Are There Economic Benefits in DFM ? - a Fabless Design House Perspective - Riko Radojcic Matt Nowak

What IS DFM to Qualcomm ? cost nDFM = everything that a designer can do to ensure that the product meets its performance, power and cost targets nTight design to process coupling for :  Litho effects =to get what you see in x-y plane  CMP effects =to get what you see in z-plane  Variability =to get what you simulate electrically  Random effects =to get the yield that you modeled  Act of God effects =to get the operating margins you need nDFM = everything that the foundry partner tells you to do ? nDFM = everything that the EDA supplier says you can do ? nDFM = everything that the ‘DFM Start Ups’ say you should do ? nDFM = everything that the designer thinks he should do ? nDFM = business as before and hope for the best ? So …. ?

What is the Big Deal about DFM ? nProcess windows are getting awfully tight nErgo, need to tighten up the design : as always !!! ~Gate Delays SPICE & STA SPICE & STA ~Wire Delays RC Extractors etc.. RC Extractors etc.. ~Power Simulators & Vt selectors Simulators & Vt selectors CONCERN / ISSUE METRIC / TOOL DESIGN CONSIDERATIONS ~Cost & Yield AREA ???? AREA ???? DFM issues are not a discontinuity However there is a challenge with metrics Area used to be a metric, but…. Yield is the obvious new metric but … Designer has no way of designing for it without a way of quantifying it

QUALCOMM Perspective Why go to next generation of Technology ? nHigher Performance  New features nCut-over as soon as possible  Allow for de-bug, system integration nYield is Boss ! nDFM very important nLower Cost  Existing (mostly) features nCut-over when cost is better  Allow for system re-qual time nDie Size is Boss ! nDFM desirable & optional DFM seems to be ‘the right thing to do’, but… Does it really make money for a Fabless COT House?

So, ARE there Economic Benefits in DFM ? nDon’t Really Know : 1.The commercial DFM space is fuzzy and confusing 2.The economic value of a DFM technology is confusing 3.The business case for DFM is unclear 4.The relative benefit of a DFM methodology is unclear nWhat we do know :  DFM is done to Optimize Costs and is hence a Business Decision  Any designer should be able to deal with the DFM trade offs

nP&R nTiming nSynthesis nArchitecture n….. Pre GDS Transformation nHot Spots nDRC/MRC nEdge Massage nComps n….. Post GDS Transformation Need to Segment DFM Universe nLitho nCMP nYield nParametric n…. Models & Simulators nTest chip nMeasurements nData massage nValidation n…. Process Characterization processdesign nSegregate DFM activities into 4 types Qualcomm has a different role with each ‘bucket’ Qualcomm role is that of a ‘champion’ or a ‘promoter’Qualcomm role is that of a ‘champion’ or a ‘promoter’ Principal differentiator is accuracy.Principal differentiator is accuracy. Qualcomm role is that of a ‘client’Qualcomm role is that of a ‘client’ Principal differentiator is use mode and integration.Principal differentiator is use mode and integration. For this segment need a 3 way relationship

$ Process maturity Benefit Assessment Need to Monetize DFM Technical Value DFM Design Decisions Process Design Kit EDA Tools Yes/No nDFM is Design to Optimize vs. Pass Go-No Go Checks  Function of process characteristics and maturity  Function of design technology and trade off vs. other variables  Function of product volume and ASP ramp Benefit of DFM Parameter 1 e.g. minimum die size - typically constant or slight ‘linear’ improvement with maturity Benefit of DFM Parameter 2 e.g. parametric yield or forbidden pitch compliance - typically high at ramp and saturates at maturity ?

Need a DFM Business Case nBusiness Model for Sharing Costs and Benefits  Foundry should reap the benefits for process improvements  Fabless design house should reap some benefit for DFM improvements Wafer Business Die Business FoundryFabless The incremental COST of DFM is mostly on the Fabless side not entirely: foundry does more characterization With Wafer Business the BENEFITS are mostly with Fabless side not entirely: foundry gets larger virtual capacity With Die Business BENEFITS are mostly on the Foundry side not entirely fabless gets larger capacity The challenge is a Business Case for DFM in a Fabless House using a Die Priced Business Mode

Need a Metric for DFM Benefits nYield or Project Cost is too Complicated  Retroactive direct measurements  Many compounding and confounding factors  Feedback loops long & often cloaked in economic factors nNeed some Relative DFM “Score”  Arbitrary normalized number  Simple and non controversial  Not necessarily a direct measures of yield or cost but intuitively related to those  Usable without a dedicated process-specific test chip nPerhaps Use Separate Scores, as makes sense ?  e.g. systematic vs. random vs. parametric phenomena

In fact Need Two DFM Metrics Match 2 kinds of costs Cost of Design n“Cost of Design”  measure the efficiency of implementing a given DFM optimization at various steps in the design flow e.g. routing, or (expensive) polishing, or (more expensive) OPC/RET Cost of Si n“Cost of Si”  measure the net effectiveness of a given DFM optimization e.g. percentage of redundant vias Cost of Design Cost of Si Parametric IntegrityAX Shape IntegrityBY Thickness IntegrityCZ Functional IntegrityDw Being a high volume shop, these are the key metrics that Qualcomm is very sensitive to These metrics are important to Qualcomm but will not make or break a product

DFM Metrics for Cost of Design ? ne.g. frequency of occurrence of undesirable constructs ?  ‘Bucketed’ in groups and Weighted by : Dimensions (e.g. at Design Rule vs. at DFM Guideline) Phenomenon (e.g. OPC unfriendly jog vs. notch vs. …. ) nOr something like this ? Undesirable construct Dimension weighing factor Process weighing factor Weighted frequency A for x<DFM rule B for x>DFM rule High = HCount * A*H Low = LCount * B*L Total Score Type 1 =  Even un-weighted numbers would be useful as it would allow comparing a design before and after some DFM transformation

DFM Metrics for Cost of Si ? ne.g. f requency of occurrence of given features  ‘Bucketed’ in groups and Weighted by : Phenomenon (e.g. via vs. CA vs. …) Layout configuration (e.g. bordered vs. borderless via…) nOr something like this ? Feature Process weighing factor Weighted frequency Single viaHigh = HCount * H Redundant viaLow = LCount * L Critical areaMed = MCount * M Total Score Type 2 =  Even un-weighted numbers would be useful as it would allow comparing a design before and after some DFM transformation

So, ARE there Economic Benefits in DFM ? nCan’t Really Know – and will not know until : 1.The DFM buzz is clarified Need a clear definition of client-supplier relationships 2.Tool ROI is clarified Need a way of estimating the benefits 3.DFM cost vs. benefit is clarified Need ability to simulate and share benefit vs. time 4.Process data and ownership is clarified Need standard Design Kits nWhat we do know :  We need DFM Score metrics  We need Process Design Kits