DARPA GTX: The GSRC Technology Extra- polation System, “A Living Roadmap” A. Caldwell, A. B....
-
date post
21-Dec-2015 -
Category
Documents
-
view
221 -
download
1
Transcript of DARPA GTX: The GSRC Technology Extra- polation System, “A Living Roadmap” A. Caldwell, A. B....
DARPADARPA
GTX: The GSRC Technology Extra-GTX: The GSRC Technology Extra-polation System, “A Living Roadmap”polation System, “A Living Roadmap”
A. Caldwell, A. B. Kahng, F. Koushanfar,A. Caldwell, A. B. Kahng, F. Koushanfar,
H. Lu, I. Markov, M. Oliver andH. Lu, I. Markov, M. Oliver and
D. StroobandtD. Stroobandt
11/3/99
2
OverviewOverview
IntroductionIntroduction
Previous roadmapping effortsPrevious roadmapping efforts
Overview of GTX: goal and structureOverview of GTX: goal and structure
Fundamental featuresFundamental features
DemonstrationDemonstration
Example use scenarios:Example use scenarios:Roadmap emulation, development, and maintenanceRoadmap emulation, development, and maintenance
Roadmap evaluation and comparisonRoadmap evaluation and comparison
11/3/99
3
Introduction: Technology ExtrapolationIntroduction: Technology Extrapolation
Evaluates impactEvaluates impact of ofdesign technologydesign technologyprocess technologyprocess technology
Evaluates impact onEvaluates impact onachievable designachievable designassociated design problemsassociated design problems
Questions to be addressedQuestions to be addressed
Sets Sets new requirementsnew requirements for CAD tools and methodologies for CAD tools and methodologies
RoadmapsRoadmaps: familiar and influential example: familiar and influential example
How and when do L, How and when do L, SOI, SER, etc. matter?SOI, SER, etc. matter?
What is the most power-efficient noise What is the most power-efficient noise management strategy?management strategy?
Will layout tools need to perform Will layout tools need to perform process simulation to efficiently process simulation to efficiently model cross-die and cross-wafer model cross-die and cross-wafer
manufacturing variation?manufacturing variation?
11/3/99
4
Introduction: RoadmappingIntroduction: Roadmapping
Roadmapping efforts drive development of design technology:Roadmapping efforts drive development of design technology: System architects, designers, CAD managers use roadmaps to determineSystem architects, designers, CAD managers use roadmaps to determine
tough problemstough problems risks, …risks, …
EDA suppliers use roadmaps to determineEDA suppliers use roadmaps to determine R&D investmentR&D investment product pipelineproduct pipeline
Government and consortia use roadmaps to determine levels of investmentGovernment and consortia use roadmaps to determine levels of investment
Roadmaps serve as a guide to the most promising directions, the Roadmaps serve as a guide to the most promising directions, the
most critical problemsmost critical problems
11/3/99
5
Roadmap Process and Its ImplicationsRoadmap Process and Its Implications
Basic TechnologicalAssumptions
Basic MethodologicalAssumptions
Implications tothe Community
Models andDiscussion
Translation to SpecificResearch Agendas
“Timing closure is a hard problem and will only get harder”
“We will fund research ontiming-aware partitioning”
Research Proposedto Implement Agenda R. Newton, ICCAD99 panel
11/3/99
6
Roadmap ProcessRoadmap Process
Basic TechnologicalAssumptions
Basic MethodologicalAssumptions
Implications tothe Community
Models andDiscussion
Couched in Terms ofRoadmap Implications
“Timing closure is a hard problem and will only get harder”
Research Proposed to Solve Hard Problem
“I can make a breakthrough in technology or methodology”
“Here’s how my work is critical for addressing your problem”
Newmodels
R. Newton, ICCAD99 panel
11/3/99
7
Introduction: RoadmappingIntroduction: Roadmapping
Difficulties of roadmappingDifficulties of roadmappingNo crystal ballNo crystal ballNew technologies, new circuitNew technologies, new circuit
techniques, new designtechniques, new design
methodologies and toolsmethodologies and toolsAlways difficult to predict achievable design, especially in the Always difficult to predict achievable design, especially in the
futurefuture
Roadmaps rely onRoadmaps rely onModels: technology projections, design attributes, design toolsModels: technology projections, design attributes, design toolsCalibrations: measurements of technology and design parametersCalibrations: measurements of technology and design parameters
11/3/99
8
Roadmapping in the pastRoadmapping in the past
Previous and ongoing effortsPrevious and ongoing efforts ITRS RoadmapsITRS Roadmaps Tools: SUSPENS, GENESYS, RIPE, BACPAC, …Tools: SUSPENS, GENESYS, RIPE, BACPAC, … Numerous tools in industryNumerous tools in industry
ObservationsObservations Predict “same” parameters but with different assumptions, inputsPredict “same” parameters but with different assumptions, inputs Lack of documentation and visibility into internal calculationsLack of documentation and visibility into internal calculations Single inference chain for a given output (hard-coded modeling)Single inference chain for a given output (hard-coded modeling) Inflexible: user cannot define studies of other, related parametersInflexible: user cannot define studies of other, related parameters Near-total duplication of effortNear-total duplication of effort Missing: models of CAD tools and optimizations (what is really “achievable”?)Missing: models of CAD tools and optimizations (what is really “achievable”?) Missing: scope, comprehensive coverageMissing: scope, comprehensive coverage
11/3/99
9
Questions To Ask About RoadmapsQuestions To Ask About Roadmaps
How do different roadmap predictions How do different roadmap predictions comparecompare??
How to How to evaluateevaluate underlying models? (sanity checks) underlying models? (sanity checks)
How do we How do we reusereuse and and extendextend models to encompass new models to encompass new
aspects of technology, new axes of achievable design?aspects of technology, new axes of achievable design?
What is the impact of modeling What is the impact of modeling choiceschoices on predictions? on predictions?
Need a new infrastructure, new concept!Need a new infrastructure, new concept!
11/3/99
10
Previous Systems versus Ideal SystemPrevious Systems versus Ideal System
Same parameters but Same parameters but
different assumptionsdifferent assumptions
Inflexible, not easy toInflexible, not easy to
add other studies add other studies
Hard-coded, no easy Hard-coded, no easy
changeschanges
No internal visibilityNo internal visibility
Duplication of effortDuplication of effort
FlexibilityFlexibility
QualityQuality
TransparencyTransparency
Prevention of Prevention of
redundant effortredundant effort
11/3/99
11
Goals of A New Technology Extrapolation SystemGoals of A New Technology Extrapolation System
FlexibilityFlexibility Interactively edit chains of relations between parametersInteractively edit chains of relations between parameters
Define new parameters and relations between themDefine new parameters and relations between them
Perform specific studies (but different studies at different times)Perform specific studies (but different studies at different times)
QualityQualityContinuous improvementsContinuous improvements
World-wide participation of expertsWorld-wide participation of experts
TransparencyTransparency
Prevention of redundant effortPrevention of redundant effort
11/3/99
12
Goals of New Technology Extrapolation (cont.)Goals of New Technology Extrapolation (cont.)
FlexibilityFlexibility
QualityQuality
TransparencyTransparencyOpen-source mechanismOpen-source mechanism
Models are visible to the userModels are visible to the user
Prevention of redundant effortPrevention of redundant effortPermanent repository of first choicePermanent repository of first choice
Adoptability and maintainabilityAdoptability and maintainability
11/3/99
13
GTX: GSRC Technology Extrapolation SystemGTX: GSRC Technology Extrapolation System
GTX is set up as a GTX is set up as a frameworkframework for technology extrapolation for technology extrapolation
Flexibility, quality, visibility allow a “living roadmap”:Flexibility, quality, visibility allow a “living roadmap”:emulateemulate existing roadmap (modeling) efforts existing roadmap (modeling) effortsdevelopdevelop new roadmaps (models) new roadmaps (models)evaluateevaluate roadmaps (models) roadmaps (models)comparecompare roadmaps (models) to each other roadmaps (models) to each other
KnowledgeData
ModelsImplementation
Derivation
Presentation
11/3/99
14
GTX StructureGTX Structure
Knowledge representation:Knowledge representation: ParametersParameters (description of technology, circuit and design attributes) (description of technology, circuit and design attributes)
RulesRules (methods to derive unknown parameters from known ones): (methods to derive unknown parameters from known ones): closed-form modelsclosed-form models executable algorithm implementationsexecutable algorithm implementations table-lookupstable-lookups
Rule chainsRule chains (serialized user-defined rules) (serialized user-defined rules) interactive specification and comparison of alternative modeling choicesinteractive specification and comparison of alternative modeling choices
ImplementationImplementation Execution by a Execution by a derivation enginederivation engine to perform studies to perform studies
Embedding into Embedding into GUIGUI for ease of use, interactivity, display of results for ease of use, interactivity, display of results
See poster for details of GTX frameworkSee poster for details of GTX framework
11/3/99
15
Knowledge RepresentationKnowledge Representation
Rules and parameters are specified separately from the Rules and parameters are specified separately from the
derivation enginederivation engine
Human-readable ASCII grammarHuman-readable ASCII grammar
#rule #rule BACPAC_dl_chipBACPAC_dl_chip#description#description#output#outputdouble {m} dl_chipdouble {m} dl_chip;;
#inputs#inputsdouble {m^2} dA_chipdouble {m^2} dA_chip;;
#body #body sqrt(dA_chip)sqrt(dA_chip)#reference#reference#endrule#endrule
#parameter #parameter dl_chipdl_chip#type #type doubledouble#units {#units {mm}}#default#default1e-21e-2
#description#descriptionchip side lengthchip side length
#reference#reference#endparameter#endparameter
11/3/99
16
Knowledge Representation (cont.)Knowledge Representation (cont.)
Rules and parameters are specified separately from the derivation Rules and parameters are specified separately from the derivation
engineengine
Human-readable ASCII grammarHuman-readable ASCII grammar
Benefits:Benefits: Easy creation and sharing of parameters / rules by multiple usersEasy creation and sharing of parameters / rules by multiple users
D. Sylvester and K. Cao: device and power modules that “drop in” to GTXD. Sylvester and K. Cao: device and power modules that “drop in” to GTX
Extensible to models of arbitrary complexity (specialized prediction methods, Extensible to models of arbitrary complexity (specialized prediction methods, technology data sets, optimization engines)technology data sets, optimization engines) Avant! Apollo or Cadence SE P&R tool: just another wirelength estimatorAvant! Apollo or Cadence SE P&R tool: just another wirelength estimator
Applies to any domain of work in semiconductors, VLSI CADApplies to any domain of work in semiconductors, VLSI CAD Transistor sizing, single wire optimizations, system-level wiring predictions,…Transistor sizing, single wire optimizations, system-level wiring predictions,…
11/3/99
17
Parameter and Rule NamingParameter and Rule Naming
Importance of consistent naming cannot be overstatedImportance of consistent naming cannot be overstated
Naming conventions for parametersNaming conventions for parameters[<preposition>] _ <principal> _ {[qualifier] _ <place>} _ {<qualifier>} _ [<adverbial>] _ [<index>] _ [<unit>][<preposition>] _ <principal> _ {[qualifier] _ <place>} _ {<qualifier>} _ [<adverbial>] _ [<index>] _ [<unit>]
Example: Example: r_int_tot_lyr_pu_dlr_int_tot_lyr_pu_dl
Benefits:Benefits:Relatively easy to understand parameter from its nameRelatively easy to understand parameter from its nameDistinguishable (no two parameters should have the same name)Distinguishable (no two parameters should have the same name)
r_int (interconnect resistance) = r_int (interconnect resistivity) ?r_int (interconnect resistance) = r_int (interconnect resistivity) ?
Unique (no two names for the same parameter)Unique (no two names for the same parameter) R_int = R_wire ?R_int = R_wire ?
Sortable (important literals come first)Sortable (important literals come first)
11/3/99
18
Additional FeaturesAdditional Features
Optimization over a collection of rules (with constraints)Optimization over a collection of rules (with constraints)Example: buffer insertion for minimal delay with area constraintsExample: buffer insertion for minimal delay with area constraints
Executables can be calledExecutables can be calledExample: various optimizations of global delay through IPEM Example: various optimizations of global delay through IPEM
(Interconnect Performance Estimation Models, J. Cong, UCLA)(Interconnect Performance Estimation Models, J. Cong, UCLA)
Internal code rules for optimizationsInternal code rules for optimizationsExample: optimization of number and size of repeaters for global Example: optimization of number and size of repeaters for global
wireswires
Storing of calibration data (e.g., “technology files”) for Storing of calibration data (e.g., “technology files”) for known process, design pointsknown process, design points
11/3/99
19
Additional Features (cont.)Additional Features (cont.)
Visualization (plotting, printing, saving to file) Visualization (plotting, printing, saving to file)
Sweeping over sets of input valuesSweeping over sets of input valuesExample: clock frequency for different Rent exponents and Example: clock frequency for different Rent exponents and
varying logic depthvarying logic depth
11/3/99
20
GTX: Open and User-friendlyGTX: Open and User-friendly
Openness in grammar, parameters and rulesOpenness in grammar, parameters and rulesEasy sharing of data in research environmentEasy sharing of data in research environment
Contributions from other groupsContributions from other groups
Allows developing of proprietary modelsAllows developing of proprietary modelsSeparation between supplied (shared) and user-defined Separation between supplied (shared) and user-defined
parameters / rulesparameters / rules
GTX offers usability behind firewallsGTX offers usability behind firewalls
Framework for sharing results instead of data is plannedFramework for sharing results instead of data is planned
Multi-platform (SUN Solaris, Windows, Linux)Multi-platform (SUN Solaris, Windows, Linux)
11/3/99
22
GTX Current StatusGTX Current Status
EmulationEmulation of ofCycle-time models of SUSPENS (with extension by Takahashi), Cycle-time models of SUSPENS (with extension by Takahashi),
BACPAC, Fisher (ITRS)BACPAC, Fisher (ITRS) Interconnect tuning studiesInterconnect tuning studiesMain modulesMain modules
Clock / powerClock / power SOISOI Domino logicDomino logic Device and PowerDevice and Power Global interconnectGlobal interconnect System-level powerSystem-level power PackagingPackaging Reliability and YieldReliability and Yield ……
11/3/99
23
GTX Current StatusGTX Current Status
EvaluationEvaluation of cycle-time models of cycle-time modelsParameter sensitivityParameter sensitivity
ComparisonComparison between cycle-time models between cycle-time modelsModel sensitivityModel sensitivity
DevelopmentDevelopment of new models of new modelsModel of via impact on required routing resources (number of Model of via impact on required routing resources (number of
layers, pitch, etc.)layers, pitch, etc.)
11/3/99
24
Change parameter values and observe resulting difference Change parameter values and observe resulting difference
in outputsin outputs
See poster on Sensitivity Analysis for further detailsSee poster on Sensitivity Analysis for further details
Evaluation: Parameter SensitivityEvaluation: Parameter Sensitivity
11/3/99
25
Comparison: Model SensitivityComparison: Model Sensitivity
Replace rule in a model’s rule chain by another model’s Replace rule in a model’s rule chain by another model’s
rule and observe the difference in outputsrule and observe the difference in outputs
See poster on Sensitivity Analysis for further detailsSee poster on Sensitivity Analysis for further details
BACPAC BACPAC with rule from Fisher
11/3/99
26
Development: Via Impact ModelDevelopment: Via Impact Model
Goal: model impact of vias on layer track utilizationGoal: model impact of vias on layer track utilization
Only taking into account area taken by via is not enoughOnly taking into account area taken by via is not enough
Stochastic model of the number of wires blocked by vias Stochastic model of the number of wires blocked by vias
used to estimate the via impactused to estimate the via impact
Via impact model improves prediction of number of layers Via impact model improves prediction of number of layers
needed for the routingneeded for the routingVerified with recent 4LM block, Cadence Silicon Ensemble P&RVerified with recent 4LM block, Cadence Silicon Ensemble P&R
See poster on Via Impact Model for further detailsSee poster on Via Impact Model for further details
11/3/99
27
ConclusionConclusion
GTX: a new framework for roadmapping models and technology GTX: a new framework for roadmapping models and technology
extrapolation effortsextrapolation efforts Flexible and extensibleFlexible and extensible
Enables easy reuse of modelsEnables easy reuse of models
Provides a common parameter base between all modelsProvides a common parameter base between all models
Provides user interactionProvides user interaction
Relies on open-source and contributions by expert usersRelies on open-source and contributions by expert users
““Living Roadmap”Living Roadmap”
Technology extrapolation becomes easierTechnology extrapolation becomes easier
More principled understanding of requirements for CAD toolsMore principled understanding of requirements for CAD tools
11/3/99
28
GTX Project InformationGTX Project Information
Design: A. Caldwell, A. B. Kahng, I. Markov, M. OliverDesign: A. Caldwell, A. B. Kahng, I. Markov, M. Oliver
Implementation: M. OliverImplementation: M. Oliver
Knowledge gathering and implementation: A. B. Kahng,Knowledge gathering and implementation: A. B. Kahng,
F. Koushanfar, H. Lu, D. StroobandtF. Koushanfar, H. Lu, D. Stroobandt
Detailed information and downloading of prototype version Detailed information and downloading of prototype version of GTX: of GTX: http://www.gigascale.org/GTX/http://www.gigascale.org/GTX/
To contact the developers, ask questions, send comments, To contact the developers, ask questions, send comments, or to contribute with models to be included in GTX, please or to contribute with models to be included in GTX, please send E-mail to send E-mail to [email protected]@cs.ucla.edu
11/3/99
29
AcknowledgementsAcknowledgements
MARCO GSRCMARCO GSRC
F.W.O. (Belgium) for D. Stroobandt’s grant to visit UCLAF.W.O. (Belgium) for D. Stroobandt’s grant to visit UCLA
Dr. Phil Fisher, Dr. Dennis Sylvester and Kevin Cao for Dr. Phil Fisher, Dr. Dennis Sylvester and Kevin Cao for
providing access to their models and helpful inputsproviding access to their models and helpful inputs
Professors Ken Rose, James Meindl, Scott Wills and Kurt Professors Ken Rose, James Meindl, Scott Wills and Kurt
Keutzer for fruitful discussionsKeutzer for fruitful discussions