Upload
race
View
35
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Net Diagnosis using Stuck-at and Transition Fault Models Master’s Defense Lixing Zhao. Thesis Advisor: Dr. Vishwani D. Agrawal Thesis Committee: Dr. Adit Singh and Dr. Victor P. Nelson. Department of Electrical and Computer Engineering Auburn University, AL 36849 USA. Outline. Motivation - PowerPoint PPT Presentation
Citation preview
Oct. 26, 2011 Lixing's MS Defense 1
Net Diagnosis using Stuck-at and Transition Fault Models
Master’s DefenseMaster’s DefenseLixing ZhaoLixing Zhao
Department of Electrical and Computer EngineeringAuburn University, AL 36849 USA
Thesis Advisor: Dr. Vishwani D. AgrawalThesis Committee: Dr. Adit Singh and Dr. Victor P. Nelson
Oct. 26, 2011 2
Outline
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
Lixing’s MS DefenseLixing's MS Defense
Oct. 26, 2011 Lixing's MS Defense 3
Motivation
Due to high logic density of modern VLSI design and manufacturing, chips on the first round of tape-out often suffer a relatively low yield that may be unacceptable.
Fault diagnosis can bring the yield up in later manufacturing rounds by indentifying the possible causes of defect in earlier tape-outs.
Net fault diagnosis is an important area of fault diagnosis. Because of the large routing area of modern VLSI devices, the routing interconnection nets are more vulnerable to certain defects.
In this work, we try to provide an effective method on solving the net-diagnosis problem.
Oct. 26, 2011 Lixing's MS Defense 4
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
Outline
Oct. 26, 2011 Lixing's MS Defense 5
How does Fault Diagnosis Work?Defective Circuit
Circuit Net-listTest Vectors
Circuit Responses
Expected Responses
Compare
Diagnosis AlgorithmPossible Defect Locations
Oct. 26, 2011 Lixing's MS Defense 6
Circuit Under Diagnosis The Circuit Under Diagnosis (CUD) can be classified
into two groups: Combinational Circuits:
Sequential Circuits:
Oct. 26, 2011 Lixing's MS Defense 7
Diagnosis Pattern Random Pattern:
Randomly or Pseudo Randomly Generated Computer program or Pattern generator (e.g., LFSR)
N-Detect Pattern: Each fault is detected by at least N different Patterns ATPG-based
Fault-Model-Based Patterns: Patterns used for diagnosing faults based on certain fault model Using certain generating algorithm and ATPG
• Yu Zhang and Vishwani D. Agrawal, "A Dianostic Test Generation System," in Proc. International Test Conf., Nov.2010, pp.1-9.• Yu Zhang and Vishwani D. Agrawal, "Reduced Complexity Test Generation Algorithms for Transition Fault Diagnosis",
International Conf. on Computer Design, Oct. 2011. pp. 96-101.
Oct. 26, 2011 Lixing's MS Defense 8
Oct. 26, 2011 Lixing's MS Defense 9
Fault ModelsFault model is an abstraction of the real defect in chip and different fault models are used to handle different types of defects in fault diagnosis.
The types of fault model that can be used on a net:
Stuck-at fault
Transition fault
Bridging fault
Open fault
Open Fault
Interconnect Open: undesired breaks and electrical discontinuities on interconnection line.
Resistive Open: narrow crack Modeled as transition delay fault
Oct. 26, 2011 Lixing's MS Defense 10
Oct. 26, 2011 Lixing's MS Defense 11
Complete Open: wide crackThe coupling capacitors between the floating node and the supply and ground. The drain voltage of the driven gates.The effects from surrounding lines.
Net Structure
Oct. 26, 2011 Lixing's MS Defense 12
A net means a connection metal wire in the circuit.
Previous Works on Multiple Faults Diagnosis
Single-Fault-Simulation(TFS)-BasedMultiple-Faults-Simulation(MFS)-BasedSingle-Location-AT-a-Time(SLAT)-BasedRegion-Model-Based
Oct. 26, 2011 Lixing's MS Defense 13
SLAT
Oct. 26, 2011 Lixing's MS Defense 14
X-Region
Oct. 26, 2011 Lixing's MS Defense 15
Diagnosis Strategies Cause-Effect method
diagnosis faults by comparing the fault simulation results with the CUD response.
Traditional SimulationMore information available but more costs.
Effect-Cause methoddiagnosis faults by tracing from erroneous primary
outputs.Back-trace simulationLost some information but fewer costs.
Oct. 26, 2011 Lixing's MS Defense 16
OUTLINE
Oct. 26, 2011 Lixing's MS Defense 17
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
Oct. 26, 2011 Lixing's MS Defense 18
Problem StatementGiven the failing response of CUD
Failing Pattern IndexIndex of Erroneous Primary Outputs (EPO)
Given the net-list of CUD Verilog file
Find out locations of faulty nets with certain defects.
Oct. 26, 2011 Lixing's MS Defense 19
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
OUTLINE
Oct. 26, 2011 Lixing's MS Defense 20
Proposed Fault Filtering System
Count Assignment:A Count is a value we assign to each fault
candidate under certain measurement method.Contribution: A more balanced count assignment
method for fault candidates filtering.The Count we use in our filtering system is a ratio-
count.Count = Matched Value
Expected Value in Fault Simulation
Oct. 26, 2011 Lixing's MS Defense 21
Oct. 26, 2011 Lixing's MS Defense 22
Oct. 26, 2011 Lixing's MS Defense 23
Failing Pattern Index Matching DEF1:The union of all the failing pattern index from the single
fault simulation of a fault candidate is defined as the Detectable Pattern Set (DPS) of this fault under the test.
DEF2:The union of all the failing pattern index of the observed CUD response is defined as the Failing Pattern Set (FPS) of the test.
DEF3:The shared part between the DPS of a fault candidate and the FPS of CUD is called the Shared Pattern Set (SPS) of this candidate.
• THEOREM: If the CUD is a circuit with multiple faults and we assume that all the multiple faults in the circuit will not totally cancel each other on the primary outputs, then the DPS of any one of the multiple faults in the circuit should be a subset of FPS of the test, in other words, the SPS of the fault candidate equals to its DPS.
• INTUITIVE ASSUPTION: The percentage of SPS takes in DPS of a fault represents the possibility that this fault be a real one in CUD.
• Count =
Oct. 26, 2011 Lixing's MS Defense 24
Number of patterns in SPSNumber of patterns in DPS
Oct. 26, 2011 Lixing's MS Defense 25
Oct. 26, 2011 Lixing's MS Defense 26
FPIM with EPO-HittingDEF: Under the same test pattern, if the
affected primary outputs of a candidate fault simulation shares at least one erroneous output with the faulty response of CUD, then we say that this fault candidate can 'hit' the EPO under this pattern and this pattern is called a Hit-Pattern of this candidate.
Count = Number of Hit-Patterns in SPSNumber of patterns in DPS
EPO-Matching DEF1: A Pattern-EPO-Pair (PEP) is a pair of failing pattern number
and an EPO associated with it. Like [P2, PO1], which indicates under pattern P2, an error is observed on PO1. PEP could be used to either represents the faulty response of CUD testing or the fault simulation results of fault candidate.
DEF2: The union of all the PEPs from CUD testing is called PEP-Set-of-CUD (PEPSC) and the union of the PEPs under all the patterns in SPS of a fault candidate is called the PEP-Set-of-Fault (PEPSF). The shared part of PEPSC of a fault candidate with PEPSF of CUD is called Shared-PEP-Set (SPEPS) of this fault candidate.
Oct. 26, 2011 Lixing's MS Defense 27
THEOREM: If CUD is a circuit with multiple faults and we assume that there is no cancelling effect among these faults, then the PEPSF of a fault candidate in single fault simulation should be the subset of of CUD and the SPEPS of the fault candidate should equal to its PEPSF.
INTUITIVE ASSUPTION: The percentage SPEPS taking in PEPSF of a fault candidate indicates the possibility that this fault be a real one in CUD.
Count =
Oct. 26, 2011 Lixing's MS Defense 28
Number of PEPs in SPEPSNumber of PEPs in PEPSF
Oct. 26, 2011 Lixing's MS Defense 29
Count of Fault1: 3/3=1 Count of Fault2: 3/4 = 0.75
• For each step, we will set one threshold value to filter the unrelated fault candidates out. These threshold values depend on the our assumption on fault density of CUD.
Oct. 26, 2011 Lixing's MS Defense 30
Oct. 26, 2011 Lixing's MS Defense 31
Filtering Results
Circuit # of Total Faults Reduction Rate Survival Rate
C432 524 0.77 0.975
C880 942 0.97 0.96
C1355 1574 0.75 0.97
C1908 1879 0.85 0.95
C2670 2747 0.925 0.97
C3540 3428 0.95 0.965
C6288 7744 0.933 0.96
C7552 7419 0.992 0.96
Oct. 26, 2011 Lixing's MS Defense 32
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
OUTLINE
Oct. 26, 2011 Lixing's MS Defense 33
Candidate Ranking SystemAfter getting a smaller list of fault candidates from
filtering stage, we need to rank the fault candidates so that we can have a better diagnosis resolution.
A structure called EPO-Tree is used in our work.
34Oct. 26, 2011 Lixing's MS Defense
Oct. 26, 2011 Lixing's MS Defense 35
Branch Ranking in EPO-Trees with Same Branch Combination
Observation: The activation situations are sometimes similar under certain test patterns, which means these patterns can activate same set of injected faults in the CUD and the observed EPO combinations from the CUD are the same.
Intuitive Assumption: Assuming we have a circuit with large enough number of primary outputs, when the failing outputs combinations are the same under different test patterns, because it is not very easy to repeat the same combinations for different injected faults in CUD, it is possible that the cause of these failures are the same. If we can find shared set of leaves between the corresponding branches in these EPO-Trees, then these shared faults are more possible than other faults in branch to be the real faults.
Oct. 26, 2011 Lixing's MS Defense 36
Branch Ranking with Counts from Reduction Stage
Branch ranking procedure we used in step two, three and five.
The candidates in each branch have already had an initial rank from previous stage, now what we have to do is to utilize the counts of each fault got from reduction stage to rank the candidates within each group.
Oct. 26, 2011 Lixing's MS Defense 37
Branch Ranking with Leaves Count in Each EPO-Tree
Rule: If several fault candidates still have the same rank after previous ranking steps, then we assume that the ones with more leaves in the EPO-Tree have more chance to be real faults in CUD, because it is much easier to activate just one or two faults than many faults together to cause the same effects.
Oct. 26, 2011 Lixing's MS Defense 38
Final Fault Ranking• We rank the fault candidates by considering
the best rank they have among all the branches in all EPO-Trees.
• A Top-Single-Fault is a single fault that has top rank in a branch. This kind of faults are the most suspicious fault candidates to us in diagnosis.
• Because we have applied many constraints in branch ranking, the earlier the TSF comes out, the more suspicious it seemed to us.
Oct. 26, 2011 Lixing's MS Defense 39
40Lixing's MS Defense
Oct. 26, 2011 Lixing's MS Defense 41
Experimental Results
Oct. 26, 2011 Lixing's MS Defense 42
Cir Our Work(1 Stuck-at) Wang’s work(1 stuck-at)Dia FHR RES T(s) Dia FHR Res T(s)
c2670 1.0 1.0 1.0 6.4 1.0 1.27 1.3 0.01
c3540 1.0 1.0 1.0 0.5 1.0 1.2 1.5 0.01
c6288 1.0 1.0 1.0 0.6 1.0 1.1 1.3 0.01
c7552 1.0 1.0 1.0 1.5 1.0 1.15 1.6 0.01
Z. Wang, M. Marek-Sadowska, and J. Rajski, "Analysis and methodology for multiple-fault diagnosis,” IEEE Tran on CAD of Integrated Circuits and Systems, March 2006, vol. 25, pp. 558-576.
Cir Our Work(2 Stuck-at) Wang’s work(2 stuck-at)Dia FHR RES T(s) Dia FHR Res T(s)
c2670 0.97 1.0 2.0 8 0.97 1.35 2.0 0.05
c3540 0.95 1.0 1.0 0.6 0.95 1.2 1.7 0.06
c6288 0.99 1.0 2.0 1 0.97 1.28 2.0 0.2
c7552 0.93 1.0 2.0 1.7 0.925 1.25 2.0 0.2
Oct. 26, 2011 Lixing's MS Defense 43
Cir Our Work(3 Stuck-at) Wang’s work(3 stuck-at)Dia FHR RES T(s) Dia FHR Res T(s)
c2670 0.94 1.0 2.0 13 0.925 1.35 2.6 0.1
c3540 0.94 1.0 2.0 0.7 0.92 1.15 2.4 0.1
c6288 0.95 1.0 2.0 2.7 0.93 1.15 2.5 0.5
c7552 0.90 1.04 2.0 2.9 0.92 1.2 2.3 0.25
Cir Our Work(4 stuck-at) Wang’s work(4 stuck-at)Dia FHR RES T(s) Dia FHR Res T(s)
c2670 0.89 1.06 2.0 17 0.92 1.3 2.6 0.2
c3540 0.91 1.0 2.0 1.8 0.89 1.25 2.5 0.2
c6288 0.92 1.02 2.0 5.7 0.82 1.15 2.8 0.8
c7552 0.88 1.1 2.0 3.2 0.91 1.2 2.4 0.5
Fault List Extension• Before we start handling the net diagnosis work,
we need to first extend the collapsed faults to uncollapsed faults.
Oct. 26, 2011 Lixing's MS Defense 44
From the net-list of the circuit, we can get the corresponding net for each fault. Then each group of equivalent faults can be transformed into a set of nets.
Oct. 26, 2011 Lixing's MS Defense 45
Oct. 26, 2011 Lixing's MS Defense 46
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
OUTLINE
Net RankingFirst, we build a net pool, which will include
all the net candidates of each rank group.Final net candidate list includes two parts:
The nets which we can find more than two members in the net pool.
The nets which can only be found once in the net pool.
Nets’ ranking are based on the group's rank they derived from.
Oct. 26, 2011 Lixing's MS Defense 47
Oct. 26, 2011 Lixing's MS Defense 48
Experimental Results
Oct. 26, 2011 Lixing's MS Defense 49
For each fault model and each benchmark circuit, we randomly constructed 20-50 faulty circuits. For each injected net fault, we randomly selected 2-4 fault sites which could be either the stem or the branches of the net. For stuck-at model, we injected one single stuck-at fault on each fault site and for transition net fault we injected a D-flip-flop on each fault site to perform transition delay behavior.
Oct. 26, 2011 Lixing's MS Defense 50
Cir One Net Fault(Stuck-at) Two Net Fault(Stuck-at)
Dia FHR Res T Dia FHR Res T
c432 1.0 1.1 2.0 0.1 0.98 1.2 3.0 0.2
c880 1.0 1.1 2.0 0.1 1.0 1.1 3.0 0.1
c1355 0.98 2.0 3.0 1.0 0.75 2.6 5.0 16
c1908 0.92 1.4 3.0 7.5 0.86 1.3 3.0 13
c2670 0.98 1.4 3.0 7.4 0.95 1.14 4.0 12
c3540 0.98 1.04 3.0 0.5 0.96 1.1 3.0 0.7
c6288 0.98 1.1 3.0 3 0.9 1.06 3.0 4
c7552 0.96 1.1 3.0 3 0.9 1.2 3.0 11
Oct. 26, 2011 Lixing's MS Defense 51
Cir One Net Fault(Transition) Two Net Fault(Transition)
Dia FHR Res T Dia FHR Res T
c432 1.0 1.05 2.0 0.48 0.9 1.2 3.0 0.7
c880 0.9 1.2 2.0 0.8 0.8 1.3 3.0 2.3
c1355 1.0 1.0 1.0 0.9 0.86 1.3 3.0 13
c1908 0.96 1.2 2.0 7.2 0.93 1.4 3.0 35
c2670 1.0 1.16 2.0 0.9 0.97 1.06 2.0 4.6
c3540 0.98 1.2 2.0 1.7 0.95 1.1 2.0 2.3
c6288 1.0 1.08 2.0 2.8 0.92 1.3 2.0 2.8
c7552 0.98 1.3 2.0 7.5 0.96 1.08 2.0 13.1
Oct. 26, 2011 Lixing's MS Defense 52
MotivationBackgroundProblem StatementContributions
Proposed Fault Filtering SystemProposed Fault Ranking SystemProposed Net Ranking System
Conclusion
OUTLINE
Advantages
Oct. 26, 2011 Lixing's MS Defense 53
Utilizing only single fault simulation, which avoids the problem of exponential searching space in multiple-fault-simulation-based works.
No requirement for the ability of testing pattern to trigger single fault at a time.
A balanced candidate filtering system which can effectively reduce the number of fault candidates and to some extent tolerate the non-stuck-at behavior caused by other types of faults.
A candidate ranking system with high First Hit Rank(FHR), diagnosis resolution and diagnosability.
Suitable for diagnosing multiple types of faults.
Oct. 26, 2011 Lixing's MS Defense 54
Conclusion Traditional gate faults are closely related to the fault models and not
necessarily to physical defects. Therefore, from a practical viewpoint it makes sense to diagnose a faulty net on a VLSI chip than to locate a `modeled' fault.
Our use of stuck-at and transition faults models is for a practical reason, i.e., availability of tools for test generation and fault simulation. These models are used only for the possibility of analysis they offer. In identifying faulty nets no assumption is made about the actual fault on them except that those nets `may' have caused the observed and simulated errors. The fault models may, or may not be, used as suggestions. We verified our work with the injected multiple-faults.
In the future, arbitrary defects such as bridges, opens, short, etc., should be examined to evaluate the presented diagnosis algorithms.
Oct. 26, 2011 Lixing's MS Defense 55
References1. Yu Zhang and Vishwani D. Agrawal, “A Dianostic Test Generation System,” In Proc.
International Test Conf., 2010, pp.1-9.2. Yu Zhang and Vishwani D. Agrawal, “Reduced Complexity Test Generation Algorithms
for Transition Fault Diagnosis,” In Proc. International Conf. on Computer Design, 2011, pp. 96-101.
3. N. Sridhar and M.S. Hsiao, “On Efficient Error Diagnosis of Digital Circuits,” In Proc. International Test Conference, 2001, pp. 678 - 687.
4. S.M. Reddy, H. Tang, I. Pomeranz, S. Kajihara and K. Kinoshita, “On Testing of Interconnect Open Defects in Combinational Logic Circuits with Stems of Large Fanout,” In Proc. Intl Test Conf. , 2002, pp. 83-87.
5. Z. Wang, M.Sadowska, and J. Rajski, “Analysis and Methodology for Multiple-fault Diagnosis,” IEEE Tran on CAD of Integrated Circuits and Systems, March 2006, vol. 25, pp. 558-576.
6. S. Venkataraman and S. B. Drummonds, “Poirot: Applications of a Logic Fault Diagnosis Tool,” IEEE Design and Test of Computers, Jan. 2001, pp. 19-29.
7. J. Segura and C. F. Hawkins, “CMOS Electronics: how it works, how it fails,” Wiley- IEEE, Apr. 2004.
Oct. 26, 2011 Lixing's MS Defense 56
Thank You . . .
Oct. 26, 2011 Lixing's MS Defense 57
Questions