mmWave IC Design in CMOS Technology
Andrea Mazzanti
June 21, 2016
Università degli Studi di Pavia
Laboratorio Circuiti Integrati Analogici
Introduction
• mmWave refers to the frequency range 30GHz – 300GHz(wavelength spans from 10mm to 1mm)
• First experiments of mmWave generation date back to 1890’s.
• In the last century, mostly used for research, with exotic technologies (waveguides, vacuum tubes, compound semiconductors).
• Did not find wide application due to cost/performance issues
• With cut-off frequency in excess of 300GHz, mainstream CMOS technology is expected to enable mmWave mass-market applications.
2
• mmWave applications & CMOS technology
• Components and building blocks•Low Noise Amplifiers
•Oscillators
•Power Amplifiers
• Conclusions
Outline
3
Automotive Radars
77-81GHz Short-range radar: parking assist, object detection Long-range radar: automatic cruise control, low visibility
(fog),object detection Long range vision: automatic driver Products in silicon (SiGe BiCMOS) already developed
4
Radar Operating Principle
5
High-Resolution imaging demonstrator96 TX+96 RX
1m
2m
2mm
ROHDE & SCHWARZ , InfineonIEEE Microwave Magazine, 2012
• 70-82GHz operation frequency. SiGe Bipolar technology. • 3072 TX + 3072 RX, 1536 ICs. ~kWatt power dissipation, ~1mW radiated power • Massive integration may reduce power/size/cost• Proposed for security screening, but other applications are possible : medical
imaging, 3D cameras… 6
Gbit/s wireless connectivity at 60GHzLeveraging the availability of wide spectrum for high-rate transfer
W-PAN
Few products already on the market
• 7 GHz of unlicensed bandwidth• Potentially more than 10Gbit/s wireless
connectivity • High O2 attenuation -> improve security
& spectrum reuse
7
Mobile Connectivity Beyond 2020Massive growth in traffic volume
Wire
less
Tra
ffic
(Exa
B/y
ear)
Massive growth in connected devices
5G mobile network expected to provide impressive growth in speed & capacity
8
mmWave is a key technology for 5G• mmWave spectrum attractive to enable high
speed access to the infrastructure. • Discussion on-going for best frequency band
(28GHz, 38GHz, 60GHz, 70GHz…)
source: IEEE Comm. Magazine, sept. 2014, & presentations from Samsung, Intel, NSN, Ericsson, Huawei
• High density of access points (small-, pico-, femto-cells)
• Multi antenna systems (MIMO) in terminals and access points
• Coverage from 20m to 200mt• Intensive use of mmWave
links for backhauling
9
mmWave 5G demonstrator
• power, size and cost reduction are now the key challenges • Hot research topics for academia and industry in the near future:
- mainstream technologies for cost & high integration- mix. of different technologies (3D, SiP) for power & performances- baseband-radio-antennas co-design- optimized circuits and architecture for dense MIMO systems
• Very complex transceivers architectures to support beam-forming & beam-steering.
• Demonstrators at 28GHz proposed by Samsung (IEEE Com.Magazine, Feb. 2012)
10
CMOS Technology
Opportunity and challenges
11
Why CMOS technology
• Speed of MOS transistors is continuously improving with scaling
• Cheaper, for large volumes, compared to SiGe BiCMOS and III-V technologies in particular
• Unprecedented integration level with power digital signal processing (DSP), analog circuits and antennas?!
• Passive devices (transmission lines, inductors, capacitors) easier to integrate (less area) at higher frequencies
• Communication and computing applications have benefited from these trends
• First mmWave CMOS circuits from Berkely at ISSCC 2004, CMOS 0.13um, 60GHz frequency 12
Cut-off Frequency
13
Supply Voltage & Output Conductance
Gate Length
gm/gds at biasing for max fT
Supply Voltage
J.Pekarik et al., IEEE CICC 2004
• in 65nm node (fT ~160GHz) gm/gds~ 6 , Vdd=1V
• gm/gds represents maximum voltage gain
• Low supply voltage limits dynamic range and linearity of amplifiers, output power and efficiency of PAs, phase noise of oscillators
• Issues common to Analog design but at mmWave:
- must use Lmin and high current density for speed (fT)
- cascodes introduce parasitic poles- Current-mode circuit techniques not
yet applicable
14
Losses of parasitic capacitors
• Large polysilicon gate resistance (rg) at small gate length• rg is responsible for power dissipation when AC current flows through Cgs
Example: 20x1um / 65nm nMOS device has rg ~ 16Ω, Cgs~ 28fF
Qg @ 60GHz = 5.9 Qg @ 6GHz = 5915
Passive ComponentsBack-End Of Line (BEOL) critical for passive components:
Shrinking of the metal layers leads to:
-Larger metals resistance (Rs, losses)
-Larger substrate parasitics (Csub, Rsub)
16
RF vs mmWave CMOS design
fo 5 GHz 60 GHz
Technology 250 nm 90 nm
fT 20 GHz (4 x fo) 120 GHz (2 x fo)
Supply Voltage 2.5V 1.2V
gm/gds 15.2 10.6
Q of gate capacitance > 40 ~6
Inductor Q 10-15 15-20
First CMOS RF products (5GHz) with 250nm tech. nodeFirst CMOS mm-Wave products (60GHz) with 90nm tech. node
17
Improvement with technology scaling?0.13 um 6 layers
IEDM 2010, S. Francisco
Continuous scaling driven by complex Systems on Chip
~ 20-30% fT improvement only per generation
aggressive scaling of metal layers• large impact of routing parasitic (layout) • passive components penalty
18
• mmWave applications & CMOS technology
• Components and building blocks•Low Noise Amplifiers
•Oscillators
•Power Amplifiers
• Conclusions
Outline
19
Device Modeling
CMOS technology Design Kits (DKs) usually do not (fully) support mm-Wave design:
- device models are not very accurate beyond 20 GHz- mmWave components not characterized or missing (e.g. very small inductors and capacitors, TLINEs)
Device customizations, optimizations, measurement and modeling is a key step for mmWave design
20
Inductors & Capacitors Modeling- frequency dependence of losses (skin effect)- differential vs single-ended excitations- substrate losses
21
Transmission Lines• Usually not available in design kits• short λ at mmWaves makes TLINES small
enough to be used on-chip • long interconnections between building blocks• confinement of EM fields minimize cross-talk,
spurious coupling between blocks• emulate lumped components (capacitors,
inductors)• matching networks, phase shifters
0.0
0.2
0.4
0.6
0.8
1.0
0 10 20 30 40 50 60 70
α [d
B/m
m]
Freq. [GHz]
Optimized geometry and shield for substrate loss demonstrated a record attenuation of 0.6dB/mm@60GHz
IEEE JSSC-2009, UniPV 22
Active Devices• Layout of transistors determines extrinsic parasitics and strongly affect
performances
• Maximum frequency of oscillations, fmax is a good figure of merit :
minimize Rg minimize Cgd maximize fT 23
Minimize Rg & Cgd larger spacing between gate (G) and drain (D) contacts may help to reduce Cgd
Gate resistance: double gate contacts and many small fingers in parallel
Rg’ = Rg / 16
duble contact, multi-fingers device
124
fmax
For given device size and layout, fT is maximized by maximizing gm
25
CMOS Building Blocks:
- Low Noise Amplifiers (LNAs)- Voltage Controlled Oscillators (VCOs)- Power Amplifiers (PAs)
26
Low Noise Amplifier
• LNA must provide low NF and high gain
• 50 ohm nominal input impedance (S11 < -10dB)
• Unconditional stability against variations of the antenna impedance• Linearity requirements less critical than for RF applications
27
Common Source Input Stage• L1 resonates at output
• Stability concerns due to Cgd
• Relatively high gain, thanks to contribution of the matching network
UniPV, CICC 2007
Matching on a noisy resistance (rg) introduces a 3dB lower bound on NFEx: rg=16Ω, Qg=6 @ 60GHz
-> Amatch= 4.6dB
• gain 9-12 dB
• NF ~ 5.5-6.6 dB
• Matching network provides voltage gain:
28
Inductively Degenerated Input Stage• Ls introduces a noiseless real part on the
input impedance at the gate:
• Eliminates the 3dB lower bound on NF but looses gain from the matching network
• More noise contribution from load and following stages
• Lower gain from matching network
Ex: 3dB penalty if ωTLs = rg29
LNA ExamplesPellerano et al,
JSSC 2008Heydari et al,
JSSC 2007Weyers et al, ISSCC 2008
Yao et al, JSSC 2007
Tech. 90nm 90nm 65nm 90nm
fo 64GHz 63GHz 60GHz 57GHz
-3dB BW 8GHz 4.5GHz 7.5GHz 6GHz
Gain 13.5 12.2dB 22.3dB 15.5dB
NFmin 6.7 6.5dB 6.1dB 5.5dB
# stages 2 2 3 3
Power 48mW 10.5mW 35mW 24mW
• NF bounded to 5-6dB. (NF down to 2dB or less demonstrated at RF)• Multi-stages required to achieve reasonable gain (5–6 dB gain/stage)• Limited bandwidth ~10%. (at RF, more than 100% demonstrated)
30
GBW in multi-stage LNAs
• The inductors between stages resonate with parasitic caps (c1,c2)
• R represents losses of LC-tank at resonance:
input match gm
V1
V2
Rc1
c2
in
50ΩV3
𝑅𝑅 =𝑄𝑄
𝜔𝜔0(𝑐𝑐1 + 𝑐𝑐2)
• -3dB bandwidth of each stage: 𝐵𝐵𝐵𝐵 =𝜔𝜔0𝑄𝑄
• Voltage gain of each stage: 𝐴𝐴 = 𝑔𝑔𝑔𝑔 𝑅𝑅 =𝑔𝑔𝑔𝑔𝑄𝑄
𝜔𝜔0(𝑐𝑐1 + 𝑐𝑐2)
• Gain-Bandwidth Product (GBW)= fixed by transistor parameters 𝑔𝑔𝑔𝑔
(𝑐𝑐1 + 𝑐𝑐2)31
GBW enhancement with coupled resonators
-10
-8
-6
-4
-2
0
40 50 60 70 80Freq. [GHz.]
Single LC Resonator
Norm
. Gai
n [d
B]
Cc ↑
• gain-bandwidth trade-off in singly-tuned inter-stage loads leads to large dissipation.
• coupled resanators are introducedto increase bandwidth and limitdissipation
32
Coupled-Resonators LNA Results
0
5
10
15
20
25
30
-20
-10
0
10
20
30
40
50 55 60 65 70
Noi
se F
igur
e (d
B)
Gai
n (d
B)
Freq. (GHz)
-25
-20
-15
-10
-5
0
50 55 60 65 70
|S11
| (dB
)
Freq. (GHz)
Gain 28 dBCenter freq. 60GHz
3dB-Bandwidth 14GHzGBW 350GHz
NF < 5.5 dBPower 20mW
33
CMOS Building Blocks:- Low Noise Amplifiers (LNAs)- Voltage Controlled Oscillators (VCOs)- Power Amplifiers (PAs)
34
Voltage Controlled Oscillator (VCO)
• Main block of the synthesizer, provides LO for mixers
• Oscillator with output frequency set by Vctrl
• Large output signal swing desirable
• Frequency tuning enough to cover channels + processing spreads
• High spectral purity (low phase noise)35
MOS VaractorTrade-off between Quality Factor (Qv)
and capacitance variation Cmax/Cmin
Measurements at 24GHz
For Lg=200nm
QV @ 24GHz = 15
QV @ 60GHz = 6
At 60GHz QV ~2.5 times lower than 24GHz
[C.Cao JSSC 2006]
36
LC-Tank VCO
Rp represents tank losses
At LC resonance:
37
LC - Tank Losses
A comparison with RF:
60 GHz: QL=15, QC = 6 → QT = 4.2 Varactors limit QT
5 GHz: QL=10, QC > 40 → QT > 8 Inductors limit QT
Two times more losses → ~ two times more gm for oscillator start-up
QL QC
38
Phase Noise
Noise of the devices perturbsrandomly the zero crossing ofthe oscillator output signal
Spectrum of LO with phase noise
39
Phase Noise Requirements
• Phase noise rotates signal constellation and impairs BER
IEEE 802.15-06-0477-01-003c [Online].
2Gbit/sec16-QAM
40
Phase Noise
Power Dissipation
• Maximize QT → trade-off with T.R.
• Maximize V2osc / Rp:
A comparison with RF:
60 GHz, QT = 4.2
5 GHz, QT > 8
60GHz VCO would require522 x power dissipation forthe same phase noise…
essentially not possible
IB
41
Tuning Range
•The tank is loaded by many fixed capacitivecontributions (parasitics, load)
•CV hardly exceed 20-30% of the total tankcapacitance
•T.R. drops dramatically as frequencyincreases
42
VCO example - I[UniPV, RFIC 2008]
• 65nm CMOS, 7.2mW from 1.2V supply• Analog + Digital Switching of Varactors• fo=53GHz, Tuning Range 11.5%• Phase Noise @ 10MHz offset -118dBc/Hz
43
VCO example - IIUniPV, ISSCC2013
• 32nm CMOS, 9.8mW from 1V• f0 40GHz, outstanding T.R of 32%• Phase Noise @ 10MHz offset -115 / -118dBc/Hz
new LC-tank topology: inductor splitting with Mswand transformer feedback to extend Tuning Range w./o. penalty on Q
44
115GHz LO generation with Injection LockingUniPV, ISSCC 2010
Tuning-Range decreases dramatically with frequency
At > 60GHz, VCO + doubler yelds better performance than a fundamental frequency VCO
55GHz VCO110GHz injection-locked oscillator
An injection-locking scheme based on a modified Clapp oscillator topology was proposed as a low-power, wideband frequency doubler
45
115GHz LO generation with Injection LockingUniPV, ISSCC 2010
Record Tuning Range (6x) with state-of-the art phase noise and power dissipation for the VCO+doubler chain
65nm CMOS
46
CMOS Building Blocks:- Low Noise Amplifiers (LNAs)- Voltage Controlled Oscillators (VCOs)- Power Amplifiers (PAs)
47
Power Amplifier
• Deliver mmWave power to the antenna
• Stability against load impedance variation
• Linearity very important (efficient modulation needs linear amplification)
• Power efficiency
• Power gain: multiple stages (PA + driver(s) )
• Device stress: long time reliability48
Power Amplifier
• Class-A for max power gain and linearity. L1 resonates with cap. at node X
• Vmax (0-pk drain swing) ≤ VDD
• Given the power (P) and Vmax, Rin is calculated from
• Drain voltage > VDD : long term reliability concerns
for a RF voltage > 60% of nominal supply, Pout degrades by 2dB over 7 years
[Ruberto, IEEE RFIC 2005] 49
Power Efficiency
Finite Q of passive components, low supply voltage and Class-A operation limit drastically the power efficiency
PX Pout
50
Numerical Example
Assume: VDD=1V, PX = 25mW, Vmax=0.6V ( Vdrain_max=1.6V )
PX Pout
Required Rin is 7.2Ω, Iin = Vmax/Rin = 83mA ( = IDC, Class A)
Pdiss= VDD x IDC = 83 mWηM1 = 30%
51
Matching Network LossesQ of passives (L,C, TLINEs) responsible for matching losses
PX Pout
Rin is 7.2Ω , Q=5: ηmatch = 66%, Pout=16.5mW
η = ηM1 x ηmatch = 19%
52
Power Combining Techniques
Efficiency degrades as output power increases
Use many low power PAs in parallel• Lumped elements power combining (transformers)• Distributed elements power combining (TLINEs)• Spatial Power Combining (Antennas) 53
High-gain, Wide-Band PA challenge
The higher the gain-per-stage, the higher is the efficiency (left plot)
Due to GBW limitation, efficiency is reduced in wide-band PAs (right plot)
Hard to achieve efficiency > 10% with 15GHz bandwidth
output match
input match
outputstagedriver
pre-driver
50Ω
Low gain at mmWave, needs several driving stages.
54
Coupled Resonators for GBW enhancement in PAsmagnetically-coupled resonators in (1) matching networks and (2) power splitter/combiner in a 2-paths differential CMOS PA
UniPV, [CICC2015, JSSC2015]
55
Coupled Resonators for GBW enhancement in PAsUniPV, CICC2015, JSSC2015
Summary & Conclusions
• mmWaves band offers a variety of new applications• Technology scaling opened the opportunity to develop CMOS
ICs working a tens of GHz• Building blocks and transceivers demonstrated• Big performance gap, compared to RF(1-10GHz) ICs• Mild improvement expected from further.• The high operating frequency and CMOS limitations introduce
many new challenges to circuits design• Space for new ideas at all levels: device, circuit and
architecture57