Chapter 15

Digital PLLs and Clock Recovery

A digital PLL replaces the charge pump and its large loop-filter capacitor by a time-to-digital converter, a few adders and a digitally controlled oscillator. This chapter builds the pieces for the course radio: an inverter-chain TDC whose resolution we simulate in SG13G2, a DCO made from the Chapter 11 tank with switched capacitor banks, and the z-domain loop that joins them. It then turns to the loops that use a single bit of phase information, bang-bang PLLs and clock and data recovery, and to the delay-locked loop. The result is an illustrative all-digital PLL that meets the course's phase-noise specification at 1 MHz with 8.9 dB to spare, and whose in-band noise is set, almost exactly at the budget, by one inverter delay.

  • Explain what an all-digital PLL replaces in a charge-pump PLL, and describe it in the phase domain with a frequency control word.
  • Derive the in-band phase-noise floor of a TDC from its resolution, and simulate the resolution of an inverter-chain TDC and its supply and temperature sensitivity.
  • Calculate the frequency step of a capacitor-bank DCO and the phase noise of its quantization, with and without delta-sigma dithering.
  • Model a type-II digital loop in the z-domain, choose its proportional and integral gains, and check the model with a time-domain simulation.
  • Explain how bang-bang phase detectors, delay-locked loops and clock-and-data-recovery loops work, and estimate the effective gain of a bang-bang detector and the jitter tolerance of a bang-bang CDR.
Contents

15.1 Why digital PLLs

The PLL chapter of The Art of Analog IC Design and Chapter 12 of this course build the classic charge-pump PLL: a phase-frequency detector turns the phase error into current pulses, a passive loop filter integrates them on a capacitor, and the capacitor’s voltage tunes the VCO. The loop filter is the awkward part. For a charge-pump loop around the course VCO with a 200 kHz bandwidth, a 100 µA charge pump, the Chapter 11 tuning gain of 108 MHz/V, a division ratio of 122.1 and the stabilizing zero at a quarter of the crossover frequency, the filter needs R = 14.2 kΩ and C1 = 224 pF: 0.149 mm² of MIM capacitor at 1.5 fF/µm² (Chapter 7), and the capacitance grows as the square of the inverse bandwidth. The capacitor’s leakage, the charge pump’s current mismatch and the varactor’s dependence on the tuning voltage all add their own errors.

An all-digital PLL (ADPLL) keeps the loop and changes every block on it (Figure 15.1a) [153, 154]. Phase is counted in units of the oscillator period, the unit interval (UI). A reference phase accumulator adds the frequency control word FCW = fDCO/fref once per reference cycle; a counter of the oscillator’s edges gives its integer phase \(R_V\) and a time-to-digital converter (TDC) measures the fraction \(\varepsilon\) at each reference edge. Their difference is the phase error,

\[\text{FCW} = \frac{f_{DCO}}{f_{ref}}, \qquad \phi_E[k] = k\cdot\text{FCW} - \big(R_V[k] + \varepsilon[k]\big), \tag{15.1}\]

a number. A digital filter turns it into an oscillator tuning word (OTW) for a digitally controlled oscillator (DCO), an LC oscillator tuned by switched capacitors instead of a varactor. The filter is a few adders and registers, so its bandwidth is set by programmable coefficients, it costs no area, it does not leak, and it can change its own bandwidth during acquisition. Fractional-N operation needs no divider or delta-sigma modulator in the loop: FCW is simply a fractional number. For the course channel at 2442 MHz the DCO runs at 4.884 GHz and FCW = 122.1. The price is quantization: the TDC measures time in steps of \(\Delta t_{res}\) and the DCO changes frequency in steps of \(\Delta f_{res}\), and both steps become noise.

Two diagrams. (a) Block diagram of a phase-domain all-digital PLL. A 40 megahertz crystal drives a reference phase accumulator, R-R of k equals k times FCW, with FCW equal to f-DCO over f-ref equal to 122.1. Its output enters a summing node, from which the phase error phi-E of k goes to a loop filter alpha plus rho over one minus z to the minus one, then to a normalization block multiplying by f-ref over K-DCO-hat, whose output OTW tunes an orange LC DCO block with PVT, acquisition and tracking capacitor banks and delta-sigma dither. The DCO output at 4.884 gigahertz goes to the divide-by-two and LO. In the feedback path the DCO clocks a DCO edge counter giving the integer phase R-V and an orange TDC block, which also receives the reference edge and gives the fraction epsilon with step delta-t-res; R-V plus epsilon is subtracted at the summing node. (b) A flash TDC: five inverters in a chain driven by the DCO signal, each output tapped into the D input of a flip-flop; an orange line labelled REF clocks all five flip-flops, whose outputs are Q1 to Q5. A note says each stage delays the DCO edge by one inverter delay; the reference edge freezes the line, and with alternate taps inverted Q1 to Qn is a thermometer code whose 1-to-0 step locates the DCO edge to within delta-t-res.
Figure 15.1. An ADPLL compares phases as numbers: a reference accumulator against a DCO edge counter plus a TDC fraction. (a) Phase-domain ADPLL after Staszewski and Balsara; everything between the TDC and the DCO is synchronous logic clocked at fref. (b) The flash (inverter-chain) TDC that measures the fraction ε.

15.2 Time-to-digital converters

The inverter-chain TDC

The simplest TDC is a delay line sampled by flip-flops (Figure 15.1b). The DCO edge runs down a chain of inverters; the reference edge clocks a flip-flop at every tap at once, freezing the line. With every other tap inverted, the flip-flops hold a thermometer code whose step marks how far the DCO edge has travelled: the time from the DCO edge to the reference edge, to within one inverter delay \(\Delta t_{res}\) [155]. Two such measurements, on a rising and a falling DCO edge, give the period \(T_V\) in units of \(\Delta t_{res}\) as well, and the fraction \(\varepsilon\) is the measured delay divided by that period. To cover a whole period the line needs

\[N_{stages} \ge \frac{T_V}{\Delta t_{res}} = \frac{1}{f_{DCO}\,\Delta t_{res}} \tag{15.2}\]

stages in the fastest corner.

We simulated the delay line in ngspice: twelve SG13G2 inverters (lv MOS, Wn/Wp = 1/2 µm, L = 0.13 µm), each driving the next stage and one sampler input (fan-out of two, FO2), at three supplies and three temperatures (Figure 15.2a). At 1.2 V and 27 °C one stage takes 43.5 ps (45.1 ps falling, 41.9 ps rising). The supply dominates the spread: from 1.08 to 1.32 V the delay falls from 54.5 to 36.4 ps, 1.74 % per 10 mV. Temperature hardly matters at this supply (0.0146 ps/°C), because the falling threshold voltage and the falling mobility nearly cancel. Over all nine corners the delay spans 35.0–54.8 ps, a ratio of 1.57. The DCO period at 4.884 GHz is 205 ps, so a nominal line spans one period in 4.71 stages and the fastest corner needs 5.85: an eight-stage line is enough, and the TDC resolves only about five levels per DCO cycle.

The spread is why an ADPLL measures its TDC’s gain continuously, by averaging the measured period, and normalizes \(\varepsilon\) with it [154]; Section 15.6 shows the alternative of locking the delay line with a DLL. A finer resolution than one gate delay needs a Vernier TDC: two delay lines with stage delays \(\tau_1 > \tau_2\), the start edge in the slow line and the stop edge in the fast one, whose difference \(\tau_1 - \tau_2\) is the resolution [156]. Loading the same inverters with two samplers instead of one (FO3) gives 61.0 ps per stage, so a Vernier pair of FO3 and FO2 lines resolves 17.5 ps, at the price of more stages, a longer conversion and a resolution that is now a small difference of two large, mismatched delays.

Quantization noise and the in-band floor

The TDC rounds the time error to a multiple of \(\Delta t_{res}\). If the error is “busy”, varying by many steps from one reference edge to the next, the rounding error is uniform over one step, white, and uncorrelated with the signal. Converted to phase of the DCO and spread over the band sampled at \(f_{ref}\), it gives

Key equation · In-band phase-noise floor of a TDC

\[\mathcal{L}_{TDC} = \frac{(2\pi)^2}{12}\left(\frac{\Delta t_{res}}{T_V}\right)^2\frac{1}{f_{ref}} = \frac{(2\pi\,\Delta t_{res}\,f_{DCO})^2}{12\,f_{ref}}, \tag{15.3}\]

referred to the DCO frequency [155]. Inside the loop bandwidth the loop copies the measured phase, noise included, to the output, so Eq. (15.3) is the ADPLL’s in-band floor.

Derivation · The TDC floor

The rounding error of the time measurement is uniform over \(\Delta t_{res}\), with variance \(\Delta t_{res}^2/12\). As a phase of the DCO it is \(2\pi\,\Delta t/T_V\), so its variance is \(\sigma_\phi^2 = (2\pi\Delta t_{res}/T_V)^2/12\) rad². One sample is taken per reference cycle, and a white sequence of variance \(\sigma_\phi^2\) sampled at \(f_{ref}\) has a two-sided spectral density \(\sigma_\phi^2/f_{ref}\) between \(-f_{ref}/2\) and \(f_{ref}/2\). The single-sideband phase noise is the two-sided density of the phase, \(\mathcal{L} = S_\phi^{(2)} = \sigma_\phi^2/f_{ref}\), which is Eq. (15.3). The floor falls by 3 dB per doubling of \(f_{ref}\) and rises by 6 dB per doubling of \(\Delta t_{res}\) or of \(f_{DCO}\).

For the simulated 43.5 ps at 40 MHz, Eq. (15.3) gives −84.3 dBc/Hz at 4.884 GHz, −90.3 dBc/Hz at the LO (Figure 15.2b). The course budget of Chapter 5 allows −90 dBc/Hz at the LO inside the loop bandwidth, −84.0 dBc/Hz at the DCO: the inverter TDC sits only 0.3 dB below it, and in the slowest corner (54.8 ps, −82.3 dBc/Hz) above it. The Vernier’s 17.5 ps gives −92.2 dBc/Hz. To stay 10 dB below the budget at 40 MHz the TDC would need 14.3 ps.

Two panels. (a) Delay per stage in picoseconds against supply from 1.08 to 1.32 volts. Three FO2 curves for minus 40, 27 and 85 degrees Celsius almost overlap, falling from about 55 picoseconds at 1.08 volts to 35 to 38 at 1.32 volts; a note marks 43.5 picoseconds at 1.2 volts. An FO3 curve at 27 degrees lies higher, from 76 to 51 picoseconds. (b) In-band phase noise at 4.884 gigahertz in dBc per hertz against TDC resolution from 0.5 to 50 picoseconds, log axis: three straight lines rising 20 dB per decade for reference frequencies of 20, 40 and 100 megahertz; a dashed red line at minus 84 marks the Chapter 5 in-band mask at the DCO; an orange dot at 43.5 picoseconds and minus 84.3 dBc per hertz lies on the mask; a teal square marks the Vernier at 17.5 picoseconds near minus 92.
Figure 15.2. One SG13G2 inverter delay, 43.5 ps, puts the TDC's in-band floor right at the course budget. (a) Delay per stage of inverter delay lines (lv MOS, 1/2 µm, L = 0.13 µm) loaded by one (FO2) or two (FO3) sampler inputs; ngspice transients, typical models. (b) Eq. (15.3) against resolution at three reference frequencies, with the simulated inverter and Vernier resolutions at 40 MHz.

Common pitfall · Treating TDC quantization as white noise

Eq. (15.3) assumes the rounding error is random. In a quiet loop it is not: the DCO phase hardly moves between reference edges, and with FCW = 122.1 the fractional phase returns to the same ten values every ten cycles. The error then repeats, and its power appears as spurs at multiples of 4 MHz (Section 15.4), not as a floor. In an integer channel it is constant, and the TDC behaves like a one-bit detector (Section 15.5).

15.3 Digitally controlled oscillators

A DCO is an LC oscillator whose tank capacitance is a sum of switched units. A step \(\Delta C\) in a tank of capacitance \(C\) moves the frequency by

\[\Delta f = \frac{f}{2}\,\frac{\Delta C}{C}. \tag{15.4}\]

The Chapter 11 tank (537 fF at 4.868 GHz) gives 4.53 kHz per attofarad. Its 4-bit bank, with an LSB of 32.6 MHz (7.19 fF), already covers process, voltage and temperature. An ADPLL adds finer banks below it [152]: an acquisition bank of 2 MHz steps (441 aF), used while the loop pulls in, and a tracking bank whose 250 kHz steps (55.2 aF) are thermometer-coded so that every code change is one small, glitch-free step. Steps of tens of attofarads are far below a MIM capacitor’s minimum; they are made from the difference between the two states of a small MOS varactor, or by placing a unit capacitor in series with a larger one. We do not design those cells here. Without an analog tuning node, the DCO also loses the varactor’s conversion of tuning-line noise into frequency noise; its supply sensitivity remains.

Quantization noise and dithering

The tracking word changes once per reference cycle, so the DCO frequency is the wanted frequency plus a rounding error uniform over \(\Delta f_{res}\), held for one period. Phase is the integral of frequency, which turns the white frequency error into phase noise falling at 20 dB per decade [152]:

\[\mathcal{L}_{DCO,q}(\Delta f) = \frac{1}{12}\left(\frac{\Delta f_{res}}{\Delta f}\right)^2\frac{1}{f_{ref}}\,\mathrm{sinc}^2\!\left(\frac{\Delta f}{f_{ref}}\right). \tag{15.5}\]

The 250 kHz tracking step gives −98.9 dBc/Hz at 1 MHz (Figure 15.3), 16.5 dB above the DCO’s own −115.4 dBc/Hz (the pessimistic end of the Chapter 11 estimate). Without help the step would have to shrink to 11.8 kHz, 2.6 aF. The help is dithering: a delta-sigma modulator clocked at \(f_d \gg f_{ref}\) toggles the tracking bank’s least significant unit so that its average carries the fractional part of the word, and shapes the rounding error to high frequencies (Chapter 14 treats the modulators). With an \(n\)-th-order modulator,

\[\mathcal{L}_{DCO,q}(\Delta f) = \frac{1}{12}\left(\frac{\Delta f_{res}}{\Delta f}\right)^2\frac{1}{f_d}\left[2\sin\frac{\pi\Delta f}{f_d}\right]^{2n}\mathrm{sinc}^2\!\left(\frac{\Delta f}{f_d}\right). \tag{15.6}\]

At \(f_d\) = 640 MHz (16 fref, about fDCO/7.63), a first-order modulator flattens the noise to −151.1 dBc/Hz; a second-order one pushes it to −191.2 dBc/Hz at 1 MHz and −171.2 dBc/Hz at 10 MHz, negligible next to the oscillator.

Phase noise in dBc per hertz at 4.884 gigahertz against offset frequency from 10 kilohertz to 20 megahertz, log axis. The free-running DCO falls at 20 dB per decade through minus 115.4 at 1 megahertz. The acquisition bank with 2 megahertz steps and no dither lies highest, near minus 81 at 1 megahertz. The tracking bank with 250 kilohertz steps and no dither lies at minus 98.9 at 1 megahertz, above the dashed Chapter 5 mask. With first-order delta-sigma dither at 640 megahertz the tracking bank's noise is flat at minus 151; with second-order dither it rises from below minus 200 to minus 165 at 20 megahertz, minus 191 at 1 megahertz.
Figure 15.3. A DCO's frequency steps are noise unless they are dithered. Eqs. (15.5) and (15.6) for the banks of the illustrative DCO (Chapter 11 tank) at fref = 40 MHz, against the free-running oscillator (Chapter 11 estimate with F = 5) and the Chapter 5 mask referred to the DCO. Open-loop: inside its bandwidth the loop suppresses all of these.

15.4 The digital loop filter and the z-domain model

A type-II loop in the z-domain

The loop filter of Figure 15.1a is a proportional–integral filter: a gain \(\alpha\) in parallel with an accumulator of gain \(\rho\). The normalization block multiplies by \(f_{ref}/\hat K_{DCO}\), the inverse of the DCO’s measured gain, so a filter output of 1 changes the DCO frequency by exactly \(f_{ref}\), one UI per reference period. With one cycle of latency between the phase measurement and the DCO update, the open-loop gain in UI per UI is [154]

\[L(z) = \left(\alpha + \frac{\rho}{1-z^{-1}}\right)\frac{z^{-1}}{1-z^{-1}}. \tag{15.7}\]

The second factor is the DCO, an integrator of frequency into phase. Two integrators make a type-II loop: like the charge-pump PLL with its capacitor, it tracks a frequency offset with zero phase error. Far below \(f_{ref}\), \(z^{-1}/(1-z^{-1}) \approx f_{ref}/s\), and the loop becomes the familiar continuous one (Chapter 12):

\[L(s) \approx \frac{\alpha f_{ref}}{s} + \frac{\rho f_{ref}^2}{s^2}, \qquad \omega_n = f_{ref}\sqrt{\rho}, \quad \zeta = \frac{\alpha}{2\sqrt{\rho}}, \quad f_c \approx \frac{\alpha f_{ref}}{2\pi}. \tag{15.8}\]

The gains are usually powers of two, so the multipliers are shifts; with \(\alpha = 2^{-k}\) and \(\rho = 2^{-(2k+1)}\) the damping is 0.707. Like the charge-pump loop, the loop low-pass filters the TDC’s noise and high-pass filters the DCO’s:

\[\mathcal{L}_{out} = \mathcal{L}_{TDC}\,\lvert H\rvert^2 + \left(\mathcal{L}_{DCO} + \mathcal{L}_{DCO,q}\right)\lvert 1-H\rvert^2, \qquad H = \frac{L}{1+L}. \tag{15.9}\]

The loop is a sampled system, and the approximation fails as the bandwidth approaches \(f_{ref}\) (Figure 15.4b). The exact phase margin of Eq. (15.7) equals the continuous 65.5° for narrow loops but falls to 52.6° at a crossover of \(f_{ref}/10\), the usual limit for any sampled PLL [12].

Two panels. (a) DCO frequency error in megahertz against time after a 4 megahertz frequency step, from 0.1 to 100 microseconds on a log axis. The tracking loop with alpha 2 to the minus 8 and rho 2 to the minus 17 starts at minus 4, overshoots to about plus 0.8 near 20 microseconds and settles by about 80 microseconds. The acquisition loop with alpha 2 to the minus 4 and rho 2 to the minus 9 settles within about 5 microseconds after a small overshoot. A dotted orange gear-shift curve follows the acquisition loop and stays settled after switching at 5 microseconds. (b) Phase margin in degrees against crossover frequency over f-ref from 2 times 10 to the minus 4 to about 0.4, log axis. The continuous-time approximation is flat at 65.5 degrees; the exact sampled loop stays at 65.5 at low frequency, falls to about 52 degrees at 0.1, marked by a dotted line, and to below 10 degrees near 0.4. An orange dot marks the course ADPLL at 27 kilohertz and 65.5 degrees.
Figure 15.4. A narrow digital loop is quiet but slow; gear shifting gives fast settling without giving up the narrow loop. (a) Linear z-domain model, Eq. (15.7), after a 4 MHz step of the frequency control word (one BLE channel at the DCO). (b) Phase margin of Eq. (15.7), with ρ = α2/2, against the crossover frequency normalized to fref.

An illustrative ADPLL for the course radio

Worked example 15.1 · Choosing the loop gains

With the simulated 43.5 ps TDC and the Chapter 11 DCO, Eq. (15.9) sets the bandwidth. A wide loop passes the TDC’s −84.3 dBc/Hz floor out to large offsets; a narrow one leaves the DCO’s 1/f² noise unsuppressed. Sweeping \(\alpha = 2^{-k}\) with \(\rho = \alpha^2/2\), the integrated phase error at the LO from 1 kHz to 8 MHz is smallest at \(\alpha = 2^{-8}\), \(\rho = 2^{-17}\): 0.807°, against 1.71° at \(\alpha = 2^{-5}\) (a 295 kHz loop) and 1.13° at \(\alpha = 2^{-10}\) (9.05 kHz). The chosen loop crosses over at 27.4 kHz, has a 3 dB bandwidth of 36.3 kHz, ζ = 0.707 and a 65.5° phase margin. The optimum lies near the offset where the TDC floor meets the free-running DCO noise, 27.9 kHz here. The 295 kHz loop would give −104.0 dBc/Hz at 1 MHz from the LO, 6 dB worse than the specification.

A 36 kHz loop is slow: it settles a 4 MHz channel step to 10 kHz in 78.3 µs (Figure 15.4a), half of the 150 µs that BLE leaves between packets [184]. Digital loops escape this trade by gear shifting: they start with wide gains (\(\alpha = 2^{-4}\), \(\rho = 2^{-9}\), a 442 kHz crossover), which settle the same step in 4.8 µs, and switch to the narrow gains once locked, here after 5 µs, keeping the accumulator’s state so that the switch disturbs nothing [153].

To check Eq. (15.9) we simulated the loop in the time domain at the reference rate: a TDC that rounds the DCO’s fractional phase to 43.5 ps, the PI filter, a tracking bank dithered by a second-order delta-sigma modulator at 640 MHz, and a DCO whose phase performs the random walk that gives −115.4 dBc/Hz at 1 MHz. This is a behavioural model, not a circuit: the TDC is ideal apart from its step, and the DCO’s gain is normalized exactly. Figure 15.5 shows the result for two cases. With a random offset added before the TDC and subtracted after it (a subtractive dither that makes the rounding error white, as Eq. (15.3) assumes), the simulated spectrum follows the linear model within 0.761 dB (median from 10 kHz to 10 MHz): −82.2 dBc/Hz in band against −80.8 dBc/Hz, and −112.8 dBc/Hz at 1 MHz in both. Without the dither, at FCW = 122.1, the pitfall of Section 15.2 appears: the in-band noise drops to −97.8 dBc/Hz, but spurs rise at multiples of 4 MHz, the largest −61.5 dBc at 4 MHz from the DCO, −67.5 dBc at 2 MHz from the LO, on top of the adjacent channel. These are the fractional spurs of Chapter 14 in a new form; real ADPLLs fight them with dither, finer TDCs, and digital cancellation of the predictable fractional phase.

Phase noise in dBc per hertz at 4.884 gigahertz against offset from 1 kilohertz to 20 megahertz, log axis. The dashed red Chapter 5 mask at the DCO is flat at minus 84 up to 100 kilohertz and falls at 20 dB per decade through minus 104 at 1 megahertz. The orange dashed TDC contribution is flat at minus 84 and rolls off above about 30 kilohertz; the teal dashed DCO contribution rises from minus 105 at 1 kilohertz to about minus 84 near 20 kilohertz and then falls at 20 dB per decade. The blue linear-model total peaks near minus 80 at 15 kilohertz, crosses the mask near 30 kilohertz and lies at minus 112.8 at 1 megahertz, 8.9 dB below the mask. A pale blue simulated spectrum with subtractive TDC dither follows the total closely from 5 kilohertz to 20 megahertz. A pale orange simulated spectrum without dither lies near minus 100 in band and shows spurs rising to about minus 85 dBc per hertz at 4, 8, 12 and 16 megahertz. The dithered DCO quantization lies below the axis.
Figure 15.5. The illustrative ADPLL meets the 1 MHz specification with 8.9 dB to spare; inside its band the inverter TDC sets the noise. Linear model, Eq. (15.9), with α = 2−8, ρ = 2−17, against a time-domain behavioural simulation at the reference rate (217 cycles, Welch estimate), with and without subtractive TDC dither. FCW = 122.1, fref = 40 MHz, typical corner.

Table 15.1 collects the design. Reference noise is left out: a 40 MHz crystal is multiplied by FCW² in band, and Chapter 12 adds it to the budget.

Table 15.1. An illustrative ADPLL for the course radio at the 2442 MHz channel (DCO 4.884 GHz, FCW = 122.1), typical corner, 27 °C. Only the TDC delay is a transistor-level simulation; the rest is the behavioural and linear models of this chapter.

Block or quantity Choice Result Basis
TDC flash, SG13G2 inverters 1/2 µm, FO2, ≥ 6 stages Δtres = 43.5 ps (35.0–54.8 ps over ±10 % VDD, −40 to 85 °C) ngspice
TDC in-band floor Eq. (15.3) at 40 MHz −84.3 dBc/Hz at the DCO (−90.3 at the LO); budget −84.0 Eq. (15.3)
DCO Chapter 11 tank; banks of 32.6 MHz, 2 MHz, 250 kHz 4.53 kHz/aF; tracking LSB 55.2 aF Eq. (15.4)
DCO dither 2nd-order ΔΣ at 640 MHz −191.2 dBc/Hz at 1 MHz (negligible) Eq. (15.6)
Loop filter α = 2−8, ρ = 2−17; acquisition α = 2−4, ρ = 2−9 fc = 27.4 kHz, 65.5°; 4 MHz step settles in 4.8 µs with gear shift z-domain model
Phase noise at 1 MHz   −112.8 dBc/Hz at the DCO (−118.9 at the LO); mask −104.0, margin 8.9 dB model and time domain
In-band (10 kHz)   −80.8 dBc/Hz at the DCO, 3.2 dB above the flat part of the mask model
Integrated phase error at the LO, 10 kHz–8 MHz   0.753° (model), 0.664° (time domain); the mask itself gives 1.11° model and time domain
Fractional spur without TDC dither FCW = 122.1 −67.5 dBc at 2 MHz from the LO time domain
Better TDC options Vernier FO3/FO2; bang-bang PD (Section 15.5) 17.5 ps: −92.2 dBc/Hz; BBPD: −123.2 dBc/Hz equivalent ngspice; Eq. (15.11)

What this means for your design

In SG13G2 at 40 MHz, one inverter delay is just good enough for the course radio: the loop meets −110 dBc/Hz at 1 MHz from the LO, and its integrated phase error is below what Chapter 5’s mask assumed, but the in-band noise has no margin and rises above the mask in the slow corner. Every improvement attacks Eq. (15.3): a finer TDC (Vernier, or interpolating between two lines), a faster reference (80 MHz buys 3 dB), or no multi-bit TDC at all, which is where the next section goes.

15.5 Bang-bang PLLs

A bang-bang phase detector (BBPD) is a single flip-flop: the reference edge samples the DCO clock, and the output says only whether the DCO is early or late. It is the extreme TDC, one bit with no dead zone, and it is fast, small and free of the delay line’s PVT spread. Its gain depends on the jitter. With no noise, a loop around it never settles: the DCO is always pushed one way or the other, and the phase error circulates in a limit cycle whose size is set by the proportional step and the loop latency [157]. With random jitter of rms value \(\sigma\) (in UI) at its input, the detector’s average output against a small phase error \(\phi\) is \(\mathrm{erf}\big(\phi/(\sqrt{2}\,\sigma)\big)\), a straight line near zero with the slope

\[K_{BB} = \sqrt{\frac{2}{\pi}}\,\frac{1}{\sigma} \tag{15.10}\]

per UI. The rest of the detector’s output, the difference between its ±1 and that straight line, behaves as a white noise source referred to its input of variance \((\pi/2 - 1)\sigma^2\) [157], which gives a floor like Eq. (15.3):

\[\mathcal{L}_{BB} = (2\pi)^2\left(\frac{\pi}{2}-1\right)\frac{\sigma^2}{f_{ref}}. \tag{15.11}\]

We simulated a bang-bang version of the course loop in the time domain: the same DCO noise, a proportional step of \(2^{-14}\) UI (12.5 fs of DCO phase per decision) and \(\rho = 2^{-20}\) (Figure 15.6). Without noise the phase error cycles with a period of 8.08 reference cycles and a peak-to-peak size of 50.4 fs. With the DCO’s random walk (55 fs per reference period) the jitter at the detector is σ = 190 fs, so Eq. (15.10) gives KBB = 861 per UI and the linearized loop a bandwidth of 443 kHz; the simulated jitter transfer has its 3 dB point at 472 kHz. With a DCO 12 dB noisier, σ grows to 3.14 ps, KBB falls to 52 and the bandwidth to 72.3 kHz (simulated: 92.2 kHz). The bandwidth of a bang-bang loop is set by its own jitter, which makes it nonlinear, and harder to specify, than a loop with a linear detector.

Two panels. (a) Phase error at the bang-bang detector in femtoseconds against reference cycle from 0 to 60: without noise, a staircase limit cycle between about minus 25 and plus 25 femtoseconds repeating every eight cycles; with DCO noise, a random-looking orange trace wandering between about minus 270 and plus 280 femtoseconds. (b) Jitter transfer in decibels against modulation frequency from 10 kilohertz to 20 megahertz: for the nominal DCO, a blue linearized curve flat near 0 dB with slight peaking, falling through minus 3 dB near 440 kilohertz to minus 30 dB at 20 megahertz, with simulated circles following it; for a DCO 12 dB noisier, an orange curve peaking at plus 7 dB near 45 kilohertz and falling steeply through minus 3 dB near 70 kilohertz, with simulated circles nearby but without the peak.
Figure 15.6. A bang-bang loop limit-cycles without noise, and its bandwidth shrinks as its jitter grows. (a) Phase error at the BBPD, proportional step 2−14 UI, ρ = 2−20, one cycle of latency. (b) Jitter transfer: circles from time-domain simulations with a small sinusoidal phase modulation on the reference; lines from Eq. (15.7) with the detector gain of Eq. (15.10) at the simulated σ.

Eq. (15.11) explains the appeal. With σ = 190 fs the bang-bang detector’s equivalent floor is −123.2 dBc/Hz, against −84.3 dBc/Hz for the 43.5 ps TDC: in a low-jitter loop, one bit placed exactly at the edge resolves better than many bits spread across the period. The catch is fractional-N. A BBPD only works within a few σ of lock, so the predictable fractional phase (here 0.1 UI per cycle) must be removed before the detector, by a digital-to-time converter that delays the reference edge by the known amount. That combination is the basis of today’s low-jitter fractional-N digital PLLs [158].

15.6 Delay-locked loops

A delay-locked loop (DLL) aligns the output of a voltage-controlled delay line (VCDL) with its input one period later (Figure 15.7a). A phase detector compares the line’s input and output edges, a charge pump and capacitor (or a counter) integrate the error, and the integrated value sets the delay. There is no oscillator: the line only delays the reference, so a disturbance passes through once and is gone. Noise does not accumulate from cycle to cycle as it does in a VCO, and the loop, with a single integrator, is first-order and easy to stabilize [159, 11]. The price is that a DLL cannot synthesize a new frequency: its output frequency is its input’s, and the reference’s jitter goes straight through.

The common hazard is false lock. A phase detector cannot tell a delay of one period from a delay of two, or a delay stuck at its minimum from one that is still too short. A DLL therefore starts at its minimum delay, which must be shorter than one period, and its maximum delay must stay below two periods. The inverter line of Section 15.2 shows how hard that can be: its delay changes by a factor of 1.57 over supply and temperature, so a line that spans one period in the fastest corner spans 1.57 periods in the slowest, and the control range must cover that without reaching two.

DLLs supply the multi-phase clocks of time-interleaved converters and memory interfaces. In an ADPLL they offer a second way to normalize a TDC: a DLL that locks an N-stage line to the DCO period makes each stage exactly \(T_V/N\) in every corner, replacing the digital gain estimate of Section 15.2 by an analog loop.

15.7 Clock and data recovery

A serial data receiver gets no clock, only data. Clock and data recovery (CDR) extracts one from the transitions of the data and samples the data with it. Random non-return-to-zero (NRZ) data has a transition in only half its bit periods, and none during long runs of equal bits, so the phase detector must work with missing edges, and say nothing when there is no transition.

Two detectors dominate. The Hogge detector is linear: an XOR gate produces a pulse from each data transition to the next sampling clock edge, whose width grows with the phase error, and a second XOR produces a reference pulse of half a bit, so the difference of the two averages is proportional to the error [161]. The Alexander detector is bang-bang (Figure 15.7b): it samples the data at the centre of each bit and at the edge between bits. If a transition occurred, the edge sample equals either the bit before or the bit after, which tells whether the clock is early or late; with no transition it gives no output [160]. The bang-bang version is the norm at high data rates, because its decisions come from the same kind of flip-flops that sample the data.

Two panels. (a) A delay-locked loop: the reference passes through a voltage-controlled delay line of N stages tau, with taps phi 1 to phi N; a phase detector in orange compares the reference with the line's output, and a charge pump with a capacitor sets the control voltage V ctrl of the line. A note says it locks the output to the reference one period later, starting at minimum delay below T with maximum below 2T, or it can false-lock. (b) A bang-bang Alexander CDR: NRZ data goes to two flip-flops, one clocked at the bit centre by clk giving B, one at the edge by clk bar giving T; a third flip-flop delays B to A. An early-late block in orange compares A, T and B: if A differs from B, T equal to A means early, T equal to B late. Its decision moves the recovered clock by plus or minus Kp through a phase interpolator; B is the recovered data.
Figure 15.7. A DLL delays the reference until the line spans one period; a bang-bang CDR samples each bit at its centre and its edge and nudges the clock by a fixed step. (a) DLL with a voltage-controlled delay line (Section 15.6). (b) First-order bang-bang CDR with an Alexander detector (Section 15.7), the loop of Figure 15.8.

A CDR is specified by three jitter properties: jitter transfer, the low-pass response from input to recovered-clock jitter; jitter generation, the jitter it adds with clean data; and jitter tolerance, the largest sinusoidal input jitter, against its frequency, at which it still receives without errors. Slow input jitter is tracked; jitter beyond the loop bandwidth must fit in the eye. A bang-bang CDR adds a third regime, slewing: the recovered phase moves by a fixed step \(K_p\) per decision, at most \(D_T K_p\) per bit with a transition density \(D_T\) = 1/2, so a sinusoid of peak amplitude \(A\) (UI) at frequency \(f_{SJ}\) is tracked only while its slope stays below that:

\[A_{tol} \approx \frac{D_T\,K_p\,f_{bit}}{2\pi f_{SJ}}. \tag{15.12}\]

Figure 15.8 shows a behavioural first-order bang-bang CDR at 10 Gb/s with \(K_p = 2^{-7}\) UI (7.81 mUI), random data and 0.01 UI rms of random jitter. Below a few megahertz the simulated tolerance follows Eq. (15.12) (median ratio 1.02; 6.42 UI against 6.22 UI at 1 MHz). Eq. (15.12) falls to half a UI at 12.4 MHz; above a few tens of megahertz the loop no longer follows and the tolerance flattens (0.358 UI, the median above 150 MHz), what is left of the half-UI eye after the random jitter and the loop’s own dither. A real CDR adds an integral path, as in Eq. (15.7), to track the frequency offset between transmitter and receiver crystals; it raises the low-frequency tolerance further [2].

Two panels. (a) Jitter tolerance in UI peak against sinusoidal-jitter frequency from 0.1 to 500 megahertz, both log axes. Blue simulated points fall from about 64 UI at 0.1 megahertz at 20 dB per decade, slightly above an orange dashed slew-rate line of Eq. 15.12, and flatten near 0.36 UI above about 50 megahertz; a dotted line marks 0.5 UI. (b) Phase in UI against time from 0 to 0.3 microseconds: a 0.3 UI sinusoid at 10 megahertz with the recovered blue phase following it closely; a 2 UI sinusoid with the recovered orange phase following a triangle wave that cannot keep up with the sinusoid's slope.
Figure 15.8. A bang-bang CDR tolerates large slow jitter up to its slew rate, and about half a UI of fast jitter. (a) Jitter tolerance of a first-order bang-bang (Alexander) CDR at 10 Gb/s, Kp = 2−7 UI: largest sinusoidal jitter received without an error in 4 × 104 bits, against Eq. (15.12). (b) Input and recovered phase for 0.3 UI and 2 UI of sinusoidal jitter at 10 MHz (2335 errors in 3000 bits for the second). Behavioural model, one decision per bit.

Try it · The digital-PLL script

Every number in this chapter comes from tools/figs/ch15_digital_pll.py: python3 tools/figs/ch15_digital_pll.py redraws everything in under a minute (the ngspice delay lines are cached; --resim reruns them, about a minute each). Three experiments: (1) Halve the inverter widths (WN, WP) and run tdc --resim: does the delay change, and why does a smaller inverter not make a faster TDC when the load scales with it? (2) Force a wide loop with NUM_LOOP.update(alpha=2**-5, rho=2**-11) before part_pn() and compare the 1 MHz phase noise with Worked example 15.1. (3) Change FCW to 122.0 and to 122.25 and rerun pn with the TDC dither off: where do the spurs go, and what happens to the in-band noise?

  • An ADPLL counts phase in DCO cycles: a reference accumulator, an edge counter plus a TDC, a PI filter and a capacitor-bank DCO replace the PFD, charge pump, loop-filter capacitor (224 pF for a 200 kHz analog loop around the course VCO) and varactor.
  • The TDC sets the in-band floor, (2π Δtres fDCO)²/(12 fref). One SG13G2 inverter delay, 43.5 ps (35.0–54.8 ps over supply and temperature), gives −84.3 dBc/Hz at 4.884 GHz, right at the course budget; a Vernier pair resolves 17.5 ps.
  • DCO steps of 250 kHz (55.2 aF in the Chapter 11 tank) would add −98.9 dBc/Hz at 1 MHz; a second-order delta-sigma dither at 640 MHz makes them negligible.
  • The type-II digital loop with α = 2−8, ρ = 2−17 (27.4 kHz) minimizes the integrated phase error, meets the 1 MHz specification with 8.9 dB of margin, and settles a channel step in 4.8 µs with gear shifting. Without dither, the TDC’s rounding makes fractional spurs (−67.5 dBc at 2 MHz from the LO).
  • A bang-bang detector’s gain is √(2/π)/σ: its loop’s bandwidth depends on its jitter, and in a quiet loop its equivalent noise (−123.2 dBc/Hz) is far below a multi-bit inverter TDC’s.
  • DLLs delay instead of oscillate, so their jitter does not accumulate, but they must avoid false lock; bang-bang CDRs track slow jitter up to their slew rate, DTKpfbit/(2πfSJ).

Further reading

  • R. B. Staszewski and P. T. Balsara [154] on the phase-domain ADPLL, and the Bluetooth transmitter of Staszewski et al. [153], the first ADPLL in a commercial radio, with the DCO architecture in [152] and the TDC in [155].
  • N. Da Dalt [157] on the nonlinear dynamics of bang-bang PLLs, and D. Tasca et al. [158] for a fractional-N bang-bang digital PLL with a digital-to-time converter.
  • P. Dudek, S. Szczepański and J. V. Hatfield [156] on the Vernier delay-line TDC; J. G. Maneatis [159] on self-biased DLLs and PLLs.
  • J. D. H. Alexander [160] and C. R. Hogge [161], the two short papers behind most CDR phase detectors.
  • B. Razavi, Design of CMOS Phase-Locked Loops [11], and S. Palermo’s ECEN 620 notes [2], for digital PLLs, DLLs and CDRs in the context of the whole PLL family; F. M. Gardner [12] for sampled loops.

Back to top

SkillSurf · RFIC Design with Open-Source Tools, draft v0.1 (November 2026). Every source is listed on the References page.

This course was written with the help of AI tools and may contain mistakes. If you find one, please tell us through SkillSurf or email skillsurfedu@gmail.com, and we will correct it.

This site uses Just the Docs, a documentation theme for Jekyll.