S-parameter Modeling and Simulation for

Signal Integrity Analysis

S-parameter Modeling

Brad Brim
Sigrity, Product Marketing Manager
[email protected]
Agenda
Ò Definition and Properties of S-parameters

Ò File Formats and Connectivity

Ò Reduced Order Models of S-parameters

Ò Application of S-parameters
Ò This material is introductory
Ò … but it is critical to remember because improper generation
and application of S-parameters is often directly attributable
to violating these basic concepts

Ò The verbiage will have an AMS chip-centric bias

Ò due to track sponsorship
Ò all concepts apply equally well to package and system
Ò applications will include package and system
Ò Much of this material is already available elsewhere
Ò technical journal archives
Ò college texts
Ò EM, circuits and distributed circuits courses
Ò trade publications
Ò conference archives
Ò SI/RF educational companies
Ò wiki
Ò http://en.wikipedia.org/wiki/Scattering_parameters

Ò The focus of this presentation is to discuss the relevance of

this previously-available material
Ò explain the most commonly misunderstood (forgotten) parts
Ò practical issues for circuit/system simulation
Where will you encounter S-parameters?
Ò package and system models
Ò commonly applied for chip/package/system simulation

Ò device/circuit models
Ò whether individual devices or larger on-chip circuits

Ò reduced-order models
Ò whether macromodel S-parameters or I.P. or
measurements
Ò Circuit Simulation
Ò Modified nodal analysis
Ò SPICE and most (if not all) circuit simulators in both the time and
frequency domain are based on modified nodal analysis
Ò S-parameters are not applied directly in any of these simulators,
but are first converted to Y-parameters for integration with the base
analysis

Ò Frequency domain data

Ò S-parameters are no different than any other frequency domain
circuit parameter data
Ò all general frequency domain data is difficult to incorporate into a
transient circuit simulation
Ò some special forms of data are more efficient, but this does not
represent the real systems you will encounter in chip/package/system
design
Ò The “wave” definition
Ò very natural for EM theory and computations
Ò voltage wave amplitude ratios
Ò reference impedance is that of the transmission line at each
port
Ò The “circuit” definition
Ò Most common definition applied is from Kurokawa
Ò Kurokawa, K., "Power Waves and the Scattering Matrix", IEEE
Trans. Micr. Theory & Tech., Mar. 1965, pp194-202
Ò Notice the complex conjugate in the definition for reference
impedance due to “power” definition instead of wave amplitude
Ò The “measurement” definition
Ò power is measured internally to a
VNA (vector network analyzer)
Ò reference impedance can be
selected based on the
“calibration” scheme
Ò measured reference, line
impedance, renormalized to Rref
Ò Calibration is a critical step
Ò Deembedding removes “feed
structure”
How the definitions differ
Ò conj(Zref)
Ò voltage waves have no reflection when Zload=Zref
Ò circuits transfer maximum power when Zload=conj(Zsource)

Ò The difference is in the transformation between S- and Y-

parameters
Ò a simple conjugate of reference impedance

Ò If Zref is complex (e.g. highly lossy lines) convert to real value Rref
in the S-parameter generation environment
Ò avoids inconsistencies between generation and application
environments
S-parameters are “differential” in nature
Ò only the local voltage difference is relevant

Ò there is no such thing as global ”ground”

Ò you can never know the voltage at any of the

S-parameter reference nodes … NEVER!
Ò all “ports” are 2-terminal for generation of S-
parameters
Ò the current into one terminal is the same as the current
out of the other terminal (for the 2N model)
“2N” and “N+1” circuit components

Ò 2N: nodelist: port1 ref1 port2 ref2 …

Ò N+1: nodelist: port1 port2 … ref

Ò Either can be applied in a circuit simulator

Ò the 2N is identical to the N+1 with ideal n=1:1 transformers wrapped around it

Ò The information contained in the S-parameter file is exactly the same in either case
Ò the circuit behavior had better be exactly the same in either case

Ò The reference node should be “common”

Ò it is typically set to node 0 (global ground)
Ò this is okay, since you never know Vref

Ò The current into port1 is equal to the current out of ref1

Ò trying to force it to be otherwise may result in an error from the circuit simulator or yield strange results
Mixed-mode S-parameters
Ò “single-ended”
Ò single uncoupled traces - Sij

Ò “mixed-mode” - {SDD, SCC, SDC, SCD}ij

Ò diff-pairs
Ò if you ignore the common mode you may be in serious trouble
Ò the two modes can mix together internal to the circuit
Ò emission is much greater from the common mode

Ò “multi-mode”
Ò N-coupled lines
Ò modes typically not considered by designers for N>2
Ò general waveguides
Ò including cutoff modes
Common language
(see the wiki web page cited earlier for much more)

Ò Return Loss: reflection, Sii

Ò Insertion Loss: thru transmission, S21

Ò Coupling, Crosstalk, etc: Sij

Ò Near-end and Far-end Crosstalk (NEXT, FEXT)

Ò distinguishes the Sij for coupled thru paths for the source and receiver ends

Ò dB: 20(log10(S))

Ò phase: phase(S)
Ò typically “wrapped” [-180,180]
Ò negative phase accumulation with delay
Common displays
Ò most common display
Ò X-axis: linear frequency
Ò Y-axis: dB for magnitude, [-180,180] degrees for phase

Ò Smith Charts
Ò simply a bilinear mapping of magnitude and phase to a
2D plot
Ò relevant to reflection S-parameters
Ò many GUIs provide readout of equivalent impedance
Ò bottom half is capacitive, top half is inductive
Ò resonances are tight loops
Ò matched condition is in the center, open/short circuit on
the outside of the circle on opposite sides
Properties of S-parameters
Ò Reciprocity
Ò Sij = Sji … except for loss and active circuits

Ò Passivity
Ò (I-SS*) >= 0

Ò Causality
Ò no response before T=0, or appropriate physical delay

Ò Smoothness
Ò not really a property so don’t insist on it
Practical issues for modeling with S-parameters
Ò Noise
Ò measurements
Ò calibration errors, repeatability, noise
Ò EM computations
Ò numerical error, convergence, design specification (geometrical representation, material properties)

Ò Phase
Ò phase is critical to get correct, often more so than amplitude
Ò phases corresponding to extremely small magnitudes can be significantly wrong

Ò Magnitude
Ò if you are looking at extremely small values and they are significantly wrong you must
sometimes ask yourself if it really matters

Ò Frequency-domain TDR
Ò amplitude and phase can help to debug without an actual TDR
File formats
Ò “Touchstone”
Ò been around for many years
Ò extension of SNP, where N=integer number of ports
Ò version 2.0 specified by IBIS
Ò mixed mode parameters, per-port Zref, etc.
Ò version 2.1 available soon
Ò sparse data, binary
Ò future versions
Ò rational functions instead of discrete frequency samples

Ò Proprietary formats
Ò Sigrity BNP – “binary network parameters”
Ò rational function representations possible, enabling extraction at “any” frequency
Ò API available upon request (HSPICE applies it)
Ò did not want to document and support with version control for the entire industry
Ò future versions of Touchstone will have many of the same capabilities and resulting benefits
Model Connectivity
Ò Two proprietary formats are available
Ò Sigrity MCP: Model Connection Protocol
Ò public domain format, multiple EDA companies read/write
Ò Apache CPP: Chip Package Protocol
Ò covered under NDA, multiple EDA companies read/write

Ò Need to connect two high pin count models

Ò netlist usage presents issues
Ò schematics typically apply individual “wires”
Ò need a “bus” connection concept
Ò need to account for
Ò coordinate transformations: flip, origin, translation, rotation
Ò misalignment
Ò connectivity can be driven by pin names or locations
Model Connectivity Example

Ò Connection Modes
Ò by name or by physical location

Ò Electrical resolution may vary

Ò May merge two models

Ò Pwr/Gnd nodes highlighted

Ò Supports both MCP and CPP

Reduced Order Models
Ò Motivation
Ò It is difficult to deal with frequency domain data in a transient
circuit simulator
Ò pin counts are increasing
Ò data bandwidth is increasing
Ò data is often not “well conditioned”

Ò Benefits relative to direct usage of S-parameters

Ò faster, more accurate transient circuit simulation
Ò higher pin counts typically supported
Ò reduced file size relative to original S-parameters
Reduced Order Models
Ò Vector Fitting
Ò the underlying technical algorithms of most (if not all) macromodel generators
Ò essentially a pole-zero fit of the data

Ò Difficulties
Ò high pin count – memory, time and numerical convergence
Ò previously 100 was considered good, now 500
Ò preserving passivity and causality
Ò tradeoffs between maintaining the original S-parameters and generating good
transient results
Ò ultimate judge is time domain circuit simulation

Ò W-element models
Ò more efficient and more accurate than general S-parameters for long TLines
Ò TL phase delay has an infinite number of poles/zeroes
Macromodel generation flow
Macromodel 4-port serial channel example
 die
Vdpg Vdsg

Applications of S-parameters
to chip-package PDN Vdie-brd

Ò Do not try to measure voltage at

reference nodes when working
with S-parameters Vbpg Vbsg

board
Applications of S-parameters to parallel buses and serial
channels
Ò Return paths
Ò be consistent amongst all your models
Ò use an IO buffer model to properly push/pull current into the PDN

Ò Power Ports
Ò model signals and power together
Ò use a common return path

Ò IBIS 5.0 buffer models

Ò power-aware IO models are now available
Ò can convert from transistor models and apply in HSPICE and Sigrity
tools
Applications of S-parameters to parallel buses
Ò Single versus dual referencing PDN noise
Ò required simultaneous signal/power modeling
Ò 15mv single vs. 45mv dual, though two extra layers in the package

Ò Memory interface
Ò silicon vendor and system house worked together to model system
Ò eye opening and jitter predictions varied significantly
Ò used different return paths
Ò one applied separate signal/power modeling
Ò with different return paths
Ò must use power-aware IO buffer models to model PDN currents
Applications of S-parameters to parallel buses (memory)
Lessons:
(1) Use a driver model to push/pull current from the real return path(s), since it
varies with rising/falling edges.
(2) Model signals, power and ground together.
Eye Diagram #1 Eye Diagram #2

(a) VSS and VDD return path (a) VSS return path
(b) PDN and signals modeled together (b) PDN (VSS and VDD) and signals
are modeled separately
Eye Mask Width = 803 pS
Jitter = 89 pS Eye Mask Width = 758 pS
Jitter = 114 pS
Differences due to:
Different return paths (VSS or VSS/VDD*)
Signals and PDN treated differently (all together or separately with different return paths)
Lessons learned … and re-learned regularly
Ò There is no such thing as global ground

Ò You never know Vref when working with S-parameters

Ò Use a consistent return path amongst all models

Ò Model signals and power together

Ò use a common return path
Ò use a power-aware IO buffer model

Ò S-parameters are often not well conditioned to support transient

circuit simulation
Thank You!