Sie sind auf Seite 1von 58

IISc Bangalore, February 2008

Compact Modeling

Josef Watts
IBM Semiconductor Research and Development Center
January 2008

© 2006 IBM Corporation


IBM Semiconductor Research and Development Center – Joe Watts

What is a Compact Model?

2
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Compact MOSFET Model


Gate

Compact Cgd=f2(Vgd, Vgs) Cgs=f3(Vgd, Vgs)


Model
Drain Source

Jds = f1(Vds,Vgs)

TCAD
Model

3
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

PowerPC
microprocessor

2 cores

10 levels of wiring

7.9E8 transistors

65nm SOI MOSFET Transistors


stress liner, halos, gate tunneling,
floating body effects, self heating.
4
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Design proceeds simultaneously


at multiple levels of abstraction
Information passes from level to
level via models
o del
g icM
L o

Tim
ing
Mod
el

odel
p a ct M
C o m

5
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Design Information
Layout Layout Schematic
Netlister
Data Extractor Data

Net list

Circuit
Circuit Simulator
Characteristics
Compact (PowerSPICE
(delay,
Model HSPICE
gain, etc)
Spectre)
Process Information
6
Circuit Behavior
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

What does the simulator do? (DC Case)

ƒ Enforce Kirkoff's Laws


– V1=V(R1)+V(R2) R1
– V(R2)=V(R3)
– I(V1) = I(R1)
– I(R1) = I(R2) + I(R3) V1
R2 R3

7
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

What does the simulator do? (DC case)


ƒ Enforce Kirkoff's Laws V1
– V1=V(R1)+V(R2)
– V(R2)=V(R3) R1
– I(V1) = I(R1)
– I(R1) = I(R2) + I(R3) V2
– Assign a voltage to every V1
node—by construction
voltage law is satisfied R2 R3
– If I know I(X) as a function of
V(X) I can test current law.
– That’s what the compact V3=0
model supplies: I=f(V)

8
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

What does the simulator do? (DC case)


ƒ Enforce Kirkoff's Current Law V1
– 0=fR1(V1-V2)-fR2(V2-V3)-fR3(V1-V2)
– Calculate current at each node R1
– This is the error vector E
– If any element is not zero the V2
simulator must adjust the voltages V1
to make all the errors sufficiently
small. R2 R3

V3=0

9
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

What does the simulator do? (DC case)


V1
ƒ Adjust voltages to zero the
errors
R1
r r
E = [J ]V V1
V2
⎛ dI ( R1) dI ( R 2) dI ( R3) ⎞
E[2] = ⎜ − − ⎟ • ΔV [2]
⎝ dV 2 dV 2 dV 2 ⎠
R2 R3
– Now I need derivatives of I
w.r.t. V
– The Compact model
V3=0
supplies those also

10
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

What does the simulator do? (AC Case)

ƒ Same thing plus


– Account for charge storage R1
in capacitors when the
voltage is not constant
– Also account for emf in
inductors when the current
V1
is not constant
R2 C3
– Compact model must
supply L & C values.

11
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Charge Conservation

ƒ Consider the gate to source Vb=0


capacitance of a MOSFET. Vb=-1

Cgs
ƒ C=C(Vgs, Vgd, Vbs)
ƒ V2
Q = ∫ C (Vgs , Vds , Vsb )dVgs 1 2
V1
ƒ Integrate around a loop and
net delta Q equals zero
3
ƒ But ΔVgs=0 on segment 2
and C on segements 1 & 3
are not equal. Vg
ƒ Charge accumulates on the The compact model must
gate
provide both Q and C
12
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Kirkoff’s current law becomes

0 = ∑ I ⋅dt + ∑ (Q(ti ) − Q(ti −1 )

To form Jacobian you need


dQ
=C
dt

Compact model must calculate:


I(V), dI(V)/dV, Q(V), dQ(V)/dV

13
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Compact model complexity

I = V/R is a compact model for a resistor

14
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Compact model complexity

I = V/R is a compact model for a resistor

I = V/(θ*(L-dL)/(W-dW))
Add Geometric Scaling

15
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Compact model complexity

I = V/R is a compact model for a resistor

I = V/((θo+TCR*(T-25))*(L-dL)/(W-dW))
Add: Geometric Scaling
Temperature Scaling

16
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Compact model complexity

I = V/R is a compact model for a resistor

TR
Jth Rth

I = V/((θo+TCR*(VTR+T-25))*(L-dL)/(W-dW)
Jth = V*I
Rth=Rth/(L*W)
Add: Geometric Scaling
Temperature Scaling
Self Heating
17
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Compact Model Anatomy


BSIMSOI3.2 FET model
equations
Process Model BR resistor Process or
Statistical
decoupling model
NFET capacitor
FET model
PFET Annular Diode parameters
Passive
AVNFET Wire_cap component models
OP
OPresistor
resistor
Technology
Model
18
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Communication Example

ƒ Goal: A digital speed boost for .25u CMOS


– Same ground rules
– Shorten Lpoly by lithography/etch change
– Shorter Leff
• Smaller gate capacitance
• Higher current
– Heavy Halo to control Vt roll-off
• Same Vt means same Ioff

19
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Scaling of the FET Ldrawn

ƒ Ldrawn is determine by Mask


lithography—what can you Lpoly
resolve on the mask
Gate
ƒ Lpoly is determined by
Source Drain
Drain
Ldrawn + develop + etch
Leff
ƒ Leff is determine Lpoly + N+
P-
hot processes
ƒ Cg_total ~ Lpoly
ƒ Id ~ Leff

20
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Short Channel Vt

ƒ Long device VT determine by 1D electrostatics

Gate
Source Drain
Drain

N+
P-

ƒ Short Channel VT is determined by 2D electrostatics

Gate
Source Drain
Drain

N+
P-

21
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Vt vs Leff
Threshold Voltage

Vtmin

Lmin Leff

22
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Add Halo Implants


ƒ Long device has Vt mostly determined Gate
by Substrate doping
Source Drain
Drain

N+
P+ P-
ƒ Halo blocks field lines from drain to
source end of channel
ƒ Short device has Vt mostly determined
Gate
by halo doping
Source Drain
Drain

N+
P+ P-

23
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Vt vs Leff

Halo
Threshold Voltage

No Halo

Lmin Leff

24
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

The plan

ƒ Lpoly 8% less
– Cg_total 8% less
ƒ Leff 12% less (at Lmin)
– Id 8% more
ƒ Build same design as original process
ƒ Microprocessor frequency 16% more

25
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Simulation results: Latches don’t work

ƒ Latch used NFET


Clock=Vdd
passgates
Data In Data Out
ƒ Output does not reach Vdd
ƒ At Vdd-Vt device shuts off
ƒ Next gate must switch
completely with a “weak
one” on input

Vsource
26
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Vt vs Leff

Halo
Threshold Voltage

No Halo

Lmin + 2 grid points

Lmin Leff

27
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Simulation results: Latches don’t work

ƒ Latch used NFET


Clock=Vdd
passgates
Data=Vdd ƒ Pass transistors were
longer than minimum
ƒ New NFET had more Vt
rollup
6s+, Lmin
ƒ New NFET had more body
Idrain

6s0, Lmin
6s+, Llatch effect
6s0, Llatch

Vsource
28
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Model Build Flow


Device
Targets
Target Model
Hardware Model Starts Here
Starts Here
Good Design
Wafer Test Manual
Targets

Measure Fit Model Instance Center Fit Process Package for


Data To Data Effects Model Model Simulator

Test Test Test Inst. Test Model Test Test in


Data Fit Quality Effects Quality Distributions Simulator

Ship Model
29
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Hardware Models, Target Models and Everything In


Between

Golden chip: near nominal


for all device specs
←Hardware Model
Hardware meets key
device specs
←Hardware
Hardware usable for roll off
Influenced Model
Output impedance, etc.

Hardware is very far


←Target Model
from targets

No hardware exist

30
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Hardware Measurements

ƒ Id for various ƒ For SOI also need


– Vg – Buried BJT
– Vd – Diode
– Vbody
– Body Resistance
ƒ Impact Ionization
– Self Heating
ƒ Igate current
ƒ For RF model
ƒ STI stress effect
– S parameters
ƒ Liner Stress
ƒ Nwell proximity
ƒ Cj
ƒ Cov

31
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Model Build Flow


Device
Targets
Target Model
Hardware Model Starts Here
Starts Here
Good Design
Wafer Test Manual
Targets

Measure Fit Model Instance Center Fit Process Package for


Data To Data Effects Model Model Simulator

Test Test Test Inst. Test Model Test Test in


Data Fit Quality Effects Quality Distributions Simulator

Ship Model
32
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Hardware Fitting

ƒ FET
– We use standard models: can’t change the equations
– Certain parameters must have physical meanings
– One set of parameters must fit all sizes
– Find values of parameters that gives the best fit.
ƒ nonFET there are no standard models
– You can change the equations
– You can change the model topology

33
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Fitting by Local Optimization


ƒ Different parameters describe the physics of different regions
ƒ Adjust small groups of parameters to fit different regions

2nd
Id

4th
L=.12

1st

L=10 3rd
Vg Vd

34
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Fitting by Local Optimization

Id

L=.12

L=10

Vg Vd

35
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Fitting by Local Optimization


ƒ As each group of parameter is adjusted that set of data is well fit
ƒ And each previously fit group gets a little worse
ƒ The result is multiple iterations to get an adquate fit

Id
L=.12

L=10
Vg Vd
36
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Global Optimization

ƒ Use a very Robust


Optimizer to fit all the data
in one pass
ƒ Requires a good fitness
function to weight all the
data properly
ƒ We use a genetic algorithm
developed in house.

37
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Genetic Algorithm

ƒ Mimics biological evolution


ƒ Operates on a “population” of
solutions
ƒ Find the fitter individual
solutions and breed the next
generation
ƒ Repeat until you have a good
enough solution

38
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Genetic Algorithm: The life cycle


Hill Climb
algorithm

Patriarch
Assign
Fitness to Each
Individual
Breed New
Population
Create
Population
by Mutation
Test for
Diversity

39
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Fitness Function

ƒMatching to currents
ƒPhysical
reasonableness
ƒMathematical
robustness
ƒMatching to derivatives

40
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Model Build Flow


Device
Targets
Target Model
Hardware Model Starts Here
Starts Here
Good Design
Wafer Test Manual
Targets

Measure Fit Model Instance Center Fit Process Package for


Data To Data Effects Model Model Simulator

Test Test Test Inst. Test Model Test Test in


Data Fit Quality Effects Quality Distributions Simulator

Ship Model
41
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Systematic, Random & Correlated

ƒ Systematic…P=P+ΔP(layout)
– Using the layout, we can predict the magnitude and direction of the
effect and
– The prediction is good for any lot though out the life of the program
ƒ Random Correlated…P=P+ΔP(layout)*global random #
– Not systematic but
– Using the layout, we can identify classes of devices which will have
the same magnitude and direction of effect.
ƒ Random Uncorrelated…P=P+ΔP(layout)*local random #
– Using the layout, we cannot know the magnitude and direction of the
effect and
– We cannot predict how predict that the effect will be the same as any
other device on the chip.
ƒ The classification depends not on the physics but how much
knowledge is available to the designer !

42
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

NW

Nwell Proximity Effect


PC

RX STI stress Effect

CA Random Dopant Fluctuation

Liner Stress

Corner Rounding

Litho & Etch Effects

Are these two FETs


Identical?
43
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

NW

Nwell Proximity Effect


PC
Systematic
RX STI stress Effect

CA Random Dopant Fluctuation

Liner Stress

Corner Rounding

Litho & Etch Effects

44
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Well Proximity Effect

Figure courtesy of P. Drennan, CICC 06


45
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

NW

PC
Nwell Proximity Effect
Systematic
RX STI stress Effect
SA
Systematic
CA Random Dopant Fluctuation
SB SA
Liner Stress
SB
Corner Rounding

Litho & Etch Effects

46
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

STI Stress
GATE

Oxide Source
Source Drain
Drain Oxide

Silicon

During processing STI is etched out and filled with oxide


at a high temperature. After cooling mechanical stress is
locked in.
47
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

NW
Nwell Proximity Effect
PC Systematic
STI stress Effect
RX Systematic
Random Dopant Fluctuation
CA
Uncorrelated
Liner Stress

Corner Rounding

Litho & Etch Effects

Major memory effect because of


small device size
48
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Random Dopant Fluctuation

ƒ Dopant atoms
are randomly
place

Vt Mismatch
ƒ For small FETs
there are few
total atoms
ƒ Random
variations are a
larger fractional
change
1/sqrt(W*Leff)[1/u]

49
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

NW

Nwell Proximity Effect


Adjacent PC PC
Systematic
RX STI stress Effect
SA
Systematic
CA Random Dopant Fluctuation
SB SA Uncorrelated
Engineered Stress
SB Systematic
Corner Rounding

Litho & Etch Effects


1 CA 2 CA

50
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Layout-dependent Stress Model


Dense Gate Array
MC or CA proximity
MC or CA strapping

Small Isolate FET


PC proximity
Longitudinal DSL Transverse DSL PFET A
Isolated PC
tension
compression tension

Parallel BP Edge
compression d

PFET B
One PC pair
0.315um pitch

Product Like
tension RX length
Parallel
edge far
away PFET C
d 19 PC pairs
0.315um pitch
tension
This dimension being
varied on the x-axis
of the plot

Reference pfet layouts


51
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

NW
Nwell Proximity Effect
PC
Systematic
STI stress Effect
RX Systematic
Random Dopant Fluctuation
CA Uncorrelated
Liner Stress
Systematic
Corner Rounding
Systematic and
Correlated
Litho & Etch Effects

52
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Corner Rounding Model


width

ƒ Rounding (webbing) of
corners makes part of the PC
channel longer
RX
ƒ Delta W is used to modify the Nominal
device behavior alignment, small
effect
– Nominal and tolerance of
delta W can be effected
ƒ This is a systematic effect—L width

always gets longer


PC
ƒ This is a random correlated
effect—size of the effect RX
depends on poly to isolation
alignment
Misalignment can
make effect bigger
or smaller

53
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Same Orientation NW
Nwell Proximity Effect
PC Systematic
STI stress Effect
RX Systematic
PC
Pitch Random Dopant Fluctuation
CA
Uncorrelated
Liner Stress
Systematic
Corner Rounding
Uncorrelated
Litho & Etch Effects
Correlated and
Uncorrelated
Close Proximity

54
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Model Build Flow


Device
Targets
Target Model
Hardware Model Starts Here
Starts Here
Good Design
Wafer Test Manual
Targets

Measure Fit Model Instance Center Fit Process Package for


Data To Data Effects Model Model Simulator

Test Test Test Inst. Test Model Test Test in


Data Fit Quality Effects Quality Distributions Simulator

Ship Model
55
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Model Centering

As Fit Model

My Chip

DM Targets

Centered
Model

56
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Model Build Flow


Device
Targets
Target Model
Hardware Model Starts Here
Starts Here
Good Design
Wafer Test Manual
Targets

Measure Fit Model Instance Center Fit Process Package for


Data To Data Effects Model Model Simulator

Test Test Test Inst. Test Model Test Test in


Data Fit Quality Effects Quality Distributions Simulator

Ship Model
57
IISc -- February 2008
IBM Semiconductor Research and Development Center – Joe Watts

Bibliography

ƒ "Modeling of Variation in Submicron CMOS ULSI Technologies", S. Springer, et al IEEE Trans. On


ED. Sept. 2006
ƒ “A simple yet accurate mismatch model for circuit simulation”, Z. Jin, et al. Nanotech 2006
ƒ “Modeling FET Variation Within a Chip as a Function of Circuit Design and Layout Choices”, Josef
Watts,et al, Nanotech 2005
ƒ “Modeling MOSFET Process Variation using PSP”, J. Watts, et al, Nanotech 2007
ƒ “A comprehensive MOSFET mismatch model”, P. Drennan and C. McAndrew, IEDM 1999, p. 167-170
ƒ “On the correlations between process parameters in statistical modeling”, Slezak, et al, Nanotech
2004, Vol 2, pp144-146
ƒ “Statistical timing of parametric yield prediction of digital integrated circuits”, Jess, et al, DAC 2003,
p932-937
ƒ Advanced Compact Models for MOSFETs, J. Watts, et al, Nanotech 2005

58
IISc -- February 2008

Das könnte Ihnen auch gefallen