Title: Realizable Parasitic Reduction Using Generalized Y Transformation
1Realizable Parasitic Reduction Using
Generalized Y-? Transformation
- Jeffrey Z. Qin, Chung-Kuan Cheng
- Computer Science and Engineering
- University of California, San Diego
2Agenda
- Motivation
- Generalize Y-? Transformation
- Elements
- Order of Reduced Model
- Order of Picking Reducible Nodes
- Numerical Problem
- Realize ?-Admittance Using Templates
- Experimental Results
- Conclusion
3Motivation
- Existing model order reduction methods
- Matrix-based reduction
- PACT, PRIMA
- Topology-based reduction
- AWE ? tree based, not realizable
- TICER ? 1st order realizable model
- Features needed ? more general topology,
higher-order realizable model
4Overview Y-? Transformation
- Advantage of traditional Y-? transformation
- Topology-based method
- Handles linear circuits of any topology
- Transforms with fidelity transfer function
evaluated by Y-? is equal to the one evaluated
by Cramers rule - Plus, after being generalized
- stable models
- realizable parasitic reduction
5Traditional Y-? Transformation
- Conductance in series
- Conductance in star-structure
n0
n1
n2
n1
n2
n1
n1
n0
n3
n2
n2
n3
6Fidelity of Y-? Transformation
Cramers rule
Y-? Transformation
7Generalized Y-? Reduction
- Improve traditional Y-? transformation
generalized
traditional
Handle RCLK-VJ? No
YES
Transformed model order? High
REDUCED
Optimize priority of picking
nodes?
No YES
Resolve common-factor numerical problems?
No YES
8Handle RCLK-VJ Elements
Handle R, C, L, and J
all coefficients are positive! because no
subtraction ops.
9Handle RCLK-VJ Elements (2)
Handle K
Nodal equations are the same.
e. g.
10Handle RCLK-VJ Elements (3)
- Handle V
- voltage sources (V) can be transformed into
current sources (J).
11Order Reduced by Truncation
Admittance truncated after certain order terms
Fidelity of Y- ? transformation is still
preserved.
12Optimize Node Ordering
- Y-? transformation Gauss elimination
- priority of picking nodes pivoting in LU
factorization
13Common Factors
Type I Common Factors
n1
not recognize
recognize type-I
n2
n3
14Common Factors (2)
Type II Common Factors
n2
n3
n2
n3
should be canceled out.
15Common Factors (3)
Why is it harmful?
- 1st order coefficients grow at about 2n rate
- 2nd order coefficients grow at about 3n rate,
- Worse when circuit is larger, preserved order is
higher. - Scaling doesnt work!
16Common Factors (4)
- Theoretical Contribution
- common-factor effects generally exist in Y-?
transformation for linear circuits of any
topology - devise algorithm to treat common factors
- ? admittance is in the simplest form, or has no
redundancy, after the treatment.
17Realizable Parasitic Reduction
- Why realize?
- Realizable reduced order models preferred by
standard parasitic extraction tools, for
stability, compatibility concerns. - Hard problem!
- No tool is able to realize 1st-order admittance
matching 1st order moments from arbitrary linear
circuit - ? admittance may not be realizable in nature
- Y-? all coefficients are non-negative!
181st order Realization
- Y-? reduction 1st-order reduced admittance
-
- can be realized, because
19High Order Realization Template
- solution
- Formulate the realization problem as an
optimization problem using geometric programming
(GP). - 0th and 1st order templates
20High-Order Realization Template (2)
21Realization Formulation
Pij and qij are the j-th product term of
22Realization Formulation (2)
- Then, given a low-order admittance
where ai and bj are positive constants, we
have - GP formulation
-
Objective
subject to
and
23Waveform Evaluation
Comparison with AWE and PRIMA
SPICE PRIMA (8th order) AWE (3rd order) Y-? (8th
order)
24Waveform Evaluation (2)
Comparison with TICER
25Pole Estimation
Exact transfer function
dominant poles
Y-? matches less moments, but more accurate!
26Natural frequency Estimation
Pole plot
Waveform plot
27Common-Factor Impacts
Growth of powers of s
Without cancellation For tree-like circuits,
the order of the input admittance grows
quadratically with the size of LC elements.
With cancellation The orders are the same as
those of original circuits.
28Common-Factor Impacts (2)
Growth of coefficients
Scaling doesnt work, as coefficients of
different orders grow at different
rate. Common-factor cancellation works great!
1st order coefficient
2nd order coefficient
29Conclusion
- A generalized Y-? reduction
- handles RCLK-VJ circuits of any topology
- Yields order-reduced models with fidelity
- Common factor cancellations
- remove fake poles/zeros
- make the reduction numerically stable.
- Parasitic realization
- Stability of the reduced model is guaranteed
- Realizable models good for parasitic reduction.