lec3 (substrate etc).ppt · why an 8-form coil?? p. andreani – system-on-chip n. itoh et al.,...
TRANSCRIPT
Noise coupling in SoC/SiP
System-on-Chip
Pietro AndreaniDept. of Electrical and Information Technology
Lund University, Sweden
• “Switching” noise generation
• Noise injection and reception
• Substrate coupling in CMOS
• Noise propagation
Overview
P. Andreani – System-on-Chip
• Biasing strategies
2Noise coupling
Vdd
Types of noise coupling
• Crosstalk • dV/dt, dI/dt
• Power line noise coupling (dI/dt)
ThicknessSpacing
Crosstalk
Vss
Cir.1 Cir.2dI/dtVdrop
P. Andreani – System-on-Chip
• Substrate noise coupling (dV/dt)
• Radiation noise coupling
λ1GHz2 =1.25cm
Substrate
3Noise coupling
Reducing crosstalk through metal wires
• No parallel lines (difficult to realize)
• Large space between lines (area penalty)
• Reduce Cm/C and Lm/L with a ground plane close to lines (speed and power penalties)
In nm CMOS technologies wire spacing is small large crosstalk must be reduced:
P. Andreani – System-on-Chip
• Place ground lines between active lines (area penalty, as well as speed)
• CAD tools to estimate the crosstalk and to limit the crosstalk within an acceptable level
• Special coding to cancel the crosstalk from two or more active “digital” lines, e.g. differential in the case of 2 active lines
• Innovative CMOS process steps are needed in general!
4Noise coupling
General power supply distribution model
P. Andreani – System-on-Chip
ChipBoard
• Internal power/ground distribution can be seen as resistive except bond wires
• Noise may be transmitted from one chip to another chip the small sensitive analog signals may easily become corrupted
• At the board level, large decoupling components can be used – more difficult to do so at the chip level
5Noise coupling
Noise due to wire resistance
Average/peak voltage drops in wires with full current densities
Wire type
Sheet resistanceCurrent density100 μm long1000 μm long
Metal-1
0.08 Ω1 mA/μm
7mV/70mV70mV/0.7V
Metal-top
0.04 Ω1 mA/μm
4mV/40mV40mV/0.4V
P. Andreani – System-on-Chip
1000 μm long10 mm long
70mV/0.7V0.7V/7V
40mV/0.4V0.4V/4V
Note: In synchronous circuits, Ipeak ≈ 10 × Iaverage
Peak voltage drop (noise) is a fundamental source of noise in SoC
6Noise coupling
Inductance is even more critical than resistance
dI/dt noise via power supply
Inductance of power lines and bonding wires often crucial
I
L
V
P. Andreani – System-on-Chip
V = L
7Noise coupling
dΙdt
Chip interface
(a) Wire bonding.
P. Andreani – System-on-Chip
(a) Wire bonding.
(b) Flip-chip bonding.
8Noise coupling
Pad circuit (with bond wire)
Pin
Bond wire inductance
large protectivediodes
I/OLb≈1-5nH
Pad50x50 μm2
Vdd
P. Andreani – System-on-Chip
Cpin > 1pF Cn+Cn > 1pFCpad≈0.25pF
Out/In In/OutL>1nH
C1>1pF C2>1pF
9Noise coupling
dI/dt noise in CMOS
• Simultaneous switching noise (SSN), also called ground bounce
• Output buffers drive large loads especially important multiple pins may be needed
• Millions of gates switching simultaneously very large current spikes
P. Andreani – System-on-Chip
spikes
• Logic families with almost constant power consumption have been proposed, similar to BJT ECL logic (i.e. enhanced source-coupled logic, folded source-coupled logic) large quiescent power consumption not popular
• On-chip or on-package decoupling capacitors are always needed
10Noise coupling
Design criteria for reducing dI/dt noise
L1R1
C1
R2
C2
VDD+-
LR
<< RC1.
2.dIdt
<< VDDL
I
L2
L= L1+L2
Inductivetime constant
Capacitivetime constant
P. Andreani – System-on-Chip
• Simultaneously switching large dI/dt
• High speed RC must be small harder to meet L << R2C
• Large designs small R harder to meet L << R2C
• During switching, power fluctuation must be kept to a small part of VDD, not only for analog but also for digital circuits.
11Noise coupling
dt
Decoupling capacitor CD reduces fluctuation
CD
C
VDD+-
L1
L2
ΔV
P. Andreani – System-on-Chip
CDVDD = (VDD + ΔV)(CD +C) ΔV =C
CD + CVDD
• If a max. 10% voltage drop is allowed CD > 10C required
• If total chip capacitance of a microprocessor is 1nF CD should be larger than 10nF (!)
• In reality, the unswitched fraction of C acts as a decoupling capacitance
12Noise coupling
Assuming all current is supplied by CD (worst case)
Technological differences of P-/P+ substrates
Wafer thickness
Epitaxial layer:4-15 μm1-50 Ωcm
p-
P. Andreani – System-on-Chip 13Noise coupling
Wafer thickness50-500 μm1-50 Ωcm
p-1-50 mΩcmp+
Low-conductive lightly-doped wafer
High-conductive heavily-doped wafer with an epitaxial layer
Substrate noise coupling in CMOS
Digital GND Analog GND
Biasing contact
Sensitive Node
Cj
Voltage transient
Biasing contact
V V I
Cj
p-
Wires & pads
12 3
P. Andreani – System-on-Chip 14Noise coupling
1. Noise injection through reverse-biased junction capacitance
2. Noise injection through contacts
3. Noise injection through wire-to-substrate capacitance
Also others: through forward-biased junction and hot carriers
VBS VT IDSp-
Noise injection via substrate contacts
Digital GND
Biasing contact
p-
Digital POWER
Biasing contactn-well
200-500 mV 200-500mV
P. Andreani – System-on-Chip 15Noise coupling
p-
• Due to simultaneous switching, noise amplitudes on power and ground wires may reach 500mV
• This noise directly affects the substrate through bias contacts
• Contacts are spread all over the substrate large impact, even more significant than that of junction capacitances
Noise injection via parasitic capacitance
PadsPower/Ground wires
RF inductor
P. Andreani – System-on-Chip 16Noise coupling
Pads RF inductor
• Lowest level of wire parasitic capacitance to substrate is another source of noise injection
• Pads, power/ground wires and RF inductors are particularly significant in producing substrate noise through their large parasitic capacitance
Why an 8-form coil??
P. Andreani – System-on-Chip
N. Itoh et al., “5.8 GHz RF Transceiver LSI Including On-Chip Matching Circuits”, 2006 IEEE BCTM, paper 14.1
1717Noise coupling
Noise injection via forward biased junction
Reverse-bias
Forward-bias
Noise on digital ground wire
Protective diodes
Negative pulses may
n-well
Digital ground wire
Source diodes
P. Andreani – System-on-Chip 18Noise coupling
Negative pulses maybe directly injected into substrate
• The pad protective diodes and the nMOS source diodes can be forward-biased by the negative noise pulses on the digital ground wires inject large noise into the substrate
• This noise is similar to that through substrate contacts
Substrate Substrate
Noise reception via Cj, Cd, and contacts
Analog GNDSensitive node
Biasing contact
Cj Cd
P. Andreani – System-on-Chip 19Noise coupling
• The substrate noise can be received by the sensitive node through Cj and Cd, where Cd is the capacitance between the channel and substrate (which are separated by the depletion layer)
• The noise can also be received by the analog ground through the substrate contact. For an analog circuit with a low PSRR, this would be a major noise source
Noise reception via body effect
Analog GNDSensitive node
Bulk (substrate)
Source
VBS
Noise at drain (V)0.0200.0180.0160.0140.0120.0100.0080.0060.0040.002
VDS = 2.5VVBS = 10mV
P. Andreani – System-on-Chip 20Noise coupling
0.0022E16 4E16 6E16 8E16 1E17
Substrate doping (cm-3)
VT = Vto+γ( 2φb-VBS - 2φb )γ =
2qεsiNA
Cox
φb = lnKTq
NAni
VBS VT IDS ( VDS )
Effect of a guard ring (P-)
No backplane contact
(less important in P-)
P. Andreani – System-on-Chip 21Noise coupling
Without a guard-ring With a guard-ring
Constant voltage contour lines simulated with the RAPHAEL
Replacing guard ring with contacts (P-)
P. Andreani – System-on-Chip 22Noise coupling
continuous trench (100μm) 5 contacts (each 10 μm)
3 contacts (each 10 μm) 2 contacts (each 10 μm)
Noise reduction techniques (ideal supply lines)
500μm
10μmNoisy NMOS NMOS
1) No measure
P-
SiO22μm
2μm
2) An oxide trench
P-
P. Andreani – System-on-Chip 23Noise coupling
3) An N-ring to Vdd
P-
+5V
N+
4) An N-ring inside an N-well
P-
+5V
N+
N-well
Noise reduction techniques – II
5) An N-ring to ground
P-
0V
N+
6) A P-ring to ground
P-
0V
P+
P. Andreani – System-on-Chip 24Noise coupling
7) A buried layer to ground
P-
0V
N+
4μm
8μm
0μm
Comparison between the different techniques
No measure (1)
Oxide trench (2)
N-ring to Vdd (3)
N-ring inside N-well (4)
P. Andreani – System-on-Chip 25Noise coupling
0 25 50 75 100 125 150 175
N-ring inside N-well (4)
N-ring to ground (5)
P-ring to ground (6)
Buried layer (7)
Peak-to-peak noise level (mV)
However, these results may be process dependent!
Noise reduction in SOI technology
P substrate
SiO2
Analog part Digital part
P. Andreani – System-on-Chip 26Noise coupling
• Silicon on Insulator (SOI) technology isolates analog and digital parts into separate substrate islands better for noise reduction.
• At high frequency, however, the two islands are still coupled through the capacitance to the p-substrate limited noise reduction, high cost SOI not very attractive in general
Conclusions
Noise vs. distance with an ideally grounded backplane
Backside contacts
Type of guard-ring
Contacts instead of ring
Length of guard ring
Position of guard ring
P- substrate
Noise attenuation increases with distance
Almost useless
P+ in P- and N+ in N- substrate
Yes, easy for wire crossing
Less important
Across the path, close to source
P. Andreani – System-on-Chip 27Noise coupling
Position of guard ring
Width of guard ring
Low impedance to ground
Across the path, close to source
Less important
Important
In general, the effectiveness of decoupling depends on the impedance of the return path guard ring contacts should have a very low impedance
Backplane contact not important in P-
Fundamental questions
• What is better for biasing the substrate, digital ground or analog ground?
• Is it better to use an exclusively dedicated substrate biasing line?What will be the consequences?
• Is a guard ring really effective? Where should the guard ring be connected?
P. Andreani – System-on-Chip Noise coupling 28
connected?
• For a given biasing strategy, is it better to use many or just a few contacts?
• Will the pad ring affect the biasing strategy?
Test circuit arrangement
• Technology: 2M N-well 1.2μm CMOS (old!!)
• Digital part: 256 noisy inverters in 16 groups, 16 inverters each group, driven by a combinational tree
• Analog part: a simple sense amplifier with short-circuited input and output at a DC voltage of Vdd/2, i.e. 2.5 V in this case, loaded by another identical sense amplifier
• Power supply: independent lines for the analog part, the digital
P. Andreani – System-on-Chip Noise coupling 29
• Power supply: independent lines for the analog part, the digital part, and the digital pads
• Pads: 30 digital and 10 analog pads
• Parasitic capacitance: extracted from the layout
• Substrate resistance: extracted by device simulator MEDICI
Preliminary floor plan of the test circuitAnalog signals
Power/ground
Power/ground
P. Andreani – System-on-Chip Noise coupling 30
Digital signals
Digital signals
Digital signals
Power/ground rings
Illustration of the test circuit
256 inverters in 16 groups
50/1.2
.
(VEP-D) (VEP-A)
(VAP)(VDP)
P. Andreani – System-on-Chip Noise coupling 31
16 groups 30/1.2
(VEG-A)(VEG-D)
(VAG)(VDG)
All biasing strategies will be evaluated
Other conditions and terminology
(3) Terminology for lines:
A: the line feeds analog section
D: the line feeds digital section
E: the line used for biasing without feeding current to any transistor
(4) Terminology for biasing:
(1) Equivalent circuit of a package pin:
P. Andreani – System-on-Chip Noise coupling 32
Figures for a large chip
X/Y: a biasing combination
X indicates the line to which digital contacts are connected
Y indicates the line to which analog contacts are connected
X/G/Y: G is the line to which the guard ring contacts are connected
(2) Clock frequency: 20 MHz
Mutual inductance between different pins is disregarded!
Biasing strategies (1)
VDP
VDG
VAP
VAG
D/D
D
D
A
A
D/A
VDP
VDG
VAP
VAG
D
D
A
A
P. Andreani – System-on-Chip Noise coupling 33
D/E
VDP
VDG
VAP
VAG
D
D
A
A
VEP
VEG
A/A
VDP
VDG
VAP
VAG
D
D
A
A
Biasing strategies (2)
E/A
VDP
VDG
VAP
VAG
D
D
A
A
VEG
VEP
E/E
VDP
VDG
VAP
VAG
D
D
A
A
VEP
VEG
P. Andreani – System-on-Chip Noise coupling 34
No SSN
VDP
VDG
VAP
VAG
D
D
A
A
SSN – Simultaneous Switching Noise
Noise level in P- substrate
D/D
D/A
D/E
A/A
E/A
E/E
No SSNNormalized noise level
66.5mV
P. Andreani – System-on-Chip Noise coupling 35
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
• A/A is the worst strategy, while D/E or E/A are the best
• D/A and E/A minimize body effects, but noise propagates through substrate, then couples to supplies through contacts/capacitances
• Notice however that D/A is even (slightly) better than E/E great, no need of dedicated lines
• With ideal pins (no SSN), noise only through substrate insignificant compared to that through power/ground contacts
Normalized noise level
A/A - the worst biasing strategy
VAPVDP
P. Andreani – System-on-Chip Noise coupling 36
Noise goes first to nearby contacts and then directly through wires to the analog section
VAGVDG
D/D - the second worst biasing strategy
VAPVDP
P. Andreani – System-on-Chip Noise coupling 37
Noise can directly go to the contacts of analogsection and reach the source node in close distance
VAGVDG
D/A - a good biasing strategy without E-line
VAPVDP
P. Andreani – System-on-Chip Noise coupling 38
The only path for noise going to analog section is the P- substrate, which will attenuate the noise due to its large resistance. D/A is better than D/D or even E/E
VAGVDG
D/E - one of the best biasing strategies
VAPVDP
VEP
P. Andreani – System-on-Chip Noise coupling 39
Compared with D/A, supply noise is decoupled with analog power supply lines
VAGVDG
VEG
E/A - one of the best biasing strategies
VAPVDP
VEP
P. Andreani – System-on-Chip Noise coupling 40
Compared to D/A, noise is decoupled from the digital power supply lines
VAGVDG
VEG
E/E - worse than D/E or E/A
VAPVDP
VEP
P. Andreani – System-on-Chip Noise coupling 41
The noise coupled to nearby digital contacts directly goes to analog section via wires and makes things worse
VAGVDG
VEG
E1/E2 - the best biasing?
VAPVDP
VEP1 VEP2
P. Andreani – System-on-Chip Noise coupling 42
It should be the best biasing at the cost of two dedicated supplies. There are no internal wires between digital and analog sections.
VAGVDG
VEG1VEG2
No SSN - attractive but practically unfeasible
VAPVDP
P. Andreani – System-on-Chip Noise coupling 43
Without inductance, the substrate is ideally biased, only device noise remains. The two power supplies may be merged. However, it is unfeasible practically.
VAGVDG
What if fewer contacts?
VD
G
VAP
VAG
VDP
10×
10× 10×
10×Unchanged
P. Andreani – System-on-Chip Noise coupling 44
Fewer contacts contact-related resistance increases less noise drained coupling between devices increases, but noise on supply lines decreases
p-
10×
10×
10×
The fewer the contacts, the higher the noise (P-)
40%
10%
8%
21%
37%
35%
D/D
D/A
D/E
A/A
E/A
E/E
No SSN
Reference circuit1/10 contact number
P. Andreani – System-on-Chip Noise coupling 45
• The “no SSN” case has the largest relative increase – this is because device coupling is maximized in this case
• The other results are less straightforward to interpret
122%No SSN
0 10 20 30 40 50 60 70 80 90 100Root-mean-square voltage over a 10ns period (mV)
The effect of guard ring biasing (p-)
D/AD/A/AD/E/A
D/ED/E/E
E/AE/A/AE/E/A
E/EE/E/E
P. Andreani – System-on-Chip Noise coupling 46
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Normalized noise level No SSN
• Little change, since the large parasitic capacitances on the supply lines are the dominating mechanism
• D/A/A and E/A/A bias the rings with analog supply, and the reason for the limited effect is that the original analog ground is already a guard ring
• E/E/A biases the ring with the same line biasing the digital section, it introduces noise (D/D/A or D/D/E are similar to D/D (i.e., bad))
The effect of pin inductance (p-)
D/D
D/A
D/E
A/A
E/A
E/E
10 nH1 nH
66.5mV
P. Andreani – System-on-Chip Noise coupling 47
E/E
No SSN
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Normalized noise level
• With small terminal inductance (almost no SSN), there is little differencebetween different biasing strategies
• For a chip with large terminal inductance, the selection of biasing strategy becomes important
• Inductance “large” or “small” depends on the noise level we can tolerate
Conclusions (P- substrate)
• Minimize impedance on supply lines
• Separate supply domains for analog/digital – use D/A or D/E for biasing (E/A ok, but may cause latch up)
• Distance between analog and digital helps always
• Guard rings help, too best if biased with dedicated
P. Andreani – System-on-Chip Noise coupling 48
• Guard rings help, too best if biased with dedicated lines
• And, implement all analog functions as differential circuits whenever possible
Buried layers with deep contacts
digital GND noisy noisy
digital Vdd
analog GND sensitive sensitive
analog Vdd
(2) Only under the sensitive devices with deep contacts:
digital GND noisy
NMOSnoisy PMOS
digital Vdd
analog GND sensitive
NMOSsensitive PMOS
analog Vdd
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
(1) Only under the noisy section with deep contacts:
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
P. Andreani – System-on-Chip
GND noisy NMOS
noisy PMOS
Vdd GND sensitive NMOS
sensitive PMOS
Vdd
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
digital GND noisy
NMOSnoisy PMOS
digital Vdd
analog GND sensitive
NMOSsensitive PMOS
analog Vdd
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cmP-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
(3) Under all the circuitry with deep contacts:
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
49Noise coupling
Buried layers without deep contacts
digital GND noisy noisy
digital Vdd
analog GND sensitive sensitive
analog Vdd
(5) Only under the sensitive devices without deep contacts:
digital GND noisy
NMOSnoisy PMOS
digital Vdd
analog GND sensitive
NMOSsensitive PMOS
analog Vdd
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
(4) Only under the noisy section without deep contacts:
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
P. Andreani – System-on-Chip
digital GND noisy
NMOSnoisy PMOS
digital Vdd
analog GND sensitive
NMOSsensitive PMOS
analog Vdd
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cmP-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
(6) Under all the circuitry without deep contacts:
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
GND noisy NMOS
noisy PMOS
Vdd GND sensitive NMOS
sensitive PMOS
Vdd
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
N-well 10 Ω⋅cmN-well 10 Ω⋅cm
50Noise coupling
Three possible arrangements of buried layers
(1)
(2)
(3)
(4)
(5)
(6)
P. Andreani – System-on-Chip
• As expected, buried layers reduce substrate noise significantly
• The deep bias contacts are essential, particularly for buried layers under the analog section – otherwise, basically same situation as P+
without backplane contacts (N/P buried layers are connected by a relatively large junction capacitance)
• The best arrangement is case (3), i.e. to place buried layers under all the circuitry with deep contacts, although case (2) already gives substantial noise reduction
51Noise coupling
How to bias buried layers
VG
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cmP-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
(1) D/D, A/A and E/E are bad biasing strategies for buried layers
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
VP
P. Andreani – System-on-Chip
VDG or VEG
P-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cmP-BL 50 mΩ⋅cm N-BL 25 mΩ⋅cm
(2) D/A and E/A are good biasing strategies for buried layers
N-well 10 Ω⋅cm N-well 10 Ω⋅cm
VAPVDP or VEP VAG
52Noise coupling
Design criterion for buried layers
Allow only capacitive junctions betweenburied layers under different sections
Never allow low-resistive paths betweenanalog and digital
Digital P-well & P-BL
Digital N-well & N-BL
Analog N-well & N-BL
Analog P-well & P-BL
Analog N-well & N-BL
Digital N-well & N-BL
Digital P-well & P-BL
Analog P-well & P-BL
or
P. Andreani – System-on-Chip
Digital P-well & P-BL
Analog P-well & P-BL
Analog N-well & N-BL
Digital N-well & N-BL
Digital N-well & N-BL
Digital P-well & P-BL
or
53Noise coupling
Available buried layers in BiCMOS
NPN bipolar transistorEmitter Base CollectorPMOSNMOS
N-well contact
P-well contact
Add buried layers in CMOS process needs at least 3 extra masks (much) more expensive – also, the density of digital cells would be lower
Buried layers are available in high performance BiCMOS process (N-BL is always needed to decrease the parasitic collector resistance in npn BJTs) – the only extra mask is for the deep contact of P-BL
P. Andreani – System-on-Chip
P-well N-well P-well N-well
P-BL N-BL N-BLP-BL
N+-con-tact
P+N+
Emitter Base Collector
P- substrate
N+P+P+
PMOSNMOS
N+N+N+P+
contactcontact
Used for insulation in high performance BiCMOS, without biasing
54Noise coupling
Triple-well nMOS in CMOS processes
N-well P-well
State-of-the-art CMOS processes provide triple-well nMOS devices for improved isolation
Deep contacts usually not available
Isolation is very good at low frequencies – decreases after a few GHz because of capacitive coupling
P. Andreani – System-on-Chip
P-well
N-well
P- substrate
NMOS
N+N+N+P+
N-well contact
P-well contact
55Noise coupling