modeling of passives at 60ghz and beyond - … · collision avoidance radar (77g) d-band (100g to...
TRANSCRIPT
Modeling passives at 60GHz and beyond
Dr. Sharad Kapur
Integrand Software, Inc.
DAC Workshop on CMOS Design at 60 GHz and Beyond: Capabilities and Challenges
Outline
Applications at 60GHz and above
Modeling of passives at high frequencies
3D EM simulation of IC structures
• Methods for accurate and efficient solution
Design issues at high frequencies
Millimeter wave passives
• Couplers, dividers, quadrature-hybrids
Novel IC examples from industry and academia
• Validation and comparison to measurements
• 60GHz to 200GHz
Challenges ahead
Conclusion
Applications at 60GHz and above
Wireless links (60G)
• Chip to Chip
• Base station to Base station
• Wireless HDTV
FMCW imaging (60G)
Collision avoidance radar (77G)
D-band (100G to 170G)
E-band, W-band: (70G-100G)
Doppler Radar (atmospheric monitoring) 180G
Medical applications: 200G+
• Wireless detection of vital signs
Challenges at high frequencies
Modeling of active devices
• Depends a lot on the type of technology, SiGe, BiCMOS, CMOS, GaAs
Modeling of passive components
• Largely dependent on the metallization and substrate
• Does not matter if it is CMOS, SiGe or GaAs
This talk will be about modeling of passive structures and components at very high frequencies
The physics
Behavior of passives described “fully” by Maxwell’s equations
Differential Form
Integral Form
Differential vs Integral
Differential formulations
Finite-element, FDTD
Flexible
Imposes no constraint on shape of metals, dielectric regions
Need to enforce Maxwell’s equations everywhere surrounding the object
Leads to large sparse matrix solve
Integral formulations
MoM, BEM, Integral formulations
Planar dielectrics, conductors
Need to enforce Maxwell’s equations only on conductors (Green’s theorem)
Leads to smaller dense matrix to solve
Many techniques developed recently
For IC passives this approach is the best
6
Tech trends
Thick metals
• 0.1um to 8um copper
High-resistivity substrates
• 10 -cm to 1000-cm
Fine feature sizes
• 0.05m width at 28nm
Many metal layers
• High density capacitors
Easier for MOM/Integral solutions that do not need to discretize dielectric interfaces and only conductors
7
IC processes offer tight tolerances, low variability
When a device can be built with reasonable quality compared to an off-chip or an LTCC structure it will be integrated
It is difficult to build off chip passives at very high frequencies (easier to build on chip).
• Measurement equipment is now using integrated passives
IC processes
8
3D planar EM simulation
EMX is a 3D EM simulator
3D volume integral formulation (time-harmonic)
Unknowns are charges and currents
• Surface charges and volume currents
Includes retardation (time of flight)
Also skin effect is critical at higher frequencies so you need a volume current and meshing of the metals
Matrix formulation
Suppose that N elements in the mesh
Conventional approach O(N3) time and O(N2) memory
• Cost is prohibitive
• Double the size of the problem 8X time
10
(continuous form)
BAx (A is a dense matrix)
Iterative methods (GMRES, Yale, 1986)
Matrix vector products instead of matrix inversion
This reduced the time to O(N2)
The Fast Multipole Method was developed in 1987
Applied to capacitance/EM solution
• FastCap, Fast Henry, (1990s). (MIT/White)
• IES3, Bell Labs, Kapur and Long, (1990s)
Used in EMX O(N)
These sorts of problems can be solved in linear time
Innovations in numerics (GMRES+FMM)
11
bAx
},...,,,{ 2 bAbAAbbK n
n
Time and Memory scaling
Single frequency simulation (including iterative solve)
Compare speed and memory for 1, 2, 4, 8, …, 64 inductors
Very important that it is a linear time solution
1 inductor 64 inductors
12
Issues at high frequencies
At high frequencies the skin effect is a significant issue
Meshes need to be finer to incorporate these effects
This leads to larger problem size (smaller structures?)
1GHz 60GHz
Issues at high frequencies
At high frequencies the skin effect is a significant issue
Meshes need to be finer to incorporate these effects
This leads to larger problem size (smaller structures?)
1um mesh 0.1um mesh
Issues at high frequencies
Coupling at higher frequencies
1GHz 60GHz
Full-wave vs Quasi-static
Full-Wave vs quasi-static
Time of flight becomes important once frequency is “large” compared to the speed of light
• Electrical length of structure is important
• Specially important for electrically long structures
Cannot compromise on using a quasi-static solution
Laplace regime…not yet Helmoltz.
Design issues
Inductors become tiny and getting high Q is easy
Capacitors become the bottle neck at higher frequencies since capacitor Q goes down because of skin effect
Courtesy: TSMC. 65nm RFCMOS, 9LM thick metal technology. Published at RFIC 2009 “Including Pattern-Dependent Effects in Electromagnetic Simulations of On-Chip Passive Components”, Integrand and TSMC
Guard rings
Use of guard rings or seal rings in layout
The ring is a loop of metal
May be a metal line close to the inductor (supply line or bias line)
Simulations
Using a guard ring shows odd behavior at higher frequencies
The ring is a metal line of about 600um long.
At 50Ghz the wavelength is 3000um.
Order of a quarter wavelength line. Couples to the inductor as an LC resonator
Removing the ring or cutting the ring fixes the issue.
Dummy fill simulation
EMX models circulating currents in dummy fill. Eddy current loss increases at higher frequencies
Passive couplers
Couple defined amount of power in transmission line to a port enabling the signal to be used in another circuit
Couple power flowing in one direction. Power entering the output is not coupled to the input port
Directional couplers constructed from two coupled transmission lines set close enough together such that energy passing through one is coupled to the other
3dB couplers, 6dB couplers, etc.
Not “lumped”. Usually ¼ wavelength
6dB coupler
Passive dividers
Wilkinson divider
Achieve isolation between the output ports while maintaining a matched condition on all ports
User also be used as a power combiner because it is made up of passive components and hence reciprocal
Wilkinson Divider
Wilkinson Dividers can be used as power splitters and power combiners
In silicon it is not ideal and all real effects need to be included (ground plane, resistors, etc.)
EM simulation needed
Alcatel Lucent
Circuit Examples at high frequencies
Experimental Validation
Measurement and Circuits for 60GHz-200GHz
1. High-resolution 60-GHz DCO with Reconfigurable Distributed Metal Capacitors in Passive Resonators
2. Amplifiers for D-Band applications
3. W-band data links, phased array receivers and transmitters
All simulations of passives done with EMX
High resolution DCO at 60GHz
60-GHz FMCW imaging (all digital PLL)
Range:
• 3-5m, resolution: 1-5cm
• Through cardboard boxes
60-GHz DCO (L-DCO, T-DCO) • 6-GHz linear freq. tuning range for
modulation
• ~1MHz fine tuning steps
• Moderate power, small area
• Interfacing with divider, PA in TX, mixer in RX
• Mm-wave frequency divider chain
W. Wu, J.R. Long, R.B.Staszewski, J.J. Pekarik (Delft University) "High-resolution 60-GHz DCOs with Reconfigurable Distributed Metal Capacitors in Passive Resonators," IEEE Radio Frequency Integrated Circuits Symposium, June 2012.
Test chip micrographs
26
L-DCODiv-64
60 GHz output
Div-64 output
Digital
control
400 um
36
0 u
m
T-DCODiv-64
Digital
control
60 GHz output
Div-64 output
400 um
38
0 u
m
• Inductor and Transformer-based fine tuning DCOs and divider-by-64 chain
Complete L-DCO tank EM simulation
Including:
• Coarse, mid-coarse tuning bank
• Fine tuning bank
• Interconnection to divider and buffer
• Mimcap for AC coupling
• Ground ring
100+ port EMX simulation
TU-Delft Confidential 27
T-line and inductor tuning
Passive structures with tuned by capacitive loading
• Transmission line with fine and coarse cap banks
• Inductor with shield strips
28
Coarse
tuning bank:
M7
Transmission line
on M8Cin
Mid-coarse
tuning bank:
M6
D0
Lin
CPFine tuning
bank: M6
01
15
Thermometer c
ode index
60-GHz DCO measurements
29
Measured vs. simulated coarse tuning curves of DCOs
-2 0 2 4 6 8 10 12 14 16 18 20
56
57
58
59
60
61
62
63
64
DC
O o
utp
ut fr
eq
ue
ncy (
GH
z)
L-DCO measured T-DCO measured L-DCO simulated T-DCO simulated
Coarse-tune thermometer code index
-2 0 2 4 6 8 10 12 14 16 18 20
-94
-93
-92
-91
L-DCO meas.
T-DCO meas.
L-DCO sim.
T-DCO sim.
DC
O P
N a
t 1
MH
z o
ffse
t (d
Bc/H
z)
Coarse tune thermometer code index
-90.5
-91.5
-92.5
-93.5
L-DCO
T-DCO
Phase noise agreement
180GHz Amplifier
D-band amplifier
6dB coupler
180GHz H2O attenuation window
Atmosphere monitoring
University of Toronto IMS 2012 (private communication) Ionnis Sarkas, Sorin Voinigescu
6dB coupler and amplifier
S-parameter measurements of 6dB coupler up to 180GHz
Silicon based radar imaging
Silicon-based radar and imaging sensors operating above 120 GHz
D-band sensor imaging
Used EMX to model supply ground distribution network up to 200GHz
The on-chip portion of the network consists
of a 200-um long and 100-um wide periodic array of mesh cells of interleaved supply and ground planes
IEEE MIKON 2012 May 21-23,
Warsaw, Poland.
University of Toronto
Silicon based radar imaging
Silicon-based radar and imaging sensors operating above 120 GHz
E-Band and W-Band Data links
Receiver and transmitter arrays designed for steerable-beam and highly spectral efficient data links in both E- and W-band.
4-channel receiver and transmitter array with integrated direct-conversion mixers.
The arrays span newly released commercial bands at 71-76, 81-86 92-95 GHz
Low-noise figure and good output power.
RFIC 2012
RFIn1
RFIn2
RFIn3
RFIn4
LNA
LNA
LNA
LNA
I/Q Down-
converter
IFIOut
IFQOut
LOIn
SPI Control
Ac
tiv
e &
Pa
ss
ive
Po
we
r C
om
bin
er
Phase
Shifter (1)
Phase
Shifter (2)
Phase
Shifter (3)
Phase
Shifter (4)
W-band chip
All passive traces modeled with EMX
Wilkinson Divider
Transformer Hybrid architecture
MM wave passive circuits. Not easy to construct a discrete component equivalent
Need full S-parameter EM simulation
All simulations in the 70GHz to 200GHz range
Measured results
-2.5
0
2.5
5
7.5
10
-15
-10
-5
0
5
10
75 80 85 90 95 100 105 110
Me
asure
d M
agnitu
de
Error (d
B)Me
asu
red
Ph
ase
Err
or
(De
gre
es)
Frequency (GHz)
Single Slice Phase/Mag Accuracy
0
5
10
15
20
25
30
35
40
70 75 80 85 90 95 100 105 110
Me
asu
red
Re
ceiv
er
Gai
n (
dB
)
Input RF Frequency (GHz)
Receiver Gain - Element 1 (H3)Receiver Gain - Element 2 (H3)Receiver Gain - Element 3 (H3)Receiver Gain - Element 4 (H3)Receiver Gain - Element 1 (H2)Receiver Gain - Element 2 (H2)Receiver Gain - Element 3 (H2)Receiver Gain - Element 4 (H2)
5
6
7
8
9
10
11
12
13
14
15
70 75 80 85 90 95 100 105 110
Me
asu
red
Re
ceiv
er
NF
(dB
)
Input RF Frequency (GHz)
Receiver NF - Element 1 (H3)Receiver NF - Element 2 (H3)Receiver NF - Element 3 (H3)Receiver NF - Element 4 (H3)Receiver NF - Element 1 (H2)Receiver NF - Element 2 (H2)Receiver NF - Element 3 (H2)Receiver NF - Element 4 (H2)
Conclusions
Maxwell’s equations used to model passive structures at high frequencies
Need fast 3D Full-wave EM solver to capture relevant effects
Linear time solver like EMX
Design issues
• Coupling to stray lines
• Capacitor Q degradation
• Retardation needs to be modeling
Circuits can and are being built for very high frequency applications
Challenges ahead…
Constant improvement of speed and memory use
Cannot do full-chip level 3D EM extraction
Better characterization of interconnect and substrate
Packaging considerations
• Simple RLC models for packages won’t suffice (seeing this for SERDES applications 30-40GHz).
Measurement difficulties at very high frequencies
Improvements in measurement equipment for D-band and W-band (higher frequencies)
Acknowledgements
TU Delft
• Providing DCO example
University of Toronto
• 6dB coupler example
• Imaging example
Alcatel Lucent
• W-band data link Rx/Tx example
Helic and TSMC for hosting workshop.
Extra Slides
Foundation of EMX
Uses the hierarchical Fast Multipole Method for rapid solution
• Nearby regions are handled directly
• Far regions are approximated to fixed precision
Solution time is O(N) in time and memory, i.e. double the problem size, double the time
• as opposed to N2 or N3
Can solve very large problems with minimal computer resources
41
Fabricated width is different from drawn width according to rules provided by foundry by a “bias” amount
Shaded regions represent original drawn geometry
Lines represents modified “grown” geometry based on local width and spacing
Modifying layout
42
CMOS VCO
43
3D mesh (inductor+ capacitor bank)
Courtesy: Wipro, TSMC90nm, 1P5M