280 likes | 510 Views
Benchmarking for [Physical] Synthesis. Igor Markov and Prabhakar Kudva The Univ. of Michigan / IBM. In This Talk …. Benchmark ing vs benchmarks Benchmarking exposes new research Qs Why industry should care about benchmarking
E N D
Benchmarking for [Physical] Synthesis Igor Markov and Prabhakar Kudva The Univ. of Michigan/ IBM
In This Talk … • Benchmarking vs benchmarks • Benchmarking exposes new research Qs • Why industry should care about benchmarking • What is (and is not) being doneto improve benchmarking infrastructure • Not in this talk, but in a focus group • Incentives for verifying published work • How to accelerate a culture change
Benchmarking • Design benchmarks • Data model / representation; Instances • Objectives (QOR metrics) and constraints • Algorithms, methodologies; Implementations • Solvers: ditto • Empirical and theoretical analyses, e.g., • Hard vs easy benchmarks (regardless of size) • Correlation between different objectives • Upper / lower bounds for QOR, statistical behavior, etc • Dualism between benchmarks and solvers • For more details, see http://gigascale.org/bookshelf
Industrial Benchmarking • Growing size & complexity of VLSI chips • Design objectives • Area / power / yield / etc • Design constraints • Timing / FP + fixed-die partitions / fixed IPs /routability / pin access / signal integrity… • Can the same algo excel in all contexts? • Sophistication of layout and logic motivate open benchmarking for Synthesis and P&R
Design Types • ASICs • Lots of fixed I/Os, few macros, millions of standard cells • Design densities : 40-80% (IBM) • Flat and hierarchical designs • SoCs • Many more macro blocks, cores • Datapaths + control logic • Can have very low design densities : < 20% • Micro-Processor (P) Random Logic Macros(RLM) • Hierarchical partitions are LS+P&R instances (5-30K) • High placement densities : 80%-98% (low whitespace) • Many fixed I/Os, relatively few standard cells • Note: “Partitioning w Terminals”DAC`99, ISPD `99, ASPDAC`00
Why Invest in Benchmarking • Academia • Benchmarks can identify / capture new research problems • Empirical validation of novel research • Open-source tools/BMs can be analyzed and tweaked • Industry • Evaluation and transfer of academic research • Support for executive decisions(which tools are relatively week & must be improved) • Open-source tools/BMs can be analyzed and tweaked • When is an EDA problem (not) solved? • Are there good solver implementations? • Can they “solve” existing benchmarks?
Participation / Leadership Necessary • Activity 1: Benchmarking platform / flows • Activity 2: Establishing common evaluators • Static timing analysis • Congestion / yield prediction • Power estimation • Activity 3: Standard-cell libraries • Activity 4: Large designs w bells & whistles • Activity 5: Automation of benchmarking
Activity 1: Benchmarking Platform • Benchmarking “platform”: a reasonable subset of • data model • specific data representations (e.g., file formats) • access mechanisms (e.g., APIs) • reference implementation (e.g., a design database) • design examples in compatible formats • Base platforms available (next slide) • More participation necessary • regular discussions • additional tasks / features outlined
Common Methodology Platform Common Model (Open Access?) Synthesis (SIS, MVSIS…) Blif Bookshelf format Placement (Capo, Dragon, Feng Shui, mPl,…) Blue Flow exists, Common model hooks: To be Done
Placement Utilities http://vlsicad.eecs.umich.edu/BK/PlaceUtils/ • Accept input in the GSRC Bookshelf format • Format converters • LEF/DEF Bookshelf • Bookshelf Kraftwerk (DAC98 BP, E&J) • BLIF(SIS) Bookshelf • Evaluators, checkers, postprocessors and plotters • Contributions in these categories are welcome
Placement Utilities (cont’d) • Wirelength Calculator (HPWL) • Independent evaluation of placement results • Placement Plotter • Saves gnuplot scripts ( .eps, .gif, …) • Multiple views (cells only, cells+nets, rows,…) • Probabilistic Congestion Maps (Lou et al.) • Gnuplot scripts • Matlab scripts • better graphics, including 3-d fly-by views • .xpm files ( .gif, .jpg, .eps, …)
Placement Utilities (cont’d) • Legality checker • Simple legalizer • Layout Generator • Given a netlist, creates a row structure • Tunable %whitespace, aspect ratio, etc • All available in binaries/PERL at http://vlsicad.eecs.umich.edu/BK/PlaceUtils/ • Most source codes are shipped w Capo
Activity 2: Creating Evaluators • Contribute measures/analysis tools for: • Timing Analysis • Congestion/Yield • Power • Area • Noise….
Benchmarking Needs for Timing Opt. • A common, reusable STA methodology • High-quality, open-source infrastructure • False paths; realistic gate/delay models • Metrics validated against phys. synthesis • The simpler the better,but must be good predictors • Buffer insertion profoundly impacts layout • The use of linear wirelength in timing-driven layout assumes buffers insertion (min-cut vs quadratic) • Apparently, synthesis is affected too
Vertical Benchmarks • “Tool flow” • Two or more EDA tools, chained sequentially(potentially, part of a complete design cycle) • Sample contexts: physical synthesis, place & route, retiming followed by sequential verification • Vertical benchmarks • Multiple, redundant snapshots of a tool flowsufficient info for detailed analysis of tool performance • Herman Schmit @CMU is maintaining a resp. slot in the VLSI CAD Bookshelf • See http://gigascale.org/bookself • Include flat gate-level netlists • Library information ( < 250nm) • Realistic timing & fixed-die constraints
Infrastructure Needs • Need common evaluators of delay / power • To avoid inconsistent / outdated results • Relevant initiatives from Si2 • OLA (Open Library Architecture) • OpenAccess • For more info, see http://www.si2.org • Still: no reliable public STA tool • Sought: OA-based utilities for timing/layout
Activity 3: Standard-cell Libraries • Libraries carry technology information • Impact of wirelength delays increases in recent technology generations • Cell characteristics must be compatible • Some benchmarks in the Bookshelfuse 0.25m and 0.35m libraries • Geometry info is there, + timing (in some cases) • Cadence test library? • Artisan libraries? • Use commercial tools to create libraries • Prolific, Cadabra,…
Activity 4: Need New BenchmarksTo Confirm / Defeat Tool Tuning • Data on tuning from the ISPD03 paper“Benchmarking for Placement”, Adya et al. • Observe that • Capo does well on Cadence-Capo, grid-like circuits • Dragon does well on IBM-Place (IBM-Dragon) • FengShui does well on MCNC benchmarks • mPL does well on PEKO • This is hardly a coincidence • Motivation for more / better benchmarks • P.S. Most differences above have been explained,all placers above have been improved
Activity 4: Large Benchmark Creation • www.opencores.org has large designs • May be a good starting point –use vendor tools to create blif files(+post results) • Note: there may be different ways to convert • A group of design houses (IBM, Intel, LSI, HP)is planning a release of new largegate-level benchmarks for layout • Probably no logic information
Activity 5: Benchmarking Automation • Rigorous benchmarking is laborious. Risk of errors is high • How do we keep things simple / accessible? • Encapsulate software management in an ASP • Web uploads for binaries and source in tar.gz w Makefiles • Web uploads for benchmarks • GUI interface for NxM simulations; tables created automatically • GUI interface for composing tool-flows; flows can be saved/reused • Distributed back-end includes job scheduling • Email notification of job completion • All files created are available on the Web (permissions & policies) • Anyone can re-run / study your experiment or interface with it
Follow-on Action Plan • Looking for volunteers to -test Bookshelf.exe • Particularly, in the context of synthesis & verification • Contact: Igor imarkov@eecs.umich.edu • Create a joint benchmarking groupfrom industry and academia • Contact: Prabhakar kudva@us.ibm.com • Regular discussions • Development basedon common infrastructure