1. Dual Graph-Based Hot Spot Detection Andrew B. Kahng 1 Chul-Hong Park 2 Xu Xu 1 (1) Blaze DFM, Inc. (2) ECE, University of California at San Diego

2. University of California, San Diego Outline Introduction of Hot Spot Detection Dual Graph Based Approach Experimental Results Conclusions

3. University of California, San Diego Why Hot Spot Detection? Hot spots = features whose CD variation > T  Form under a variety of conditions  Reduce manufacturing yield  Should be detected and solved in the early stage Commercial tools: ORC (Mentor) and LRC (Synopsys) Hot spot

4. University of California, San Diego Previous Methods Park et al. (SPIE 1999) proposed rule based detection with look-up tables  Number of parameters increase for complex patterns Speed merit of rule-based approach is reduced  Inaccurate Simulation-based approach has been a mainstream  Detect hot spots accurately  Hot spots can be changed according to process conditions  Model generations are significant overhead Key Question Can we detect the hotspots fast and accurately?

5. University of California, San Diego How We Think About Hot Spot Hotspot is a 2-dimensional function of line and space with traditional parameters of DOF and Exposure   Detect too many hot spots to classify the real hot spots Our approach: more topological / graph-oriented  Practical methodology: Filter the chip layout down to a small candidate set of hotspots, which can then be checked using the golden ORC/LRC tool

6. University of California, San Diego (a) (b)(c) Nominal CD Lithography Simulation Different complexity leads to different CD variation CD variation is affected by different process condition More complex pattern, higher probability of hot spot  Probability: Pattern(c) > Pattern(b) > Pattern(a) Simulation Condition: C-1: NA=0.85, σ=0.96/0.76, C-2: NA=0.75, σ=0.75/0.55, C-3: NA=0.75, σ=0.75/045 DOF=0.2um, Exposure=+10% of nominal exposure

7. University of California, San Diego Outline Introduction of Hot Spot Detection Dual Graph Based Approach Experimental Results Conclusions

8. University of California, San Diego Hot Spot Detection Problem Given: Layout L simulation conditions hot spot definition Detect: Hot spots whose CD variation >T To Minimize: Number of un-detected and falsely detected hot spots

9. University of California, San Diego “ Bad ” Patterns Lead to Hotspots Corner effectProximity effect In general, single effect does not lead to hot spots. Hot spots are accumulative effects. 4 proximity effects, 2 corner effects

10. University of California, San Diego Proposed Hot Spot Detection Flow Layout Layout Graph Construction Graph Planarization Three-Level Detection Local Pattern Density Filter Output Hot Spots

11. University of California, San Diego Layout Graph Construction Corner effect Proximity effect Feature  node Two features with corner/proximity effects  edge

12. University of California, San Diego Edge Weighting Scheme Closed-form formula based approach  Weights of corner edges: constant  Weights of proximate edges: f(w 1, w 2, l, d)= (w 1 ’w 2 ’l’) /d Here w 1 ’= w 1 when w 1 <c 0 = c 0 otherwise Lookup table based l w1w1 w2w2 d

13. University of California, San Diego Graph Planarization Delete one edge of any pair of crossing edges Convert the layout graph into its dual graph (face  dual node) PlanarizationDual graph

14. University of California, San Diego Three-Level Hot Spot Detection For each edge  If (its weight > T 0 )  report hot spot For each face (dual node)  If (the total weight > T 1 )  report hot spot Sort all dual nodes according to weights Iteratively merge two dual nodes with max merged weight For each merged face (dual node)  If (the total edge weight > T 2 )  report hot spot EdgeFaceMerged Face

15. University of California, San Diego Local Pattern Density Filter Hot spot Not Hot spot Hot spots depend on the local pattern density A hot spots filtering based local pattern density to reduce falsely detected hot spots

16. University of California, San Diego Outline Introduction of Hot Spot Detection Dual Graph Based Approach Experimental Results Conclusions

17. University of California, San Diego Experimental Setup Testcase: alu128 core  8.7K instances  90nm technology  Chip size is 335 um X 285 um  The netlists from OpenCores. CalibreOPC, CalibreORC from Mentor Graphics are used for model-based OPC, and optical rule check (ORC) Our algorithms are implemented in C++

18. University of California, San Diego An Example of Hotspot Filtering 2D function (width, space) finds too many hotspots to classify the real hotspots Real hotspot can be detected by dual graph based approach with weighted cost function  Detect hotspots which missed by rule-based approach  Result is similar to simulation-based approach (b) Hotspot (a) No Hotspot

19. University of California, San Diego Experimental Results Simulation Condition Number of Hot SpotsRun time(s) ORCDetectedFalse Detected ORCOur DOF=0.1 ET=0.36 17 136901.37 DOF=0.1 ET=0.37 21 226901.52 DOF=0.1 ET=0.38 25 466902.32 DOF=0.2 ET=0.38 152 12916904.38 Total215 137227609.59 Runtime of our method is more than 287X faster compared to ORC Achieves 100% hot spot detection with small falsely defected hot spots overhead

20. University of California, San Diego Outline Introduction of Hot Spot Detection Dual Graph Based Approach Experimental Results Conclusions

21. University of California, San Diego Conclusion A novel fast dual graph based hot spot detection algorithm Our method can detect hot spots with small false detected overhead Runtime improvement is more than 287X compared with ORC Future works  Fast hot spot detection engine in detailed router  Cool spot detection: a pattern that is known to be ORC/LRC-clean through the OPC

22. University of California, San Diego