280 likes | 366 Views
Rules vs. Tools. Lou Scheffer. Lars Liebmann, , Riko Radojcic , David White ISPD Austin, March 2006. What’s the problem?. What are designers allowed to design? Design rules are the traditional solution But modern processes are very complex
E N D
Rules vs. Tools Lou Scheffer. Lars Liebmann, , Riko Radojcic, David White ISPD Austin, March 2006
What’s the problem? • What are designers allowed to design? • Design rules are the traditional solution • But modern processes are very complex • Rule books become huge (>600 pages) • Too much to easily remember and use • Still do not tell enough • Two possible solutions • Restricted rules - ‘Rules’ • Model based design – ‘Tools’
What’s wrong with the current system • In the beginning, there was a rule for contact spacing…
What’s wrong with the current system • Then there were different same-net rules • This was pretty easy to understand
What’s wrong with the current system • Then the spacing started to depend on the number of neighbors > 2 neighbors 0, 1, or 2 Neighbors
What’s wrong with the current system • Now need a rule on what constitutes a neighbor These are officially neighbors
What’s wrong with the current system • So a 3x3 array uniformly needs the ‘big’ rule
What’s wrong with the current system • And some clever designer notices this Now you can use the 2 neighbor rule!
What’s wrong with the current system • But the fab does not think this should be allowed No, you cannot use the 2 neighbor rule!
What’s wrong with the current system • So now there is a rule stating ‘It is not legal to remove the middle via of a 3x3 array in order to use the 2 neighbor rule’ • Hard to check in DRC • Hard to remember • Is this a good use of everyone’s time? • Does it really make a yield difference?
So what could you do? • ‘Rules’ – here is a list of the legal via arrays. You must use one of these. • ‘Tools’ – here is a test that tells which via arrays are manufacturable. Any via array this passes this test is OK. • ‘Extreme tools’ – here is a model that predicts via fail rate in manufacturing. You decide.
Can do this at many levels • Layout • Example : Restricted design rules • Circuit • Example: No dynamic logic • Architecture • Example: Must have latches on block boundaries
Problem not unique to IC design • 3 cases from other fields • One where both approaches are used, with extreme size differences • One where rules are used • One where models are used
Law vs ethics • What is allowed behavior? • Rules based – legal systems • Tools based – honor codes • Specify by laws (info from Wikipedia, USA federal only) • Divided into 50 ‘titles’; each 1 or more printed volumes • Titles may optionally be divided into subtitles, parts, subparts, chapters, and subchapters. • For example, privacy act of 1974 is “Title five, United States Code, section five hundred fifty-two A."
Law vs Ethics • The basic intent of all these laws can be replaced by an ‘honor code’, such as • A cadet will not lie, cheat, steal, or tolerate those who do • Much, much shorter (5 orders of magnitude?) but requires interpretation
Example where rules are the accepted way • Civil engineering – building codes • Detailed code not needed in principle • Quite conservative (6x safety factors) • Reduces the chance of error • Designed to be easy to inspect • Reduces the consequence of error • Big safety factor • Makes sure common failures are addressed • Multiple exits, exit lighting, width of doorways, etc.
Example where models are the rule • Food – both rules and info are possible: • Rule: Eat your vegetables before dessert • Tool: each food lists calories and ingredients; you decide what to eat. • Different people have different tradeoffs • Disclose, then let people decide for themselves
Benefits from restricted rules? • Problem: Even a simple wire is hard to predict • Lithography is a complex operation • Designer draws a polygon • OPC is applied (very non-linear) • Exposure through a complex optical system • Develops a non-linear resist, then etch • Need to add CMP to model vertical dimension • Then need to compute range of variation • Can all be modelled, but it’s not easy
Restricted design rules: Do this • One direction, one width, one spacing, all wires on grid, all empty spaces filled
Tried this as an experiment • Forbid wrong way in router • Wire length only 1% greater • Routability unchanged • 10% more vias • But if this lets you print smaller, it’s a (big) net win!
Implications for Designer For designer/CAD, looks like lots of restrictions • Power routes must be wires in parallel • Wide signal nets must be parallel wires • Delay computation must handle loops • Routers cannot jog – more vias • IP blocks cannot be rotated • Litho folks often phrase this as ‘This is needed, or you won’t get 45 (or 32 or 22 nm…) • Designers don’t believe this for historical reasons • Looks like the fab is saying “Eat your vegetables”
Implications for Fab • Litho is very easy; can be tuned for direction • Illumination andpolarization can be optimized • CMP is very easy • No width effects, small density variations • Can help with non-modelled effects • These are nice for the fab, but designer does not care • Can produce a smaller pitch • Variation can be (much) less • These are things designers value
Political implications • Fab needs to ‘share’ the benefits, e.g. • 90 nm pitch with restricted rules • Smaller variation • Or • 100 nm pitch with arbitrary geometries but larger variation • Then designers could see the benefit, and decide for themselves…
And now our panelists • Each with a different perspective • Lar Liebmann • Litho guy • David White • CMP modeller • Riko Radojcic • User perspective
Questions to keep in mind • Can the average engineer use complex models productively? • Can the EDA tools use these models? • Can the fabs provide these models, keep them updated, and stand behind them? • Won't this keep fabs from making improvements to their process?
More questions • What's the relative importance of time to market and getting the most out of the silicon? • How big is the penalty, or benefit, for using restricted rules? • Of the many problems (litho, CMP, lifetime degradation, etc.) which should be treated by rules and which by models? • Should this be design dependent?