13.12.2012 Views

Benchmark 5 Example

Benchmark 5 Example

Benchmark 5 Example

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

ICCAD-2012 CAD Contest in<br />

Fuzzy Pattern Matching for<br />

Physical Verification and<br />

<strong>Benchmark</strong> Suite<br />

J. Andres Torres<br />

ICCAD 2012 (November 5-8 th )


2<br />

Outline<br />

� The evolution of physical verification<br />

� State of the art results<br />

� Contest evaluation criteria<br />

ICCAD – November 2012<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


3<br />

Evolution in Physical Verification<br />

Percent of Total PV Activity<br />

100%<br />

90%<br />

80%<br />

70%<br />

60%<br />

50%<br />

40%<br />

30%<br />

20%<br />

10%<br />

0%<br />

ICCAD – November 2012<br />

Migration of Mfg Requirements Into Design<br />

90nm 65nm 45nm 28nm 20nm<br />

Technology Node<br />

Double Patterning (DP)<br />

Litho Simulation (LFD)<br />

Smart Fill<br />

CMP Simulation (CMPA)<br />

Critical Area Analysis (CAA)<br />

Recommended Rules/Scoring<br />

Mfg Required<br />

Mfg Recommended<br />

Traditional DRC<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


4<br />

Requirements on physical verification<br />

� Be aware of anything that can go wrong in the fab and<br />

prevent certain structures to appear in the layout to<br />

maximize product yield.<br />

� Many types of errors come from the FAB: Rapid thermal<br />

anneal, etch and litho to name a few.<br />

� Geometric checks have been the workhorse in physical<br />

verification, and because they are intrinsically range<br />

constraints they are able to cover (and eliminate) a large<br />

number of difficult patterns.<br />

� Models have been used when traditional rules would need<br />

to be overly aggressive and thus open up more valid<br />

topologies that otherwise would need to be constrained.<br />

ICCAD – November 2012<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


5<br />

Runtime Vs accuracy requirements<br />

� Geometric based rules are the fastest but may limit the<br />

design space too aggressively.<br />

� Model based verification is most accurate but it is limited<br />

to model availability and runtime.<br />

� Is there a way to reach a compromise to breach the gap<br />

between the two?<br />

ICCAD – November 2012<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


6<br />

State-of-the-art results<br />

Drmanac et<br />

al, DAC09<br />

Torres et<br />

al, SPIE09<br />

Ding et al,<br />

ICICDT09<br />

MLK – ANN<br />

GD GD+LR<br />

� Desired target performance for a 28nm layout<br />

— Low false alarm: 80%<br />

— Portability: General calibration strategy for different types of<br />

patterns.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com<br />

MLK – SVM<br />

GD GD+LR<br />

Avg Hhit 88% 80% 87% 88% 82% 89% 83%<br />

Avg Nhit 94.523% 99.985% 99.809% 99.847% 99.994% 99.895% 99.998%<br />

False alarm / mm 2 300K 1.1K 13.5K 10K 0.45K 7.5K 0.13K<br />

Avg CPU hour / mm 2 356 30 10 1.5 1.5 2.0 2.0<br />

Avg real hour / mm 2 100 8.50 2.80 0.40 0.40 0.52 0.52<br />

ICCAD – November 2012


7<br />

The importance of accuracy and portability<br />

Left: simulating “Ding et al, ICICDT09” ;<br />

Middle: reported from “Torres et al, SPIE09”;<br />

Right: current state of the art.<br />

� Open issues<br />

› Calibration is tedious and not general enough<br />

› Performance about twice as slow as minimum requirement.<br />

› Still large number of badly classified errors for other testcases<br />

ICCAD – November 2012<br />

1mm<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


8<br />

Calibration test case description<br />

� For IP reasons only clips and areas around the hotspots<br />

and selected non-hotspot locations were selected and<br />

provided for calibrating the process.<br />

ICCAD – November 2012<br />

� The clip input layout as well<br />

as the frames around the<br />

“core” and frame context of<br />

the hotspot were provided.<br />

� The core was generated by a<br />

one ambit box around the<br />

hotspot.<br />

� The frame provided one and<br />

half ambits of context.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


9<br />

Test case description<br />

� Five test cases were provided for various metal layers<br />

which have a high content of 2D structures. Ranging from<br />

32 to 28nm processes.<br />

Testcase Calibration/<br />

Blind size*<br />

[MB]<br />

Technology<br />

Type 1<br />

count**<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com<br />

Type 2<br />

count**<br />

<strong>Benchmark</strong>1 0.1/6.4 Interconnect 32nm 60/133 39/90<br />

<strong>Benchmark</strong>2 3.0/248.5 Interconnect 28nm 173/504 1/4<br />

<strong>Benchmark</strong>3 3.4/216.6 Interconnect 28nm 892/1741 17/22<br />

<strong>Benchmark</strong>4 1.0/74.6 Interconnect 28nm 90/168 5/7<br />

<strong>Benchmark</strong>5 0.5/47.6 Interconnect 28nm 25/40 1/1<br />

*Blind size cases are in OASIS format in flat format<br />

** The count corresponds to the calibration and blind set<br />

ICCAD – November 2012


10<br />

Expected challenges<br />

� Find an efficient way to sample the layout to indentify the<br />

potential locations that cause problems<br />

� Calibration of locations where because of the hotspot<br />

shapes are not small, the underlying calibration pattern is<br />

wider (non square).<br />

� There a hundreds of millions of structures in the layout<br />

and yet the problem is that the amount of “bad” data in<br />

real production is really small. However is significant<br />

enough and the variants in layout so different that the<br />

contest is a test to evaluate the potential of pushing the<br />

state of the art to a region where methods can be useful.<br />

ICCAD – November 2012<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


11<br />

Scoring: Ideal area<br />

False count<br />


12<br />

Scoring: Typical signatures in blind testcases<br />

False count<br />

Exact<br />

Pattern<br />

Matching<br />

Time<br />

ICCAD – November 2012<br />

Hit count<br />

Simulations<br />

� Exact pattern matching is very<br />

fast < 0.1 CPU-HRS/mm 2 with<br />

low false counts but for a<br />

complete blind testcase the hit<br />

count can be zero<br />

� Model based simulation is most<br />

accurate (golden reference<br />

which means hit count = 100%<br />

and false count zero), but with<br />

performance of the order of<br />

100 CPU-HRS/mm 2<br />

� The contest’s challenge is to<br />

move from the exact pattern<br />

matching range towards the<br />

ideal area.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


13<br />

Contest results<br />

ICCAD – November 2012<br />

� Some entries met the<br />

requirement of runtime less<br />

than 1 CPU-hr/mm 2<br />

� None of the contest entries met<br />

the requirement of greater than<br />

80% accuracy.<br />

� Most entries matched the Extra<br />

requirement of < 100 Counts<br />

per mm 2 but at the expense of<br />

heavy accuracy loss.<br />

� For that reason the following<br />

criterion was selected:<br />

— Best average accuracy (highest)<br />

— Followed by best extra (lowest)<br />

— The highlighted region shows the<br />

ideal zone which remains empty<br />

� Second and third place teams<br />

are defined by their average<br />

accuracy<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


Team<br />

14<br />

Average method performance<br />

Runtime<br />

[CPU-hrs/mm 2 ]<br />

Hit Ratio<br />

[Hit/Total]*100%<br />

Extra<br />

[Count/mm 2 ] [Hit/Extra]<br />

21 0.74 29.05 562.95 0.052<br />

40 8.10 21.39 214.20 0.100<br />

31 0.56 19.80 362.30 0.055<br />

4 0.31 15.52 47.11 0.329<br />

17 1.48 10.87 0 N/A<br />

6 0.18 10.65 147.92 0.072<br />

32 25.60 0.41 186.08 0.002<br />

2 0.15 0.09 0.27 0.333<br />

� While team 40 has a better average [Hit/Extra] ratio, Team 21’s average accuracy and<br />

average runtime was better.<br />

� Team 17 decided to go for what appears an exact pattern matching approach. Meaning<br />

that they had difficulties identifying hotspots in the blind sections of the benchmarks.<br />

ICCAD – November 2012<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


15<br />

<strong>Benchmark</strong> 5 <strong>Example</strong>: Team 21 (1 st place)<br />

Section containing<br />

calibration<br />

ICCAD – November 2012<br />

Full blind set<br />

� The primary goal of<br />

the contest was to<br />

investigate the<br />

possibility to predict<br />

patterns outside of<br />

the calibration set.<br />

� Team 21 did the<br />

best job in that<br />

regard.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


16<br />

<strong>Benchmark</strong> 5 <strong>Example</strong>: Team 40 (2 nd place)<br />

Section containing<br />

calibration<br />

ICCAD – November 2012<br />

Full blind set<br />

� Assuming the same<br />

hit rate the<br />

reduction of extras<br />

is of paramount<br />

importance when<br />

using this method in<br />

real production.<br />

� Team 40 provided<br />

the best average<br />

hit/extra ratio of a<br />

blind dataset.<br />

� Unfortunately the<br />

accuracy was low.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


17<br />

<strong>Benchmark</strong> 5 <strong>Example</strong>: Team 31 (3 rd place)<br />

Section containing<br />

calibration<br />

ICCAD – November 2012<br />

Full blind set<br />

� Team 31 was the<br />

next best accuracy<br />

with a ratio<br />

comparable to team<br />

21.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


18<br />

<strong>Benchmark</strong> 5 <strong>Example</strong>: Team 17 (best extra)<br />

Section containing<br />

calibration<br />

ICCAD – November 2012<br />

Full blind set<br />

� Special mention to<br />

team 17 Whose<br />

method provided<br />

zero extras but at<br />

the expense of not<br />

predicting any<br />

configurations in<br />

the full blind areas<br />

of the benchmarks.<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


19<br />

Acknowledgements<br />

� To all the teams for your hard work and dedication. Do<br />

not feel discouraged, this is an important problem looking<br />

for a solution, and I will be more than happy to test your<br />

code again.<br />

� To the Contest organizers for giving me this opportunity,<br />

it was a lot of work, but it was fun and worth doing.<br />

ICCAD – November 2012<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


ICCAD – November 2012<br />

20<br />

THANK YOU<br />

© 2010 Mentor Graphics Corp. Company Confidential<br />

www.mentor.com


3 rd Place Winner<br />

cada031: Extinguisher<br />

Chi-Yuan Liu, Sheng-Yen Chen,<br />

Iou-Jen Liu, and Prof. Yao-Wen Chang<br />

National Taiwan University


2 nd Place Winner<br />

cada040: UTDetector<br />

Bei Yu, Jhih-Rong Gao,<br />

and Prof. David Z. Pan<br />

The University of Texas at Austin


1 st Place Winner<br />

cada021: Iris’ Excalibur<br />

Geng-He Lin and Yen-Ting Yu<br />

National Chiao Tung University

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!