/centre for analysis, scientific computing and applications Introduction Multirate time-integration Model order reduction Applications Conclusions Redundancy reduction of IC models by Multirate time-integration and Model order reduction A. Verhoeven 1,2 E.J.W. ter Maten 1,2 R.M.M. Mattheij 1 [email protected]1 Eindhoven University of Technology (CASA) 2 NXP Semiconductors (Design Methods) CASA AIO-day, Dorint Hotel, Eindhoven, November 13 2007
33
Embed
Redundancy reduction of IC models · 11/13/2007 · 3 Model order reduction 4 Applications 5 Conclusions 15. GF NPY_]PQZ]LYLWd^T^ ^NTPY_TQTNNZX[`_TYRLYOL[[WTNL_TZY^ Introduction
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Redundancy reduction of IC modelsby Multirate time-integration and
Model order reduction
A. Verhoeven1,2 E.J.W. ter Maten1,2 R.M.M. Mattheij1
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Circuit simulation
The derived equations can be written as a system ofdifferential-algebraic equations (DAE):
ddt
[q(t , x)] + j(t , x) = 0. (1)
A transient analysis computes the solution x : [0, T ] → Rd
for a given initial solution x(0) = x0.
In Pstar this DAE is discretised by the Backward DifferenceFormula (BDF). The linear systems at each Newtoniteration are solved by an hierarchical type of Gausselimination.
6
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Limits of default transient analysis
The default single-rate time-integration algorithm for thetransient analysis is not efficient if the continuous ornumeric model contains redundancy, i.e. if large parts havea low activity level or even stay constant.
Then it is possible to decrease the simulation costs whilethe accuracy is maintained. This can be done by
Efficient simulation of theoriginal model: Multiratetime-integration, dynamicalpartitioning, bypassing,etc.
Model order reduction: anew model of smaller size(and complexity) iscreated. 7
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Outline
1 Introduction
2 Multirate time-integration
3 Model order reduction
4 Applications
5 Conclusions
8
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Multirate time-integrationThe circuit model is partitionedin an active (A) and a latent (L)part that are integrated at thefine and coarse time-grids,respectively.
ddt
[qA(t , xA, xL)] + jA(t , xA, xL) = 0, (2)
ddt
[qL(t , xA, xL)] + jL(t , xA, xL) = 0. (3)
s
s
ssss
ss
L A
xL ∈ RdL xA ∈ RdA
�
�
-
Hn
Tn
Tn+1
6Interface
���hn,1tn,0
tn,1
tn+1,0
9
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Multirate for hierarchical circuit models
For circuit models the partitioning should coincide with theexisting hierarchical structure.
We can only refine the internal part x(A) for given terminalpart x(A) = v(t) (interpolation-based voltage source).
ddt
[q(A)(t , x(A))] + j(A)
(t , x(A)) = 0, x(A) =
[v(t)x(A)
]. (4)
It is preferable to refine the complete xA for given terminalcurrents iL→A = j(t) (interpolation=based current source).
ddt
[q(A)(t , x(A))] + j(A)(t , x(A)) = j(t). (5)
New trend is to use controlled current sources instead, e.g.
iL→A = j(t) + Gx(A) or iL→A =ddt
[q(A)(t , x(A))] + j(A)
(t , x(A)). 10
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Error analysis of multirate BDF method
Weighted error norm at the coarse time-grid (compoundphase), where 0 ≤ τ � 1
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Multirate stepsize control
Error constraints
{rnC ≤ TOL,
rn−1,mA ≤ TOL.
Error model
rnC = φn
CHK+1n
rn−1,mA = ˆrn−1,m
A + rnI
ˆrn−1,mA = φn−1,m
A hk+1n−1,m
rnI = φn
I HK+1n
Coupled tolerance levels forw ∈ (0, 1):{
TOLA = (1− w)TOL,TOLI = wTOL.
This parameter w can be chosensuch that the expected workload isminimised.Independent stepsize control of Hnand hn−1,m such that{
ˆrn−1,mA ≤ TOLA,
hmax rnI ≤ TOLI
⇒ rn−1,mA ≤ TOL.
12
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Adaptive partitioning control
For (digital) circuits with moving active areas a staticpartitioning is not sufficient. Therefore the multiratepartitioning is adaptively controlled such that the expectedlocal speed-up factor is optimised.
The partitioning is controlled after each refinement phasebased on a local efficiency analysis at t = Tn.
It is also allowed to modify the partitioning just after thecompound step if it was not accepted. 13
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Semi-optimal partitioning techniques
1 By means of d all possible transitions of the most activelatent element and the most latent active element arecompared and optimised, iteratively.
2 Tolerance level εrel < 1 for relative local error per element:
|di | > εrel‖d‖. (6)
3 Tolerance level εabs>TOL for absolute local error perelement:
|di | > εabs. (7)
4 From all needed stepsizes per element the largest gap isdetected to separate the system in a fast and a slow part.
14
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Outline
1 Introduction
2 Multirate time-integration
3 Model order reduction
4 Applications
5 Conclusions
15
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Model order reduction (MOR)
Consider the following nonlinear DAE system{ ddt [q(x)] + j(x) = Bu , x(0) = x0,
y = h(x),
where x ∈ Rd , u ∈ Rm, y ∈ Rp with d � m, p. We are onlyinterested in the relationship between u and y in thetime-domain. With offline MOR the model is replaced by alow-order model for z ∈ Rr{ d
dt [q(z)] + j(z) = Bu , z(0) = z0,
y = h(z).
Onesided methods only consider the function u → z, whiletwosided methods really consider u → y. 16
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Linear Time-Invariant (LTI) models
{x = Ax + Bu , x(0) = x0,
y = Cx.(8)
The observability and controllability functions are
Lc(x0) = min{12
∫ 0
−∞‖u(t)‖2dt : u ∈ L2(−∞, 0), x(−∞) = 0},
Lo(x0) =12
∫ ∞
0‖y(t)‖2dt ,∀τ∈[0,∞)u(τ) = 0.
For linear systems we have Lc(x0) = 12xT
0 W−1x0 andLo(x0) = 1
2xT0 Mx0, where W, M ∈ Rd×d are the
controllability and observability Gramians.
The (energy) ratio Lo(x)Lc(x) should be balanced.
17
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Truncated balanced realization (TBR)
The Gramians W, M satisfy the Lyapunov equationsAW + WAT = −BBT , (9)
AT M + MA = −CT C. (10)
The system is balanced w.r.t. basis V if W = VΣVT andM = V−T ΣV−1 are simultaneously diagonalised. such that
Lo(x)
Lc(x)=
xT Mx
xT W−1x=
xT V−T ΣV−1x
xT V−1Σ−1V−T x=
zT Σ2zzT z
.
The singular values of Σ often converge rapidly to zero. Areduced model can be derived by truncation.{
z = V−1AVz + V−1Bu , z(0) = z0,y = CVz.
(11)
There exist many other MOR techniques for LTI systems,like PRIMA, PVL, PMTBR, etc. 18
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Trajectory PieceWise Linear (TPWL)
−1 0 1 2 3 4−1
−0.5
0
0.5
1
1.5
2
2.5
3
x1
x 2
B
A
C
D
E
The nonlinear system is linearisedat {(t1, x(ts)), . . . , (ts, x(ts))} along agiven trajectory. Each linearisedsystem is reduced by LTI modelreduction techniques. A globalbasis V is computed by a svd of alllocal reduced basisvectors. Finallythe global model is constructed bya weighted sum of all locallyreduced linearised systems.
{ ∑si=1 wi(z)
[VT CiVz + VT GiVz− VT B iu(t)
]= 0,
y = h(z).(12)
19
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Empirical Gramians
The observability and controllability Gramians satisfy
W =
∫ ∞
0eAtBBT eAT
tdt , M =
∫ ∞
0eAT
tCT CeAtdt . (13)
Consider [x1, . . . , xm] and [y1, . . . , yn], where x i and y j satisfy{ddt [q(x i)] + j(x i) = b iδ(t),
x i(0) = 0.
{ddt [q(t , x j)] + j(t , x j) = 0, x j(0) = ej ,
y j = h(x j).
Then W, M can be numerically integrated as follows
W =m∑
i=1
1N
N∑k=1
x i(tk )x i(tk )T , M =n∑
i=1
1N
N∑k=1
y i(tk )T y i(tk ). (14)
For LTI systems we have that W → W, M → M if N →∞. 20
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Galerkin projection
Empirical balanced truncation (EBT) uses these formulasfor W, M with a larger sets of inputs and initial values fornonlinear systems.
TBR is used to balance W, M by solving a system ofLyapunov equations. Thus a basis V can be constructedby truncation.Proper Orthogonal Decomposition (POD) approximatesW, M by using only one trajectory.
W = M =1N
N∑k=1
x1(tk )x1(tk )T = VΣVT . (15)
The reduced model for z ∈ Rr is constructed by Galerkinprojection.{
ddt
[VT q(t , Vz)
]+ VT j(t , Vz) = VT Bu , z(0) = z0,
y = h(Vz).(16)
21
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Missing Point Estimation (MPE)For Galerkin projection the reduced functions, e.g.VT q(t , Vz), need the complete evaluation of q, etc.
Let V ∈ Rd×r be a given basis and P ∈ {0, 1}g×d aselection matrix with PPT = Ig , then V, VT areapproximated by
V ≈ TPV, VT ≈ VT TP.Here T ∈ Rd×g is an interpolation matrix that can beoptimised in a Least Squares sense.
Define V = PV ∈ Rg×r , W = TT V ∈ Rg×r , then we canapproximate x ≈ TVz, VT q(Vz) ≈ WT Pq(TVz), etc.The original reduced model is replaced by{
ddt
[W
TPq(t , TVz)
]+ W
TPj(t , TVz) = W
TPBu , z(0) = z0,
y = h(TVz).(17) 22
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Outline
1 Introduction
2 Multirate time-integration
3 Model order reduction
4 Applications
5 Conclusions
23
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Inverter chain model
Equations
u1 = Uop− u1 −Υf (uin, u1, 0)
uk = Uop− uk −Υf (uk−1, uk , 0)
for k = 2, . . . , n
uin =
t − 5 5 ≤ t ≤ 105 10 ≤ t ≤ 15(17− t)5/2 15 ≤ t ≤ 170 otherwise
24
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
The higher accuracy ofPMTBR can be used to getsmaller models than PRIMAgets.
Reduced models are stillvalid for different inputs.
Linearisation points aredirectly computed during arough transient simulationbefore.
POD
The POD basis V is found bysolving the eigenvalueproblem for the correlationmatrix.
The POD basis V comprisesthe eigenvectorscorresponding to 20 largesteigenvalues and captures thedynamics on [0, 20ns] verywell.
With MPE it is possible toremove 70% of the equationswhich reduces the evaluationcosts. 29
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Outline
1 Introduction
2 Multirate time-integration
3 Model order reduction
4 Applications
5 Conclusions
30
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Conclusions for multirate
Normal transient simulation can be inefficient for circuitswith large slow part. Multirate algorithms can be muchmore efficient for these applications while the accuracy ispreserved.
A multirate error control mechanism has been developedwhich allows much larger steps at the coarse time-grid.Also several partitioning algorithms have been investigatedand implemented.
Current prototype in Pstar shows good results and apotential to improve. Search for the realistic designexamples to determine in which areas the multiratepotential is largest.
31
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Conclusions for MOR
TPWL: behaves very well. Small balls may need longerextraction time.
Empirical Balanced Truncation: approximates the systemu → y but needs a very high extraction time.
POD: cheap version that is only accurate if x stays close tosnapshot-subspace.
Galerkin projection methods based on empirical Gramianswork nicely but have high evaluation costs. Theseevaluation costs can be reduced by techniques like MissingPoint Estimation.
Promising results of nonlinear MOR for inverter chain anddiode chain models.
32
/centre for analysis, scientific computing and applications
Introduction Multirate time-integration Model order reduction Applications Conclusions
Future plans
Publication of paper about MOR in IFIP proceedings
PhD thesis will be sent to press
PhD defence at January 8th, 2008
Next year I will start as software engineer at VORtech BV(Delft).