1. Practical Macromodels for Digital I/O
Paul Franzon
Madhavan Swaminathan, Michael Steer
Acknowledgements: Ambrish Varma, Bhyrav Mutnury

2. Outline Background Macromodeling Techniques Proposal : The Way Forward
Macromodeling Needs
Requirements for Successful Macromodels
Macromodeling Techniques
IBIS
Numerical Models
Physical
Proposal : The Way Forward

3. I/O Macromodeling Replace full Spice driver model Macromodel with
Hides proprietary information
Speeds up simulation

4. I/O Macromodeling Required Goals: Vendor-independent format
Have to capture TX, RX, and package parasitic essentials
Sufficiently accurate to be useful
Easy to automatically generate from Spice and Measurements
Easy to verify
Easy and fast to simulate

5. I/O Macromodels Desirable Goals Unique solution Orthogonality
Multiple Verification Approaches
Physical basis
Human readability
Monotonic
Simulator Independent
Extendable to core noise SSN

6. IBIS IBIS http://www.eda.org/pub/ibis/
Input Output Buffer Information Specification
EIA Standard 656 -A
Behavioral model of I/O buffers
Model is represented as set of VI and VT curves.
Fast simulations
Protects proprietary information contained in the IC

7. Basic IBIS Model A Basic model consist of:
Four I-V curves: - pullup and pulldown
- PWR and GND Clamp
Two Ramps: -dv/dt_rise, dv/dt_fall
Die Capacitance and
Packaging information

8. Spice2Ibis Automatically converts a Spice deck to an IBIS file
Over 1,000 active users
Over 3,000 lines of code
Was critical to success of IBIS language
S2IBIS
Command File
-Header information
-Component Description
Parser
-Extracts all relevant
information
from Command file
SPICE
-HSPICE,PSPICE,
SPICE2
SPICE3,SPECTRE
-Calls Spice,
analyzes data
Prints results
IBIS Model
Circuit Layout

9. IBIS Limitations IBIS hard to scale for use with high speed I/O
Up to 100 ps delay error
Subtle inflections missed
SSN over-predicted

10. IBIS Alternatives Current Work: Numerical black-box methods
“Parametric” Models:
Stievano, Maio, and Canavero, Politecnico di Torino
Spline/Finite Time Difference
Swaminathan, Georgia Tech

11. Numerical Methods “Black Box” Modeling Static Term
istatic(k) = w1(k)f1(k) + w2(k)f2(k)
Captures output impedance of driver
IBIS:
f1 and f2 are fitted VI tables. i.e. I(Vout)
Spline :
f1 and f2 are numerically fitted power series, i.e. I(Vout, ….)
Parametric:
Basis functions: Gaussian or Sigmoid

12. Numerical Methods Black Box Modeling Dynamic Term
IBIS : i(k) = w1(k)V1(k)f1(k) + w2(k)V2(k)f2(k)
+ output physical Ccomp
Spline :
f1, f2 dynamic by numerically fitted load capacitance (captures di/dt)
Parametric
Calculated as a function of past values, using Basis functions
e.g. RBF using Gaussians
Vn(T)

13. Physical Model Circuit Level: Captures staged turn-on drivers M1(T)
Gain loss during
switching event (Vgs)
Second order effects = f(Vds)
w=step function
only in break before
make drivers
CML reduces CM dependence
Circuit Level:
M1(T) S G
Ids=k(Vgs-Vt)2 D
M2(T) or
etc.

14. Comments on Models IBIS Spline/FTD Parametric
No numerical flexibility to capture subtle physical effects
Most physical (but very first order)
Easy to automate
Spline/FTD
Input waveform dependence
Less physical
More accurate
Relatively hard to automate
Numerical fitting of power series
Parametric
Least physical
Most accurate
Harder to automate model production
More complex numerical procedures

15. Proposal Collaborative Effort: NC State University Georgia Tech
Politecnico di Torino

16. The Way Forward
Find the compromise between complexity and automation, while considering other goals
Spline
RBF
IBIS 4.0
RBF Sigmoid
Power Series
Reduced Order Power Series
Wavelet basis fns
Neural net fitting
Etc.
I=f(Vout)
ErrorCurrentSource*
I=f(Vout,Vdd/Vss)
M=f(T,Vdd)
Etc.
*e.g. Numerically fitted error fn

17. The Way Forward Test on standard driver set: Evaluate
Conventional, LVDS, DDR, Other CML, Emerging
Test Spice model formats in Freeda
Evaluate
Metrics:
Accuracy in predicting delay, peak SSN, xmitted SSN, Xtalk, refn noise
Accuracy outside range where fitted
Macromodel utility factors listed earlier
Esp. Complexity of model fit procedure

18. The Way Forward 2. Promulgate alpha standard
Develop SpicetoMacromodel
3. Evaluate and propagate more broadly

19. Proposed Consortium Set up industry consortium to fund this work
Benefits to consortium members
IC vendor companies
Macromodel tuned to their drivers
Early access to successful macromodel formats
First to market
CAD companies
Early access to successful formats
Influence macromodel for simulator compatibility

20. Discussion (on the day)
Acknowledgement and contribution of the highly accurate “black box” techniques contributed by Prof. Flavio Canavero and his group.
George Katopis:
1. Study and development of a tool that helps the user with the selection of black box option based on "expected" accuracy and time.
2. Automatic generation of the black box models

21. Conclusions Ease of Use is just as important as model accuracy
All macromodels are numerical black box format
Key question is complexity and type of functions used