1. DARPA GTX: The GSRC Technology Extra- polation System, “A Living Roadmap” A. Caldwell, A. B. Kahng, F. Koushanfar, H. Lu, I. Markov, M. Oliver and D. Stroobandt

2. 2 11/3/99 Overview u Introduction u Previous roadmapping efforts u Overview of GTX: goal and structure u Fundamental features u Demonstration u Example use scenarios: s Roadmap emulation, development, and maintenance s Roadmap evaluation and comparison

3. 3 11/3/99 Introduction: Technology Extrapolation u Evaluates impact of s design technology s process technology u Evaluates impact on s achievable design s associated design problems u Questions to be addressed u Sets new requirements for CAD tools and methodologies u Roadmaps: familiar and influential example How and when do L, SOI, SER, etc. matter? What is the most power-efficient noise management strategy? Will layout tools need to perform process simulation to efficiently model cross-die and cross-wafer manufacturing variation?

4. 4 11/3/99 Introduction: Roadmapping u Roadmapping efforts drive development of design technology: s System architects, designers, CAD managers use roadmaps to determine t tough problems t risks, … s EDA suppliers use roadmaps to determine t R&D investment t product pipeline s Government and consortia use roadmaps to determine levels of investment u Roadmaps serve as a guide to the most promising directions, the most critical problems

5. 5 11/3/99 Roadmap Process and Its Implications Basic Technological Assumptions Basic Methodological Assumptions Implications to the Community Models and Discussion Translation to Specific Research Agendas “Timing closure is a hard problem and will only get harder” “We will fund research on timing-aware partitioning” Research Proposed to Implement Agenda R. Newton, ICCAD99 panel

6. 6 11/3/99 Roadmap Process Basic Technological Assumptions Basic Methodological Assumptions Implications to the Community Models and Discussion Couched in Terms of Roadmap 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” New models R. Newton, ICCAD99 panel

7. 7 11/3/99 Introduction: Roadmapping u Difficulties of roadmapping s No crystal ball s New technologies, new circuit techniques, new design methodologies and tools s Always difficult to predict achievable design, especially in the future u Roadmaps rely on s Models: technology projections, design attributes, design tools s Calibrations: measurements of technology and design parameters

8. 8 11/3/99 Roadmapping in the past u Previous and ongoing efforts s ITRS Roadmaps s Tools: SUSPENS, GENESYS, RIPE, BACPAC, … s Numerous tools in industry u Observations s Predict “same” parameters but with different assumptions, inputs s Lack of documentation and visibility into internal calculations s Single inference chain for a given output (hard-coded modeling) s Inflexible: user cannot define studies of other, related parameters s Near-total duplication of effort s Missing: models of CAD tools and optimizations (what is really “achievable”?) s Missing: scope, comprehensive coverage

9. 9 11/3/99 Questions To Ask About Roadmaps u How do different roadmap predictions compare? u How to evaluate underlying models? (sanity checks) u How do we reuse and extend models to encompass new aspects of technology, new axes of achievable design? u What is the impact of modeling choices on predictions? Need a new infrastructure, new concept!

10. 10 11/3/99 Previous Systems versus Ideal System - Same parameters but different assumptions - Inflexible, not easy to add other studies - Hard-coded, no easy changes - No internal visibility - Duplication of effort FlexibilityQualityTransparency Prevention of redundant effort

11. 11 11/3/99 Goals of A New Technology Extrapolation System u Flexibility s Interactively edit chains of relations between parameters s Define new parameters and relations between them s Perform specific studies (but different studies at different times) u Quality s Continuous improvements s World-wide participation of experts u Transparency u Prevention of redundant effort

12. 12 11/3/99 Goals of New Technology Extrapolation (cont.) u Flexibility u Quality u Transparency s Open-source mechanism s Models are visible to the user u Prevention of redundant effort s Permanent repository of first choice s Adoptability and maintainability

13. 13 11/3/99 GTX: GSRC Technology Extrapolation System u GTX is set up as a framework for technology extrapolation u Flexibility, quality, visibility allow a “living roadmap”: s emulate existing roadmap (modeling) efforts s develop new roadmaps (models) s evaluate roadmaps (models) s compare roadmaps (models) to each other Knowledge Data Models Implementation Derivation Presentation

14. 14 11/3/99 GTX Structure u Knowledge representation: s Parameters (description of technology, circuit and design attributes) s Rules (methods to derive unknown parameters from known ones): t closed-form models t executable algorithm implementations t table-lookups s Rule chains (serialized user-defined rules) t interactive specification and comparison of alternative modeling choices u Implementation s Execution by a derivation engine to perform studies s Embedding into GUI for ease of use, interactivity, display of results u See poster for details of GTX framework

15. 15 11/3/99 Knowledge Representation u Rules and parameters are specified separately from the derivation engine u Human-readable ASCII grammar #rule BACPAC_dl_chip #description#output double {m} dl_chip; #inputs double {m^2} dA_chip; #body sqrt(dA_chip) #reference#endrule #parameter dl_chip #type double #units {m} #default1e-2#description chip side length #reference#endparameter

16. 16 11/3/99 Knowledge Representation (cont.) u Rules and parameters are specified separately from the derivation engine u Human-readable ASCII grammar u Benefits: s Easy creation and sharing of parameters / rules by multiple users t D. Sylvester and K. Cao: device and power modules that “drop in” to GTX s Extensible to models of arbitrary complexity (specialized prediction methods, technology data sets, optimization engines) t Avant! Apollo or Cadence SE P&R tool: just another wirelength estimator s Applies to any domain of work in semiconductors, VLSI CAD t Transistor sizing, single wire optimizations, system-level wiring predictions,…

17. 17 11/3/99 Parameter and Rule Naming u Importance of consistent naming cannot be overstated u Naming conventions for parameters [ ] _ _ {[qualifier] _ } _ { } _ [ ] _ [ ] _ [ ]  Example: r_int_tot_lyr_pu_dl u Benefits: s Relatively easy to understand parameter from its name s Distinguishable (no two parameters should have the same name) t r_int (interconnect resistance) = r_int (interconnect resistivity) ? s Unique (no two names for the same parameter) t R_int = R_wire ? s Sortable (important literals come first)

18. 18 11/3/99 Additional Features u Optimization over a collection of rules (with constraints) s Example: buffer insertion for minimal delay with area constraints u Executables can be called s Example: various optimizations of global delay through IPEM (Interconnect Performance Estimation Models, J. Cong, UCLA) u Internal code rules for optimizations s Example: optimization of number and size of repeaters for global wires u Storing of calibration data (e.g., “technology files”) for known process, design points

19. 19 11/3/99 Additional Features (cont.) u Visualization (plotting, printing, saving to file) u Sweeping over sets of input values s Example: clock frequency for different Rent exponents and varying logic depth

20. 20 11/3/99 GTX: Open and User-friendly u Openness in grammar, parameters and rules s Easy sharing of data in research environment s Contributions from other groups u Allows developing of proprietary models s Separation between supplied (shared) and user-defined parameters / rules s GTX offers usability behind firewalls s Framework for sharing results instead of data is planned u Multi-platform (SUN Solaris, Windows, Linux)

21. 21 11/3/99 Demonstration

22. 22 11/3/99 GTX Current Status u Emulation of s Cycle-time models of SUSPENS (with extension by Takahashi), BACPAC, Fisher (ITRS) s Interconnect tuning studies s Main modules t Clock / power t SOI t Domino logic t Device and Power t Global interconnect t System-level power t Packaging t Reliability and Yield t …

23. 23 11/3/99 GTX Current Status u Evaluation of cycle-time models s Parameter sensitivity u Comparison between cycle-time models s Model sensitivity u Development of new models s Model of via impact on required routing resources (number of layers, pitch, etc.)

24. 24 11/3/99 u Change parameter values and observe resulting difference in outputs u See poster on Sensitivity Analysis for further details Evaluation: Parameter Sensitivity

25. 25 11/3/99 Comparison: Model Sensitivity u Replace rule in a model’s rule chain by another model’s rule and observe the difference in outputs u See poster on Sensitivity Analysis for further details BACPACBACPAC with rule from Fisher

26. 26 11/3/99 Development: Via Impact Model u Goal: model impact of vias on layer track utilization u Only taking into account area taken by via is not enough u Stochastic model of the number of wires blocked by vias used to estimate the via impact u Via impact model improves prediction of number of layers needed for the routing s Verified with recent 4LM block, Cadence Silicon Ensemble P&R u See poster on Via Impact Model for further details

27. 27 11/3/99 Conclusion u GTX: a new framework for roadmapping models and technology extrapolation efforts s Flexible and extensible s Enables easy reuse of models s Provides a common parameter base between all models s Provides user interaction s Relies on open-source and contributions by expert users s “Living Roadmap” u Technology extrapolation becomes easier u More principled understanding of requirements for CAD tools

28. 28 11/3/99 GTX Project Information u Design: A. Caldwell, A. B. Kahng, I. Markov, M. Oliver u Implementation: M. Oliver u Knowledge gathering and implementation: A. B. Kahng, F. Koushanfar, H. Lu, D. Stroobandt u Detailed information and downloading of prototype version of GTX: http://www.gigascale.org/GTX/ u To contact the developers, ask questions, send comments, or to contribute with models to be included in GTX, please send E-mail to GTX@cs.ucla.edu

29. 29 11/3/99 Acknowledgements u MARCO GSRC u F.W.O. (Belgium) for D. Stroobandt’s grant to visit UCLA u Dr. Phil Fisher, Dr. Dennis Sylvester and Kevin Cao for providing access to their models and helpful inputs u Professors Ken Rose, James Meindl, Scott Wills and Kurt Keutzer for fruitful discussions