20.03.2015 Views

Design of a High-Speed 12-bit Differential Pipelined A/D Converter

Design of a High-Speed 12-bit Differential Pipelined A/D Converter

Design of a High-Speed 12-bit Differential Pipelined A/D Converter

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

DESIGN OF A HIGH-SPEED <strong>12</strong>-BIT DIFFERENTIAL<br />

PIPELINED A/D CONVERTER<br />

Diploma Project<br />

Thomas Liechti<br />

February 2004<br />

Assistant: Zeynep Toprak (LSM)<br />

Pr<strong>of</strong>essor: Yusuf Leblebici (LSM)<br />

Microelectronic Systems Laboratory (LSM)<br />

Swiss Federal Institute <strong>of</strong> Technology Lausanne


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

Table <strong>of</strong> Contents<br />

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1<br />

1.1 Performance measures <strong>of</strong> A/D converters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2<br />

1.2 A/D-<strong>Converter</strong> Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3<br />

1.3 A/D <strong>Converter</strong> Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4<br />

1.4 ADC Pipeline Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5<br />

1.4.1 4-Stage <strong>Converter</strong> Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5<br />

1.4.2 Analog Pipeline Stage Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6<br />

1.4.3 <strong>Pipelined</strong> A/D Conversion and Digital Error Correction . . . . . . . . . . . . . . . . . . . . . . 6<br />

1.4.4 Analysis <strong>of</strong> accuracy requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8<br />

1.4.5 Clocking scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8<br />

2 4-<strong>bit</strong> Flash Analog-to-Digital <strong>Converter</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9<br />

2.1 4-<strong>bit</strong> Flash A/D <strong>Converter</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9<br />

2.1.1 Flash ADC Floorplan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11<br />

2.2 <strong>Differential</strong> Comparator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11<br />

2.2.1 Choice <strong>of</strong> Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11<br />

2.2.2 Comparator circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13<br />

2.2.3 Comparator Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15<br />

2.3 Performance verification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15<br />

2.3.1 Comparator Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16<br />

2.3.2 Flash ADC Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17<br />

3 4-<strong>bit</strong> Digital-to-Analog <strong>Converter</strong>. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20<br />

3.1 Current-steering D/A <strong>Converter</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20<br />

3.1.1 Continuous DAC Gain Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20<br />

3.1.2 DAC Floorplan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />

3.2 Unit current cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />

3.2.1 Fournier-Senn [15] Current cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />

3.2.2 Regulated Cascode Current Cell. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24<br />

3.3 Simulated Performance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24<br />

4 Residue Amplifier and Sample-and-Hold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24<br />

4.1 Switched Capacitor Residue Amplifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25<br />

4.1.1 Circuit Topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25<br />

4.1.2 Charge Injection <strong>of</strong> MOS Switches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26<br />

4.1.3 Capacitor and Switch Sizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27<br />

4.2 <strong>Differential</strong> OTA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28<br />

4.2.1 Mirrored cascode with class AB input stage with preamplifier . . . . . . . . . . . . . . . 28<br />

5 Top-Level Floorplanning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />

5.1 Analog Pipeline Stage Floorplan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />

5.2 Floorplan <strong>of</strong> complete pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />

6 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32<br />

i


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33<br />

ii


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

1 Introduction<br />

The goal <strong>of</strong> this diploma project is to redesign an existing pipelined <strong>12</strong>-<strong>bit</strong> 200-MS/s<br />

single-ended analog-to-digital converter [4] to make its analog signal path fully differential.<br />

The converter has four 4-<strong>bit</strong> pipeline stages, each stage consisting <strong>of</strong> a Flash<br />

analog-to-digital converter, a digital-to-analog converter, and a residue amplifier (Figure 2).<br />

Only the blocks in the analog signal path have to be redesigned. The digital part consisting<br />

<strong>of</strong> thermometric to binary encoders and digital error correction does not need to be<br />

modified to make the converter differential. It is taken as is from the design presented in [4].<br />

A prototype <strong>of</strong> a converter stage will be implemented on silicon. Some blocks will need to<br />

be finished after this report has been written.<br />

Making the analog signal path <strong>of</strong> the converter fully differential has several advantages:<br />

• Increased signal dynamic range, which is especially important for low-voltage analog<br />

designs.<br />

• Even harmonics introduced by circuit non-linearities are cancelled, improving the<br />

harmonic distortion characteristics <strong>of</strong> the system.<br />

• Immunity to common mode noise coming from the power supply or digital parts<br />

residing on the same chip for example.<br />

• Errors due to MOS-switch charge injection and clock feedthrough can more easily be<br />

cancelled as these errors are <strong>of</strong>ten common-mode signals. This is especially important<br />

in high-precision switched-capacitor circuits.<br />

These advantages make a fully differential signal path especially useful for mixed-signal<br />

designs [1], where analog and (noisy) digital circuitry have to coexist on the same die, and<br />

where the low supply voltage is imposed by the digital process.<br />

The ADC is designed as a block in a conventional logic 0.18µ CMOS process so it can easily<br />

be integrated into a digital system. As this process does not provide high precision resistors<br />

and capacitors the design should rely as little as possible on the precise matching <strong>of</strong> these<br />

elements. Thus, using precisely matched resistors and capacitors has to be avoided where<br />

possible. Special care has to be taken when drawing the layout <strong>of</strong> matched elements.<br />

The design also has to cope with low-voltage and deep-submicron technology issues. To<br />

analog circuits, the down-scaling <strong>of</strong> minimum feature size is not as beneficial as to digital<br />

circuits [8]:<br />

• The signal-to-noise ratio (SNR), i.e. the dynamic range between signal amplitude and<br />

noise floor is inherently limited by the low supply voltage (imposed by low<br />

breakdown voltages) because a thermal noise floor is always present, and the supply<br />

voltage imposes an upper bound on (voltage) signal amplitude. For very low voltage<br />

designs, current mode processing can thus be an interesting alternative to voltage<br />

mode processing as current is not directly limited by the supply voltage (and the<br />

oxide breakdown voltage <strong>of</strong> the process).<br />

• Many low voltage circuit topologies are inherently slower that their high-voltage<br />

counterparts. Also, as oxide layers get thinner and features move closer together with<br />

technology scaling, parasitic capacitances get larger.<br />

• Short channel effects such as very high g ds (drain-to-source conductance) degrade<br />

transistor performance.<br />

1


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

The challenge <strong>of</strong> this project is thus to find and dimension fully differential<br />

implementations <strong>of</strong> the functional blocks <strong>of</strong> the single-ended converter that satisfy the very<br />

am<strong>bit</strong>ious speed specification (Section 1.3) despite the technological limits.<br />

Applications <strong>of</strong> ADC with performance specifications in this range are used for high<br />

bandwidth applications such as digital video and wireless communication devices.<br />

This report first gives a short summary <strong>of</strong> commonly used performance measures for A/D<br />

converters. Only a very short overview on converter architectures is given because a<br />

pipelined architecture has already been chosen for this design. In Section 1.4 the converter<br />

structure is described in detail. The design <strong>of</strong> the Flash ADC is detailed in Section 2,<br />

Section 3 describes the DAC design, and the Residue Amplifier is described in Section 4.<br />

Finally top level floorplanning is discussed in Section 5. Section 6 summarizes the work that<br />

has been done and indicates the following steps in the development <strong>of</strong> the converter.<br />

1.1 Performance measures <strong>of</strong> A/D converters<br />

An analog-to-digital converter (Figure 1) transforms an analog signal (continuous in time<br />

and amplitude) to a digital signal which is discrete in time and amplitude. First, the input<br />

waveform is sampled at discrete time intervals (assumed equidistant) by a sample-and-hold<br />

(S/H) circuit. The S/H output is a continuous-amplitude discrete-time signal proportional<br />

to the input signal’s amplitude at the sampling instant. The n-<strong>bit</strong> A/D converter quantizes<br />

this signal into 2 n discrete amplitude levels, each one <strong>of</strong> which is described by a n-<strong>bit</strong><br />

codeword.<br />

The amplitude quantization introduces a quantization error. For input signals with<br />

frequency content only below half the sampling rate, the system’s accuracy ideally is only<br />

limited by this error. However, in practical converters, other sources <strong>of</strong> errors such as circuit<br />

element mismatches and random noise add to the total error, and limit the effective<br />

accuracy that can be achieved.<br />

The definitions used in this work <strong>of</strong> different performance measure <strong>of</strong> ADC systems are<br />

summarized next. More performance measures are described in [4].<br />

Digital Output<br />

Analog<br />

Input<br />

S/H<br />

Sampled<br />

Input<br />

n−<strong>bit</strong><br />

A/D <strong>Converter</strong><br />

B 0<br />

B 1<br />

B<br />

2<br />

B n−3<br />

B n−2<br />

B n−1<br />

Figure 1: Block diagram <strong>of</strong> an Analog-to-Digital <strong>Converter</strong> [5]<br />

• The <strong>Differential</strong> Nonlinearity (DNL) error describes the difference between the ideal<br />

step size (1 LSB) and the effective step sizes <strong>of</strong> the converter. For an A/D converter,<br />

2


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

the DNL error is thus the difference between two adjacent converter thresholds,<br />

normalized by 1 LSB.<br />

• The Integral Nonlinearity (INL) error measures the difference between ideal and real<br />

code midpoints (see Figure 7) <strong>of</strong> the converter DC characteristic.<br />

• Sampling time uncertainty (aperture jitter) measures the deviation <strong>of</strong> the effective<br />

sampling instant from the ideal sampling instant. If the input signal is time varying,<br />

this uncertainty introduces an effective amplitude error in the S/H output because<br />

the sampled value is implicitly assigned to the ideal sampling instant. The magnitude<br />

<strong>of</strong> the introduced error increases with the rate <strong>of</strong> change in the input signal, and hence<br />

with input signal frequency.<br />

• The signal-to-noise ratio (SNR) measures the ratio between signal power and total<br />

noise power. If (harmonic) distortion is included in the noise, the SNR is also called<br />

SINAD (signal-to-noise-and-distortion ratio). The SNR <strong>of</strong> an A/D converter is<br />

intrinsically limited by the quantization noise. This upper limit on the SNR is<br />

approximately given by (1), where N is the number <strong>of</strong> <strong>bit</strong>s <strong>of</strong> the converter.<br />

S ⁄ N = 6.02 ⋅ N + 1.76dB<br />

(1)<br />

• The effective number <strong>of</strong> <strong>bit</strong>s is the value <strong>of</strong> N in (1) that results in the effectively<br />

measured SNR.<br />

• The effective resolution bandwidth (ERBW) is the input signal range, for which the<br />

SNR at the converter output stays within 3 dB <strong>of</strong> the low frequency SNR value.<br />

1.2 A/D-<strong>Converter</strong> Architectures<br />

Analog-to-digital converters can be classified according to the sequence <strong>of</strong> operations<br />

performed to determine the digital value corresponding to the analog input sample value.<br />

At the highest level, converters can be divided into oversampling converters and Nyquist<br />

rate converters. Nyquist rate converters convert each analog sample at maximum precision,<br />

thus minimizing quantization noise for each sample. This allows conversion <strong>of</strong> analog<br />

signals approaching the Nyquist rate, although in practice signals are usually sampled at 3<br />

to 20 times the input signal’s bandwidth. Strictly speaking, only converters whose input<br />

bandwidth is at least f s /2 are Nyquist rate converters [10]. Oversampling converters highly<br />

oversample (typically by a factor <strong>of</strong> 20 to 5<strong>12</strong>) the input signal to spread the quantization<br />

noise over a large frequency band. Noise outside the signal band can then be filtered to<br />

improve the SNR. The quantization error power is minimized for a sequence <strong>of</strong> samples<br />

rather than for single samples.<br />

For high-speed applications, Nyquist-rate converters have to be used because the required<br />

sampling rate for oversampling is orders <strong>of</strong> magnitude higher than signal bandwidth. For<br />

high bandwidth applications, the required sampling rate cannot be realized with current<br />

technology. (New very high speed technology permits ∆Σ-<strong>Converter</strong>s to be used for RF<br />

applications [10].)<br />

3


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

There is a large variety <strong>of</strong> architectures for Nyquist rate converters:<br />

• Flash (parallel) converter are very fast but the number <strong>of</strong> required comparators<br />

increases exponentially with the number <strong>of</strong> <strong>bit</strong>s, thus entailing large ICs (high cost,<br />

difficult device matching), high power consumption and high input capacitance.<br />

• Time-interleaving converters: To or more converters work in parallel with shifted<br />

clocks. <strong>High</strong> power dissipation.<br />

• Serial and successive approximation converters: Convert an input sample using a<br />

number <strong>of</strong> sequential steps. These converters can be very accurate and small, but slow<br />

because several conversion steps need to be performed for each sample.<br />

Because the architecture to be used for the converter is given, no study <strong>of</strong> other converter<br />

architectures has been carried out. For high-speed Nyquist-rate A/D converters a pipelined<br />

architecture is very well suited because it allows to decouple conversion rate (sampling<br />

rate) from conversion time. That is, throughput can be increased by extending the latency<br />

between the time the analog sample is taken and the time when the corresponding digital<br />

value is available at the converters output (in many applications latency is not as critical as<br />

throughput.) The idea is to split the conversion into a number <strong>of</strong> serially executed<br />

low-resolution conversions. In one sampling interval, only a low resolution conversion has<br />

to be achieved instead <strong>of</strong> a full resolution conversion. The low resolution conversion can be<br />

much faster because the accuracy requirement <strong>of</strong> the comparators (in the flash ADC) is<br />

relaxed, which allows faster comparator architectures to be used (trade accuracy for speed).<br />

Another advantage is the reduced number <strong>of</strong> comparators: The pipelined ADC uses less,<br />

and less accurate comparators than a Flash ADC with the same resolution. The other<br />

elements in the pipeline stage (DAC and residue amplifier, see Figure 3), however, need full<br />

resolution accuracy, not just the per-stage resolution accuracy (see Section 1.4.4).<br />

1.3 A/D <strong>Converter</strong> Specifications<br />

The A/D <strong>Converter</strong> specifications are summarized in Table 1. Note that no power spec is<br />

given. The first objective is to reach the 200 MHz sampling speed, and not low power<br />

consumption. The design may thus trade power for speed and accuracy.<br />

Table 1: ADC Specifications<br />

Stated Resolution<br />

Voltage swing<br />

Sampling Rate<br />

Architecture<br />

<strong>12</strong> <strong>bit</strong>s<br />

1V pp (differential)<br />

200 MHz<br />

4 pipelined 4-<strong>bit</strong> flash stages with <strong>bit</strong> overlapping<br />

Technology UMC 0.18µ logic CMOS (1.8 V)<br />

The specifications in Table 1 are comparable to the performance <strong>of</strong> current state <strong>of</strong> the art<br />

converters [10][4].<br />

The 1 Volt peak-to-peak input signal swing has been set to 0.6 to 1.6 V, resulting in an analog<br />

ground level <strong>of</strong> 1.1 V. The lower bound <strong>of</strong> 0.6 V allows using NMOS differential input pairs<br />

(the nominal threshold voltage being 0.5 V) while leaving 200 mV headroom for PMOS<br />

current mirrors. The analog ground level stays the same throughout the analog pipeline.<br />

4


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

1.4 ADC Pipeline Architecture<br />

1.4.1 4-Stage <strong>Converter</strong> Structure<br />

Figure 2 shows the structure <strong>of</strong> the complete 4-stage converter pipeline. It consists <strong>of</strong> a<br />

(external)<br />

ANALOG<br />

PIPELINE STAGE 1<br />

ANALOG<br />

PIPELINE STAGE 2<br />

ANALOG<br />

PIPELINE STAGE 3<br />

ANALOG<br />

PIPELINE STAGE 4<br />

vin<br />

Front−End<br />

S/H<br />

vin’<br />

ADC 1<br />

DAC 1<br />

ADC 2<br />

DAC 2<br />

ADC 3<br />

DAC 3<br />

ADC 4<br />

4<br />

A1 B1<br />

3<br />

A2 B2<br />

ENCODER 1 ENCODER 2 (smart)<br />

ENCODER 3 (smart) ENCODER 4 (smart)<br />

3<br />

A3 B3<br />

3<br />

B1<br />

0<br />

FA<br />

DFF DFF DFF DFF<br />

FA FA FA FA<br />

A1<br />

B1<br />

B2<br />

DFF<br />

FA<br />

DFF DFF DFF DFF<br />

FA FA FA FA<br />

DFF DFF DFF<br />

FA FA FA<br />

A2<br />

B2<br />

B3<br />

DFF<br />

FA<br />

DFF DFF DFF DFF<br />

FA FA FA FA<br />

DFF DFF DFF<br />

FA FA FA<br />

DFF DFF DFF<br />

FA FA FA<br />

A3<br />

B3<br />

DFF<br />

DFF DFF DFF DFF<br />

DFF DFF DFF<br />

DFF DFF DFF DFF DFF DFF<br />

OVERFLOW MSB LSB<br />

DISCARDED<br />

Figure 2: Block diagram <strong>of</strong> four-stage pipelined A/D converter [4]<br />

horizontal analog pipeline and a vertical digital pipeline performing digital error correction<br />

and assembling the digital outputs <strong>of</strong> the four analog ADC stages. The same signal flow<br />

structure is used in the layout <strong>of</strong> the converter (Section 5.2). It allows easy slicing <strong>of</strong> the<br />

system into four almost identical parts that can simply be abutted to form the whole<br />

pipeline, thus reducing routing length between the stages. Another important benefit <strong>of</strong> this<br />

arrangement <strong>of</strong> circuit blocks is that it clearly separates the analog part from the digital part.<br />

This will minimize noise injected from the digital circuitry into the analog signal path.<br />

Once the first stage is completed, the whole pipeline can be assembled very easily. Only<br />

small modifications are needed in the encoder and digital part <strong>of</strong> the stages.<br />

A front-end sample-and-hold circuit required at input <strong>of</strong> first pipeline stage to hold the<br />

input stable during the conversion cycle. For the following stages the residue amplifier (see<br />

Section 1.4.2) in the previous pipeline stage will act as sample and hold circuit. The<br />

front-end sample-and-hold block has very stringent requirements on sampling time<br />

uncertainty (aperture jitter) because it samples an time-varying analog signal. Inside the<br />

pipeline stages, settled signals are sampled, and consequently sampling time jitter is less<br />

critical there.<br />

In [10] it is suggested that aperture jitter is the dominant limiting factor for the SNR <strong>of</strong><br />

current high-performance ADCs. Front-end sampling is thus very critical. In the prototype,<br />

the front-end sample-and-hold circuit will be external to the chip and its design is not part<br />

5


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

<strong>of</strong> this project. For the measurements <strong>of</strong> DC characteristics, slow varying analog inputs can<br />

be used to render the front-end S&H circuit unnecessary.<br />

The circuits in the digital pipeline are directly taken from [4]. Only their layout has to be<br />

redrawn.<br />

Note that the encoder delay is not part <strong>of</strong> the total analog stage delay as a DAC topology<br />

directly using the Flash thermometric output is employed. The digital pipeline thus works<br />

in parallel with the analog one; its timing is less critical than that <strong>of</strong> the analog pipeline.<br />

1.4.2 Analog Pipeline Stage Structure<br />

Figure 3 shows the structure <strong>of</strong> an analog pipeline stage. It consists <strong>of</strong> a Flash A/D<br />

converter, a D/A converter, and a residue amplifier. The unit gain buffer is needed because<br />

<strong>of</strong> the poor output driving capability <strong>of</strong> the DAC topology used (Section 3.1)<br />

The DC transfer characteristic <strong>of</strong> the 4-<strong>bit</strong> DAC (Figure 4) has been chosen such that the<br />

(ideal) residue voltage always stays within +/- 0.5 LSB.<br />

Analog 4−Bit Pipeline Stage<br />

4−Bit Flash<br />

A/D <strong>Converter</strong><br />

15<br />

4−Bit<br />

D/A <strong>Converter</strong><br />

Unit Gain<br />

Buffer<br />

Residue Amplifier<br />

and<br />

Sample & Hold<br />

8x<br />

Clock<br />

to Encoder<br />

Clock<br />

Figure 3:<br />

Structure <strong>of</strong> an Analog Pipeline Stage<br />

1.4.3 <strong>Pipelined</strong> A/D Conversion and Digital Error Correction<br />

Digital error correction by <strong>bit</strong> overlapping uses an extra <strong>bit</strong> per stage to detect possible overor<br />

underflow <strong>of</strong> the amplified residue from the previous stage. Error correction thus<br />

requires halving the input range <strong>of</strong> stages 2, 3, and 4 to 500 mV peak-to-peak. Each stage<br />

except the first one only contribute 3 <strong>bit</strong>s to the digital output code.<br />

In the absence <strong>of</strong> any other error sources, <strong>bit</strong> overlapping allows for a +/-0.5 LSB integral<br />

nonlinearity <strong>of</strong> the 4-<strong>bit</strong> A/D converter. It relaxes the requirements for comparator <strong>of</strong>fset<br />

and reference ladder precision. However, an effort must still be made to keep <strong>of</strong>fset and<br />

reference ladder errors small in order to keep the <strong>bit</strong>-overlapping as a “last resort”. Note<br />

that gain and linearity errors in the D/A converter and the residue amplifier are not<br />

corrected by the <strong>bit</strong> overlapping.<br />

Two different types <strong>of</strong> encoders are needed for this error correction scheme: the first stage<br />

has a normal thermometric to binary encoder, all following stages have use smart encoders<br />

to detect over- or underflow conditions (Table 2). The shaded cells highlight the codes<br />

resulting from over- or underflow <strong>of</strong> the amplified residue.<br />

6


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

vout<br />

+1V<br />

1111<br />

1110<br />

1101<br />

1100<br />

1011<br />

1010<br />

1001<br />

−1V<br />

1000<br />

0111<br />

+1V<br />

vin<br />

0110<br />

0101<br />

0100<br />

0011<br />

0010<br />

0001<br />

0000<br />

−1V<br />

Figure 4:<br />

DC Characteristic <strong>of</strong> 4-<strong>bit</strong> Flash ADC and DAC connected in series<br />

Table 2: Thermometric to binary code conversion table<br />

Thermometer code<br />

(decimal representation)<br />

Binary code<br />

(stage 1 encoder)<br />

Binary code <strong>of</strong><br />

“smart” encoders<br />

(stages 2-4)<br />

Overflow<br />

(a)<br />

Underflow<br />

(b)<br />

15 1111 011 1 0<br />

14 1110 010 1 0<br />

13 1101 001 1 0<br />

<strong>12</strong> 1100 000 1 0<br />

11 1011 111 0 0<br />

10 1010 110 0 0<br />

9 1001 101 0 0<br />

8 1000 100 0 0<br />

7 0111 011 0 0<br />

6 0110 010 0 0<br />

5 0101 001 0 0<br />

7


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

Table 2: Thermometric to binary code conversion table<br />

Thermometer code<br />

(decimal representation)<br />

Binary code<br />

(stage 1 encoder)<br />

Binary code <strong>of</strong><br />

“smart” encoders<br />

(stages 2-4)<br />

Overflow<br />

(a)<br />

Underflow<br />

(b)<br />

4 0100 000 0 0<br />

3 0011 111 0 1<br />

2 0010 110 0 1<br />

1 0001 101 0 1<br />

0 0000 100 0 1<br />

Bubble errors in the thermometric code (due to metastability or noise in the comparators for<br />

example) are not corrected, but could cause gross errors (including short circuit) depending<br />

on the encoder. The encoder implementation proposed in [4] could cause short circuits by<br />

connecting a <strong>bit</strong> line simultaneously to Vdd and ground because more than one decoder<br />

row is activated.<br />

1.4.4 Analysis <strong>of</strong> accuracy requirements<br />

Thanks to <strong>bit</strong> overlapping the Flash ADC theoretically only needs to be 4 <strong>bit</strong> accurate. The<br />

DAC and the Residue Amplifier, however, need full <strong>12</strong>-<strong>bit</strong> accuracy, at least in first stage.<br />

Because all stages use the same building blocks, the blocks all have to fulfill the precision<br />

requirements <strong>of</strong> the first stage.<br />

Errors are most significant in first stage: DAC errors in the first stage are amplified 8 3 = 5<strong>12</strong><br />

times before reaching the last pipeline stage. There they should still be smaller than 0.5 LSB<br />

<strong>of</strong> the Flash converter. The error at the input <strong>of</strong> the first stage’s Residue Amplifier thus has<br />

to be < 0.24 mV. Note that the last stage has an LSB <strong>of</strong> 250 mV (instead <strong>of</strong> <strong>12</strong>5 mV like the<br />

other stages) because the LSB <strong>of</strong> its 3 output <strong>bit</strong>s can be discarded for a <strong>12</strong>-<strong>bit</strong> output (see<br />

Figure 1).<br />

The maximum allowable gain error <strong>of</strong> the first stage Residue Amplifier can be estimated as<br />

follows: Everything in the converter is assumed ideal except for the first stage Residue<br />

Amplifier which has a gain <strong>of</strong> 8(1+ε). In the worst case the residue will be 0.5 LSB = 62.5 mV.<br />

Hence 8*62.5*ε*8 2 < <strong>12</strong>5 mV => ε < 2/8 3 = 0.004. The maximum allowable gain error is thus<br />

estimated to be a few per mils.<br />

In the presented design, calibration is only used where it is easily implementable and does<br />

not add much circuit complexity, or need many additional I/O pins. A continuous<br />

calibration feedback is used to adjust DAC gain, but there is no calibration mechanism for<br />

the Residue Amplifier gain.<br />

1.4.5 Clocking scheme<br />

The clocking scheme is kept as simple as possible. Because <strong>of</strong> switched-capacitor Residue<br />

Amplifier, more clock phases are needed than in the single-ended design in [4]. Figure 5<br />

shows the clocking scheme <strong>of</strong> the pipeline stage. The signal names refer to the clock phases<br />

used in the comparator (Figure 11) and the Residue Amplifier (Figure 26).<br />

8


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

input/output stable<br />

input/output stable<br />

PHI2<br />

comprator<br />

reset<br />

1a,b<br />

sample input<br />

sample input<br />

R<br />

residue amp<br />

reset<br />

2a<br />

2b<br />

sample DAC output<br />

t=0<br />

DAC settled<br />

t=Ts<br />

SR latch switched<br />

Residue Amp settled<br />

time<br />

Figure 5:<br />

Clocking Diagram for Pipeline Stages<br />

There are three critical phases that cannot overlap and that have to fit inside the 5 ns<br />

sampling interval: (1) reset <strong>of</strong> the Residue Amplifier, (2) sampling <strong>of</strong> the DAC output and<br />

settling <strong>of</strong> the Residue Amplifier output, and (3) sampling and settling <strong>of</strong> input voltage. The<br />

input to the stage (and thus to the comparators) can change as soon as the SR-latches <strong>of</strong> the<br />

comparators (Figure 11) have switched. The comparator will be fully unbalanced, and its<br />

stage can only change after it has been reset.<br />

Four external triggers are needed, all other clocks can be triggered by other clock edges<br />

(indicated in Figure 5 by dashed arrows): the beginning <strong>of</strong> the regeneration phase <strong>of</strong> the<br />

comparator, the end <strong>of</strong> the sampling <strong>of</strong> the input voltage, the end <strong>of</strong> the Residue Amplifier<br />

reset, and the end <strong>of</strong> the DAC output sampling. Making these four events evenly spaced<br />

allows generating all clocks from two 90-degree shifted 200MHz clocks.<br />

Note that the comparator reset timing is not critical. To reduce possible hysteresis due to<br />

incomplete reset between two cycles, the comparator is reset for as long as possible.<br />

<strong>Design</strong>ing a self-contained clock generation circuit (using a PLL or DLL) is not necessary for<br />

the first prototype. In fact, inputting two shifted 200 MHz clocks from the outside gives<br />

more degrees <strong>of</strong> freedom (frequency, duty cycle) on the timing <strong>of</strong> the internal clock signals.<br />

2 4-<strong>bit</strong> Flash Analog-to-Digital <strong>Converter</strong><br />

2.1 4-<strong>bit</strong> Flash A/D <strong>Converter</strong><br />

A Flash topology is used for the 4-<strong>bit</strong> ADC as this topology is very fast and quite compact<br />

for a small number <strong>of</strong> <strong>bit</strong>s. An N-<strong>bit</strong> Flash ADC needs 2 N -1 comparators, hence, for 4 <strong>bit</strong>s,<br />

15 comparators are needed. The 15 differential reference voltages are generated using a<br />

9


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

resistive ladder as shown in Figure 6 [6]. N+Polysilicon resistors are used for the resistor<br />

vin1<br />

vin2<br />

vref1<br />

vref2<br />

VBIAS<br />

PHI2<br />

PHI1<br />

Q15<br />

Q15<br />

Q14<br />

Q14<br />

Q13<br />

Q13<br />

Q<strong>12</strong><br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

R R R R<br />

R R R<br />

R<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

Q1<br />

Q1<br />

Q2<br />

Q2<br />

Q3<br />

Q3<br />

Q4<br />

Q<strong>12</strong><br />

vref+<br />

vref−<br />

vref+<br />

vref−<br />

Q4<br />

Q11<br />

Q11<br />

Q10<br />

Q10<br />

Q9<br />

Q9<br />

PHI1<br />

PHI2<br />

VBIAS<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

R<br />

R R<br />

R<br />

R R<br />

R R<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

vin+<br />

vin−<br />

vref+<br />

vref−<br />

Q5<br />

Q5<br />

Q6<br />

Q6<br />

Q7<br />

Q7<br />

Q8<br />

Q8<br />

Figure 6:<br />

Flash <strong>Converter</strong><br />

ladder because they provide reasonable matching and linearity.<br />

The resistor ladder value has been set to 500 Ω, resulting in a static current <strong>of</strong> <strong>12</strong>5 µA. A large<br />

resistor area (W*L) improves matching and helps stabilize the reference voltages thanks to<br />

the large parasitic capacitance. The recovery time <strong>of</strong> the reference voltage nodes should be<br />

short enough for the nodes potentials to fully recover from coupling noise between two<br />

sampling instants.<br />

Two external pins (vref_plus and vref_minus) are used to define signal range.<br />

10


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

output code<br />

15<br />

14<br />

13<br />

<strong>12</strong><br />

11<br />

ideal threshold<br />

10<br />

−1V<br />

levels<br />

9<br />

8<br />

+1V<br />

7<br />

vin<br />

6<br />

5<br />

4<br />

3<br />

ideal code<br />

midpoints<br />

2<br />

1<br />

0<br />

Figure 7:<br />

Ideal DC transfer characteristic <strong>of</strong> the 4-<strong>bit</strong> Flash ADC<br />

2.1.1 Flash ADC Floorplan<br />

Figure 8 shows the floorplan <strong>of</strong> the 4-<strong>bit</strong> Flash ADC. The folded arrangement <strong>of</strong> resistor<br />

ladder causes mismatches to be symmetrical to the input range midpoint (first and last<br />

resistor etc. are closely matched because adjacent). The floorplan indicates a possible way<br />

<strong>of</strong> laying out and connecting the 16 reference ladder resistors. The proposed arrangement<br />

<strong>of</strong> 32 resistor elements can be made wide to stack to about the same height as the 15<br />

comparators (comparator height is 15µm). Wider resistors will improve matching and the<br />

increased parasitic capacitance will make the nodes less prone to capacitive coupling.<br />

2.2 <strong>Differential</strong> Comparator<br />

A differential comparator compares a differential input voltage to a differential reference<br />

voltage, i.e. it implements the following inequality:<br />

( v in1 – v in2 ) v ref 1 – v ref 2<br />

– ( ) > 0<br />

(2)<br />

One state <strong>of</strong> the binary comparator output will indicate that (2) valuates to true, the other<br />

one that it evaluates to false.<br />

2.2.1 Choice <strong>of</strong> Topology<br />

Inequality (2) indicates that a differential comparator can be built as a differencing circuit<br />

followed by a single-ended comparator.<br />

11


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

reference input<br />

analog input<br />

01<br />

01<br />

Resistor ladder<br />

01<br />

01<br />

00 11<br />

00 11<br />

00 11<br />

00<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

reference voltages<br />

11<br />

Comparator 1<br />

Comparator 15<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

Comparator 2<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

01<br />

01<br />

00 11<br />

00 11<br />

00 11<br />

01<br />

01<br />

01<br />

01<br />

01<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

Comparator 14<br />

01<br />

01<br />

01<br />

01<br />

01<br />

01<br />

Comparator 7<br />

digital outputs to encoder and DAC<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

Comparator 9<br />

00 11<br />

00 11<br />

01<br />

01<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

Comparator 8<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

00 11<br />

Clock<br />

Figure 8:<br />

Floorplan <strong>of</strong> 4-<strong>bit</strong> Flash ADC<br />

A topology based on two cross-coupled flip-flops [11] and two differential pairs [<strong>12</strong>] has<br />

finally been chosen (briefly mention other topologies that have been considered?). The cross<br />

coupled flip-flops (Figure 9(b)) provide fast decision: The small input difference (output <strong>of</strong><br />

the differencing circuit) is quickly regenerated to a rail-to-rail signal by the high (nonlinear)<br />

gain <strong>of</strong> the regeneration flip-flops. The two differential pairs (Figure 9(a)) implement a<br />

(nonlinear) differential difference amplifier:<br />

f( I a , v in1 – v ref 1 ) – f( I b , v in2 – v ref 2 ) > 0<br />

(3)<br />

Because f(I,∆v) is monotonic in ∆v (and in this particular case also I), (2) and (3) always<br />

evaluate to the same logic value if I a and I b are the same.<br />

Note that (2) can also be written as<br />

( v in1 – v ref 1 ) –( v in2 – v ref 2 ) > 0<br />

(4)<br />

The second view (4) is better for the input range requirements <strong>of</strong> the differential pairs. When<br />

crossing thresholds, the differential inputs should be as close as possible to the origin <strong>of</strong> the<br />

transfer curves <strong>of</strong> the differential pairs because there the gain is highest, leading to smaller<br />

<strong>12</strong>


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

resolvable voltage differences. Far away from the origin, the differential gain <strong>of</strong> the pairs<br />

goes to zero, no decision can be made.<br />

Vdd<br />

Vdd<br />

phi1<br />

phi1<br />

iout1<br />

iout2<br />

vout1<br />

vout2<br />

S<br />

phi1<br />

R<br />

Q<br />

vin1<br />

vref1<br />

vin2<br />

vref2<br />

iin1<br />

iin2<br />

phi2<br />

Q<br />

vbias<br />

(a) (b) (c)<br />

Figure 9:<br />

Comparator Elements: NMOS Input Pairs (a), Regenerative Flip-Flops (b), SR-Latch (c)<br />

A topology consisting <strong>of</strong> the three parts shown in Figure 9 was finally adopted: (a) two<br />

differential input pairs (differencing circuit), (b) a clocked cross-coupled latch (the actual<br />

comparator), and (c) an SR-latch to hold the comparator output until the next clock cycle.<br />

Three different versions <strong>of</strong> the comparator were examined. All have NMOS input pairs<br />

since PMOS transistors would have to be very large to allow a 0.6 to 1.6 V input range.<br />

1. PMOS pull-up, inverters, NAND-based SR latch<br />

2. current mirrors, NMOS pull-down, NAND-based SR latch<br />

3. PMOS pull-up, NOR-based SR latch<br />

NMOS pull-down (for setting or resetting the SR-latch) provides faster response time than<br />

PMOS pull-up. Inverters at the output for the regenerative flip-flops add delay but at the<br />

same time buffer the cross-coupled latch’s output. Finally, the NAND-based SR latch is<br />

faster than NOR-based equivalent because the NOR version has PMOS transistors in series,<br />

while in the NAND version, the NMOS transistors are stacked; also, the NOR causes low<br />

output crossing point, which is not useful when using NMOS current switches in the DAC<br />

(see Section 3.1).<br />

Version 2 <strong>of</strong> the comparator proved to perform the best. A useful side effect <strong>of</strong> mirroring the<br />

current from the differencing circuit before injecting it into the regeneration stage is that<br />

there is less switching induced noise injected into reference ladder.<br />

2.2.2 Comparator circuit<br />

The comparator circuit in Figure 11 works as follows: During reset, the switches M5a and<br />

M5b disconnect S and R <strong>of</strong> the SR latch from the sensing nodes (aa and bb). The inputs <strong>of</strong> the<br />

latch are pulled up to Vdd by M6a and M6b, causing the latch to keep it’s state. M7 is closed,<br />

equalizing the sensing node voltages. A mismatch between the two differential input<br />

voltages causes an unequal amount <strong>of</strong> current to be injected into the sensing nodes. When<br />

switch M7 is released, the first regeneration phase starts, and the small current imbalance<br />

will cause the cross coupled transistors M3a and M3b to pull down one <strong>of</strong> the sensing nodes.<br />

Then M5a and M5b are opened, and either SorR is pulled to ground, switching the state <strong>of</strong><br />

13


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

the SR-latch. The comparator can then be reset again without disturbing it’s output state.<br />

The two non-overlapping clock phases controlling the comparator are shown in Figure 10.<br />

Making the first regeneration phase longer speeds up the second phase as the second phase<br />

starts with larger difference voltage. However, since the first regeneration phase also adds<br />

to the total comparator response time, making it too long will actually slow down<br />

comparator response. A value <strong>of</strong> 200ps has been chosen.-<br />

Table 3: Transistor aspect ratios and fingering for the comparator circuit (Figure 11)<br />

Transistor W (total) [µm] L [µm] Number <strong>of</strong> Fingers a<br />

M0a, M0b, M0c, M0d 1.5 1 2<br />

M1a, M1b 6 2.5 4<br />

M2a, M2b, M2c, M2d 3.6 0.18 4<br />

M3a, M3b 3 0.18 2<br />

M4a, M4b 1 0.18 2<br />

M5a, M5b 1 0.18 2<br />

M6a, M6b 0.24 0.18 1<br />

M7 0.5 0.18 1<br />

M8a, M8b 2 0.18 1<br />

M9a, M9b 2.5 0.18 1<br />

M10a, M10b 3 0.18 1<br />

M11a, M11b 0.24 0.18 1<br />

a. number <strong>of</strong> fingers per transistor in layout <strong>of</strong> the comparator cell<br />

PHI1<br />

PHI2<br />

Reset<br />

100 ps<br />

200 ps<br />

5ns<br />

Figure 10:<br />

Comparator Timing<br />

The complete comparator circuit it given in Figure 11. A bias current <strong>of</strong> 10 µA has been<br />

chosen. Trade-<strong>of</strong>fs exist for the switch sizing: making M5a and M5b large helps pulling<br />

down the active branch quickly, but increases the glitch size when switching on. The size <strong>of</strong><br />

the transistors has to be kept small enough to prevent the glitches from feeding through the<br />

14


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

Vdd<br />

M2a<br />

M2b<br />

M2c<br />

M2d<br />

phi1<br />

M6a M4a<br />

M4b M6b<br />

phi1<br />

M10a M11a<br />

M11b M10b<br />

S<br />

a<br />

b<br />

bb<br />

aa<br />

M5a<br />

phi1<br />

M5b<br />

R<br />

Q<br />

Q<br />

vin1<br />

M0a<br />

M0b<br />

vref1<br />

vin2<br />

M0c<br />

M0d<br />

vref2<br />

phi2<br />

M9a<br />

M9b<br />

M7<br />

vbias<br />

M1a<br />

M1b<br />

M3a<br />

M3b<br />

M8a<br />

M8b<br />

Figure 11:<br />

Circuit schematic <strong>of</strong> the fully differential comparator<br />

SR-latch to the comparator output. The size <strong>of</strong> the resetting switch influences the required<br />

resetting time. Since the clocking scheme <strong>of</strong> the pipeline (Figure 5) allows a long reset phase,<br />

this switch can be kept small, reducing parasitic capacitance on the regeneration nodes, and<br />

increasing the gain during the sensing phase (end <strong>of</strong> reset) because the current difference<br />

flowing through M7 causes a larger voltage imbalance <strong>of</strong> the sensing nodes.<br />

2.2.3 Comparator Layout<br />

The complete comparator layout is given in Appendix A. The silicon area is approximately<br />

450 µm 2 . To reduce parasitic capacitances <strong>of</strong> interconnects, minimum width metal lines have<br />

been used. Also, care has been taken to balance the parasitic caps <strong>of</strong> the sensing nodes.<br />

Unequal sensing node capacitance could lead to increased comparator <strong>of</strong>fset.<br />

2.3 Performance verification<br />

The extracted netlist <strong>of</strong> the comparator has been used to compare pre- and post-layout<br />

performance. The number <strong>of</strong> fingers in the comparator schematic is made equal to the<br />

effective number <strong>of</strong> fingers in the layout, so that the effect due to interconnect parasitics can<br />

be distinguished from effects due to different transistor fingering, which has a large impact<br />

on the source and drain junction capacitances.<br />

The three investigated simulation corners are given in Table 1. For the simulations, each<br />

comparator output has been loaded with 50 fF to simulate the input capacitance <strong>of</strong> the DAC<br />

current switches and the thermometric-to-binary encoder.<br />

Table 4: Simulation Corners for Performance Verification<br />

Worst Case Typical Case Best Case<br />

Technology Corner SS TT FF<br />

Temperature <strong>12</strong>5 C 27 C -25 C<br />

Vdd -10% nominal +10%<br />

Bias current -10% nominal +10%<br />

15


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

2.3.1 Comparator Performance<br />

Figure <strong>12</strong> to Figure 16 give an overview <strong>of</strong> the obtained simulation results. The comparator<br />

<strong>of</strong>fset plots reveal that a 1 ns reset time is too short, especially when layout parasitics are<br />

taken into account. Making the comparator reset as long as possible will fix the hysteresis<br />

problem. Figure 15 shows that already an extension to 1.5 ns almost completely removes the<br />

hysteresis for typical simulation conditions. Figure 16 shows that typical comparator<br />

response time is below 600 ps, even for input differences <strong>of</strong> only a few millivolts. In the<br />

worst case (Figure 17), the comparator does not respond within the 5ns clock period. Such<br />

a condition will lead to <strong>bit</strong> errors in the Flash output.<br />

The mean comparator power consumption under typical conditions is about 140 µW.<br />

15<br />

10<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

Schematic, Typical case<br />

20<br />

15<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

Post−layout, Typical case<br />

10<br />

5<br />

5<br />

Offset [mV]<br />

0<br />

Offset [mV]<br />

0<br />

−5<br />

−5<br />

−10<br />

−10<br />

−15<br />

−15<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−20<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure <strong>12</strong>:<br />

Comparator Offsets in Typical Case (1 ns Reset Time)<br />

30<br />

20<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

Schematic, Worst case<br />

50<br />

40<br />

30<br />

Post−layout, Worst case<br />

Offset [mV]<br />

10<br />

0<br />

Offset [mV]<br />

20<br />

10<br />

0<br />

−10<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

−10<br />

−20<br />

−20<br />

−30<br />

−40<br />

−30<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−50<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure 13:<br />

Comparator Offsets in Worst Case (1 ns Reset Time)<br />

16


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

6<br />

4<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

Schematic, Best case<br />

2<br />

1.5<br />

Post−layout, Best case<br />

1<br />

2<br />

Offset [mV]<br />

0<br />

Offset [mV]<br />

0.5<br />

0<br />

1.1 V<br />

−2<br />

−0.5<br />

−4<br />

−1<br />

−6<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−1.5<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure 14:<br />

Comparator Offsets in Best Case (1 ns Reset Time)<br />

15<br />

10<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

Post−layout, Typical case, 1.5 ns reset<br />

25<br />

20<br />

15<br />

Post−layout, Worst case, 1.5 ns reset<br />

5<br />

10<br />

5<br />

Offset [mV]<br />

0<br />

Offset [mV]<br />

0<br />

−5<br />

−5<br />

−10<br />

−10<br />

−15<br />

−20<br />

1.0 V<br />

1.05 V<br />

1.1 V<br />

1.15 V<br />

1.2 V<br />

−15<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−25<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure 15:<br />

Comparator Offset Improvement for extended reset time <strong>of</strong> 1.5 ns<br />

2.3.2 Flash ADC Performance<br />

Mismatch in the resistor ladder has been modeled by using normally distributed resistor<br />

values with mean 500 Ω and 5% standard deviation (σ). The resistor matching report<br />

indicates that N+Poly resistor values will have a σ <strong>of</strong> less than 1%. We used 5% to<br />

compensate for the fact that this model assumes that the values <strong>of</strong> adjacent resistors are<br />

uncorrelated, which is certainly not true on silicon.<br />

Resistor ladder mismatch (σ = 5%) has not been included in the INL and DNL plots, as it is<br />

dominating the INL and DNL due to the comparators. Its effect on INL and DNL is shown<br />

as dashed lines in Figure 18.<br />

To construct the DNL and INL plots <strong>of</strong> the 4-<strong>bit</strong> Flash the mean <strong>of</strong> the rising and falling<br />

<strong>of</strong>fset with 1 ns reset time has been used as effective <strong>of</strong>fset. The hysteresis is assumed to<br />

introduce a constant difference between rising and falling threshold. The mean value <strong>of</strong> the<br />

17


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

650<br />

Comparator response time (Typical Case)<br />

−0.875 mV<br />

0 mV<br />

0.875 mV<br />

<strong>12</strong>00<br />

1150<br />

Comparator response time (Worst Case)<br />

−0.875 mV<br />

0 mV<br />

0.875 mV<br />

1100<br />

600<br />

1050<br />

Response time [ps]<br />

Response time [ps]<br />

1000<br />

950<br />

900<br />

550<br />

850<br />

800<br />

750<br />

500<br />

−100 −50 0 50 100<br />

∆in [mV]<br />

700<br />

−100 −50 0 50 100<br />

∆in [mV]<br />

Figure 16:<br />

Comparator Response Time for post-layout simulations (1.5 ns Reset Time)<br />

two <strong>of</strong>fsets represents the comparator <strong>of</strong>fset for a long enough reset phase). The input<br />

common mode voltage has been set to 1.1V.<br />

8 x 10−3 Flash DNL (schematic simulation)<br />

0.5 x 10−3 Flash INL (schematic simulation)<br />

6<br />

0<br />

4<br />

2<br />

−0.5<br />

Typical case<br />

Worst case<br />

Best case<br />

DNL [LSB]<br />

0<br />

INL [LSB]<br />

−1<br />

−2<br />

−1.5<br />

−4<br />

−6<br />

Typical case<br />

Worst case<br />

Best case<br />

−2<br />

−8<br />

2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−2.5<br />

2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure 17:<br />

4-<strong>bit</strong> Flash DNL and INL (schematic simulations)<br />

Figure 19 shows a Signal-to-Noise Ratio (SNR) plot for the 4-<strong>bit</strong> Flash. It has been obtained<br />

by simulating the Flash converter with the extracted comparator netlist. An 11 MHz<br />

sinusoidal input signal is sampled at 200 MHz. The dashed line indicates the maximum<br />

SNR that can be obtained by a 4-<strong>bit</strong> ADC. This limit is imposed by the quantization noise.<br />

The simulated SNR should approximately stay constant for input signal frequencies up to<br />

the Nyquist rate. Figure 19 shows a significantly lower SNR already for frequencies well<br />

below the Nyquist rate. This is probably due to the very small number <strong>of</strong> ADC output<br />

samples (91) that have been used to calculate the SNR. A small number <strong>of</strong> samples has been<br />

taken because <strong>of</strong> long simulation times.<br />

18


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

0.06<br />

0.05<br />

0.04<br />

Flash DNL (post layout simulation)<br />

Typcial case<br />

Worst case<br />

Best case<br />

Resistor ladder<br />

0.05<br />

0.04<br />

0.03<br />

Flash INL (post layout simulation)<br />

0.03<br />

0.02<br />

0.02<br />

0.01<br />

DNL [LSB]<br />

0.01<br />

INL [LSB]<br />

0<br />

0<br />

−0.01<br />

−0.02<br />

−0.01<br />

−0.02<br />

−0.03<br />

Typcial case<br />

Worst case<br />

Best case<br />

Resistor ladder<br />

−0.03<br />

−0.04<br />

−0.04<br />

2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−0.05<br />

2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure 18:<br />

4-<strong>bit</strong> Flash DNL and INL (post-layout simulation)<br />

To find the maximum sampling frequency <strong>of</strong> the 4-<strong>bit</strong> DAC, the sampling frequency has to<br />

be swept for a fixed input frequency (which has to be lower than half the lowest sampling<br />

frequency tested). The sampling frequency at which the SNR decreases by 3 dB can be<br />

regarded as the converter’s maximum sampling speed. Because the comparator response<br />

time (Figure 16) and reset time sum up to almost 2 ns, this maximum frequency is expected<br />

to be around 500 MHz. In the pipeline, however, only the response time is critical, as the<br />

comparators can be reset while the DAC and the Residue Amplifier are working (Figure 5).<br />

For a more complete characterization <strong>of</strong> the converter, SNR as a function <strong>of</strong> input signal<br />

amplitude (at constant input and sampling frequencies) does also have to be simulated.<br />

These simulations have not yet been done.<br />

26<br />

SNR <strong>of</strong> Flash ADC<br />

25<br />

SNR <strong>of</strong> ADC<br />

Quantization Noise Limit<br />

24<br />

SNR [dB]<br />

23<br />

22<br />

21<br />

20<br />

19<br />

11 22 33 44 55 66 77 88 99<br />

Input frequency [MHz]<br />

Figure 19:<br />

4-<strong>bit</strong> Flash SNR for Typical case (post-layout)<br />

19


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

3 4-<strong>bit</strong> Digital-to-Analog <strong>Converter</strong><br />

Because there are no accurate capacitors available, the DAC is not implemented as an<br />

MDAC as in [4], but as a current-steering DAC. Using a current steering DAC will greatly<br />

increase the power consumption <strong>of</strong> the converter compared to an implementation using a<br />

capacitive MDAC (which has no static current consumption).<br />

Matching <strong>of</strong> unit-current cells is very critical because DAC linearity directly depends on the<br />

matching <strong>of</strong> the unit currents. Special care has to be taken when layout out the current cells<br />

(see Section 3.1.2). Since the gain is controlled by the absolute values <strong>of</strong> the unit current and<br />

the load resistors, a calibration feedback is absolutely needed.<br />

Using non-weighted current cells simplifies design as the flash output can directly be used<br />

for controlling the current cells. Depending on the current cell, a simple deglitching circuit<br />

may have to be employed: when switching the current between the output branches<br />

(Figure 20), there must always be a path for the current drawn by the current source<br />

transistor. If the current path is blocked, the transistor will leave saturation; reestablishing<br />

the current will then take some time, causing a glitch in the output voltage.<br />

The DAC output needs buffering because it must drive the large input capacitance <strong>of</strong> the<br />

Residue Amplifier. The buffer needs precise gain <strong>of</strong> 1 and large output swing. Since the<br />

OTA(s) in the buffer will be used in unit gain configuration, the OTA(s) will also need large<br />

input signal swing.<br />

A resistor ladder DAC was also tested, but then rejected because <strong>of</strong> insufficient resistor<br />

matching accuracy, and because it would have needed to introduce an encoder into the<br />

analog signal path, thus increasing delay.<br />

3.1 Current-steering D/A <strong>Converter</strong><br />

The schematic <strong>of</strong> the current-steering DAC is given in Figure 20. It’s output voltage is given<br />

by:<br />

v od = v out1 – v out2 = RI 0 ( 2n–<br />

N)<br />

v ------------------------------ v out1 + v out2<br />

oc 2<br />

V NI +<br />

= = – ---------------------------<br />

2RI 0 1<br />

dd 2<br />

(5)<br />

(6)<br />

where n is the thermometric output value <strong>of</strong> the Flash, R the load resistance, I 0 the unit<br />

current, and N = 15. I 1 is a DC current used to adjust the common mode level <strong>of</strong> the output<br />

voltage. R and I 0 have to be chosen to obtain the required gain. The values used here are R<br />

= <strong>12</strong>50 Ω, I 0 = 50 µA, and I 1 = 185 µA.<br />

3.1.1 Continuous DAC Gain Calibration<br />

Continuous calibration only needs to regulate a DC level and can thus be slow. However,<br />

the slow feedback will also take a long time to recover from glitches injected on the reference<br />

node.<br />

Calibration is done using a unit current cell and dummy resistor identical to load resistors<br />

to set the nominal voltage <strong>of</strong> the vref control pin to Vdd-0.5LSB (1.7375 V). Due to voltage<br />

20


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

R dummy<br />

Unit Current Cells<br />

R<br />

R<br />

vout1<br />

vout2<br />

vref<br />

+<br />

−<br />

1 0<br />

I0<br />

va<br />

000000000<br />

111111111<br />

Q1<br />

I0<br />

Q1<br />

Q2<br />

I0<br />

Q2<br />

0000000000<br />

1111111111 000000000<br />

111111111<br />

1100<br />

01<br />

Q15<br />

I0<br />

Q15<br />

I1<br />

I1<br />

Figure 20:<br />

Current-steering DAC with continuous gain calibration<br />

drops in the power rails and feedback amplifier <strong>of</strong>fset, this voltage will have to be adjusted<br />

to obtain a gain <strong>of</strong> precisely 8.<br />

No high gain needed in the feedback amplifier since Vdd seen by dummy resistor is<br />

different from the external Vdd due to resistive voltage drops. Vref will have to be adjusted<br />

from the outside anyway, so static <strong>of</strong>fset at amplifier input not a problem. Because the inputs<br />

<strong>of</strong> the amplifier are very close to Vdd, an NMOS input pair with folded cascode has been<br />

chosen, which does not need a diode connected transistor as load at drains <strong>of</strong> input<br />

transistors. Only one stage is used to simplify stabilizing the feedback loop: The current<br />

source gate node va has a large capacitive load (4-5pF), and thus has to be the dominant pole<br />

node. A two stage folded cascode opamp would provide higher gain but would be hard to<br />

compensate.<br />

M4<br />

M7a<br />

M7b<br />

M8a<br />

vcasc<br />

M8b<br />

vout<br />

Ibias<br />

vin<br />

M0a<br />

M0b<br />

vref<br />

M6a<br />

vcasc<br />

M6b<br />

M3<br />

M2<br />

M1<br />

M5a<br />

M5b<br />

Figure 21:<br />

Feedback amplifier for continuous DAC calibration<br />

21


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

A bias current <strong>of</strong> 10µA is used. The bias voltage vcasc can be generated using a simple MOS<br />

Table 5: DAC calibration feedback amplifier transistor sizes<br />

Transistor W (total) [µm] L [µm]<br />

M0a, M0b 10 0.5<br />

M1 6 0.5<br />

M2 1.5 0.5<br />

M3 1.5 0.5<br />

M4 2 0.18<br />

M5a, M5b 1 0.18<br />

M6a, M6b 2.5 0.5<br />

M7a, M7b 8 0.18<br />

M8a, M8b 10 0.5<br />

(diode-connected) voltage divider. Varying vcasc by 10% does not degrade performance <strong>of</strong><br />

the calibration feedback loop.<br />

3.1.2 DAC Floorplan<br />

The current source transistor array can be laid out compactly because all transistors share<br />

the same source and gate node. The array is laid out such that transistors 1 through 16 all<br />

share the same geometry centroid (common centroid layout).<br />

The regulated cascode feedback and the current switch for each cell is put outside the<br />

matched capacitor array (Cells 1 to 15). This way the matched transistors can be put together<br />

very closely, improving matching. On the other hand, the digital inputs from the Flash ADC<br />

do not need to be routed across the analog parts <strong>of</strong> the DAC. Because the RGC feedback<br />

regulates the voltage at the source <strong>of</strong> the cascoding transistor, the voltage drop between the<br />

source <strong>of</strong> the cascode transistors and the drains <strong>of</strong> the current-source transistors should be<br />

matched, i.e. the routing collecting the current from the transistor array will have to be<br />

balanced.<br />

Once the exact size <strong>of</strong> the layout <strong>of</strong> one current cell (feedback and switch) is known, the<br />

floorplan may have to be adjusted.<br />

3.2 Unit current cell<br />

The unit current cell must have a minimum output resistance in the order <strong>of</strong> 100 MΩ, and<br />

provide fast switching without introducing large current glitches. The following sections<br />

discuss some switching current cells that have been examined.<br />

3.2.1 Fournier-Senn [15] Current cell<br />

The advantages <strong>of</strong> this cell topology is it’s speed and low minimum output voltage. Because<br />

the current switch transistors also serve as cascoding transistors only two transistors need<br />

22


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

digital<br />

inputs<br />

Cell 1<br />

Cell 2<br />

Cell 3<br />

Cell 4<br />

Cell 15 Cell 14<br />

Cell 13 Cell <strong>12</strong><br />

1<br />

2 3 4 5 6 7 8<br />

Load resistors<br />

2 1 4<br />

3 6 5 8 7<br />

9 10 11 <strong>12</strong> 13 14 15 16<br />

10 9 <strong>12</strong> 11 14 13 16 15<br />

15 16 13 14 11 <strong>12</strong> 9 10<br />

analog<br />

output<br />

16 15 14 13 <strong>12</strong> 11 10 9<br />

7 8 5 6 3 4 1 2<br />

8 7 6 5 4 3 2 1<br />

Cell 8 Cell 9 Cell 10 Cell 11<br />

digital<br />

inputs<br />

Ref. Cell<br />

Cell 7<br />

Cell 6<br />

Cell 5<br />

Figure 22:<br />

Floorplan <strong>of</strong> 4-<strong>bit</strong> current-steering DAC<br />

S<br />

S<br />

vb<br />

vb<br />

Q<br />

Q<br />

Q<br />

va<br />

Q<br />

Figure 23: Current-cell based on the configuration proposed in [15]<br />

to be stacked. The drawback is that this cell needs a deglitching circuit to generate the<br />

control signals from the Flash output. Also, the required 100 MΩ output resistance could not<br />

be reached with a simple cascoding <strong>of</strong> two transistors.<br />

23


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

3.2.2 Regulated Cascode Current Cell<br />

To improve the output resistance <strong>of</strong> the current cell we tried to include a local feedback in<br />

the Fournier-Senn current cell (Figure 23). This failed because the feedback circuit has to<br />

charge cascode transistor gate capacitance to turn on the switch, as the gate is always<br />

completely discharged for switching <strong>of</strong>f. This makes the feedback slow and causes long a<br />

settling time <strong>of</strong> after switching the current. Also, the feedback aggravated the glitch<br />

problem <strong>of</strong> the current cell.<br />

out1<br />

out2<br />

Q<br />

M2a<br />

M2b<br />

Q<br />

vbias<br />

M2<br />

M4<br />

vref +<br />

−<br />

M1<br />

vin−<br />

M0a<br />

M0b<br />

vin+<br />

vout<br />

va<br />

M0<br />

M1a<br />

M1b<br />

M3<br />

Figure 24:<br />

(a) Unit current cell<br />

Regulated cascode unit current cell<br />

(b) RGC feedback amplifier<br />

Next, the regulated cascode principle has been applied to the simple cascode current source<br />

with a stacked current switch (Figure 24). This circuit has finally been adopted although it<br />

is slower than Fournier-Senn current cell. Simulations show that output resistances <strong>of</strong> 100<br />

MW and more can easily achieved with this circuit.<br />

The circuit is slower for low output voltages because the RGC feedback needs to make a<br />

large output excursion to regulate current when output voltage is close to 0.6 V.<br />

3.3 Simulated Performance<br />

Figure 25 shows the simulated DNL and INL plots for the DAC using the RGC current<br />

source described in the precious section. The load on the DAC output nodes is 100 fF.<br />

Simulations show that the worst case settling time is 2.3 ns, and the largest glitch size at the<br />

output is <strong>12</strong> mV.<br />

4 Residue Amplifier and Sample-and-Hold<br />

Instead <strong>of</strong> making the Residue Amplifier and the interstage sample-and-hold circuit two<br />

separate circuits, we chose to combine them in a single switched capacitor circuit. The idea<br />

is thus to build a sample-and-hold circuit with a gain <strong>of</strong> 8 and a differencing circuit at its<br />

input. This approach allows subtraction <strong>of</strong> the input voltage from the DAC output and<br />

amplification <strong>of</strong> the resulting residue in one step without using more matched capacitors<br />

than would be needed for a differential sample-at-hold, which in any case needs four<br />

matched elements.<br />

Matching <strong>of</strong> the four capacitors in the Residue Amplifier is extremely critical for accurate<br />

interstage gain. Special layout techniques will have to be employed achieve sufficient<br />

24


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

<strong>12</strong> x 10−4 DAC DNL<br />

10<br />

8<br />

Typical case<br />

Worst case<br />

Best case<br />

20 x DAC INL<br />

10−3<br />

Typical case<br />

Worst case<br />

Best case<br />

15<br />

INL [LSB]<br />

6<br />

4<br />

2<br />

DNL [LSB]<br />

10<br />

5<br />

0<br />

−2<br />

0<br />

−4<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

−5<br />

1 2 3 4 5 6 7 8 9 10 11 <strong>12</strong> 13 14 15<br />

Thermometric code<br />

Figure 25:<br />

DNL and INL <strong>of</strong> current-steering DAC<br />

matching. No gain calibration mechanism is implemented. A simple mechanism using<br />

capacitor banks would require too many external pins for the first prototype<br />

implementation <strong>of</strong> the converter stage.<br />

4.1 Switched Capacitor Residue Amplifier<br />

4.1.1 Circuit Topology<br />

R<br />

1b<br />

vbias<br />

v1<br />

1a<br />

00 11<br />

00 11<br />

01<br />

01<br />

vdac1<br />

2a<br />

01<br />

01<br />

C1a<br />

01<br />

01<br />

2b<br />

00 11<br />

00 11<br />

+<br />

C2a<br />

−<br />

01<br />

01<br />

vout1<br />

v2<br />

1b<br />

01<br />

01<br />

2b<br />

−<br />

+<br />

00 11 01<br />

vout2<br />

vdac2<br />

2a<br />

C1b<br />

1a<br />

00 11<br />

C2b<br />

01<br />

vbias<br />

R<br />

Figure 26:<br />

Topology <strong>of</strong> Switched-Capacitor Residue Amplifier and Sample-and-Hold Circuit<br />

The circuit topology in Figure 26 has been chosen because it allows holding the output<br />

voltage while sampling the first differential voltage. This is important because the output <strong>of</strong><br />

25


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

the Residue Amplifier is connected to the analog input <strong>of</strong> the next pipeline stage, while one<br />

<strong>of</strong> the inputs <strong>of</strong> the Residue Amplifier has to sample the output value <strong>of</strong> the previous stage.<br />

It it thus necessary to sample the input voltage while holding the output voltage constant<br />

for the next stage.<br />

The differential and common-mode output voltages under ideal conditions (no charge<br />

injection) are given in Equations (7) and (8). v 0 represents the analog ground voltage used<br />

while sampling v 1 and v 2 (vbias in Figure 26), and v 0 ’ represents the virtual analog ground<br />

voltages imposed by the negative feedback around the OTA. Ideally v0 and v0’ should be<br />

equal. Two separate voltages have been introduced for the derivation <strong>of</strong> Equations (7) and<br />

(8) in order to study the effect <strong>of</strong> mismatch between vbias and the OTA output common<br />

mode voltage.<br />

v od = v out1 – v out2 =<br />

C 1a<br />

C<br />

-------- 1b<br />

( v<br />

C 1 – v dac1 + v 0 ′ – v 0 )–<br />

-------- ( v<br />

2a<br />

C 2 – v dac2 + v 0 ′ – v 0 )<br />

2b<br />

(7)<br />

v<br />

v out1 + v out2 C 1a<br />

C<br />

oc ------------------------------ v<br />

2<br />

0 ′ ----------- 1b<br />

= = + ( v<br />

2C 1 – v dac1 + v 0 ′ – v 0 ) + ----------- ( v<br />

2a<br />

2C 2 – v dac2 + v 0 ′ – v 0 )<br />

2b<br />

(8)<br />

If capacitors C1a and C1b, and C2a and C2b are perfectly matched, Equations (7) and (8)<br />

simplify to the expressions given by (9) and (10). It is interesting to note that under these<br />

conditions, a mismatch between v 0 and v 0 ’ only has an influence on the common-mode<br />

output signal only. The differential output voltage is unaffected. This means that the<br />

common mode feedback <strong>of</strong> the OTA (see Section 4.2) doesn’t not need to align the output<br />

common mode to analog ground with very high precision. This greatly simplifies the<br />

high-speed common-mode feedback design, as high gain is not required.<br />

v od = v out1 – v out2 =<br />

C<br />

----- 1<br />

( v<br />

C 1 – v dac1 – v 2 + v dac2 )<br />

2<br />

(9)<br />

v<br />

v out1 + v out2 C 1<br />

oc ------------------------------ v<br />

2<br />

0 ′<br />

C ----- v1 + v v<br />

---------------- 2 dac1 + v<br />

= = + ⎛ – ------------------------------- dac2<br />

+ v<br />

2 2 2<br />

0 ′ – v ⎞<br />

⎝<br />

0 ⎠<br />

(10)<br />

The circuit in Figure 26 uses a conventional differencing circuit to form the difference <strong>of</strong> the<br />

two differential input voltages. As seen in (10), this circuit does not reject the differential<br />

common-mode input voltage (i.e. the difference between the common-modes <strong>of</strong> the two<br />

differential inputs). A configuration based on the differencing circuit proposed in [14]<br />

would provide better common-mode rejection but require more switches and clock phases.<br />

Because the common modes <strong>of</strong> the input signals are approximately aligned, and the<br />

following stage (Flash ADC) provides some common mode-rejection, common-mode<br />

rejection is not critical in the Residue Amplifier. The simple conventional circuit is thus<br />

used. The clocking scheme for this circuit is included in Figure 5.<br />

4.1.2 Charge Injection <strong>of</strong> MOS Switches<br />

Equations (7) and (8) are derived using charge conservation in the sampling capacitors. The<br />

switches are assumed not to absorb or release any charge when turned on or <strong>of</strong>f. For<br />

switches implemented with MOS transistors, however, this assumption is not true. The<br />

charge forming the transistor’s channel when in on-state is released when to source, drain<br />

and bulk (substrate) node <strong>of</strong> the transistor, potentially disturbing the voltage levels on the<br />

sampling capacitors. How the charge is distributed between these nodes depends on the<br />

26


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

impedance <strong>of</strong> these nodes seen by the transistor, the rise or fall time <strong>of</strong> the gate voltage, and<br />

on the source and drain potential <strong>of</strong> the switch. Charge injection is thus difficult to predict,<br />

and because it is signal dependent, it is difficult to cancel its effects even using fully<br />

differential circuits.<br />

Note that high speed (small time constant C/g) entails large amounts <strong>of</strong> injected charge<br />

because the amount <strong>of</strong> injected charge and the switch on-resistance are linked (Equation<br />

(11)).<br />

µQ<br />

g on = –-------<br />

L 2<br />

(11)<br />

Note that ideally, the switches should only introduce common mode charge injection errors<br />

which will be rejected by the input stage <strong>of</strong> the comparators <strong>of</strong> the next pipeline stage.<br />

However, because the amount <strong>of</strong> injected charge is signal dependent, differential errors will<br />

still occur.<br />

Techniques for charge injection cancellation:<br />

• Use shorted dummy transistors on high impedance node side to absorb the injected<br />

charge. The dummy transistor has to be turned on when the switch transistor is<br />

turned <strong>of</strong>f.<br />

• Use bootstrapped switches [8]: this makes the injected charge almost independent <strong>of</strong><br />

signal level. Thus, if the circuit topology is such that all injected charge results in a<br />

common mode error, this may allowing at least partial error cancellation; however,<br />

bootstrapped switches are complex to implement<br />

• Choose large capacitors to reduce effect <strong>of</strong> injected charge on voltage. However, to<br />

keep the speed <strong>of</strong> the circuit constant, this also requires scaling <strong>of</strong> the switch<br />

transistors, resulting in more injected charge.<br />

• Bottom plate sampling (series sampling): sample against a constant potential to<br />

reduce signal dependency <strong>of</strong> error [8].<br />

4.1.3 Capacitor and Switch Sizing<br />

From Equation 9 it can be seen that the ratio <strong>of</strong> C 1 /C 2 needs to be 8 for a gain <strong>of</strong> 8. The<br />

capacitors should be large enough to allow precise matching <strong>of</strong> C1 and C2, and small<br />

enough to keep the Residue Amplifier settling time short enough for 200 MHz sampling<br />

rate (ideally less than 2 ns only). A value <strong>of</strong> 100 fF has been chosen for C 1 , requiring in C 2 =<br />

800 fF.<br />

CMOS switches are used to implement all switches in Figure 26. This allows the switches to<br />

work properly over the whole signal range. The gate overdrive <strong>of</strong> the switch transistors in<br />

on-state is small due to the relatively low Vdd <strong>of</strong> 1.8 V. This results in large on-resistance <strong>of</strong><br />

the switches, which in turn requires the switch transistors to be large. Large transistors,<br />

however, will aggravate charge injection errors and increase the parasitic capacitances in the<br />

switched-capacitor circuit. Bootstrapped switches [8] could ease this problem but add<br />

complexity. The idea <strong>of</strong> bootstrapped switches is to boost the gate voltage <strong>of</strong> the switch<br />

transistor so its gate overdrive is always Vdd-V th when in on-state. This reduces the signal<br />

dependency <strong>of</strong> on-resistance and injected charge, and allows use <strong>of</strong> smaller switch<br />

transistors for a given required on-resistance (the amount <strong>of</strong> injected charge is not reduced<br />

by the bootstrapping technique, as can be seen from (1)).<br />

27


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

The effects <strong>of</strong> different capacitor and switch sizes will have to be investigated further, as<br />

these parameters are very critical for Residue Amplifier speed and accuracy.<br />

4.2 <strong>Differential</strong> OTA<br />

Switched capacitor residue amplifier needs fully differential Opamp with high GBW and<br />

slew rate. A class AB topology is thus used instead <strong>of</strong> instead <strong>of</strong> class A amplifier. However,<br />

the conventional class AB output stage push-pull source followers do not work because <strong>of</strong><br />

the low supply voltage. Since the nominal output range <strong>of</strong> the residue amplifier is limited<br />

to half the 1 V peak-to-peak range (due to <strong>bit</strong> overlapping), the differential OTA does not<br />

necessarily need a rail-to-rail output stage, but can use cascodes, allowing high gain and<br />

large GBW with few stages. However, especially PMOS cascodes are critical, since they need<br />

to be large because there is only about 400 mV headroom to Vdd, and since the carrier<br />

mobility in PMOS transistors is much lower than in NMOS transistors.<br />

Single stage amplifier are an interesting choice because they can have only one high<br />

impedance node at the output. The load capacitance will then slow down the dominant<br />

pole, improving stability. In a “classical” two stage op-amp, the load capacitance affects the<br />

non-dominant pole that one wishes to push to high frequency.<br />

To speed up a switched-capacitor circuit one has to increase both G m and I slew (or decrease<br />

C, which is bad for noise). An OTA GBW requirement estimation formula is presented see<br />

[8]. The simple first order model used there predicts a required GBW <strong>of</strong> around 10GHz,<br />

which is not practical. With a more complex circuit topology, it is hoped to achieve fast<br />

settling even with a significantly lower GBW <strong>of</strong> only approximately 1 GHz. Because the<br />

OTA in the Residue Amplifier is used with a feedback factor f < 1, the phase margin required<br />

for stability is not the phase at the unit gain frequency, but at the frequency corresponding<br />

to a gain <strong>of</strong> 1/f. This simplifies achieving a sufficient phase margin.<br />

The slew rate requirement on the OTA is tightened by the fact that the output is reset in each<br />

cycle. The output voltage may have to change by up to half the output peak-to-peak value<br />

(here: ∆V max = 250mV).<br />

k ⋅ Vmax I C L<br />

SR<br />

= --------------------------<br />

T S<br />

(<strong>12</strong>)<br />

The required slewing current can be estimated from (<strong>12</strong>), where I SR is the slewing current<br />

available, T s is the time available for settling, and V max the voltage swing. If a third <strong>of</strong> the<br />

settling time is used for slewing, k=3.<br />

4.2.1 Mirrored cascode with class AB input stage with preamplifier<br />

Simple Folded Cascode and Mirrored Cascode topologies have been evaluated, but found<br />

to either provide insufficient slewing current or a too small GBW. Finally, a topology using<br />

a class AB input stage [9] was adopted since the required voltage swing at the OTA input is<br />

small.<br />

A low-gain high-speed input amplifier (Figure 29) is used as a preamplifier to the OTA to<br />

take advantage <strong>of</strong> AB input stage. The input signal has very limited swing, which permits<br />

28


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

vin+<br />

vin−<br />

Residue Amplifier OTA<br />

+<br />

−<br />

−<br />

pre−<br />

Amplifier<br />

+<br />

+<br />

−<br />

−<br />

class−AB<br />

OTA<br />

cmfb<br />

+<br />

Common−mode<br />

feedback amplifier<br />

in2<br />

vin1<br />

vout+<br />

vout−<br />

Figure 27:<br />

Overall structure <strong>of</strong> the Residue Amplifier differential OTA<br />

M7a<br />

M3a<br />

M3b<br />

M7b<br />

M6a<br />

vout+<br />

M5a<br />

M4a<br />

vp<br />

M11b<br />

vn<br />

M9a<br />

vbias2<br />

vbias2<br />

M10a<br />

M10b<br />

vp<br />

vin+<br />

vin−<br />

M11b<br />

M1a<br />

M0a<br />

M0b<br />

M1b<br />

vn<br />

cmfb<br />

M8a<br />

M8b<br />

vbias1 vbias1<br />

M2a<br />

M2b<br />

M9b<br />

M6b<br />

vout−<br />

M5b<br />

M4b<br />

Figure 28:<br />

Low-Voltage differential class-AB mirrored cascode OTA<br />

vout−<br />

M1a<br />

M1d<br />

M1c<br />

M1b<br />

vout+<br />

M4<br />

M2a<br />

M2b<br />

Rc<br />

Rc<br />

Cc<br />

Cc<br />

vcmref<br />

vin+<br />

M0a<br />

M0b<br />

vin−<br />

vin2<br />

M0a<br />

M0b<br />

M0c<br />

M0d<br />

vin1<br />

vout<br />

M3<br />

vbias<br />

M1a<br />

M1b<br />

Figure 29:<br />

(a) Preamplifier<br />

(c) Common-mode feedback amplifier<br />

Preamplifier and common-mode feedback amplifier for differential class-AB OTA<br />

not folding the amplifier, but also doesn’t fully unbalance the input pair to take advantage<br />

<strong>of</strong> the class AB structure.<br />

Figure 29 (b) shows the implementation <strong>of</strong> the common-mode feedback amplifier.<br />

29


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

5 Top-Level Floorplanning<br />

Since the design is only a first prototype, no self-contained biasing circuitry has been<br />

designed. An external pin is used to adjust the bias current levels, another pin is used to<br />

observe the current. Since all bias currents used are multiples <strong>of</strong> 10uA, they can easily be<br />

derived from the externally controlled reference current.<br />

5.1 Analog Pipeline Stage Floorplan<br />

Figure 30 shows the placement <strong>of</strong> the different blocks in the analog pipeline stage. Note that<br />

the encoder is not in the analog signal path, but has been put between Flash ADC and DAC<br />

to minimize routing.<br />

analog<br />

input<br />

ref. input<br />

1<br />

DAC Output Buffer<br />

Encoder<br />

analog<br />

output<br />

Flash ADC<br />

Clock<br />

binary<br />

output<br />

Current−steering DAC<br />

SC Residue Amplifier<br />

Clock<br />

Figure 30:<br />

Floorplan <strong>of</strong> analog pipeline stage<br />

5.2 Floorplan <strong>of</strong> complete pipeline<br />

Figure 31 shows the floorplan <strong>of</strong> the complete converter pipeline. The structure from [4] has<br />

been preserved. Note that analog and digital I/O’s lie on opposite sides <strong>of</strong> the block.<br />

30


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

ANALOG I/O<br />

Analog Pipeline Stage 1<br />

Analog Pipeline Stage 2<br />

Analog Pipeline Stage 3 Analog Stage 4<br />

DAC Output Buffer<br />

DAC Output Buffer<br />

DAC Output Buffer<br />

Flash ADC<br />

Encoder<br />

Current−steering DAC<br />

Flash ADC<br />

Encoder<br />

Current−steering DAC<br />

Flash ADC<br />

Encoder<br />

Current−steering DAC<br />

Flash ADC<br />

Encoder<br />

Residue Amplifier<br />

Residue Amplifier<br />

Residue Amplifier<br />

POWER and CLOCK<br />

Digital Error Correction Block<br />

Digital Error Correction Block<br />

Clock distribution and buffering<br />

Digital Error Correction Block<br />

Digital Error<br />

Correction Block<br />

POWER and CLOCK<br />

DIGITAL I/O<br />

Figure 31:<br />

Overall floorplan <strong>of</strong> the pipelined ADC<br />

31


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

6 Conclusions<br />

In the limited time available for this project only a part <strong>of</strong> the ADC could be redesigned.<br />

Circuit design <strong>of</strong> the 4-<strong>bit</strong> ADC and DAC has been finished, and floorplans for the layout <strong>of</strong><br />

these blocks have been elaborated. The layout <strong>of</strong> the comparator used in the Flash has been<br />

completed and post-layout simulations have been carried out to verify the design.<br />

A switched-capacitor topology for the Residue Amplifier has been chosen, but the required<br />

OTA is still in the design phase. The correct switch and capacitor sizes also yet have to be<br />

found. First simulation results on the Residue Amplifier circuit indicate that the sampling<br />

speed specification may have to be relaxed.<br />

For the prototype <strong>of</strong> the pipeline stage, at least the Flash and ADC converter will be<br />

finished. If a Residue Amplifier is to be included, a DAC output buffer has to be designed<br />

as well.<br />

As only one stage will be implemented, only the thermometric-to-binary encoder will have<br />

to be included, the error correction using at least two pipeline stages.<br />

The full design <strong>of</strong> a high-speed pipelined differential ADC clearly exceeded the amount <strong>of</strong><br />

work that could be done in only 4 months time. Especially because a large portion was<br />

needed to learn how to use the s<strong>of</strong>tware tools. Many important aspects <strong>of</strong> a thorough design<br />

have not been addressed. For example, no noise analysis for the different circuit blocks has<br />

been carried out.<br />

Lausanne, February 20, 2004<br />

Thomas Liechti<br />

32


<strong>Design</strong> <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Differential</strong> <strong>Pipelined</strong> A/D <strong>Converter</strong><br />

References<br />

1 Maloberti F. Analog <strong>Design</strong> for CMOS VLSI Systems. Kluwer Academic Publishers, 2001.<br />

2 Carvajal R.G., Galan J., Ramirez-Angulo J., Torralba A. “New low-power low-voltage<br />

differential class-AB OTA for SC circuits”. Proceedings <strong>of</strong> the 2003 International Symposium<br />

on Circuits and Systems, vol. 1, 2003.<br />

3 Mallya S.M., Nevin J.H. “<strong>Design</strong> Procedures for a Fully <strong>Differential</strong> Folded-Cascode<br />

CMOS Operational Amplifier”. IEEE Journal <strong>of</strong> Solid-state Circuits, vol. 24, no. 6, December<br />

1989.<br />

4 Toprak Z., <strong>Design</strong> and Realization <strong>of</strong> a <strong>High</strong>-<strong>Speed</strong> <strong>12</strong>-<strong>bit</strong> <strong>Pipelined</strong> Analog/Digital <strong>Converter</strong><br />

Block, Master Thesis, Sabanci University, 2001<br />

5 Van de Plassche R., Integrated Analog-to-Digital and Digital-to-Analog <strong>Converter</strong>s, Kluwer<br />

Academic Publishers, 1994.<br />

6 Razavi B., Wooley A. “A <strong>12</strong>-b 5-MSample/s Two-Step CMOS A/D <strong>Converter</strong>”, IEEE<br />

Journal <strong>of</strong> Solid-state Circuits, vol. 27, no. <strong>12</strong>, December 1992.<br />

7 Yotsuyanagi M., Etoh T., Hirata K. “A 10-b 50-MHz <strong>Pipelined</strong> CMOS A/D <strong>Converter</strong> with<br />

S/H”, IEEE Journal <strong>of</strong> Solid-state Circuits, vol. 28, no. 3, March 1993.<br />

8 Waltari M. E., Halonen A. I. Circuit Techniques for Low-Voltage and <strong>High</strong>-<strong>Speed</strong> A/D<br />

<strong>Converter</strong>s. Kluwer Academic Publishers, 2002.<br />

9 Elwan H., Gao W., Sadkowski R., Ismail M. “CMOS low-voltage class-AB operational<br />

trasconductance amplifier”. Electronics Letters, vol. 36, no. 17, August 2000.<br />

10 Walden R. H., “Analog-to-Digital <strong>Converter</strong> Survey and Analysis”, IEEE Journal on<br />

Selected Areas in Communications, vol. 17, no. 4, April 1999.<br />

11 Yin G. M., Op’t Eynde F., Sansen W., “A <strong>High</strong>-<strong>Speed</strong> CMOS Comparator with 8-b<br />

Resolution”, IEEE Journal <strong>of</strong> Solid-state Circuits, vol. 27, no. 2, February 1992.<br />

<strong>12</strong> Uyttenhove K., Steyaert M. S. J., “A 1.8-V 6-Bit 1.3-GHz Flash ADC in 0.25-µm CMOS”,<br />

IEEE Journal <strong>of</strong> Solid-state Circuits, vol. 38, no. 7, July 2003.<br />

13 Shih T., Der L., Lewis S. H., Hurst P.J., “A Fully <strong>Differential</strong> Comparator Using a<br />

Switched-Capacitor Differencing Circuit with Common-Mode Rejection”, IEEE Journal <strong>of</strong><br />

Solid-state Circuits, vol. 32, no. 2, February 1997.<br />

14 Der L., Lewis S. H., Hurst P. J., “A Switched-Capacitor Differencing Circuit with<br />

Common-Mode Rejection for Fully <strong>Differential</strong> Comparators”, Proceedings <strong>of</strong> the 36th<br />

Midwest Symposium on Circuits and Systems, August 1993.<br />

15 Fournier J. M., Senn P. “A 130-Mhz 8-b CMOS Video DAC for HDTV Applications”, IEEE<br />

Journal <strong>of</strong> Solid-state Circuits, vol. 26, no. 7, July 1991.<br />

33

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!