A High-resolution Vernier Gated-Ring-Oscillator TDC
                    in 90-nm CMOS
                                                    Ping Lu, Pietro Andreani
                         Department of Electrical and Information Technology, Lund University, Sweden
                                         Email: {Ping.Lu, Pietro.Andreani }@eit.lth.se

 Abastract—A Vernier Gate-Ring-Oscillator (GRO) Time
 to Digital Converter (TDC) is proposed and implemented
 in 90-nm CMOS process technology. It utilizes two GRO
 chains as the delay lines. The time resolution is
 determined by the difference between two delays, so not
 limited by the process. Moreover, the quantization noise
 can be first-order shaped by the gated behavior in the
 oscillators, which further improves the in-band TDC noise
 contribution for an ADPLL. Operating at 1.2-V supply
 with 250MHz clock, the chip achieves a less-than-10ps
 coarse resolution (varies with digital control bits) and
 consumes only 3.6-mA.
                          INTRODUCTION
    All Digital Phase-Locked Loop (ADPLL) is now a hot
 research topic for more flexibility and programmability                  Fig.1 GRO principle and noise shaping
 compared with its analog counterpart. In an ADPLL,
 Digitally Controlled Oscillator (DCO) phase noise and           then developed [5].
 TDC quantization noise dominate the in-band and                    Noise shaping is another method for reducing in-band
 out-band noise respectively. Thus in some ADPLL                 TDC noise contribution [6]. This type of TDC has still a
 applications which need a wide bandwidth, the                   coarse resolution of an inverter delay, but the
 quantization noise of TDC is important to the total noise       corresponding large quantization noise can be first-order
 performance.                                                    shaped. It pushes most of the noise to high-frequency
    The TDC noise contribution, within the loop bandwidth,       region which is then filtered by the loop filter in ADPLLs.
 at the ADPLL RF output is [1]                                   As shown in Fig.1, the cyclic behavior of GRO helps
                      (2π ) 2 Δtdelay 2 1                        reuse the staggered clocks circularly and the enable
                STDC =         (          )
                          12       TDCO       f REF              transistors (NMOS and PMOS) in the delay cell hold the
                                                          (1)
                                                                 oscillation node state between measurements. At next
 where Δtdelay denotes the delay time of an delay cell in the
                                                                 measurement interval, the oscillator starting phase
 TDC chain, TDCO denotes the period of RF output, and fREF
                                                                 corresponds to the stopping phase of the current
 is frequency of the reference clock. As can be seen, a
                                                                 measurement interval. The variable starting phase
 smaller Δtdelay leads to a smaller TDC quantization noise.
                                                                 effectively scrambles the quantization noise across the
 In a traditional TDC where the time resolution provided
                                                                 different measurement intervals and also introduces a
 by a self-loaded inverter, the quantization noise is limited
                                                                 first-order noise shaping.
 by a single inverter, so is limited by a process technology.
                                                                    Considering that the coarse resolution in GRO based
 Some advanced approaches like vernier structure [2],
                                                                 TDC is still limited by an inverter delay, a new TDC
 pulse shrinking [3] and sub-exponent TDC [4] have
                                                                 combining vernier and GRO structures is proposed in this
 therefore been proposed to achieve much shorter
                                                                 work. It uses two GROs in a classic vernier TDC to
 resolvable time than an inverter itself.
                                                                 achieve both process-independent high time resolution and
    A vernier TDC uses two delay lines with respective
                                                                 first-order shaped noise characteristic.
 delay of τ 1 and τ 2 instead of a single line. The time            This paper is arranged as follows. Section II describes
 resolution Δtdelay becomes a delay difference τ 1 − τ 2 , not   the architecture of the vernier GRO based TDC. The
 an inverter delay any more. This delay difference can be        circuit implementation is detailed in Section III. Section
 very small (down to 0) theoretically regardless of              IV gives the layout and simulation results. Conclusions are
 thermo-noise of devices. Thus TDC noise contribution can        drawn in Section V.
 be reduced greatly according to (1). Since the time
                                                                           VERNIER GATED-RING-OSCILLATOR TDC
 resolution is determined by a very small delay difference
 in the vernier structure, a large number of inverter stages        Fig.2 shows the structure of the proposed TDC. It
 are required to cover a large detection range, incurring        includes the GRO core, Phase and Frequency Detector
 large silicon overhead. A ring-oscillator vernier TDC was       (enable generator) and a multi-clock counter.
978-1-4244-8971-8/10$26.00 c 2010 IEEE
Fig.2 Block Diagram of proposed vernier GRO based TDC

   The GRO core consists of two 9-stage ring oscillators          rising edges in SGRO will lag their corresponding rising
and a sampling block. As in a vernier TDC, the two                edges in FGRO after passing a certain amount of
oscillators have different delay cells. The delay of the first    staggered phases. Once the lag happens, RBO will reset
oscillator cell is a little longer than that of the second one.   PFD and the measurement window will be closed
The difference between them indicates the time resolution.        immediately (EN_REF/ EN_CKV = ”0”).
Use SGRO (slow) and FGRO (fast) to denote those two
                                                                                         CIRCUIT DETAILS
oscillators. EN_REF and EN_CKV are enable signals for
them respectively. The oscillators work at the enable state       A. Gated-Ring-Oscillator core
(EN_REF/ EN_CKV=”1”) and hold their states at the                    Fig.3 shows the detailed GRO core. Two identical ring
disable state (EN_REF/ EN_CKV=”0”). The sampling                  oscillators are used. Delay difference comes from different
block reads the phase information of the 9 pairs of               control inputs (c<1:15>) to them. EN and ENB are enable
one-one correspondence clocks and feeds back a control            signals for NMOS and PMOS gate transistors in delay
signal. At initialization, SGRO always leads FGRO and             cells. c<1:15> are thermometer-code bits which are
the sampled value RBO is “0”. When any phase in SGRO              connected to 15 unit MOS capacitors directly. The delay
lags its counterpart phase in FGRO, RBO sends a positive          gain is about 1.28ps/bit. For a vernier TDC application,
pulse to reset PFD. Then EN_REF and EN_CKV return to              the inputs c<1:15> are always different for SGRO and
“0” to hold the oscillators’ state. All clock (ck<0:8>)           FGRO because a small time difference is needed to
edges from SGRO are also used by a multi-clock counter            quantize phase error.
to calculate the digitized time error in each measurement            The GRO delay cell is shown at the top right corner in
interval. Since the gate transistors are used in the GROs,        Fig.3. NMOS and PMOS enable transistors are controlled
quantization error residue will be accumulated in each            by EN and ENB respectively. When both enable transistors
measurement. That means a first-order noise shaping is            are in the conducting condition, the delay cell works like a
achieved at the same time, as shown in Fig.1.                     traditional inverter. Then the inverter ring oscillates as a
   It should be noticed that in this work only rising edge of     common ring oscillator. However, when EN and ENB cut
all oscillator clocks is used for the sampling block to get       off the enable transistors, the ring will stop oscillating and
rid of rising and falling edge mismatch. It suggests the          keep the output nodes’ charge unchanged because no
coarse time resolution is double difference between two           current path is connected to power supply or ground any
inverter delays because non-inversion neighboring phases          more. The oscillation frequency of each GRO can be tuned
are preferred to use.                                             discretely by 15 thermometer-code switches c<1:15>. For
   Multi-clock counter needs to count at each rising edge         SGRO, these 15 switches are always connected to ground
of ck<0:8>. The count enable signal is SGRO’s gate                suggesting a fixed cell delay (also a fixed frequency).
control EN_REF. When EN_REF = ”1”, each rising edge               Then the delay difference between SGRO and FGRO are
triggers the accumulator to add “1” to the sum. When              determined only by the 15 switches in FGRO. Actually, to
EN_REF = ”0” (the measurement is finished), the counter           reduce off-chip controls, the 15 thermometer-code
saves the sum into a latch and then resets the accumulator        switches      are     generated      by      an     additional
preparing for the next measurement.                               binary-to-thermometer coder.
   PFD generates two enable signals (EN_REF, EN_CKV)                 Sampling flip-flops detect the phase error between each
and their reversal signals (ENB_REF, ENB_CKV) to                  pair of staggered clocks. Once one of those sampled value
control NMOS and PMOS enable transistors in GRO                   becomes “1” (that is the staggered lagged clocks catch up
delay cells. EN_REF/EN_CKV is supposed to be a                    with those leading ones), RBO becomes “1” to reset PFD.
rectangular signal suggesting the phase error information            The key point for GRO core is the absolute delay in
between input clocks REF and CKV. When REF rising                 each oscillator. When phase ck<0(n+1)> catches up with
edge arrives first, for example, EN_REF becomes “1”               phase ck<n+1>, as illustrated in lower right of Fig.3,
(ENB_REF becomes “0”) forcing SGRO to start                       o<n+1> needs transmission time to become “1”. The
oscillating first. When CKV rising edge arrives after a           situation for RBO and EN_REF/EN_CKV is the same. Due
while, EN_CKV also becomes “1” forcing FGRO to work.              to this kind of gate delay, two GROs keep oscillating for a
In this example, EN_REF leads EN_CKV. It means that               short while after ck<n+1> is already lagged. The critical
the first rising edge in SGRO leads the first rising edge in      feedback path has a typical delay of 100ps~200ps
FGRO during the measurement interval. If these two                (obtained by post layout simulation). No more rising edges
oscillators have the same frequency, the phase error will         are supposed to be generated during this additional time
be always kept. However, due to the delay difference, the         interval, otherwise wrong sampling may happen in the
Fig.3 GRO core and sampling/reset delay




                                                   Fig.4 Adaptive PFD

next measurement. That means the absolute delay of each
vernier line (two-inverter delay) should be longer than
200ps. Here 400ps is adopted for some margin.
B. Phase & frequency detector (PFD)
   In the GRO core, the SRGO (driven by EN_REF) cell
delay is always longer than FGRO (driven by EN_CKV)
cell delay. It requires EN_REF to lead EN_CKV always.
However for a normal PFD in which EN_REF/EN_CKV is
corresponding to REF/CKV, this precondition cannot be
guaranteed because CKV frequency and phase are adjusted
dynamically in ADPLLs. A SR-latch is therefore used as
an arbiter, as shown in Fig.4. It senses the leading edge
between REF and CKV and gives a sign signal which
                                                                              Fig.5 Multi-clock counter
controls four multiplexers. When REF is leading, sign is
“0” meaning EN_REF corresponds to REF. On the
                                                              clock frequency and hardware consumption. In each group,
contrary, EN_REF corresponds to CKV.
                                                              clocks should have a large interval between rising edges,
C. Multi-clock counter                                        which eases the strain of accumulator. The two-inverter
   The digitized phase error can be calculated using          delay is about 400ps in this work. To make the counter
multiple clocks ck<0:8> from SGRO in each measurement         work with a large margin, the nine clocks are grouped into
interval. One approach is to assign one individual            5 groups with each containing two. The two clocks with
accumulator for each clock, which consumes more adders        the longest rising interval (e.g. ck<0> and ck<8>) share
and flip-flops. If all clocks could be combined to one that   one accumulator. Each clock rising edge produces a
contains all edges, only one accumulator is enough. But       narrow pulse by inserting a delay in the reversion-input
the new clock frequency is inevitably very high bearing on    path of the first NAND gate, as shown in Fig.5. Then the
the accumulator speed. To tackle the above problem, we        second NAND gate combines these two positive pulses to
propose to group those clocks to get a tradeoff between       one to clock an accumulator. The sum of each accumulator
multi-clock counter



                                         GRO1
             PFD
                                         GRO2                                                           (a)

                                    dummy load
                       Fig.6 Layout of TDC




                                                                                                       (b)
                                                                                   Fig.8 Simulated transistor-level TDC output
              (a)                        (b)                                   (a) Bit stream (coded decimal)     (b) FFT analysis
      Fig.7 Simulated PSD of TDC output in Simulink
       (a) Proposed TDC    (b) Classic vernier TDC                         quantization noise due to coarse resolution is well
                                                                           first-order shaped (see Fig.8(b)).
should be synchronized and then added together. The
above operation runs only in the measurement interval                                               CONCLUSIONS
(EN_REF = ”1”).                                                               This paper describes a Vernier-GRO TDC which uses
                     LAYOUT AND SIMULATION                                 vernier principle to get a high time resolution and GRO’s
                                                                           to achieve first-order shaped quantization noise. The
   The above presented TDC has been implemented using                      first-order noise shaping pushes much noise to the
90-nm CMOS technology. Fig.6 shows the layout of the                       high-frequency region which can be filtered by a low-pass
proposed TDC (without I/O pads). The active area is                        noise transfer function in an ADPLL. The chip is
0.18mm*0.15mm. The multi-clock counter uses staggered                      implemented in 90-nm CMOS technology and occupies
clocks from SGRO as referred above. Since FGRO has no                      0.18mm*0.15mm area.
identical load capacitances as SGRO, some dummy gate
loads are used to compensate delay mismatch.                                                        REFERENCES
   We compare the proposed TDC and the classic vernier                     [1] R. B. Staszewski and P. T. Balsaras, “All-Digital Frequency
TDC in Fig.7 by simulating in a simulink platform.                         Synthesizer in Deep-Submicron CMOS”, John Wiley & Sons,
Sinusoidal delay is used as error input. These two TDCs                    2006
has the same Δtdelay = 6.4 ps . For a classic vernier TDC,                 [2] P. Dudek, S.Szczepanski and J.V. Hatfield, “A
                                                                           High-Resolution CMOS Time-to-Digital Converter Utilizing a
the quantization noise has a “white” noise floor. However                  Vernier Delay Line”, IEEE Journal of Solid-State Circuits,
in the proposed TDC, most of the quantization noise is                     vol.35 No. 2, Feb. 2000, pp. 240-247.
pushed to high-frequency region (first-order shaped). A                    [3] Y. Liu, U. Vollenbruch, Y.J. Chen, etc., “A 6ps Resolution
250MHz sampling clock is used in this example. Noise                       Pulse Shrinking Time-to-Digital Converter as Phase Detector in
performance below 20MHz can be well improved, as seen                      Multi-Mode Transceiver”, IEEE radio and wireless symposium,
in the fig.7 (a).                                                          pp. 163-166, 2008
   Fig.9 shows the transistor-level simulation of TDC                      [4] S.K. Lee, Y.H. Suh, H.J. Park, J.Y. Sim, “A 1GH ADPLL
                                                                           with a 1.25ps Minimum-Resolution Sub-Exponent in 0.18um
output bit stream and its FFT analysis. A dc-delay of 22ps
                                                                           CMOS”, in 2010 IEEE Int. Solid-State Circuits Conf. Dig. Tech.
is used as input error. If ΔT denotes the input time error                 Papers, pp.482-483
between REF and CKV, the digitized error should be equal                   [5] H.H. Chang, P.Y. Wang, J.H. Zhan, etc, “A Fractional
to ΔT / Δtdelay . Δtdelay = 6.4 ps (5 unit capacitor controls              Spur-Free ADPLL with Loop-Gain Calibration and Phase-Noise
                                                                           Cancellation for GSM/GPRS/EDGE”, in 2008 IEEE Int.
are set to “1”) is set in the simulation. For a regular vernier            Solid-State Circuits Conf. Dig. Tech. Papers, pp.200-201
TDC, it should give a result of 3 by simply truncating the                 [6] Chun-Ming Hsu, Matthew Z. Straayer and Michael H. Perrott,
fractional error. But for this TDC, the output varies                      “A Low-Noise Wide-BW 3.6GHz Digital Δ∑ Fractional-N
between 3 and 4 with an average value of 3.41 (see Fig.8                   Frequency Synthesizer With a Noise-Shaping Time-to-Digital
( a) ) . T h i s s h o ws a n o t h e r b e n e f i t o f a c c ur a c y   Converter and Quantization Noise Cancellation”, IEEE Journal
overwhelming the classic vernier TDCs. Moreover,                           of Solid-State Circuits, vol. No. Dec. 2008, pp. 2776-2786

39

  • 1.
    A High-resolution VernierGated-Ring-Oscillator TDC in 90-nm CMOS Ping Lu, Pietro Andreani Department of Electrical and Information Technology, Lund University, Sweden Email: {Ping.Lu, Pietro.Andreani }@eit.lth.se Abastract—A Vernier Gate-Ring-Oscillator (GRO) Time to Digital Converter (TDC) is proposed and implemented in 90-nm CMOS process technology. It utilizes two GRO chains as the delay lines. The time resolution is determined by the difference between two delays, so not limited by the process. Moreover, the quantization noise can be first-order shaped by the gated behavior in the oscillators, which further improves the in-band TDC noise contribution for an ADPLL. Operating at 1.2-V supply with 250MHz clock, the chip achieves a less-than-10ps coarse resolution (varies with digital control bits) and consumes only 3.6-mA. INTRODUCTION All Digital Phase-Locked Loop (ADPLL) is now a hot research topic for more flexibility and programmability Fig.1 GRO principle and noise shaping compared with its analog counterpart. In an ADPLL, Digitally Controlled Oscillator (DCO) phase noise and then developed [5]. TDC quantization noise dominate the in-band and Noise shaping is another method for reducing in-band out-band noise respectively. Thus in some ADPLL TDC noise contribution [6]. This type of TDC has still a applications which need a wide bandwidth, the coarse resolution of an inverter delay, but the quantization noise of TDC is important to the total noise corresponding large quantization noise can be first-order performance. shaped. It pushes most of the noise to high-frequency The TDC noise contribution, within the loop bandwidth, region which is then filtered by the loop filter in ADPLLs. at the ADPLL RF output is [1] As shown in Fig.1, the cyclic behavior of GRO helps (2π ) 2 Δtdelay 2 1 reuse the staggered clocks circularly and the enable STDC = ( ) 12 TDCO f REF transistors (NMOS and PMOS) in the delay cell hold the (1) oscillation node state between measurements. At next where Δtdelay denotes the delay time of an delay cell in the measurement interval, the oscillator starting phase TDC chain, TDCO denotes the period of RF output, and fREF corresponds to the stopping phase of the current is frequency of the reference clock. As can be seen, a measurement interval. The variable starting phase smaller Δtdelay leads to a smaller TDC quantization noise. effectively scrambles the quantization noise across the In a traditional TDC where the time resolution provided different measurement intervals and also introduces a by a self-loaded inverter, the quantization noise is limited first-order noise shaping. by a single inverter, so is limited by a process technology. Considering that the coarse resolution in GRO based Some advanced approaches like vernier structure [2], TDC is still limited by an inverter delay, a new TDC pulse shrinking [3] and sub-exponent TDC [4] have combining vernier and GRO structures is proposed in this therefore been proposed to achieve much shorter work. It uses two GROs in a classic vernier TDC to resolvable time than an inverter itself. achieve both process-independent high time resolution and A vernier TDC uses two delay lines with respective first-order shaped noise characteristic. delay of τ 1 and τ 2 instead of a single line. The time This paper is arranged as follows. Section II describes resolution Δtdelay becomes a delay difference τ 1 − τ 2 , not the architecture of the vernier GRO based TDC. The an inverter delay any more. This delay difference can be circuit implementation is detailed in Section III. Section very small (down to 0) theoretically regardless of IV gives the layout and simulation results. Conclusions are thermo-noise of devices. Thus TDC noise contribution can drawn in Section V. be reduced greatly according to (1). Since the time VERNIER GATED-RING-OSCILLATOR TDC resolution is determined by a very small delay difference in the vernier structure, a large number of inverter stages Fig.2 shows the structure of the proposed TDC. It are required to cover a large detection range, incurring includes the GRO core, Phase and Frequency Detector large silicon overhead. A ring-oscillator vernier TDC was (enable generator) and a multi-clock counter. 978-1-4244-8971-8/10$26.00 c 2010 IEEE
  • 2.
    Fig.2 Block Diagramof proposed vernier GRO based TDC The GRO core consists of two 9-stage ring oscillators rising edges in SGRO will lag their corresponding rising and a sampling block. As in a vernier TDC, the two edges in FGRO after passing a certain amount of oscillators have different delay cells. The delay of the first staggered phases. Once the lag happens, RBO will reset oscillator cell is a little longer than that of the second one. PFD and the measurement window will be closed The difference between them indicates the time resolution. immediately (EN_REF/ EN_CKV = ”0”). Use SGRO (slow) and FGRO (fast) to denote those two CIRCUIT DETAILS oscillators. EN_REF and EN_CKV are enable signals for them respectively. The oscillators work at the enable state A. Gated-Ring-Oscillator core (EN_REF/ EN_CKV=”1”) and hold their states at the Fig.3 shows the detailed GRO core. Two identical ring disable state (EN_REF/ EN_CKV=”0”). The sampling oscillators are used. Delay difference comes from different block reads the phase information of the 9 pairs of control inputs (c<1:15>) to them. EN and ENB are enable one-one correspondence clocks and feeds back a control signals for NMOS and PMOS gate transistors in delay signal. At initialization, SGRO always leads FGRO and cells. c<1:15> are thermometer-code bits which are the sampled value RBO is “0”. When any phase in SGRO connected to 15 unit MOS capacitors directly. The delay lags its counterpart phase in FGRO, RBO sends a positive gain is about 1.28ps/bit. For a vernier TDC application, pulse to reset PFD. Then EN_REF and EN_CKV return to the inputs c<1:15> are always different for SGRO and “0” to hold the oscillators’ state. All clock (ck<0:8>) FGRO because a small time difference is needed to edges from SGRO are also used by a multi-clock counter quantize phase error. to calculate the digitized time error in each measurement The GRO delay cell is shown at the top right corner in interval. Since the gate transistors are used in the GROs, Fig.3. NMOS and PMOS enable transistors are controlled quantization error residue will be accumulated in each by EN and ENB respectively. When both enable transistors measurement. That means a first-order noise shaping is are in the conducting condition, the delay cell works like a achieved at the same time, as shown in Fig.1. traditional inverter. Then the inverter ring oscillates as a It should be noticed that in this work only rising edge of common ring oscillator. However, when EN and ENB cut all oscillator clocks is used for the sampling block to get off the enable transistors, the ring will stop oscillating and rid of rising and falling edge mismatch. It suggests the keep the output nodes’ charge unchanged because no coarse time resolution is double difference between two current path is connected to power supply or ground any inverter delays because non-inversion neighboring phases more. The oscillation frequency of each GRO can be tuned are preferred to use. discretely by 15 thermometer-code switches c<1:15>. For Multi-clock counter needs to count at each rising edge SGRO, these 15 switches are always connected to ground of ck<0:8>. The count enable signal is SGRO’s gate suggesting a fixed cell delay (also a fixed frequency). control EN_REF. When EN_REF = ”1”, each rising edge Then the delay difference between SGRO and FGRO are triggers the accumulator to add “1” to the sum. When determined only by the 15 switches in FGRO. Actually, to EN_REF = ”0” (the measurement is finished), the counter reduce off-chip controls, the 15 thermometer-code saves the sum into a latch and then resets the accumulator switches are generated by an additional preparing for the next measurement. binary-to-thermometer coder. PFD generates two enable signals (EN_REF, EN_CKV) Sampling flip-flops detect the phase error between each and their reversal signals (ENB_REF, ENB_CKV) to pair of staggered clocks. Once one of those sampled value control NMOS and PMOS enable transistors in GRO becomes “1” (that is the staggered lagged clocks catch up delay cells. EN_REF/EN_CKV is supposed to be a with those leading ones), RBO becomes “1” to reset PFD. rectangular signal suggesting the phase error information The key point for GRO core is the absolute delay in between input clocks REF and CKV. When REF rising each oscillator. When phase ck<0(n+1)> catches up with edge arrives first, for example, EN_REF becomes “1” phase ck<n+1>, as illustrated in lower right of Fig.3, (ENB_REF becomes “0”) forcing SGRO to start o<n+1> needs transmission time to become “1”. The oscillating first. When CKV rising edge arrives after a situation for RBO and EN_REF/EN_CKV is the same. Due while, EN_CKV also becomes “1” forcing FGRO to work. to this kind of gate delay, two GROs keep oscillating for a In this example, EN_REF leads EN_CKV. It means that short while after ck<n+1> is already lagged. The critical the first rising edge in SGRO leads the first rising edge in feedback path has a typical delay of 100ps~200ps FGRO during the measurement interval. If these two (obtained by post layout simulation). No more rising edges oscillators have the same frequency, the phase error will are supposed to be generated during this additional time be always kept. However, due to the delay difference, the interval, otherwise wrong sampling may happen in the
  • 3.
    Fig.3 GRO coreand sampling/reset delay Fig.4 Adaptive PFD next measurement. That means the absolute delay of each vernier line (two-inverter delay) should be longer than 200ps. Here 400ps is adopted for some margin. B. Phase & frequency detector (PFD) In the GRO core, the SRGO (driven by EN_REF) cell delay is always longer than FGRO (driven by EN_CKV) cell delay. It requires EN_REF to lead EN_CKV always. However for a normal PFD in which EN_REF/EN_CKV is corresponding to REF/CKV, this precondition cannot be guaranteed because CKV frequency and phase are adjusted dynamically in ADPLLs. A SR-latch is therefore used as an arbiter, as shown in Fig.4. It senses the leading edge between REF and CKV and gives a sign signal which Fig.5 Multi-clock counter controls four multiplexers. When REF is leading, sign is “0” meaning EN_REF corresponds to REF. On the clock frequency and hardware consumption. In each group, contrary, EN_REF corresponds to CKV. clocks should have a large interval between rising edges, C. Multi-clock counter which eases the strain of accumulator. The two-inverter The digitized phase error can be calculated using delay is about 400ps in this work. To make the counter multiple clocks ck<0:8> from SGRO in each measurement work with a large margin, the nine clocks are grouped into interval. One approach is to assign one individual 5 groups with each containing two. The two clocks with accumulator for each clock, which consumes more adders the longest rising interval (e.g. ck<0> and ck<8>) share and flip-flops. If all clocks could be combined to one that one accumulator. Each clock rising edge produces a contains all edges, only one accumulator is enough. But narrow pulse by inserting a delay in the reversion-input the new clock frequency is inevitably very high bearing on path of the first NAND gate, as shown in Fig.5. Then the the accumulator speed. To tackle the above problem, we second NAND gate combines these two positive pulses to propose to group those clocks to get a tradeoff between one to clock an accumulator. The sum of each accumulator
  • 4.
    multi-clock counter GRO1 PFD GRO2 (a) dummy load Fig.6 Layout of TDC (b) Fig.8 Simulated transistor-level TDC output (a) (b) (a) Bit stream (coded decimal) (b) FFT analysis Fig.7 Simulated PSD of TDC output in Simulink (a) Proposed TDC (b) Classic vernier TDC quantization noise due to coarse resolution is well first-order shaped (see Fig.8(b)). should be synchronized and then added together. The above operation runs only in the measurement interval CONCLUSIONS (EN_REF = ”1”). This paper describes a Vernier-GRO TDC which uses LAYOUT AND SIMULATION vernier principle to get a high time resolution and GRO’s to achieve first-order shaped quantization noise. The The above presented TDC has been implemented using first-order noise shaping pushes much noise to the 90-nm CMOS technology. Fig.6 shows the layout of the high-frequency region which can be filtered by a low-pass proposed TDC (without I/O pads). The active area is noise transfer function in an ADPLL. The chip is 0.18mm*0.15mm. The multi-clock counter uses staggered implemented in 90-nm CMOS technology and occupies clocks from SGRO as referred above. Since FGRO has no 0.18mm*0.15mm area. identical load capacitances as SGRO, some dummy gate loads are used to compensate delay mismatch. REFERENCES We compare the proposed TDC and the classic vernier [1] R. B. Staszewski and P. T. Balsaras, “All-Digital Frequency TDC in Fig.7 by simulating in a simulink platform. Synthesizer in Deep-Submicron CMOS”, John Wiley & Sons, Sinusoidal delay is used as error input. These two TDCs 2006 has the same Δtdelay = 6.4 ps . For a classic vernier TDC, [2] P. Dudek, S.Szczepanski and J.V. Hatfield, “A High-Resolution CMOS Time-to-Digital Converter Utilizing a the quantization noise has a “white” noise floor. However Vernier Delay Line”, IEEE Journal of Solid-State Circuits, in the proposed TDC, most of the quantization noise is vol.35 No. 2, Feb. 2000, pp. 240-247. pushed to high-frequency region (first-order shaped). A [3] Y. Liu, U. Vollenbruch, Y.J. Chen, etc., “A 6ps Resolution 250MHz sampling clock is used in this example. Noise Pulse Shrinking Time-to-Digital Converter as Phase Detector in performance below 20MHz can be well improved, as seen Multi-Mode Transceiver”, IEEE radio and wireless symposium, in the fig.7 (a). pp. 163-166, 2008 Fig.9 shows the transistor-level simulation of TDC [4] S.K. Lee, Y.H. Suh, H.J. Park, J.Y. Sim, “A 1GH ADPLL with a 1.25ps Minimum-Resolution Sub-Exponent in 0.18um output bit stream and its FFT analysis. A dc-delay of 22ps CMOS”, in 2010 IEEE Int. Solid-State Circuits Conf. Dig. Tech. is used as input error. If ΔT denotes the input time error Papers, pp.482-483 between REF and CKV, the digitized error should be equal [5] H.H. Chang, P.Y. Wang, J.H. Zhan, etc, “A Fractional to ΔT / Δtdelay . Δtdelay = 6.4 ps (5 unit capacitor controls Spur-Free ADPLL with Loop-Gain Calibration and Phase-Noise Cancellation for GSM/GPRS/EDGE”, in 2008 IEEE Int. are set to “1”) is set in the simulation. For a regular vernier Solid-State Circuits Conf. Dig. Tech. Papers, pp.200-201 TDC, it should give a result of 3 by simply truncating the [6] Chun-Ming Hsu, Matthew Z. Straayer and Michael H. Perrott, fractional error. But for this TDC, the output varies “A Low-Noise Wide-BW 3.6GHz Digital Δ∑ Fractional-N between 3 and 4 with an average value of 3.41 (see Fig.8 Frequency Synthesizer With a Noise-Shaping Time-to-Digital ( a) ) . T h i s s h o ws a n o t h e r b e n e f i t o f a c c ur a c y Converter and Quantization Noise Cancellation”, IEEE Journal overwhelming the classic vernier TDCs. Moreover, of Solid-State Circuits, vol. No. Dec. 2008, pp. 2776-2786