp3 1,60E-003 0,5716 1,1556 TfnspRtypCtyp4 1,96E-003 0,6320 1,3508 TsnfpRtypCtyp1 3,24E-003 0,5732 1,4668 TsnfpRtypCtyp2 3,53E-003 0,6335 1,6700 TsnfpRtypCtyp3 3,89E-003 0,5739 1,1380 TsnfpRtypCtyp4 5,00E-003 0,6350 1,3356 TssRtypCtyp1 2,80E-003 0,5728 1,4488 TssRtypCtyp2 3,07E-003 0,6331 1,6516 TssRtypCtyp3 1,94E-003 0,5719 1,1092
Table 6.9: stage 7 CM offset Since the booster attenuates the common mode amplification, the CMR is high, approximatively 75dB in the worst case:
Fig 6.20: stage 3 CMR - 100 -
Chapter 6: Amplifier Characterization Also the PSR is good, around 30dB
Fig 6.21: stage 3 PSR 6.3.3 Max dynamic configuration BW @maxSwing Gain @maxSwing phaseMargin(°) @maxSwing Corners Alltyp1 3,12E+008 86,39 61,55 Alltyp2 3,71E+008 88,71 61,27 Alltyp3 4,09E+008 90,53 62,85 Alltyp4 2,44E+008 79,29 61,04 Alltyp5 2,73E+008 82,24 61,78 TffRtypCtyp1 3,67E+008 85,34 62,81 TffRtypCtyp2 4,05E+008 87,23 64,56 TffRtypCtyp3 2,41E+008 75,11 63,35 TffRtypCtyp4 2,71E+008 77,91 64,03 TfnspRtypCtyp1 3,63E+008 84,02 61,36 TfnspRtypCtyp2 4,00E+008 86,69 63,17 TfnspRtypCtyp3 2,38E+008 74,83 61,6 TfnspRtypCtyp4 2,67E+008 78,75 62,39 TsnfpRtypCtyp1 3,74E+008 90,69 61,27 TsnfpRtypCtyp2 4,13E+008 92,2 62,82 TsnfpRtypCtyp3 2,47E+008 79,39 60,77 TsnfpRtypCtyp4 2,76E+008 81,88 61,54 TssRtypCtyp1 3,71E+008 91,88 60,05 TssRtypCtyp2 4,08E+008 93,58 61,7 TssRtypCtyp3 2,46E+008 82,6 59,29
Table 6.10: stage 3 behaviour at max dynamic - 101 -
6.3.4 Noise The noise figure and the relative integral can be easily extracted using the tools provided by the design software: Corners Alltyp1 Alltyp2 Alltyp3 Alltyp4 Alltyp5 TffRtypCtyp1 TffRtypCtyp2 TffRtypCtyp3 TffRtypCtyp4 TfnspRtypCtyp1 TfnspRtypCtyp2 TfnspRtypCtyp3 TfnspRtypCtyp4 TsnfpRtypCtyp1 TsnfpRtypCtyp2 TsnfpRtypCtyp3 TsnfpRtypCtyp4 TssRtypCtyp1 TssRtypCtyp2 TssRtypCtyp3 TssRtypCtyp4
Input_ref (nV/sqrt(Hz)) Output noise (Vrms, 100Hz-10GHz) 6,05E-009 4,67E-004 5,87E-009 4,53E-004 5,92E-009 4,57E-004 6,21E-009 4,79E-004 6,25E-009 4,82E-004 6,24E-009 4,82E-004 6,30E-009 4,86E-004 6,39E-009 4,93E-004 6,44E-009 4,97E-004 5,77E-009 4,45E-004 5,82E-009 4,49E-004 6,14E-009 4,74E-004 6,18E-009 4,77E-004 5,96E-009 4,60E-004 6,02E-009 4,65E-004 6,26E-009 4,83E-004 6,30E-009 4,86E-004 5,59E-009 4,31E-004 5,64E-009 4,35E-004 6,06E-009 4,68E-004 6,09E-009 4,70E-004
Table 6.11: stage 3 noise characterization 6.3.5 Montecarlo
Fig 6.22: phase margin and offset variations in Montecarlo simulation - 102 -
Chapter 6: Amplifier Characterization
Fig 6.23: bandwidth and noise variations in Montecarlo simulation
- 103 -
- 104 -
Chapter 7: layout
7.0 Introduction Once the tests to check against process variation are performed, the second step is to make the design robust against manufacturing effects. This step relies mostly on the layout design. Several aspects must be taken into account: in fact, with the scaling of the devices, the subwavelength gap widens, making it harder to print most structures; some structures are even harder to print, leading to lithographical distortions which in some cases result in yield loss as well as performance degradation; interconnect manufacturing issues represent the largest yield detractor in nano-CMOS processing. A design put together without design for manufacturability in mind can result in copper erosion and dishing, changing the designed characteristics affecting electromigration and timing. Certain wiring patterns can result in high yield loss due to shorts. Open via is another major yield detractor in copper technology. Interconnect density variation causes interlayer dielectric thickness variation, resulting yield loss due to underpolish metal shorts as well as unexpected timing due to variation of capacitive parasitics.
- 105 -
Fig 7.1: wire density variation Antenna problems can lead to yield loss due to gate damage and in some cases, degrade transistor performance by inducing early negative bias temperature instability or Vth shifts. So the use of antenna diodes becomes mandatory, and this implies an additional parasitic capacitance as well as increasing the risk of latch-up.
7.1 Main sources of variations The technology scaling enables exponential improvement of digital circuit performance and functions on a chip, but on the other hand, has made analog design more challenging on many fronts. Table 1 in appendix C summarizes the modelling challenges that can affect analog designs. Analog circuits require good device matching; listed below are the main sources of matching problems. - Asymmetry (leads to misalignment sensitivity) - Small geometries (narrow-width effects; short-channel effects; larger Vth variation) - Proximity effects (well proximity; poly proximity; microloading etch effects) - Position of well and ground taps (body effect differential) - Horizontal and vertical effects - Temperature differential - STI stress effects - Diffusion and poly flaring (strong design influence in the nano-CMOS regime) - Mirror layout effects (capacitance; Rsd; misalignment) - Random dopant fluctuation - Dopant channelling through gate - Poly-L variation; Leff variation - 106 -
Chapter 7: layout - Degradation due to antenna effect - Hot-carrier injection - Metal density variation (thickness variation; capacitance variation) So there are several effects to take into account; also, since the number of devices is quite high, the more performing solution seems to use a “greedy” strategy, or better, to apply local optimizations and then, at the end, optimize the global layout.
7.2 Interdigit structure Where the matching is critical, the use of interdigit structure is a good choice. It consists in alternating wherever possible same length, same width, fingered transistors. In particular, the input pair in a differential architecture is extremely sensitive to any unbalanced deviation; in fact any mismatch which can modify the behaviour of a transistor in comparison to the other, would appear at the output amplified by the gain of the amplifier itself.
Fig 7.2: Input pair layout
- 107 -
As can be seen in fig 7.2, the design is oriented towards symmetry; the left part of a component must be as much similar as possible to the right part. The goal is to make the mechanical stress introduced by the STI oxide and by the metal routing the same on both the devices of the input pair. In particular, since defects can exist in singular spot of the silicon crystal, the interdigiting strategy allows to minimize the variation, since the defect is better distributed among the two transistors. This technique is applied also on current mirrors, where matching is also fundamental.
Fig 7.3: particular of intedigited mirror layout 7.3 Antenna effect and antenna diodes The gate oxide underneath the poly is thin, between 1 and 2nm. If the charges accumulated on the poly is sufficiently large, the charges accumulated can damage the gate oxide. This is known as process antenna effect. The maximum amount of charges that can be accumulated on the poly is proportional to the area of the poly, or better, the charges are accumulated on the perimeter side-wall area of the poly, which can be calculated as the perimeter of the poly multiplied by the thickness of the poly. Thus, an effective layout practice to prevent process antenna violation is to stay within the antenna ratio design rule of the given technology. In particular:
- 108 -
Chapter 7: layout • Minimize the use of poly for routing • Minimize the use of poly to connect the gates together A different widely used solution is to place diodes to protect the poly from antenna ratio violation. Antenna diode is only effective in preventing antenna violation from metal routing, and does not help in antenna violation due to poly. The reason is simple. The diodes are made from diffusions, but the poly is deposited onto the wafer before the diffusions are implanted into the wafer. An other advantage coming from the use of antenna diodes arises during the silicon fabrication process, instead. In fact, several steps in the foundry are related on plasma (for instance etching and deposition). Charge from the electrons and ions can be collected by conductive material on the wafer, and a net charge accumulation could lead to a change of the potential of the conducting material – until that potential itself is big enough to open up the “charge drainage” path to balance out the collection from plasma. If the drainage path is through gate oxide, charge can be trapped in the oxide, leading to many side effects, including shift of device threshold, creation of interface states which leads to earlier breakdown of the oxide, mobility degradation, worse sub-threshold slope, etc. So the simplest solution is to provide an alternative path for the “drainage path”, a so called antenna diode; in normal operation the diodes are reversely biased and since they have minimum size, they have only a minuscule impact on total capacitance.
Fig 7.4: cross section of antenna diode protection
- 109 -
During processing, even being reversely biased, because of the elevated wafer temperature (200°C plus) and of the reduced breakdown voltage, they can eventually provide a discharge path.
Fig 7.5: antenna diode layout
7.4 Dummy transistors As long as the geometry is reduces, the device behaviour gets more and more sensitive on the mechanical stress. This is why the technological library is provided of different features to take into account second order effects during the design. One of the more relevant in an analog environment is the LOD effect: it simulates the effect of the STI on the matching of the transistors. The STI (shallow trench isolation) introduces a silicon spot which is in a non-uniform state of superficial stress; this implies an impact on the device performance, adding a I dsat and a Vth offset, which are a strong function of the layout. In fig 7.6 an axial representation of the stress is depicted:
- 110 -
Chapter 7: layout
Fig 7.6: STI stress This stress is function of the geometry, and can be qualitatively described by Sa and Sb, which are the distance from the gate to the edge of the OD (oxide definition) on both sides of the device: Stress=
This effect is very important when designing current mirrors, differential pairs or any other structure based on ratio and symmetry. The solution chosen relies on multifinger devices and on the use of dummy transistors; in fact, it can be shown by simulations that the use of a single dummy device appears to be very effective, but the use of two fingers dummy transistors even better since it offers marginal differential return from the values expected and the ones experienced. So the adopted methodology is based on the use of doubly fingered dummy blocks placed on both sides of every critical device. - 111 -
An additional improvement consists in merging different devices together, using same-width fingers, and placing the most sensitive devices in the middle of the array, as shown in fig 7.7:
Fig 7.7: array of fingered devices in a current mirror 7.5 Electromigration The electromigration consists in a mass transport of electrons in metals where these metals are stressed at high current densities. This may result in a change of the conductor dimension, eventually causing the creation of either voids or hillocks in the affected regions. This process is intrinsically linked to time: in an industrial design, one of the specifications to take care of is the lifetime of the device.
Fig 7.8: effects of electromigration
- 112 -
Chapter 7: layout Physically, the copper or aluminium interconnects are polycrystalline; while conducting current, the electrons interact with the atoms in the lattice, forcing them to migrate in the direction of the main flow. At the end this process consists in a material transport, which mainly occurs at the metal-dielectric interface and at the boundaries of the grains. After a sufficient amount of time, atoms are deposited, leading to the generation of hillhocks and the build-up of mechanical stress around the hillhock area. While these hillhocks can cause shorts with neighbouring interconnects, the build-up of mechanical stress can lead to cracks in the surrounding insulation layers. Subsequently, material migrations towards these cracks can generate the so called whiskers which may also introduce shorts to neighbouring wires. Also, voids can reduce the conductivity over time, which can lead increase the resistivity or to interconnect failures. It may be noted that these processes are self-accelerated effect cycle. The foundry provides a documentation reporting study and experimental studies concerning these effects. It also provides the constants to calculate, through the “Black's Law”, the mean time to failure (MTTF):
Ea A MTTF= n⋅exp kT j
where A is a cross section area dependent constant, j is the current density, Ea is the effective activation energy of the electromigration process and n is a scaling factor. Calculating power rail for metal1 in tsmc90 TechConstant Time [hours] Temperat [°K] AverageCurrent [mA] 2,89E-007 1,00E+005 3,83E+002 1,00E-001 Jmax for M1 [mA/um] Wmin[um] 2,00E+000
Irms[mA]X5°C Irms[mA]X2°C Irms[mA]X1°C 1,40E+000 8,82E-001 6,24E-001 8,54E-002 6,93E-002 6,46E-002 Ipeak,max acceptable Peak duration[s] [mA] Period[s] 1,38E+001 2,10E-011 1,00E-009
Table 7.1: routine to calculate the interconnections lifetime To optimize the human time dedicated a self calculating routine was created: it extracts - 113 -
the minimum width of a connection to ensure a desired lifetime, express as the maximum augment of resistance acceptable.
7.6 General consideration for layouting Here is reported a general vademecum to take care of during the layout process. - Minimum-spaced and minimum-width wires must be avoided wherever possible to minimize erosion distortion of the signal lines, which increases resistivity and degrades timing that is not comprehended by the tools. - Wide wires may require more space, since the walls of wide trenches have a tendency to collapse, causing shorts. The sidewall incline of wider wires is also greater and can result in shorts to neighbouring wires. - Nwell proximity effects can cause as much as a 50-mV Vth shift for NMOS and a 20mV Vth shift for PMOS. Attention must be paid to the placement of matched devices where the orientation and space to the well are identical. - Limiting the degrees of freedom in a layout, such as by having all transistors oriented the same way, can dramatically improve process control and optimization. - Design uniformity and the use of tiled devices guarantee identical devices, which helps in device matching. - Constraining poly pitch and the use of dummy devices to guarantee the neighbourhood desired makes the lithographic processes easier and results in better poly-CD control. - Symmetry in critical layout and the use of precision rules will help to ensure that the end caps have ample diffusion overlap. - The use of multiple contacts and vias has a considerable impact on yield. It is better to use more structured design methodology where random layout patterns are not allowed. - Precision or analog design rules should be used with analog cells.
- 114 -
Chapter 7: layout
Fig 7.9: common mode regulator layout
- 115 -
Fig 7.10: differential amplifier layout
- 116 -
Chapter 8: Conclusion
8.0 Introduction All the specifications required were met. Table 8.1 shows the power consumption among corners of both the amplifiers discussed, in particular the power used by the whole device and the one taken by the common mode regulator. CORNER st3Power_total (W) st3Power_DC_CMdriver st5Power_total (W) st5Power_CMdriver Alltyp1 5,27E-003 8,75E-004 2,73E-003 1,24E-004 Alltyp2 4,69E-003 7,69E-004 2,43E-003 9,41E-005 Alltyp3 5,84E-003 9,70E-004 2,99E-003 1,27E-004 Alltyp4 4,73E-003 7,87E-004 2,48E-003 1,21E-004 Alltyp5 5,89E-003 9,95E-004 3,05E-003 1,59E-004 TffRtypCtyp1 4,82E-003 8,05E-004 2,46E-003 1,15E-004 TffRtypCtyp2 6,02E-003 1,02E-003 3,03E-003 1,55E-004 TffRtypCtyp3 4,96E-003 8,41E-004 2,52E-003 1,54E-004 TffRtypCtyp4 6,20E-003 1,07E-003 3,10E-003 2,02E-004 TfnspRtypCtyp1 4,68E-003 7,57E-004 2,44E-003 1,07E-004 TfnspRtypCtyp2 5,84E-003 9,64E-004 3,01E-003 1,45E-004 TfnspRtypCtyp3 4,72E-003 7,77E-004 2,49E-003 1,40E-004 TfnspRtypCtyp4 5,90E-003 9,95E-004 3,07E-003 1,84E-004 TsnfpRtypCtyp1 4,70E-003 7,78E-004 2,42E-003 8,40E-005 TsnfpRtypCtyp2 5,85E-003 9,76E-004 2,98E-003 1,13E-004 TsnfpRtypCtyp3 4,73E-003 7,92E-004 2,46E-003 1,07E-004 TsnfpRtypCtyp4 5,90E-003 9,94E-004 3,03E-003 1,41E-004 TssRtypCtyp1 4,61E-003 7,41E-004 2,40E-003 7,99E-005 TssRtypCtyp2 5,73E-003 9,32E-004 2,96E-003 1,07E-004 TssRtypCtyp3 4,61E-003 7,49E-004 2,45E-003 9,97E-005 TssRtypCtyp4 5,74E-003 9,45E-004 3,01E-003 1,31E-004
Table 8.1: power consumption
- 117 -
8.1 Future Work Today, only MDAC from stage 3 up to stage 6 have been completely designed. Still remain stage 1 and 2 amplifiers, since the other components for the MDAC have been already designed. In table 8.1 the whole specifications set is reported. It is clear that designing the amplifiers for the first stages will be challenging, since the gain required is nearly 100dB for a GBW of 2.5GHz, but probably the most stringent requirement is the 1.58 nV / Hz of input referred noise. Previously it was described as the boosting techniques spoil the noise behaviour; moreover, if the booster are applied at the second stage of a two stages amplifier, it could
be possible that, having enough gain in the first stage, the noise specification will be met.
Supply voltage (V) Dynamic range (Vpdiff) Sampling rate (Mspl/s) num of bits Total num of bits max out error (mV) Guard time (ps) Max INL Rel INL contribution (%) Rel noise contr to total noise kT/C noise (mVrms) Ron noise (mVrms) Max op-amp out noise (mVrms) Total stage out noise Capacitor unit value (pF) Cap load seen by the output (pF) Min op-amp DC gain (dB) Min GBW (MHz) Iout capability (mA) noise BW (MHz) Input ref noise (nV/Hz^0.5) Op-amp input capacitance (pF)
Input buffer Stage 1 Stage 2 Stage 3 1,2 0,55 n.a. 0 2 2 2 0 2,81 2,81 2,81 0,061 0,244 0,977 3,906 500 0,5 n.a. 16,67 16,67 16,67 2,84 87,34 6,55 1,75 n.a. 0,24 0,26 0,55 n.a. 0,08 0,09 0,1 0,02 0,24 0,55 0,26 0,02 0,4 0,44 0,92 n.a. 0,523 0,438 0,143 4,66 4,85 1,48 0,58 linearity 97,24 85,09 70,79 2480 2160 1415 4,07 1,11 0,42 314 658 580 492 1,03 1,58 1,86 5,53 n.a. 0,25 0,15 0,055
Stage 4 Stage 5 Stage 6 Stage 7
200 2 2,81 15,625
2 2,81 62,5
2 2,81 250
2 2,81 1000
16,67 1,09 0,78 0,16 2,26 2,89 0,072 0,52 59,42 1290 0,33 413 22,7 0,55
16,67 0,22 0,78 0,15 4,16 5,16 0,072 0,53 48,29 1177 0,28 334 41,19 0,1
16,67 0,22 0,78 0,1 16,85 20,65 0,072 0,24 34,25 746 0,11 249 227,25 0,04
0 0 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
Otherwise, a different option is a three stages amplifier, since it was previously demonstrate that a 30dB and more of gain per stage is easily achievable; so at this point the problems seem to be related mainly on bandwidth.
- 118 -
Chapter 8: Conclusion The technology shows severe limitations for the output resistance achievable; at the same time there are limitations also on the maximum current density, which implies the need of parallelism for several applications, adding further capacitance which can spoil the intrinsic speed of short channel technologies.
- 119 -
- 120 -
APPENDIX A: Matlab routine The routine used in Matlab to simulate the INL for each MDAC are reported. MDAC.m function [ve_out]=MDAC(ve_in,ve_Qstate,ve_C,ve_Vref,sc_A0,sc_Gain_CL_var_rel, sc_Cp) % model : MDAC without nonidealities, except the capacitor mismatch (if taking in account in vector C) % and the AGND error. %ve_out : output signal of the MDAC %ve_in : input signal of the MDAC %ve_C : vector with all the capacitor values (the last capacitor is the feedback capacitor) %ve_Qstate : state coming out from the quantizer (values from 0 to Nstate-1). %ve_Vref : reference voltage value (VrefP = +Vref; AGND = 0; VrefN = -Verf). %sc_A0 : OPAM open loop gain %sc_Gain_CL_var_rel : Close Loop Gain error (relative). Vout = Vout_ideal * (1 + sc_Gain_CL_var_rel) %sc_Cp : OPAM input parasitic capacitance. Not yet implemented. if nargin < 3 error('MADC : not enough input arguments') elseif nargin == 3 ve_Vref = 1; sc_A0 = 1e100; sc_Gain_CL_var_rel = 0; sc_Cp = 0; elseif nargin == 4 sc_A0 = 1e100; sc_Gain_CL_var_rel = 0; sc_Cp = 0; elseif nargin == 5 sc_Gain_CL_var_rel = 0; sc_Cp = 0; elseif nargin == 6 sc_Cp = 0; end if length(ve_Vref) == 1 ve_Vref = ve_Vref * ones(1,length(ve_in)); elseif length(ve_Vref) ~= length(ve_in) error('MADC : length(ve_Vref) ~= length(ve_in)') end
- 121 -
%transform the vector Qstate on a line vector if necessary [l,c]=size(ve_Qstate); if l>c ve_Qstate=ve_Qstate'; end if min(l,c)~=1 %test if it is a vector error('MDAC : input state is not a vector') end %transform the vector C on a line vector if necessary [l,c]=size(ve_C); if l>c ve_C=ve_C'; end if min(l,c)~=1 %test if it is a vector error('MDAC : input C is not a vector') end %transform the vector ve_in on a line vector if necessary [l,c]=size(ve_in); if l>c ve_in=ve_in'; end if min(l,c)~=1 %test if it is a vector error('MDAC : input ve_in is not a vector') end % Switch matrix construction : the matrix SW give the states of the switches connected to the capacitors % in function of the quantizer output state (vector Qstate). The line i correspond to the quantizer state i-1. % The first column give the states of the switches connected to the capacitor 1... A value of -1 means that % it is the switche connected to -Vref that is closed, 0 it is the switche connected to AGND and 1 it is the % switche connected to +Vref. sc_Nstate = 2*size(ve_C,2)-1; if log(sc_Nstate+1)/log(2) ~= round(log(sc_Nstate+1)/log(2)) %number of different state possible. Only 3,7,15,31,63... % are possible. This is due to the error correction and to the fact that in digital all number are in base 2 % (gain must be a power of 2). warning('MDAC : the number of states does not correspond to a normal value. The vector C (capacitor) is not correctly sized.') end ma_SW = zeros(sc_Nstate,size(ve_C,2)-1); %initialisation of the Switch matrix. In this - 122 -
Appendix A matrix the Qstate k correspond % to line k+1 (first line have the index 1 and not 0) ma_SW(1,:)=-1; for i=2:sc_Nstate ma_SW(i,1:(size(ve_C,2)-floor(i/2)-1)) = -1; ma_SW(i,(size(ve_C,2)-floor(i/2)):(size(ve_C,2)-1)) = 1; if round(i/2)==i/2 % i is even ma_SW(i,size(ve_C,2)-i/2)=0; end end ve_C_sampling=ve_C(1:size(ve_C,2)-1); %generate a capa vector without the feedback capacitor ve_Qstate=ve_Qstate+1; %add 1 to Qstate in order to have a correspondance between the SW line and the state ve_out=((sum(ve_C)*ve_in'(ma_SW(ve_Qstate',:)*ve_C_sampling'.*ve_Vref'))/ve_C(size(ve_C,2)))'; %this formula is a direct application of the charge conservation. %Error due to finite OPAM gain : ve_out=ve_out*sc_A0/(length(ve_C)+sc_A0); ve_out=ve_out*(1 + sc_Gain_CL_var_rel);
ADC_Quan.m function ve_out_decimal = ADC_Quan(ve_in_quantizer,sc_Nstate,ve_Vref,ve_offset,sc_sigma_noise,sc_hysteresis) %Model : ideal flash quantizer foreseen for an architecture with error correction. Nonidealities added : comparator %offset, comparator noise. Hysteresis not yet implemented. %ve_in_quantizer : input signal. Must be greater than -ve_Vref and lower than ve_Vref (hard cliping above these limits). %sc_Nstate : number of different output state possible for the quantizer. The only value possible are 3,7,15,31.... = 2^n - 1. % This is due to the fact that the stage gain must be a power of 2 (on each stage we resolve/assign a integer % number of bits n => we must amplify the residue by 2^n) and this model is valid only for a structure % like the classical 1.5 bits (i.e. with 2 reference and 1 analog ground). cf Lewis in JSSC March 92. % The output states are 0,1 .... ,sc_Nstate-1 %ve_Vref : Reference value. Here we make the assumption (without losing any generality) that the references - 123 -
% are -ve_Vref, +ve_Vref and 0. %ve_offset : input differential offset of the comparator (coming from mismatch). One offset per comparator. % The first comparator is this of the lowest comparison threshold. It is supposed that there is Nstate - 1 comparator. %sc_hysteresis : half value of the misdecision interval (the same for each comparator). % Hysteresis not yet implemented. %sc_sigma_noise : scalar given the %Example 1.5 bits % % ^ OUTPUT_MDAC [-ve_Vref,+ve_Vref] % | % / % / % /| /| / % /| /| / -> INPUT [-ve_Vref,+ve_Vref] % / |/ |/ % / |/ |/ % / % / % % First segment code 00, second segment 01, third 10. % %First state is in = [-ve_Vref,-ve_Vref+3*ve_Vref/(sc_Nstate+1)[ %Second state is in = [-ve_Vref+3*ve_Vref/(sc_Nstate+1),-ve_Vref+5*ve_Vref/ (sc_Nstate+1)[ %Third state is in = [-ve_Vref+5*ve_Vref/(sc_Nstate+1),-ve_Vref+7*ve_Vref/ (sc_Nstate+1)[ %..... %sc_Nstate state is in = [-ve_Vref+((sc_Nstate*2)-1)*ve_Vref/(sc_Nstate+1),+ve_Vref] if nargin < 2 error('ADC_Quan : not enough input arguments') elseif nargin == 2 ve_Vref = ones(1,length(ve_in_quantizer)); sc_hysteresis = 0; ve_offset = zeros(1,sc_Nstate-1) sc_sigma_noise = 0; elseif nargin == 3 sc_hysteresis = 0; ve_offset = zeros(1,sc_Nstate-1) sc_sigma_noise = 0; elseif nargin == 4 sc_hysteresis = 0; sc_sigma_noise = 0; elseif nargin == 5 - 124 -
Appendix A sc_hysteresis = 0; end if length(ve_Vref) == 1 ve_Vref = ve_Vref * ones(1,length(ve_in_quantizer)); end if log(sc_Nstate+1)/log(2) ~= round(log(sc_Nstate+1)/log(2)) %number of different state possible. Only 3,7,15,31,63... are possible. %This is due to the error correction and to the fact that in digital all number are in base 2 (gain must be a power of 2). warning('ADC_Quan : the number of states does not correspond to a normal value. The variable sc_Nstate is not correctly sized.') end %transform input to a line vector in the case where it is a column vector [l,c]=size(ve_in_quantizer); if l>c % column vector ve_in_quantizer=ve_in_quantizer'; end if min(l,c)~=1 %test if it is a vector error('ADC_Quan : input ve_in_quantizer is not a vector') end %transform offset vector to a line vector in the case where it is a column vector [l,c]=size(ve_offset); if l>c % column vector ve_offset=ve_offset'; end if size(ve_offset,2) ~= sc_Nstate - 1 error('ADC_Quan : offset vector have not the correct number of element') end % ma_threshold = -ve_Vref' * ones(1,sc_Nstate-1) + ve_Vref' * [3/(sc_Nstate+1):2/ (sc_Nstate+1):(2*sc_Nstate-1) ... % /(sc_Nstate+1)]; ma_threshold = -ve_Vref' * ones(1,sc_Nstate-1) + ve_Vref' * [3/(sc_Nstate+1):2/ (sc_Nstate+1):2*sc_Nstate ... /(sc_Nstate+1)]; ma_threshold = ma_threshold + ones(length(ve_Vref),1) * ve_offset; %add constant offset (mismatch + systematic) to % the ideal treshold. ma_threshold = ma_threshold + randn(length(ve_Vref),sc_Nstate-1)*sc_sigma_noise; %add a noise to the threshold % (represent the comparator input noise) ve_out_decimal = sum(floor((sign((ve_in_quantizer' * ones(1,sc_Nstate-1)) ma_threshold) + 1)/2),2)'; - 125 -
INL.m function ve_INL=INL(se_filename_out,se_filename_in, sc_resolution, sc_QuantState, Vref, i) %se_filename_out : is simulation result filename. The results must be in 2 column. % The first column is the time and the second the MDAC output %se_filename_in : is simulation result filename. The results must be in 2 column. % The first column is the time and the second the MDAC input %The input and the output are supposed include between -Vref and +Vref. %sc_resolution : is the resolution in bit needed at the output of the MDAC. %sc_QuantState : is the number of state of the Quantizer. Must 2^n-1 where n is an integer. %Vref : is the reference value (VrefP = +Vref and VrefN = -Vref) % single data -------------------------------------------------[ve_time,ve_MDAC]=textread(se_filename_out,'%f, %f','headerlines',1); [ve_time1,ve_in1]=textread(se_filename_in,'%f, %f','headerlines',1); ve_in=ve_in1(1:length(ve_MDAC)); %Corner routine------------------------------------------------%M = csvread(se_filename_out); %N = csvread(se_filename_in); code=[0:1:256]*4; code(257)=1023; %ve_in=(code/2^10-0.5)*2; %plot(ve_in) ve_quant = ADC_Quan(ve_in,sc_QuantState,Vref); %plot(ve_quant) ve_capa = ones(1,(sc_QuantState+1)/2); ve_outideal = MDAC(ve_in,ve_quant,ve_capa,Vref) figure(1) %clf; hold on; plot(ve_outideal,'r'); grid on; plot(ve_MDAC,'g'); %transform the vector ve_in on a line vector if necessary - 126 -
Appendix A [l,c]=size(ve_MDAC); if l>c ve_MDAC=ve_MDAC'; end ve_INL=(ve_MDAC(1:length(ve_outideal))-ve_outideal)*2^(sc_resolution-1)/Vref; figure(2); %clf hold on; plot(ve_INL(3:length(ve_INL)),'r'); grid on; xlabel('code'); ylabel('INL (LSB)'); end
- 127 -
- 128 -
APPENDIX B: Native MOSFET
A transistor that has not undergone the channel doping process is termed a native transistor, and has a lower threshold voltage because it must rely on the intrinsic background or body of the transistor to set the threshold voltage. The typical native transistor threshold voltage can range from 0.1V to 0.3V Low-Vth natural threshold voltage transistors can be fabricated as an option in a dual poly gate CMOS process. A process designed for 1.2V operation may be further simplified by eliminating the steps required for hot-carrier reduction i.e., LDD (Lightly doped Drain) implant and formation. Figure B.1 shows the measured ID and gm versus gate voltage for a 20/0.4um nMOSFET biased at VDS = 0.1V.
Figure B.1: Measured ID(A) and gm(A/V) versus VGS(V) for “natural” nMOSFET
- 129 -
In Fig B.2 are illustrated the measured IDS-VDS characteristics for a natural threshold nMOSFET.
Figure B.2: Measured nMOSFET IDS-VDS characteristics. IDS ~ (VGS-Vth)1.27
- 130 -
APPENDIX C: Summary of modeling issues
Here is reported a summary of the main sources of effects coming from the miniaturization of the devices. Parameter
DITS Early voltage and output resistance Poly depletion
Gate tunnel current Mobility-dopant dependence Linear proximity effects Nonlinear proximity effects GIDL
Diffusion and poly flaring
Well proximity STI stress
Reason for effect
Synopsis of effect
Reverse short channel effect due to lateral Halo implants (technology, nonuniform doping; when channel length varies, physical device effect) Vth varies Drain induced threshold-voltage shift, due to Halo implants (technology, change in DIBL for long channel length devices when the halo implant's influence on the physical device effect) channel diminishes Halo implants (technology, Change in DIBL for long channel device similar physical device effect) to above Poly depletion is getting significant for ultrathin Ultrathin gate oxide gate oxide, which accounts for about 8nm (technology, physical device increase in equivalent oxide thickness for most effect) devices, less for predoped poly Ultrathin gate oxide Direct tunneling from gate to channel occurs (technology, physical device due to ultrathingate oxide effect) Halo implants (technology, Mobility improves with reduction in dopants physical device effect) Partly due to lithographic effects and partly to etch microloading effects, also due to dopant scattering from the poly, causing systematic Dense, isolated dopant variation as a function of poly-line space of the design Subwavelength lithography requires resolution Optical proximity correction extension High field in the drain to gate causes band-toBand-to-band tunneling band tunneling, due to high junction doping and abrupt junction Subwavelength lithography causes flaring of diffusion and poly, causing device variations of Technology and layout small geometry devices and proximity of poly effects contact pads to diffusion edge Lateral scattering of well implant atoms out of Devices at the edge of the the resist, which leads to threshold voltage well increase for device close to the well edge Proximity effects to STI to STI stress reduces electron mobility but device channel increases hole mobility, thus effecting Idsat
- 131 -
- 132 -
1. A. Cester, “Nanoelettronica”, appunti dalle lezioni, 2008 2. A. Gerosa, “Circuiti Integrati per l'eleborazione dei Segnali”, appunti dalle lezioni, 2008 3. A. Bevilacqua, “Progettazione di circuiti integrati analogici”, appunti dalle lezioni, 2007 4. H. W. L. R. Jacob Baker and D. E. Boyce, “CMOS Circuit Design, Layout and Simulation” NY: IEEE Press on Microelectronic Systems, 1998. 5. D. A. Jones and K. Martin, “Analog Integrated Circuit Design” NY: John Wiley & Sons, Inc., 1997 6. J. Vital, Pipeline Analogue-to-Digital Converters, lectures at ChipIdea Conference 7. P.R. Gray and R.G. Meyer, "MOS Operational Amplifier Design - A Tutorial Overview", IEEE Journal of Solid State Circuits, Vol. SC-17, No. 6, pp. 969-982, DIC 82 8. B. Razavi, “Design of Analog CMOS Integrated circuits”, McGrawHill, 2001 9. Gray, Hurst, Meyer, Lewis, “Analysis and design of analog integrated circuits”, Wiley, 4th ed. 10. Wong, Mittal, Cao, Starr, “Nano-CMOS Circuits and physical design”, WileyInterscience, 2005
- 133 -
11. L. Wu, “Low-Voltage Pipeline A/D Converter”, Thesis, Oregon State University 12. L. Sumanen, “Pipeline Analog-to-Digital Converters for Wide-Band Wireless Communications”, Thesis, Helsinki University of Technology, 2002 13. V. Abramzon, “Analog to digital converters for high speed links”, PhD Thesis, Stanford University, 2008 14. K. Sockalingam and R.Thibodeau, “10-Bit 5MHz Pipeline A/D Converter”, Thesis, ECE Umaine, 2002 15. T. Khanna, “Equation-based Hierarchical Optimization for a Pipelina ADC”, Thesis, Cornell University, 2005 16. Andrew M. Abo and Paul R. Gray, “A 1.5V, 10-bit, 14 MS/s CMOS pipeline analogto-digital converter” IEEE J. Solid-State Circuits, vol. 34, no. 5, pp. 599-605, May 1999 17. G. Moore, “No exponential is forever, keynote” IEEE International Solid-State Circuits Conference, 2003 18. Cline and Paul R. Gray, “A power optimized 13-b 5Msamples/s pipelined analog-to digital converter in 1:2m CMOS” IEEE J. Solid-State Circuits, vol. 31, no. 3, pp. 294303, March 1996 19. F. D'Agostino, D. Quercia, “Short-Channel Effects in MOSFETs”, project, 2000 20. A. Lakhlef, A. Benfdila, H. Achour and K. Bennamane, “Characterization of the Short Channel Effect on the Threshold Voltage in Deep Sub-micron MOSFETs”, African Physical Review, pp 18-21, 2008 21. K. Bult, G. Geelen, “A Fast Settling CMOS Op-Amp for SC Circuits with 90dB DC Gain”, IEEE J. Solid-State Circuits, vol. 25, no. 6, pp. 1379-1384, December 1990
- 134 -