ICSOC Workshop

Download Report

Transcript ICSOC Workshop

Fast BEM Algorithms for 3D Interconnect Capacitance and Resistance Extraction Wenjian Yu

EDA Lab, Dept. Computer Science & Technology, Tsinghua University

[email protected]

Direct BEM to solve Laplace Equ.

  Physical equations A cross-section view  Laplace equation within each subregion 

u

 Same boundary assumption as Raphael RC3 

2

 Bias voltages set on conductors        

u q

2  

u

u

 0

u

 2 

x

 ,

n

u

2  

q

 2 

y

0

u

2   0 , Green’s Identity:  

z

2

u

2   0  , In On On  

u q

(

u

 2

v

i

(

i

 Direct boundary element method 

v

 2

u

) 1,  ,

(u is potential) (q is normal electric field intensity)

d

  

M

 (

u

) 

v

n

v

Freespace Green’s function as weighting function 

u

n

)

d

conductor

1

q

 Laplace equation is transformed into BIE:

c s u s

   

i q s

*

u d

    

i u s

*

q d

s is a collocation point

2

Discretization and integral calculation

A portion of dielectric interface:  Discretize domain boundary • Partition quadrilateral elements with constant interpolation • • Non-uniform element partition Integrals (of kernel 1/r and 1/r 3 ) in discretized BIE: • •

c s u s

j N

   1 

j q s

*

d

 )

u j

j N

   1 

j u s

*

d

 )

q j

s P 4 (x 4 ,y 2 ,z 2 ) P 3 (x 3 ,y 2 ,z 2 ) Y

Singular integration Non-singular integration • • 

j

P 1 (x 1 ,y 1 ,z 1 ) t

Dynamic Gauss point selection

Z O

Semi-analytical approach improves computational speed and accuracy for near singular integration

P 2 (x 2 ,y 1 ,z 1 ) X

3

Locality property of direct BEM

 Write the discretized BIEs as:

H

i

u

i

G

i

q

i

, (

i

=1, …,

M

)

Ax

Compatibility equations along the interface

u a a

  

u

u b a

n

a

  

b

 

u b

n

b

Non-symmetric large-scale matrix A

f

Use GMRES to solve the equation

Charge on conductor is the sum of q Med1 Medium 1 Med2 Interface [0] Interface

A

=

[0] Medium 2 [0] Conductor [0] For problem involving multiple regions, matrix A exhibits sparsity!

4

Quasi-multiple medium method

 Quasi-multiple medium (QMM) method   Cutting the original dielectric into

m

x

n

fictitious subregions, to enlarge the

Environment Conductors

matrix sparsity in BEM computation

z

With iterative equation solver, sparsity brings actual benefit

y

Master Conductor Master Conductor

x

A 3-D multi-dielectric case within finite domain, applied 3  2 QMM cutting

Non-uniform element partition on a medium interface Strategy of QMM-cutting:

Uniform spacing

Empirical formula to determine (m, n)

Optimal selection of (m, n)

5

Efficient equation organization

    Too many subregions produce complexity of equation organizing and storing Bad scheme makes non-zero entries dispersed, and worsens the efficiency of matrix-vector multiplication in iterative solution We order unknowns and collocation points correspondingly; suitable for multi-region problems with arbitrary topology Example of matrix population

s 11 v 11 s 12 s 21 s 22 s 23 s 32 s 33 u 12 q 21 v 22 u 23 q 32 v 33

Three stratified medium 12 subregions after applying 2

2 QMM

Efficient GMRES preconditioning

 Construct MN preconditioner    [

Vavasis, SIAM J. Matrix,1992

]

PA

T

A P

T

Neighbor set of variable i:

T

A p

l

1

l

2

l

3

A p

i

e

i L i

 

e

i

,

l l

1 2

i

Var.

i

1, ...,

l n

Solve, and fill 

P

N l

1

l

2

l

3

A

Reduced equation

A

T

p

i

= 0 1 0

P i

 Our work: for multi-region BEA, propose an approach to get the neighbors, making solution faster for 30% than original Jacobi preconditioner 7

A practical field solver - QBEM

 Handling of complex structures     Bevel conductor line; conformal dielectric Structure with floating dummy fill Multi-plane dielectric in copper technology Metal with trapezoidal cross section  3-D resistance extraction     Complex 3-D structure with multiple vias Improved BEM coupled with analytical formula Extract DC resistance network Hundreds/thousands times fast than Raphael, while maximum error <3% 8