a 0.84ps-lsb 2.47mw time-to-digital converter using charge ......a 0.84ps-lsb 2.47mw time-to-digital...
TRANSCRIPT
A 0.84ps-LSB 2.47mW Time-to-Digital
Converter Using Charge Pump and
SAR-ADC
Zule Xu, Seungjong Lee, Masaya
Miyahara, and Akira Matsuzawa
Tokyo Institute of Technology, Japan
2Outline
• Motivation
• Issues of conventional techniques
• Proposed TDC
• Measured performance
• Conclusion
3Motivation
• A fine resolution TDC contributes low in-band phase
noise to a digital PLL
• Example: fv = 4GHz, fref = 40MHz, PN = -120dBc/Hz
tres = 0.87ps !
• Delay-chain’s resolution is limited to its unit delay
fv
fref
...td td td
Dout
...
Encoder
Counter
TDC
DLF DCO
Σ
FCW fv
fref
4Motivation
• For finer resolution, more energy, more area, or
more conversion times are traded off
• Is there a solution for the best balance?
0
100
200
300
400
0 0.02 0.04 0.06 0.08 0.1
Fo
M [
fJ/b
it]
Area [mm2]
5.5ps
1.25ps [JSSC`12]3.75ps
4.7ps
4.8ps Noise shaping1ps [JSSC`09]
5Issues of Recent Techniques
• Vernier chain
– PVT and jitter effects
– Arbiter’s metastability
(unacceptable when
the input <= 1ps)
CK1
CK2
...td1 td1
...
...td2 td2
...D[0]
Delay
chain
CK1
CK2
Encoder
...
Dout
TADelay
chain
• Pipeline
– Nonlinearity of the time
amplifier (TA)
– Mismatch
6Issues of Recent Techniques
• Stochastic
– Short linear range
– Highly dependent on
layout and process
N arbiters w/ mismatches
CK1
CK2
Dout
• Noise shaping
– Low input signal
bandwidth
– Requires a fast clock
for the counter
Σ
(Counter)
Digital-to-time
converter
QTin
Dout
CKfast
7Issues of a Previous Technique
• Time-to-amplitude conversion has achieved
fine resolution but…
• Fast clock necessary
• Large capacitance
• Susceptible to
leakage
• Low speed
Dout
Counter
CKfastRST V1
Tin
I
RST
I/N
EN
Fa
lling
ed
ge
trigg
ere
d
C
M*C
V2
V1
V2
...
Tin
EN
CKfast
tres = Tckfast/(M*N)
[E.R.Ruotsalainen, et.al,
pp.1507-1510, JSSC 2000]
tres = 32ps, 0.5um BiCMOS
8Time-to-Charge Conversion
• Time-to-charge conversion suggests the
potential for extremely fine resolution
ADCDout
Tin
Vout
CI RST
C 1pF
I 1mA
Vlsb 1mV
tres 1ps
tres = C*Vlsb/I
• Thermal noise restricts C and I
9Thermal Noise
• Noise increases with integration time (Δt)
• Trade-off exists between power (I) and area (C)
d
m
res
lsbvn
I
g
t
V
C
tkT
4
2
A constant for transistor sizing
Crdsin2
σvn2
ΔtΔt
256 512 768 1024
0.1
0.2
0.3
0.4
0.5
t [ps]
vn [
LS
B]
tres
= 1ps, 10-bit, Vlsb
= 1mV, gm
/Id = 9
0.5pF, 0.5mA
1pF, 1mA
2pF, 2mA
10Proposal
• Time-to-charge conversion on a SAR-ADC
• Short Ton is required to suppress the noise; Ton ≈
200ps in this design
SAR-ADC
Dout
...
...
Cu 2n*Cu
...
...
Cu 2n*Cu
Charge pump
...
CKadc
Vp
Vn
Ton
UP
CK2
CK1R
RDN
QD
QD
PFD
11Noise and Speed
1) Thermal noise from the charge pump accumulated
during (Tin+Ton)
2) SAR-ADC should be faster than 40MS/s
Tin
Ton
CK1,CK2
UP, DN
Vp, Vn
Tin
ΔV
CKadc
Sampling Conversion Sampling
...
... ...
...
Dout
< 24ns
Dn Dn+1
~25ns
12SAR-ADC
• 12-bit topology
– 10-bit ENOB@40MS/s
– [email protected] power supply
[S. Lee, SSDM 2013, to be presented]
• Dynamic comparator
– Low power
• Metal-Oxide-Metal capacitor
– High density
• Same pitch of cap. and switch
– Better matching and
scalabilityUnit capacitor
and switch
Top
Bottom
MOS
SW
Top
Bottom
MOS
SW
...
13Scaling of the SAR-ADC
• Down scaling lower power and smaller area
shorter range but not harmful for an integer-N PLL
• Resolution and intrinsic SNR are not degraded since
the required charges are not changed
oninpmnm
cpreslsb
diffvn
lsb
TTgg
ItCV
kT
V
1
2
12
,
2
Other
LSBs8Cu 32Cu 64Cu16Cu
Other
LSBs8Cu 16Cu
Other
LSBs
12-bit
10-bit 8-bit Unchanged Q2
Intrinsic SNR:
14Charge Pump
• The CMFB matches the currents of PMOS and
NMOS
Vbp_fb
Vcp
RSTUP UP
DN
RST
UP
DNDN
DN
Vcn
Vbn VbpVcm
Φ2
Φ1
Vp Vn
UPVm
Φ1 Φ2
Settled
voltage
• Conventional
– Amplifier-based
CMFB
• Proposal
– Switched-capacitor
CMFB
– For low power and
low voltage
Reset mode
15Charge Pump
• Other mismatches contribute to a static offset
• It can be canceled in the digital domain
Vbp_fb
Vcp
RSTUP UP
DN
RST
UP
DNDN
DN
Vcn
Vbn VbpVcm
Φ2
Φ1
Vp Vn
UPVm
Φ1 Φ2
Held voltage
Charging mode
-4 -3 -2 -1 0 1 2 3 40
10
20
30
40
50
(Vp-Vn) [mV]
Co
un
t
σ ≈ 1mV
Monte-Calo simulation
of the output voltage
(Δt=0, 1pF load cap.)
16Implementation
• CMOS 65nm, core area = 0.06mm2
• Measurement setups
SG1
SG2
TDCLogic
analyzer
40MHz
40MHz + 5Hz
Pulse
generator
Splitter TDCLogic
analyzerΔt
~100ps rising time
17Measurement
• DNL and performance summary
0 32 64 96 128 160 192 224 256
-1
0
1
DNL and INL in 8-bit with 0.84ps/LSB
DN
L [
LS
B]
0 32 64 96 128 160 192 224 256
-2
0
2
Code
INL
[L
SB
]
Performance Value
CMOS [nm] 65
Supply [V] 1.0
Conv. Rate [MS/s] 40
Power [mW] 2.47
Resolution [ps] 0.84
Range [bits] 8
DNL [LSB] -0.7/1.0
INL [LSB] -2.7/1.7
18
113 114 115 116 1170
0.5
1
1.5
2
2.5x 10
5
Code
Co
un
ts
67 68 69 70 71 72 73 74 75 76 770
0.5
1
1.5
2
2.5x 10
5
Code
Co
un
ts
Measurement
• Single-shot precision: < 1LSB
• The thermal noise increases with longer input
time interval
Δt=11ps
σ=0.24LSBΔt=47ps
σ=0.45LSB
19Performance Comparison
• Best balance is achieved
JSSC’10 VLSI’11 VLSI’12 ESSIRC’10 This workImproved
(Simulated)
Type Vernier Pipeline Noise Shaping Stochastic Charge Charge
CMOS [nm] 65 130 130 65 65 65
Supply [V] 1.2 1.3 1.2 1.2 1.0 1.2
Resolution [ps] 4.8 0.63 3 3 0.84 1
Range [bits] 7 11 11 4 8 10
DNL [LSB] <1 0.5 N/A 1.4 -0.7/1.0 -0.2/0.2
INL [LSB] 3.3 2 N/A 1.5 -2.7/1.7 -2.7
Frequency
[MHz]50 65
90
(OSR:16)40 40 100
Power [mW] 1.7 10.5 3.2 8 2.47 4
Area [mm2] 0.02 0.32 0.43 0.04 0.06 0.018
20Energy and Area Efficiencies
• Best balance is achieved
NFrequencyPowerFoM 2//
Energy efficiency:NBWPowerFoM 2//2/or
0
100
200
300
400
0 0.02 0.04 0.06 0.08 0.1
Fo
M [
fJ/b
it]
Area [mm2]
5.5ps
1.25ps [JSSC`12]3.75ps
4.7ps
4.8ps Noise shaping1ps [JSSC`09]
0.84ps [This work]
1ps[Improved, simulated]
21Conclusion
• The proposed TDC has achieved 0.84ps resolution,
2.47mW power consumption, and 0.06mm2 area
• The proposed TDC suggests the best balance
among resolution, energy, area, and conversion
times
• The proposed TDC has no issues from delay chains,
TAs, or arbiters
• The proposed TDC can be a practical solution for
digital PLLs
22Acknowledgement
• This work was partially supported by
HUAWEI, Berkeley Design Automation for the
use of the Analog Fast SPICE(AFS) Platform,
and VDEC in collaboration with Cadence
Design Systems, Inc.