device simulation for single-event effects
DESCRIPTION
Device Simulation for Single-Event Effects. Mark E. Law Eric Dattoli, Dan Cummings NCAA Basketball Champions - University of Florida SWAMP Center. Objectives. Provide SEE device simulation environment Address SEE specific issues Physics - strain Numerics - automatic operation Long term: - PowerPoint PPT PresentationTRANSCRIPT
Device Simulation for Single-Event Effects
Mark E. LawEric Dattoli, Dan Cummings
NCAA Basketball Champions - University of FloridaSWAMP Center
Objectives• Provide SEE device simulation environment
• Address SEE specific issues– Physics - strain– Numerics - automatic operation
• Long term:– Simulate 1000’s of events to get statistics
– With SEE appropriate physics– Without extensive human intervention
Outline• Background - FLOODS Code• Numeric Issues and Enhancements– Grid Refinement– Parallel Computing Platforms
• Physical Issues and Enhancements– Transient / Base Materials– Mobility– Coupling to MRED / GEANT
FLOOPS / FLOODS• Object-oriented codes• Multi-dimensional• P = Process / D = Device 90% code shared
• Scripting capability for PDE’s - Alagator
• Commercialized - ISE / Synopsis– Sentaurus - Process is based on FLOOPS
• Licensed at over 200 sites world-wide
What is Alagator?
• Scripting language for PDE’s• Parsed into an expression tree• Assembled using FV / FE techniques
• Stored in hierarchical parameter data base
• Models are accessible, easily modified
What is Alagator?
• Example use of operators for diffusion equation
• Fick’s Second Law of Diffusion– ddt(Boron) - 9.0e-16 * grad(Boron)– ∂C(x,t) / ∂t = D ∂2C(x,t) / ∂x2
Operator Description
“ddt” Time derivative
“grad” Spatial derivative
“sgrad” Scharfetter / Gummel Discretization Operator
“diff”
Returns the magnitude of the derivative of the argument parallel to the edge of evaluation – electric field applications for mobility in a device – returns a scalar
“trans”
Returns the magnitude of the derivative of the argument perpendicular to the edge of evaluation – electric field application for mobility in a device – returns a scalar
“elastic”
Compute elastic forces - FEM balance
Basic Upgrades• FLOODS has been used for:
– Bipolar devices (SiGe)– GaN based heterostructures MEM’s
•Coupled H diffusion to device operation•4 equations , n, p, H
– Noise simulations for RF bipolar devices
• Enhancements for modern MOS– More flexible contacting options (transients)
– Accurate mobility - transverse field– Alternate channel materials
Outline• Background - FLOODS Code• Numeric Issues and Enhancements– Grid Refinement– Parallel Computing Platforms
• Physical Issues and Enhancements– Transient / Base Materials– Mobility– Coupling to MRED
Adaptive Refinement
• Charge Deposition is not on grid lines
Charge Spreads in timeFine grid at zero timeCoarser grid as time goes
Simulate many hits, we can’t have user defined grid
Object Oriented
• Modular - Grid / Operators / Fields
• Code written for elements works in all dimensions
• Example - every element can compute Size
Element Class
Volume
Face
EdgeNode
Example - Isotropic Refinement
• Local Error Estimate - Bank Weiser Based
• Remove– Replace an edge w/ a node– Dose Stays Constant– Position new node at optimal quality position
• Addition– Subdivide an edge– Find effected volumes (Voronoi)– Centroidal positioning
SRC Supported
Anisotropic Grid - Initial• Rectangular region created at the command line
• Remainder of the silicon is smoothed
• Silicon Elements 478
• Joint Quality 0.936
• Average Quality 0.944
SRC Supported
Anisotropic Grid
• Refinement of both extension and deep source / drain
• LevelSet Spacer• Note - etch onto rectangular regions
• Silicon Elements 1150• Joint Quality 0.937• Average Quality 0.961
• Improved Quality on Add!
SRC Supported
Good for Process Simulation
• Device Simulation is Different!– Channel Needs Anisotropic refinement
– Unrefinement difficult– Global Operations and Data Structures
Device Simulation Driven Refinement
• All brick elements (2D example)• Refine and terminate• Unrefinement easier to track
– Glue elements together– Remove excess discretization nodes
• Requires Multi-point Templates– 4, 5, and 6 point square discretization (2D)
– Virtual functions in an Object Oriented Scheme
Object Oriented• Derived Specific Geometry Elements• Working on refinement• Working on Discretization
Element Class
Volume
Face
EdgeNode
2 -Edge
3 -Edge
Face
Quad
Tri
Parallel Computing
• 3D Transient is time consuming
• What can be done to accelerate?
Numerical Approximations• Discretization
– Replace continuous functions w/ piecewise linear approximations
– Grid Spacing, Time
• Linearization– Reduce nonlinear
terms using multi-dimension Newton’s method
– Mobility, Statistics, …
• Linear Matrix Problem– Number of PDE’s x
number of nodes square
– Direct Solver
Nonlinear set of PDENonline
ar algebra
ic equatio
ns
Linear Matrix Problem
Temporal and SpatialDiscretization
PoissonCarrier ContinuityLattice Temperature
Multi-dimensional Newton Linearization
Flux = (n1 - n2) / x12
1
x12
1 1
1 1
n1
n2
F
F
CPU Effort and Time• Assembly of Matrix
– Calculate the large, linear system
– Lots of Data read– Potential for
Overlapping writes– Lots of Parallel
Potential– Linear in number of
elements
• Solution of Matrix– Large Sparse System– Established means for
parallel solve– Leverage Argonne Nat’l
Lab Code– Low power of equations
n1.5
Nonlinear set of PDENonline
ar algebra
ic equatio
ns
Linear Matrix Problem
Temporal and SpatialDiscretization
PoissonCarrier ContinuityLattice Temperature
Multi-dimensional Newton Linearization
Flux = (n1 - n2) / x12
1
x12
1 1
1 1
n1
n2
F
F
Alagator Assembly• Equations are split
– Edge pieces (current, electric field)– Node pieces (recombination, time derivative)
– Element pieces (perpendicular field)
• Pieces are vectorized– 128 pieces in tight BLAS loops for performance
– Operations are broken down in scripting
• Overall CPU linear in # of pieces
Parallel Assembly
• Two Options• High Level Parallel
– Assemble Different PDE’s on Different CPU’s
– Limited Parallel Speedup
• Low Level Parallel– Split Grid, assemble pieces
• Match to Linear Solve
Parallel Assembly
• Partition the work on different processors
• Assemble pieces on processor that will solve
Parallel Performance - Assembly
• High Level Partition
• Poisson on Node 1
• Electrons on Node 0
Edge Assembly Time (s)
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Serial Node 0 Node 1
Parallel
(s)
Linear Solve Speedup - PETSC Package
• Amdahl’s Law Clearly Visible
3-D Dopant Diffusion, Linear Solve Times
0
102030
40
50
6070
80
0 1 2 3 4 5
Number of Nodes
(s) IBM SP/2 system
Linux, AMD Opteron Cluster
Speedup of Linear Solve
1
1.5
2
2.5
3
3.5
4
4.5
1 2 3 4 5
Number of Nodes
Sp
ee
du
p
IBM SP/2 MeasuredPerformance
Ideal Performance
AMD Opteron Cluster
Linear Solve Speedup - Options
• Ordering Algorithms are not helpful
• Some Parallel Methods increase solve time
Effect of Matrix Ordering on Serial Performance
0102030405060708090
100
Natural Quotient MinimumDegree
Reverse Cuthill-McKee Nested Dissection
Parallel GMRES Performance Problems
0
50
100
150
200
250
300
350
400
0 1 2 3 4 5
Number of Nodes
Se
co
nd
s
Outline• Background - FLOODS Code• Numeric Issues and Enhancements– Grid Refinement– Parallel Computing Platforms
• Physical Issues and Enhancements– Transient / Base Materials– Mobility– Coupling to MRED
Today’s TransistorScaled MOSFETS and alternate materials to extend Moore’s Law
• Technology scaling is driven by cost per transistor
• Channel length scaling is slowing in bulk planar devices
• Limited by leakage current
• Strained Si devices
S. Thompson et al., IEEE EDL. 191-193, 2004.
S. Thompson et al., IEDM Tech. Dig. 61-64, 2003.
Enable Transients for Devices
• Added transient device command• Extended Contacts to allow switching• Contact Templates Available Now
Example NMOSSwitching TransientGate Ramped from3V to 0V in 1ps
Enable Transients for Devices
• 1D Diode• Charge added to depletion region at time 0• Simplest possible SEE
Mobility Modeling• Combination of terms
– Ionized Dopants– Carrier-Carrier– Surface Roughness– Strain
• Combined using Mathiessen’s rule
1
1
b
1
s
Low-Field Mobility• Lots of models - implemented Phillips unified
model• Includes
– Dopant (dependent on dopant type)– Carrier - Carrier scattering– Minority carrier scattering
n 1
N scN sceff
N refN scf
2
n pN sceff
N sc ND NA p N sceff NA GND p
F
0
200
400
600
800
1000
1200
1016 1017 1018 1019 1020 1021 1022
Ele
ctro
n M
obili
ty (
cm2 /V
s)
Doping (cm -3)
Low-Field Mobility - Carrier-Carrier
• In single event simulation• Dominant term can be carrier - carrier• Serious mistakes by ignoring these terms
200
400
600
800
1000
1200
1016 1017 1018 1019 1020 1021
Ele
ctro
n M
obili
ty (
cm2 /V
s)
Hole Concentration (cm -3)
Donor Density of 1016
Surface Scattering• Acoustic Phonons• Surface Roughness• Both depend on
perpendicular field
• Decay factor applies only in channel
• Tuned to measured MOS results
• In progress!
ac B
E
C(N /N0)
E1/ 3(T /T0)k
sr E /E ref A
E
3
1
1
1
b
D
ac
D
sr
D e x /
100
120
140
160
180
200
220
240
260
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Mob
ility
(cm
/Vs)
Normal Electric Field (MV/cm)
Normal Field Computation• Requires element
assembly– Increased computation– More complex matrix
• Compute field perpendicular to an interface– Fixed geometry– Might interact w/ single
event
• Field perpendicular to current flow– Convergence difficulties
at low current– Assumes current is
perpendicular…..– Make sure it doesn’t
apply in bulk
100
120
140
160
180
200
220
240
260
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Mob
ility
(cm
/Vs)
Normal Electric Field (MV/cm)
Current
Field
SiO2
Channel Materials• Heterostructure Boundaries • Fairly Easy, since we had heterostructure experience in FLOODS before
• Development of Ge channel simulations
500Å Ge Channel30Å Gate NitridePoly Gate Bias Swept Up0.1m Channel LengthIdeal Doping Profiles
Note: Concentration Discontinuity at interface
Boundary Conditions• Commercial simulators only allow BC at contacts• FLOODS has large flexibility at boundaries• Example - Sink on sides• pdbSetString ReflectLeft Equation “1.0e-3*(Elec-Doping)
• Simulation as function of device simulation size
• Reflecting boundaries at edges and back change current collected at contacts
Courtesy of Ron Schrimpf, Andrew Sternberg
Finite Element Method Mechanics
• Theory of Elasticity – linear elastic materials - Silicon is modeled as an isotropic material for simplicity
• Enhanced Alagator– Added elastic operator for displacement– Added source term operators
• Elastic(displacement) + BodyStrain(Boron*k)
σ
ε
SRC Supported
(μm)
(μm
)45 nm
140
nm
120
nm
30 nm
Si0.83Ge0.17Si0.83Ge0.17
STI STI
-536
-83403
95
31
Source FLOOPS
MPa
Stress Contours
Future - Strain and SEU Upgrades
• Anisotropic operators– Current direction, strain interaction
– Mobility has an orientation
• Density of States• Recombination• Driving Forces?
Connection to Thompson
Trajectory Read
• Trajectory Read Command
Summary• Numerics
– Started Developing refinement appropriate to SEE
– Parallel Port, Begun Testing
• Physics– Built some basic capability for SEE– Read Tracks
• Next Year– Demonstrate link, run demos on parallel machines