Title: Model Order Reduction for Large Scale Dynamical Systems Lecture 10: Analog Simulation in IC Industry
1Model Order Reduction for Large Scale Dynamical
SystemsLecture 10 Analog Simulation in IC
Industry Implementation of the Multirate
Algorithm
NXP Semiconductors
Bratislav Tasic
2Content
- Introduction
- In-house NXP Simulator Pstar
- Speeding-up Analog Simulation
- Multirate Concept and Prototype
- Limitations
- Architecture
- Dynamical Partitioning
- Current Version and Future Plans
- Example
- Conclusions
3IC Business
4Introduction
- Modern circuit simulation
- Larger and more complex designs (hundreds to
millions unknowns) - Various typical dynamical behaviour (various
frequencies, analog, digital, thermal, etc.) - Fast yet accurate analyses needed (transient in
particular) - Typically more than one time-scale involved
Partitions (e.g. high/low frequencies, analog and
digital) - Professional software EDA vendors or in-house
tools
5Direction of Top 10 Semiconductor Houses
- In top-10 semiconductor houses gt70 has in-house
analogsimulation competence - Intel, TI, Freescale, NEC, ST, Infineon, Qimonda,
NXP, IBM - In top-10 analog semiconductor houses all have
in-houseanalog simulator - Infineon/Qimonda Titan, ST Eldo, Intel Lynx
and Mynx, NXP Pstar - In-house analog sim. competence offers
- Proven track record
- Predictable costs
- Competences adapted to business needs
- Tightly integrated organisation
- Shared knowledge of design andmanufacturing
issues - For nano-scale design trend is for morein-house
EDA development
6Pstar is Robust Design
- Statistics
- More capacity than any commercial tool
- 2x overall faster statistical analysis
- Statistical results directly available
- no manual post-processing needed
- World class statistical modeling
- Stability analysis
- For Voltage Regulators, Automotive, Power Amps
- Using best-in-class pole-zero analysis methods
- Assure stability for varying load conditions
- Reliability simulation
- Alliance NBTI solution built on Pstar experience
7Pstar is System Design
- Top-Down System Design
- Run system-level simulation togetherwith
circuit-level implementation - Using MATLAB or Simulink for system-level
- System can be mixed domain(electrical,
mechanical, thermal) - MEMS, automotive
- Time-accurate solution for Simulink-circuit
simulation - No other (commercial) solution can handle
feedback - Analog Behavioral Modelling
- Verilog-A standard analog HDL
8Pstar is Analog Design
- Parametric simulations
- Including design-of-experiments (process-case)
- Integrated post-processing
- Simply build standardized tests for circuit
performance - Simple pass/fail analysis
- Reuse of test-benches for similar designs
- Analog diagnosis
- Fast and easy access to all simulation data
9Pstar is Innovation
- Enables NXP Research efforts for world-class
device models - CMC world standards (PSP, Mextram, Juncap2)
- Proprietary models (FinPSP)
- Invaluable in fast characterization
10Transient Simulation
- Transient analog simulation
- Most expensive
- Solving systems of DAEs
- Highly involved, complex and expensive simulation
- Circuits
- Becoming larger daily
- More difficult to simulate
- Application areas
- Analog circuits
- Mixed systems (for simulating the analog part)
- Statistical runs
11Transient Simulation
- Network models of electrical circuits
- Network models consisting of nodes and branches
- Kirchhoffs Laws and constitutive branch
equations - ib, vb branch currents/voltages vn
nodal voltages - Modified Nodal Analysis
- One voltage has to be grounded
- Hierarchical structure
12Transient Simulation
- Solving IVP
- Implicit A(?)-stable numerical time-integration
scheme, e.g. variable order BDF (costly) - How to increase the simulation speed preserving
the accuracy?
13Transient Simulation
- High-speed simulators
- HSIM, UltraSim, etc.
- Fast, yet sometimes not accurate enough
- Switching from on to other simulators is not easy
- Golden simulators
- Pstar, Spectre, HSPICE, ADS, Titan, etc.
- Requests from IC designers
- Speedup the transient analysis runs
- Preserve the accuracy and the robustness
Preserve the golden standard - No significant extra effort required to run the
simulation
14Mathematical Background
- Speedup approaches
- Table modeling
- Model order reduction
- Parallel computation
- Isomorphic matching
- Multirate numerical integration methods
- Table modeling
- Interpolation technique to avoid evaluation of
expensive models - Model order reduction
- Still under theoretical development and
validation especially for nonlinear problems
15Mathematical Background
- Parallel computation
- More and more popular (prices of clusters are
becoming reasonable and some math. libraries
already available) - Seems suitable for hierarchical solver
- It could be perfect for statistical runs
- Isomorphic matching
- Storing repeated instances of the same circuit
- Strong advantage of the hierarchical solver
- Power behind Hsim
- Not always accurate (e.g. wakeups)
16Mathematical Background
- Multirate numerical integration methods
- Suitable for circuits where large part of the
circuit is latent (slow) during the simulation
(sub)intervals - The best candidate to achieve significant speedup
- The accuracy and robustness can be preserved
- Still under development improvements are still
to be expected in future - Compound-Fast algorithm
- New and fully developed algorithm that shows
improved stability and robustness results as
compared to standard multirate technique (no
extrapolation involved) PhD work of Arie
Verhoeven
17Multirate Methods
- Multiple time scales Multiple time steps
- Latent parts Coarse time grid
- Active parts Refined time grid
- Faster overall progression in time Speedup
factor
18Concept
Transientanalysis
Multirateprototype
19Concept
20Concept
Standard transient
21Concept
DEV
_
4
DEV
_
2
DEV
_
1
DEV
_
3
Model
3
Model
5
Model
1
SSM
_
5
SSM
_
9
SSM
_
1
SSM
_
2
SSM
_
10
SSM
_
6
DEV
_
7
DEV
_
5
Model
2
Model
4
Model
6
SSM
_
3
SSM
_
7
SSM
_
11
SSM
_
4
SSM
_
12
SSM
_
8
DEV
_
6
DEV
_
8
22Concept
DEV
_
1
DEV
_
2
DEV
_
3
DEV
_
4
Model
3
Model
5
Model
1
SSM
_
5
SSM
_
9
SSM
_
1
SSM
_
2
SSM
_
10
SSM
_
6
DEV
_
7
DEV
_
5
Model
2
Model
4
Model
6
SSM
_
3
SSM
_
7
SSM
_
11
SSM
_
4
SSM
_
12
SSM
_
8
DEV
_
6
DEV
_
8
23BDF Compound-Fast Multirate Method
- Mappings necessary to partition the circuit
- Solving modified IVP
24BDF Compound-Fast Multirate Method
- Algorithm
- Compound step (relaxed Newton process)
- Interpolation of the interface
- Refinement phase
25Adaptive Multirate Step Size Control
- Local truncation errors (LTEs) should satisfy
- Computing LTEs using extrapolation polynomials
based on BDF history and computed solutions
26Adaptive Multirate Step Size Control
- Computing Hn1 (interpolation error included)
- Computing hn-1,m1
27Examples
Non-linear chain
Speed up factor 11 Ratio 99.93 - 0.07
28Examples HSOTA
Speed up factor 1.6 - 2 Size Ratio 50
29Examples Charge Pump
Speed up factor 5 Ratio 12
FAST
30Examples
31Dynamical Partitioning
Prototype static partitioning
Multirate with dynamical partitioning
32Dynamical Partitioning
33Dynamical Partitioning
t T2
34Dynamical Partitioning
- Speedup factor estimation
- Optimization problem maximizing S
35Example
- The nonlinear chain with more FAST models
36Example
S unknown F1 unknown F2 unknown Refine S Refine
F1 Refine F2
37Example
- Performance comparison
- RR models 5000, f1/f2 500
Results obtained for the static test cases are
almost exactly the same regarding the CPU time
38Conclusions
- Multirate transient analysis based on dynamical
partitioning shows great potential highly user
friendly - Current prototype shows good results and a
potential to improve - Several partitioning algorithms have been
investigated and implemented - Search for the realistic design examples to
determine in which areas the multirate potential
is largest
39Thank you