Model Order Reduction Wil Schilders NXP Semiconductors & TU Eindhoven November 26, 2009 Utrecht University Mathematics Staff Colloquium
Outline Introduction and motivation Preliminaries Model order reduction basics Challenges in MOR Conclusions 2
The work of mathematicians may not always be very transparant. 3
But it is present everywhere in our business Collecting info Partition 1 Mathematics Collecting info Partition 2 Mathematics Transient Multirate Transient Multirate Partitioning Partitioning Mathematics Invisible contribution, visible success 4
Moore s law also holds for numerical methods! 5
Very advanced simulations Subgridding mechanisms reduce the number of mesh points in a simulation. In this example (a) the full grid is 20 times larger (35e6 mesh nodes) than (b) the subgridded version (1.75e6 mesh nodes). 6
The impossible made possible.. A complete package layout used for full wave signal integrity analysis 8 metallization layers and 40,000 devices The FD solver used 27 million mesh nodes and 5.3 million tetrahedrons The transient solver model of the full package had 640 million mesh cells and 3.7 billion of unknowns 7
Avoiding brute force The behaviour of this MOS transistor can be simulated by solving a system of 3 partial differential equations discrete system for at least 30000 unknowns Insight? Even worse: an electronic circuit consists of 10 4-10 6 MOS devices Solution: compact device model (made by physicists/engineers based on many device simulations, ~50 unk) Can we construct such compact models in an automated way? 8
Another example Integrated circuits need 3-D structure for wiring 10 years ago 1-2 layers of metal, no influence on circuit performance Present situation: 8-10 layers of metal Delay of signals, parasitic effects due to high frequencies 3-D solution of Maxwell equations leads to millions of extra unknowns Gradual Can we construct a compact model for the interconnect in an automated way? circuit 9
Model Order Reduction is about capturing dominant features Strong link to numerical linear algebra 10
And, talking about motivation.. Mathematical Challenges! 11
Outline Introduction and motivation Preliminaries Model order reduction basics Challenges in MOR Conclusions 12
Overall picture Physical System Data Modeling ODEs PDEs MOR Reduced system of ODEs Simulation Control 13
Dynamical systems 14
Problem statement 15
Approximation by projection 16
Special case: linear dynamical systems 17
Singular value decomposition 18
Optimal approximation in 2-norm 19
Operators and norms Impulse response Transfer function 20
Important observation Model Order Reduction methods approximate the transfer function In other words: not the entire internal characteristics of the problem, but only the relation between input and output (so-called external variables) No need to approximate the internal variables ( states ), but some of these may need to be kept to obtain good representation of input-output characteristics matching some part of the frequency response Original log H ( ω ) log ω 21
Outline Introduction and motivation Preliminaries Model order reduction basics Challenges in MOR Conclusions 22
Overview of MOR methods 23
Balanced truncation 24
Interpretation of balanced truncation 25
Properties of balanced truncation 26
Simple example Mesh: 50 x 100 x 50 cells @ 1u x 1u x 1u cell dimensions Termination (R = 10 K) 1u x 1u Voltage step (R = 50 ohm) over 1u gap 80 micron Al strip Ideal conductor SiO 2 100 micron 50 micron Voltage step: 0.05 ps delay + 0.25 ps risetime SINSQ Pstar T = 5 ps 27
Mapping PDE s (Maxwell Eqs.) to equivalent high order (200) ODE system 10-th order linear dynamic system fit to FDTD results Singular value plot 28
Moment matching (AWE, 1991) Arnoldi and Lanczos methods 29
Asymptotic Waveform Evaluation (AWE) Direct calculation of moments suffers from severe problems: As #moments increases, the matrix in linear system becomes extremely illconditioned at most 6-8 poles accurately Approximate system often has instable poles (real part > 0) AWE does not guarantee passivity But: basis for many new developments in MOR! 30
Moment matching (AWE, 1991) Arnoldi and Lanczos methods 31
32
The Arnoldi method 33
The Arnoldi method (2) 34
The Lanczos method 35
Two-sided Lanczos 36
Moment matching using Arnoldi (1) 37
Moment matching using Arnoldi (2) 38
Example: MOR for transmission line (RC) Original Reduced Netlist size 0.53 0.003 (Mb) # ports 22 22 # internal 3231 12 nodes # resistors 5892 28 # capacitors 3065 97 39
Outline Introduction and motivation Preliminaries Model order reduction basics Challenges in MOR Conclusions 40
Start of a new era! The Pade-via-Lanczos algorithm was published in 1994-1995 by Freund and Feldmann, and was a very important invention Not only did the method (PVL) solve the problems associated with AWE It also sparked a multitude of new developments, and a true explosion regarding the field of Model Order Reduction AWE 1990 PVL 1995 Lanczos methods Arnoldi methods Laguerre methods Vector fitting 1995-2009 41
Topic 1: passivity A linear state space system is passive if its transfer function H(s) is positive real H has no poles in C + H is real for real s Real part of z * H(s)z non-negative for all s, z Passivity is important in practice (no energy generated) Lanczos-based methods such as PVL turned out not to be passive! problem!!! Search started for methods that guarantee (provable) passivity Passive MOR: PRIMA (Passive Reduced-order Interconnect Macromodeling Algorithm) SVD-Laguerre (Knockaert & Dezutter) Laguerre with intermediate orthogonalisation (Heres & Schilders) 42
Recent developments w.r.t. passivity Several passivity-preserving methods have been developed However: if the original discretized system is not passive, or if the data originate from experiments, passivity of the reduced order model cannot be guaranteed This happens frequently in practical situations: EM software generating an equivalent circuit model that is not passive In the frequency domain, this may not be too problematic, just an accuracy issue (Sparameters may exceed 1) However, it may become really problematic in time domain simulations Passivity enforcement methods Often: trading accuracy for passivity 43
Topic 2: Large resistor networks Obtained from extraction programs to model substrate and interconnect Networks are typically extremely large, up to millions of resistors and thousands of inputs/outputs Network typically contains: Resistors Internal nodes ( state variables ) External nodes (connection to outside world, often to diodes) Model Order Reduction needed to drastically reduce number of internal nodes and resistors 44
Reduction of resistor networks 45
Reduction of resistor networks 46
General idea: exploit structure 47
69 71 152 67 5 4 63 73 212 11 1 149 59 57 13 61 55 15 53 17 51 19 21 49 23 25 47 64 215 27 29 31 45 33 35 123 37 39 218 43 220 101 41 219 134 75 86 139 81 91 153 221 223 225 227 211 229 231 209 217 184 188 186 182 233 190 180 207 273 271 192 235 194 178 205 269 196 237 203 267 198 176 80 201 82 265 239 200 174 199 263 202 241 261 113 197 112 106 118 172 204 114 111 259 195 257 243 206 170 255 193 253 245 90 251 168 249 247 208 210 92 154 166 191 164 189 150 162 160 151 187 158 156 102 100 157 185 93 155 183 110 74 181 179 177 24 22 20 175 26 28 16 173 30 66 32 18 2 171 14 34 3 12 169 10 36 167 7 9 6 165 62 8 38 163 84 83 40 78 60 159 161 77 85 42 58 76 79 44 56 88 46 54 87 89 48 50 52 216 127 274 140 119 131 128 272 141 270 120 121 122 143 124 135 146 97 94 95 268 142 107 144 138 98 266 125 115 96 147 108 103 104 99 264 145 262 116 109 148 68 126 260 117 105 256 258 254 72 70 250 252 248 246 244 242 240 238 236 234 232 230 228 65 226 222 224 Simple network 274 external nodes (pads, in red) 5384 internal nodes 8007 resistors/branches Can we reduce this network by deleting internal nodes and resistors, still guaranteeing accurate approximations to the path resistances between pads? 213 214 NOTE: there are strongly connected sets of nodes (independent subsets), so the problem is to reduce each of these strongly connected components individually 136 137 132 133 129 130 48
Two-Connected Components 48 53 49 2 50 52 51 1 3 42 43 44 4 5 6 41 38 40 39 47 46 45 29 28 54 30 27 26 25 9 12 8 11 7 10 can delete all internal nodes here 16 33 34 32 24 35 31 23 22 36 37 18 17 15 19 14 20 13 21 49
How to reduce the strongly connected components Graph theoretical techniques, such as graph dissection algorithms (but: rather time consuming) Numerical methods such as re-ordering via AMD (approximate minimum degree) Commercially available tools like METIS and SCOTCH (not satisfactory) 50
Border Block Diagonal Form 51
Block bordered diagonal form Algorithm was developed to put each of the strongly connected components into BBD form (see figure) Internal nodes can be deleted in the diagonal blocks, keeping only the external nodes and crucial internal nodes The reduced BBD matrix allows extremely fast calculation of path resistances (work of Duff et al) 52
Reduction results 53
Results are exact No loss of accuracy! Spectre New Remaining problem: how to drastically reduce the number of resistors 54
Topic 3: Nonlinear MOR One of the most popular methods is proper orthogonal decomposition (POD) It generates a matrix of snapshots (in time), then calculates the correlation matrix and its singular value decomposition (SVD) The vectors corresponding to the largest singular values are used to form a basis for solutions Drawback of POD: there is no model, only a basis for the solution space Researchers are also developing nonlinear balancing methods, but so far these can only be used for systems of a very limited size (<10) E. Verriest, Time variant balancing and nonlinear balanced realizations J. Scherpen, SVD analysis and balanced realizations for nonlinear systems 55
Classification of circuits Engineer Mathematician 56
Fast and accurate simulation of oscillators Xtal Test CLK PA MIXER Pins + Matching DAC Digital Core Modem and control Tripler LNA + Matching IF Amp RC Cal ΣΔ ADC VCO + PLL PLL Filter Challenges: Linear modeling of oscillators does not provide accurate solutions and in most cases is not able to capture subtle nonlinear dynamics of oscillators (injection locking, jitter, etc) Nonlinear models are necessary which are fast, accurate and generic 57
Nonlinear modeling of perturbed oscillators Process & Key Observations: 1. PSS of oscillator: d/dt [q(x PSS )]+ j(x PSS )=0, T=T OSC 2. u 1 (t)= d/dt(x PSS )(t) ( right Floquet eigenfunction) 3. v 1 (t) solves a linearized adjoint system ( left Floquet eigenfunction) 4. Perturbed oscillator: d/dt [q(x)]+ j(x)= b(t) has solution x(t)=x PSS (t+α(t))+x n (t), [α(t) phase noise] 5. α(t) satisfies a non-linear scalar differential equation d/dt (α)(t) = v 1 (t+α(t)).b(t), α(0)=0 Because of orthogonality only the component of b(t) along u 1 (t+α(t)) is important Involves: Time integration, Floquet Theory, Poincaré maps, Nonlinear eigenvalue methods, Model Order Reduction! 58
Example- three stage ring oscillator 153kHz ring oscillator Unlocked osc: i inj = 6 * 10-5 *sin(1.04ω 0 * t) Locked osc: i inj =6 * 10-5 * sin(1.03ω 0 * t) 59
MOR: Fast and accurate modeling of VCO pulling VCO pulling due to PA and other blocks/oscillators needs to be analyzed before production Full system simulation is CPU intensive or infeasible Behavioural model order reduction gives fast and accurate insight in pulling/locking and coupling Full mathematical theory of locking/unlocking mechanisms and conditions lacking! PSD(dB) 0 10 20 30 40 50 60 simulation f in 70 0.4 0.6 0.8 1 1.2 1.4 1.6 frequency x 10 9 60
Outline Introduction and motivation Preliminaries Model order reduction basics Challenges in MOR Conclusions 61
Conclusions regarding MOR Model Order Reduction is a flourishing field of research, both within systems & control and in numerical mathematics Strong relation to numerical linear algebra Future developments need mathematical methods from a wide variety of fields (graph theory, Floquet theory, combinatorial optimization, differential-algebraic systems, stochastic system theory,..) 62
Some recent books on MOR 63