Semiconductor Memory Diagnostics

Semiconductor Memory Semiconductor Memory DiagnosticsDiagnostics

Cheng-Wen Wu (吳誠文)

IC Design Technology Center (DTC)National Tsing Hua University

Hsinchu, Taiwan

meca1.10/cww 2

OutlineOutlineIntroductionFault models and March testsMemory error catch and analysis (MECA) systemDiagnostic algorithm generationDiagnostic data compressionConclusions

meca1.10/cww 3

IntroductionIntroductionMemory testing is more and more important• Memories are key components

Represent about 30% of the semiconductor marketDominate the chip area/yield

Memory testing is more and more difficult• Growing density, capacity, and speed• Emerging new architectures & technologies• Growing need for embedded memories

Why diagnostics?• Yield improvement

Repair and/or design/process debugging

meca1.10/cww 4

Memory Functional Model ExampleMemory Functional Model Example

meca1.10/cww 5

RAM Fault ModelsRAM Fault ModelsAddress-Decoder Fault (AF)Stuck-At Fault (SAF)Transition Fault (TF)Stuck-Open Fault (SOF)Coupling Fault (CF)• State Coupling Fault (CFst) • Inversion Coupling Fault (CFin)• Idempotent Coupling Fault (CFid)

Disturb Fault (DF)Data Retention Fault (DRF)

meca1.10/cww 6

March TestsMarch Tests

March C [Marinescu 1982]• For AF, SAF, TF, & all CFs---redundant

March C- [Goor 1991]

• Also for AF, SAF, TF, & all CFs---irredundant

)}0();0,1();1,0();0();0,1();1,0();0({

rwrwrrwrwrw

cc

c

⇓⇓

⇑⇑

)}0();0,1();1,0();0,1();1,0();0({

rwrwrwrwrw

c

c

⇓⇓

⇑⇑

meca1.10/cww 7

RAMSES Simulation ResultsRAMSES Simulation Results

meca1.10/cww 8

Fault Model SubtypesFault Model Subtypes

meca1.10/cww 9

NTHUNTHU--FTC BIST ArchitectureFTC BIST Architecture

meca1.10/cww 10

Test ModeTest ModeIn Test Mode it runs a fixed algorithm for production test and repair• Only a few pins need to be controlled, and

BGO reports the result (Go/No-Go)

meca1.10/cww 11

Fault Analysis ModeFault Analysis ModeIn Fault Analysis Mode, we can apply a longer March algorithm for diagnosis• FSI captures the error information of the faulty cells

EOP format:

meca1.10/cww 12

Error Catch and AnalysisError Catch and AnalysisLocate the faulty cellsIdentify the fault types

meca1.10/cww 13

How to Identify Fault Type?How to Identify Fault Type?Tester/BIST OutputRAM Circuit/Layout

meca1.10/cww 14

March DictionaryMarch DictionaryMarch 11N

E0 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10

meca1.10/cww 15

March Signature and Error MapMarch Signature and Error MapMarch Signature (Syndrome)

Error Map

meca1.10/cww 16

MECA SystemMECA System

meca1.10/cww 17

Error AnalyzerError Analyzer

Data log parser

Tester/BIST data log

Fault analysis

Error maps

Fault maps

March Dictionary

meca1.10/cww 18

Fault AnalysisFault AnalysisDerive analysis equations from the fault dictionary

• Convert error maps to fault maps by the equations

meca1.10/cww 19

Test Algorithm GenerationTest Algorithm Generation• Start from a base test: generated by TAGS or

user-specified• Generation options reduced to read-insertions

meca1.10/cww 20

Diagnostic ResolutionDiagnostic ResolutionDiagnostic resolution

faults detectable of#faults habledistinguis of # resolution Diagnostic =

meca1.10/cww 21

Experimental ResultsExperimental ResultsProposed diagnosis framework has been applied to commercial embedded SRAMsResults for a 16Kx8 embedded SRAM (FS80A020) are shownTester log from Credence SC212 is examinedAddress remapping (logical to physical) is applied

meca1.10/cww 22

The Total Error BitmapThe Total Error Bitmap

meca1.10/cww 23

Fault BitmapsFault BitmapsIdempotent Coupling Fault Stuck-at 0

meca1.10/cww 24

Diagnostic Data CompressionDiagnostic Data CompressionDiagnostic data usually are exported serially to the tester• Excessive tester time• Increased tester storage requirementCompression can solve this problemWe propose• Compression by syndrome accumulation• Tree-based Hamming syndrome compression

meca1.10/cww 25

Syndrome AccumulationSyndrome AccumulationConventional approach for exporting diagnostic data:• Each time a fault is detected, the BIST circuit

exports the faulty-cell address and March syndrome

There is redundancy since a fault can be detected by many Read operations during the test process

meca1.10/cww 26

Syndrome Accumulation ExampleSyndrome Accumulation Example

000 001 010 011 100 101 110 111

000

001

010

011

100

101

110

111

Column address

Row

address

Fault Diagnosis data

SAF(1)

CFst(H,0,0)

SAF(0)

CFin(L, , )

D1=(100010, 100011100011)SAF(1)

CFin(L, , )

SAF(0)

CFst(H,0,0)

D1=(101110, )

D1=(100010, )

D1=(100010, )

100000001100

011100011100

011100000100

meca1.10/cww 27

Syndrome Accumulation Circuit (SAC)Syndrome Accumulation Circuit (SAC)

Faulty-cell Address March Syndrome

CAM

Data Register

Mask Register

AddressingM

echanism

01

01

Vdd

S

01

01

Vdd

S

01

01

Vdd

S

Up/dow

n shift register

Mk-1

Wk-1

M1

W1

M0

W0

HitClkRst SData Register

W0

M0W1

M1

Wk-1

Mk-1

meca1.10/cww 28

Compression RatioCompression RatioWe define the compression ratio as

R = compressed bits/original bits

Comparison:RAM sizeMarch CdModified March CMarch-26NMarch-12NMarch-17N

256K36.84%26.19%30.30%45.71%28.89%

512K36.52%25.79%29.95%45.45%28.55%

1M36.23%25.42%29.62%45.21%28.24%

meca1.10/cww 29

Hardware OverheadHardware OverheadWe use the register-bit equivalent (rbe) modelHardware overhead (HO) vs. largest number of faults (K)

meca1.10/cww 30

Compression for Word-Oriented RAMDiagnostic data for word-oriented RAM

Faulty-cell addressMarch syndromeHamming syndrome

Hamming syndromeThe modulo-2 sum of the expected fault-

free data output vector and the output vector from the memory under test

meca1.10/cww 31

MotivationEmbedded RAM usually has long words

Diagnostic data dominated by Hamming syndrome

8kx256-bit RAM & March Cd (72 Read operations, extended for word-oriented RAM with 9 data backgrounds):

13-bit faulty address7-bit March syndrome256-bit Hamming syndrome

About 0.93 of the diagnostic data

meca1.10/cww 32

Basic IdeaA source message is regarded as a sequence of symbols with fixed sizeFor example, (00101001001001000100) can be represented by {0010 1001 0010 0100 0100}

Only three unique symbols {0010 1001 0100} appear in this exampleWe thus can compress the message by removing the redundancy

meca1.10/cww 33

Hamming Syndrome CompressionTree-basedCodeword

10

0

0

0 1

1

11 0

0

124

83 5 6

7 9 A B CD E F

Symbol Hexadecimal

0123456789ABCDEF

101101000000011001000001010000011000000111000100001001000010100000101100001100000011010000111000001111

0000000100100011010001010110011110001001101010111100110111101111

meca1.10/cww 34

EvaluationA 128Kb RAM with 2kx64, 1kx128, and 512x256 configurations are consideredThree different symbol sizes (B=4, 8, and 16) are used to estimate R

meca1.10/cww 35

Experimental ResultsR for 64-bit HS R for 128-bit HS R for 256-bit HSRF

100150200250500100150200250500100150200250500

B=4 B=8 B=16 B=4 B=8 B=16 B=4 B=8 B=16

26.8% 15.1% 9.6%26.9% 15.2% 9.7%26.9% 15.2% 9.9%27.0% 15.3% 10.0%27.2% 15.7% 10.7%

25.9% 13.8% 8.0%26.0% 13.9% 8.1%26.1% 14.0% 8.2%26.1% 14.1% 8.2%26.4% 14.5% 9.0%

Without Other faults

With two Faulty rows

With 10% cluster faults

28.6% 17.5% 12.9%28.7% 17.6% 13.1%28.7% 17.7% 13.2%28.8% 17.7% 13.3%29.0% 18.2% 14.1%32.3% 20.4% 15.3%31.1% 19.5% 14.5%30.5% 19.4% 14.2%30.1% 18.9% 14.3%29.8% 18.8% 14.6%

31.7% 22.4% 20.5%31.9% 22.3% 20.7%31.9% 22.4% 21.0%31.9% 22.5% 21.2%32.4% 23.2% 22.3%

29.8% 17.1% 10.1%28.5% 16.1% 10.0%28.1% 15.7% 9.7%27.8% 15.5% 9.7%27.5% 15.4% 9.7%

30.6% 18.1% 12.3%29.3% 17.1% 11.6%28.8% 16.9% 11.3%28.6% 16.5% 11.2%28.1% 16.3% 11.2%

29.0% 18.0% 14.2%29.1% 18.1% 14.4%29.1% 18.2% 14.5%29.3% 18.4% 14.8%29.6% 19.0% 15.8%

27.5% 15.9% 10.9%27.7% 16.0% 11.1%27.7% 16.1% 11.3%27.8% 16.2% 11.5%28.3% 17.0% 12.7%

meca1.10/cww 36

ConclusionsConclusionsOur diagnosis framework provides error catch and analysis• For design/process verification and debugging• For yield improvement

Our syndrome accumulation method compresses faulty-cell address and March syndrome to about 28% of the original size• Hardware overhead is about 0.9% for a 1Mb SRAM

with 164 faultsOur tree-based method compresses the Hamming syndromes to about 10% for a 128Kb RAM with a 16-bit symbol size and 100-500 random faults