Title: Test Technology Trends
1Chapter 12
Test Technology Trends in Nanometer Age
2What is this Chapter About?
- Introduce the test technology roadmap
- Focus on a number of difficult challenges and
test solutions - Delay testing, Physical failures and Soft errors,
FPGA testing, MEMS testing, High-speed I/O
testing, and RF testing - Concluding remarks
3Section 12.1
Test Technology Roadmap
4Moores Law and Test Challenges
- Moores law the number of transistors integrated
per square inch will double approximately every
18 months. - To keep track of Moores law die size , feature
size , gate delay , interconnect delay - To reduce interconnect delay, interconnects are
made taller and taller, and this causes crosstalk
noises between adjacent lines due to capacitive
and inductive coupling (called signal integrity
problem). This is very difficult to test.
5Moores Law and Test Challenges
- Power integrity clock frequency , supply
voltage , power supply voltage can drop by
L(di/dt). This is very difficult to test. - Process variation precise control of silicon
process is becoming more and more difficult. For
example, it is hard to control effective channel
length of a transistor. This makes power and
delay exhibit large variability. This is hard to
detect. - Low-power design faults low-power design
circuits might result in fault models that are
difficult to test. For example, drowsy cache
design by reducing power supply will cause drowsy
faults.
6Fabrication Capital versus Test Capital
7International Technology Roadmap for
Semiconductors (ITRS)
- ITRS identifies technological challenges and
needs facing the semiconductor industry over the
next 15 years - ITRS test and test equipment near-term challenges
SIA 2004 - High-speed device interfaces
- Highly integrated designs
- Reliability screens
- Manufacturing test costs
- Modeling and simulation
8International Technology Roadmap for
Semiconductors (ITRS)
- ITRS test and test equipment long-term challenges
SIA 2004 - DUT (device under test) and ATE (automatic test
equipment) interfaces - Test methodologies
- Defect analysis
- Failure analysis
- Disruptive device technologies
9International Technology Roadmap for
Semiconductors (ITRS)
- ITRS design test near-term challenges SIA 2004
- Effective speed test with increasing core
frequencies and widespread proliferation of
multi-GHz serial I/O protocols - Capacity gap between DFT/test generation/fault
grading tools/design complexity - Quality and yield impact due to test process
diagnostic limitations - Signal integrity testability and new fault models
- SOC and SIP test
- ITRS design test long-term challenges SIA 2004
- Integrated self-testing for heterogeneous SOCs
and SIPs - Diagnosis, reliability screens and yield
improvement - Fault tolerance and on-line testing
10Sections 12.2
Delay Testing
11Why Delay Testing?
- Three sources of yield loss
- Random defects causing both logical and timing
failures - Systematic failures causing both logical and
timing failures - Parametric variations more likely causing
timing failures
Yrandom
Ysystematic
Yparamtric
12Fault Models Path-Delay Gate-Delay Faults
- Path-delay fault
- Propagation delay of path exceeds clock interval
- of paths grows exponentially with number of
gates - Only consider long paths or a subset of paths
- Tests can detect small distributed failures
- Tests for longest paths also useful for
speed-sorting - Gate-delay fault
- A logic model for a defect that delays a rising
or a falling transition - Small distributed timing failures could be missed
- of modeled faults is much smaller and manageable
13Transition Faults Small Gate-Delay Faults
- Transition fault (Gross gate-delay fault)
- The extra delay caused by the fault is assumed to
be large enough to prevent the transition from
reaching any PO at the time of observation - Can be tested along any path from fault site to
any PO - The test is a vector pair that creates a
transition at the fault site and the second
vector is a test for the stuck-at fault at the
fault site - Small gate-delay fault
- Is tested along the longest propagation delay path
14Path Delay Faults - Type of Tests
- Single-path-sensitization test
- Guarantee that DUT will fail if and only if the
path under test has excessive delay - Fully characterize the timing of the path and
is ideal for delay fault diagnosis - All side inputs of gates along the given path
must be stable
Single-path-sensitization test conditions (for
AND gate)
Must be stable 1
Must be stable 1
V1 V2
V1 V2
Target path
Target path
15Path Delay Faults - Type of Tests (Contd)
- Non-robust test
- Test may be invalidated in presence of other
path delay faults
Non-robust test conditions (for AND gate)
Could be either 1 or 0
Could be either 1 or 0
V1 V2
V1 V2
Target path
Target path
16Path Delay Faults - Type of Tests (Contd)
- Robust test
- Guarantees DUT will fail if the path under test
has excessive delay
Robust test conditions (for AND gate)
Could be either 1 or 0
Must be stable 1
V1 V2
V1 V2
Target path
Target path
17Application of Delay Tests
- Require application of a vector pair to the
combinational logic portion and the circuit being
clocked at speed
input latches
output latches
Combinational Circuit
Input clock
output clock
Rated clock period
Input clock
Output clock
V1 applied
V2 applied
Output latched
- An arbitrary vector pair may not be applied to a
sequential circuit under full-scan, partial-scan
or non-scan methodology
18At-Speed Test
- At-speed test means application of test vectors
at the rated-clock speed. - Methods of at-speed test
- External, functional test
- Functional vectors applied by high-speed testers
- At-speed scan test
- Built-in self-test (BIST)
- Software-based self-test
At-speed test does not necessarily guarantee
high-quality delay testing unless tests are
designed to detect delay faults!
19Applying At-Speed Scan Tests
V2 states generated, (A) by one-bit scan shift of
V1, or (B) by
V1 applied in functional mode.
Scan chain length L
Test application scheme (B)
Test application scheme (A)
20Classifications of Paths
There are untestable paths even if full-scan is
used to deliver tests! But do we really need to
test them, if defects/variations on them do NOT
degrade circuit performance in functional mode?
21A Cost-Effective Test Strategy
- Use functional vectors
- Functional vectors can be applied at-speed and
should catch some delay defects. - Functional vectors should be evaluated for
transition fault coverage. - Derive and apply tests for undetected transition
faults - Derive and apply tests for long path-delay faults
22Delay Test/Speed Binning Challenges for Nanometer
Devices
- Delay variability increases due to process,
circuit, temperature, power, and noise factors. - No. of critical paths increases due to speed and
power saving techniques. - Clock is increasingly susceptible to
faults/variations creating test inaccuracy and
escapes. - Conventional transition and path delay models and
test methodologies are severely challenged!
23Variability of Path Delay
- Noise-induced variability
- Coupling cap -- pattern (excitation/propagation)/t
iming specific - Power grid fluctuation -- pattern specific
- Circuit induced -- leakage, charge sharing
(pattern specific) - Process-induced variability
- Spatial temporal parametric variability lot to
lot, wafer to wafer, die to die - Limitations in lithography
- CMP induced variability
- Thermal-induced variability
- Power-induced variability
Source TM Mak, Intel
24More Critical Paths Slowing Down Non-Critical
Paths
- Severe power constraint drove power optimization
everywhere. - Slowed-down paths sped-up paths all crowded
around required period. - More critical paths make it easier for
crosstalk-slowdown to propagate. - Bus coupling effect over local wires may be more
likely frequent.
25Potential Solution Go Statistical!
- Circuit delays can be modeled as correlated
random variables to take various local global
factors into account - Noise, process variations, pattern dependency,
temp. variations, etc. - Global effects can be modeled by correlations
factors between delay random variables.
Mean/variance of pin-to-pin delay or interconnect
delay
a
26Notion of Critical Path
The most critical path can be different
based upon which delay model you have in mind!
27Critical Path Varies from Chip Instance to
Instance
P1 a, e, g P2 b, e, g P3 c, f, g P4
d, f, g
Suppose 10000 chip instances are produced
P1 P2 P3 P4
43.6 19.1 23.6 13.7
28Statistical Delay Test Diagnosis Framework
- Need to consist of five major components
- Statistical timing analysis
- Statistical critical path selection
- Selecting statistical long and true paths whose
tests maximize the detection of DSM delay defects - Path coverage metric
- Estimating the quality of a path set
- Generation of high quality tests for target paths
- Identifying tests that activate longest delay
along the target path - Path delay is highly pattern dependent
- Delay fault diagnosis based on statistical timing
model
29Statistical Timing Analyzer
- Gate/Cell level
- Correlated delay vs.
- Cell delay library
- Interconnect model
- Monte Carlo Based
- Automatically determine convergence condition
- Static and dynamic
- Vector-less or vector-dependent
Arrival times
(V1V2.)
Estimate signal arrival time as random variable
30Statistical Critical Paths
Arrival time of O
O
I
- A critical path can be defined as the one with
greater than P probability of exceeding a cut-off
period T - Adjusting P and T to limit the size of critical
path set
31Considering Path Correlation for Path Selection
Output arrival times
25/3
A
overlap
24/3
B
C
22/2
After selecting path A, should path B or C be
selected?
32Considering Path Independence for Path Selection
Paths selected for test generation
- Defects on selected paths can be captured.
- However, a (small) defect falls beyond the
selected paths may not be captured. - Even with transition fault tests, path
independence can still be an important factor for
path selection.
33Statistical Critical Path Selection
- A new method achieving four objectives
- Select statistical long paths
- Consider path correlation
- Achieve path independence
- Eliminate statistical false paths
- Results indicating that selecting statistical
long paths considering correlation and
independence simultaneously - Achieves higher test quality with the same number
of selected paths - Selects fewer paths to achieve same level of test
quality
J.-J. Liou, et al., "Experience in Critical Path
Selection For Deep Sub-Micron Delay Test and
Timing Validation," ASPDAC 2003. J.-J. Liou,
et al., "False-Path Aware Statistical Timing
Analysis and Efficient Path Selection for Delay
Testing Timing Validation," DAC 2002.
34Section 12.3.1
Signal Integrity and Power Supply Noise
35Coping with Signal Integrity
36Signal Integrity
- Motivation
- Modeling
- Test Methodologies
- Enhanced BIST
- Enhanced Scan
37Motivation
- Test cost will be dominant in this decade
ITRS01
38Result of Technology Scaling
Source ITRS01 Roadmap
Factors Technology Technology Technology Technology Technology
Technology nm 0.35 0.25 0.18 0.13 0.10
Coupling C pF 41.59 49.73 56.93 64.17 70.54
Ground C pF 12.89 10.06 9.65 7.30 6.42
Mutual L nH 0.80 0.84 0.88 0.93 0.97
Self L nH 1.17 1.17 1.18 1.21 1.23
39Testing for Signal Integrity
Physical Defects
Process Variations
40Fatal Problems on First Spin
- Overall 61 of new ICs require at least one
re-spin
www.deepchip.com
41The Bottom Line
- Signal integrity loss occurs due to process
variations, manufacturing defects, the parasitic
and coupling C/L. Integrity loss leads to
failure. - Signal integrity problem is both design test
issue. A systematic approach for testing is
needed.
42Interconnect Model
- Signal integrity problems originate from
interconnects. - Distributed RLC model is too complicated.
43Integrity Loss Model
- Excessive delay degrades performance and causes
functional error. - Ringing causes functional error.
- Overshoot contributes to noise, delay, hot
carrier, time-dependent dielectric breakdown, and
electromigration.
44Prior Works
- Fault model and test pattern generation
- W. Chen, S. Gupta and M. Breuer ITC98
- M. Cuviello, S. Dey, X. Bai and Y. Zhao ICCAD99
- A.Attarha, M.Nourani VTS02
- Self-test methods for testing interconnects
- X. Bai, S. Dey and J. Rajski DAC00
- M. Nourani and A. Attarha DAC01 JETTA02
- I. Rayane, J. Medina and M. Nicolaidis VTS99
- Modified boundary scan
- J. Shin, H. Kim and S. Kang DATE99
- K. Lofstorm ITC96
- C. Chiang and S. Gupta VTS97
- S. Yang, C. Papachristou and M. Tabib-Azar
DAC01 - M. Tehranipoor, N. Ahmed, M. Nourani TCAD04
45Method 1 Enhanced BIST
- The adverse effects of integrity loss will appear
only at the working frequency. - The effects of integrity loss are usually
transient and intermittent. - At-speed testing requires high-performance ATEs.
- External test of signal integrity is limited due
to speed, access and probing difficulties.
Test Pattern Generator
Output Response Analyzer
Interconnect Under Test (IUT)
Test Controller
46On-Chip Noise Detection
- The internal Noise Detector (ND) and Skew
Detector (SD) cells sample signals and record
skew and delay violations. - Our BIST-based methodology can be integrated
within conventional BIST environments with 20 to
50 more overhead.
T P G R
Interconnect Under Test (IUT)
ND Cell
M I S R
SD Cell
BIST Controller
47Noise Detector (ND) Cell
- The ND cell detects voltage violations, e.g.
overshoot and ringing.
signal
Core i
Core j
IUT
Signal noise
T6
T3
T4
To read-out circuit
c
y
T7
x
T1
T2
Test_mode
T5
48Behavior of the ND Cell
- The noise detector (ND) cell shows a hysteresis
property and can detect two threshold voltages.
49Skew Detector (SD) Cell
XNOR
Sensor
Level restorer
c
To flip-flop
Interconnect signal (Signal Delay)
a
PDN
Inverter 2
TCK
TCK
b
b
b
Inverter 1
ADR
50Behavior of the SD Cell
ADR
Violation
51Readout Architecture
SI SO
ND Cell Vb Vc
0
1
Test Controller
52Method 2 Enhanced JTAG
- Boundary scan provides easy access to the
interconnects - The boundary scan cells and TAP controller need
modification to - generate and apply the test patterns (PGBSC)
- capture and read out the integrity violations
(OBSC)
53Maximum Aggressor (MA) Model
54Pattern Analysis in MA Model
55Pattern Generation BSC (PGBSC)
TDO/next cell
Input pin/ Core output
Output pin/ Core input
0
1
0
1
D2
Q2
D1
0
Q1
1
Mode
Q2
FF1
FF2
TDI/ previous cell
Q1
ClockDR
0
1
ShiftDR
Q3
T
FF3
SI
UpdateDR
56Operational Modes of PGBSC
PGBSC Mode Q1 SI
Victim 1 1
Aggressor 0 1
Normal X 0
Victim mode
Aggressor mode
UpdateDR
CLK-FF2
Q2
57Encoded Data for Victim Line
Victim-Select Victim Line
10000 1
01000 2
00100 3
00010 4
00001 5
58Observation BSC (OBSC)
TDO/next cell
ND/SD
Output pin/ Core input
ND FF
1
0
Input pin/ Core output
0
SD FF
0
1
D2
Q2
D1
Q1
1
0
sel
1
Q2
Mode
FF2
FF1
UpdateDR
ClockDR
ShiftDR
TDI/ previous cell
SI
59Operational Modes of OBSC
Modes ND/SD SI
NDFF 1 1
SDFF 0 1
Normal X 0
SI ShiftDR sel
1 0 0
1 1 1
0 X 1
60Operation of OBSC
TCK
Controller State
Shift-DR
Capture-DR
ClockDR
ShiftDR
SI1
Select ND/SD cell
(form the scan chain)
selSIShiftDR
61Test Architecture
BSC
OBSC
PGBSC
1
1
2
2
Standard IEEE1149.1 Interface
k
m
62New Test Instructions
- Two new instructions are added to the IEEE1149.1
instruction set - G-SITEST Instruction
- Facilitates test pattern generation based on the
MA fault model - PGBSCs are enabled with signal SI1
- ND/SD cells become active (CE1) to capture the
signal integrity information - O-SITEST Instruction
- Is used to capture and scan out the ND/SD FFs
data - Is loaded after the G-SITEST instruction
63Concluding Remarks - SI
- Signal integrity failures are intermittent
therefore, new test pattern generation, detection
and readout strategies are required. - Enhanced BIST Methodology
- Is capable of at-speed testing.
- Is relatively expensive unless limited to long
buses. - Provides data for test/reliability analysis and
diagnosis. - Enhanced JTAG Architecture
- Requires noise/skew detector cells on
interconnects. - Needs modified boundary scan cells.
- Provides a cost effective solution to test
interconnects for integrity loss.
64Coping with Power Supply Noise
65Power Supply Noise (PSN)
- Noise in high speed design
- Sharp rise and fall times (small dt)
- Changes in current drawn from Vdd (large di)
- VPSN L di(t)/dt R i(t)
- Large voltage fluctuation
- Intermittent malfunctions
- Functional failure
- Reliability problem
66Motivation
- Power supply noise analysis captures noise more
globally in high speed circuits. - PSN analysis can be useful for
- Pre synthesis noise/performance estimation
- Sensitivity analysis
- Power supply network design
- Accurate estimation of PSN in a core based SoC
without exhaustive simulation.
67Prior Work
- Genetic algorithms
- Y. Jiang, K. Cheng and A. Krstic CICC97
- G. Bai, S. Bobba and I. Haji ICECS01
- S. Zhao, K. Roy and C. Koh ICCD00
- Estimation based on modeling
- Y. Chang, S. Gupta and M. Breuer VTS97
- L. Zheng, B. Li, and H. Tenhunen ISCAS00
- M. Nourani, M. Tehranipoor, N. Ahmed VTS05
- Application power distribution and floor
planning - N. Pham, M. Cases, D. Araujo and E. Matoglu
VTS04 - S. Zhao, K. Roy and C. Koh ASP-DAC02
68Power Distribution Wire and Pin Model
- Vnoise(t) (Vdd,pin - Vdd,block(t)) (Vss,pin -
Vss,block(t))
69PSN Analysis Metrics
- Level A level of a node (distance from primary
input) implies how difficult or restrictive it is
to switch that node. - Fan-Out Switching time (tP) is inversely
proportional to fan-out. Switching low fan-out
gates will reduce dt ? increase di/dt. - Fan-In Large fan-in gates are resized (more
wide) to allow proper pull-up and pull-down.
Large width allows drawing more current.
70Effect of Level
71Effect of Fan-Out
72Effect of Fan-In
73Different Methods
Method M1 M2 M3 M4 M5 M6
Order of Metrics ILO IOL OLI OIL LOI LIO
- Ordering of metrics gives priority to switching
one gate over the other. - Level is given the highest priority due to ease
of control. - M5 (LOI) and M6 (LIO) methods are used.
74Test Pattern Generation
- Exhaustive
- 2n(2n-1) 22n possible transitions
- Random
- Cannot guarantee peak power in short time.
- Instead of exhaustive or random search,
heuristic-based approach should be used.
75Basic Approach
- Use conventional s-a-f pattern generation method
and tools to stimulate 0?1 or 1?0 transition. - Determine test pattern pairs by estimating
maximum PSN according to three PSN metrics.
76Concluding Remarks (PSN)
- Power supply noise captures the effect of noise
globally. - One goal is often to generate a pattern pair to
stimulate maximum PSN in non-embedded cores. - Combining individual PSN curves strategically
provides a way to estimate PSN of SoC without
full simulation.
77Section 12.3.2
Parametric Defects, Process Variation, and Yield
78Defects and Physical Defects
- Physical defects occur during manufacturing, and
can cause static or timing physical failures - Defects can be random or systematic, and can be
functional or parametric - Traditional work is more on functional random
spots - Other three types of defects need to be
researched - Defects can also be caused by process variations
and random imperfections
79Defect-Based Test
- Defect-based test can be done by enumerating
likely defect sites from layout - At-speed tests (path-delay tests and transition
tests) must be used
80Supplements of Conventional Stuck-at Tests
- Bridging tests enumerate likely bridging fault
sites (interconnects) by layout simulation - N-defect stuck-at tests detect every stuck-at
fault N times by targeting different sensitive
paths - TARO (transition fault propagation to all
reachable outputs) generate transition tests one
for each reachable output, for each given
transition fault. - IDDQ tests test by measuring current flow
- Functional Testing must be added to supplement
structural tests - Key issue how to generate these defect-based
tests in a timely manner to meet test goals.
81Sections 12.3.3 -12.3.4
Soft Errors and Fault Tolerance
82Cosmic Ray/Radiation Mechanics
high energy neutrons
lighter Particles
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
some particles also may pass through all material
without colliding with any atoms
83Mechanism of Neutron SER
84Neutron Environment
Normand et al.
1,000,000 feet 330 km
Shuttle
? Primary Cosmic Rays ?
Neutrons ?????Secondary Cosmic Rays
150,000 feet 50,000 m
Top of Atmosphere
Peak Neutron Flux
60,000 feet 20,000 m
?
35,000 feet 10,000 m
N,O
Aircraft
?
???
Ground 1/500 of Peak Flux
85electrons gammas
varies with altitude and geography
n0-Si interaction can result in a short range,
intense burst of charge
86Mechanism of Alpha Particle SER
Host material (package, Si, oxide, etc.)
- Impurities disintegrate in alpha decay
- Observed alpha flux 0.01/h/cm2
Impurity in ppb amounts (238U, 232Th, etc.)
87Sources of alpha
- alpha decay from radioactive isotope Pb 210
- travel short distance, a few centimeters
- but they are located at the surface of the dies
with C4 mounting - Traditional solutions, e.g., epoxy die coat, are
not effective
88Shrinking Process Decrease Charge per Node
Soft error is a function of stored charge at
sensitive nodes QCV i.e., Cnode and Vcc
How low will it go?
89RAM cell sensitive area
SER Critical area is drain/source junction rather
than metal interconnect
Not all diffusion area are sensitive Different
junction have different sensitivity
characteristics
90Caches on CPU chip
Itanium2
Neutron Scaling Trends Neutron Scaling Trends Neutron Scaling Trends Neutron Scaling Trends Neutron Scaling Trends
Tech. Scaling Factor Scaling Factor Scaling Factor Scaling Factor
Tech. ADIFF CGATE Vcc SER / bit
l 1 1 1 1
0.7x l 0.54 0.75 0.7 1.08-1.45
0.7x l 0.54 0.75 0.8 0.92-1.18
Itanium
This may lead to false security
Doubling of devices every generation will double
soft error rate as well !!
91Logic Circuit Subjected to SEU
Feedback loop exists
- Logic is not immune to SER
- All feedback nodes are susceptible
- Typical hardening techniques will pose
performance penalty - Latch is susceptible when in the holding phase
(50 of cycle) - F/F Master/Slave
- Each is sensitive during its inactive phase
typical D-latch
Disturbance change logic state permanently
Focuses on meta-stable window results in low
probability event
92SER Cannot be Screened During Manufacturing
- SER is recoverable error (if recomputed)
- Most memory elements (with feedback) are
susceptible to SER (maybe to different degree) - Errors in compute elements may become SDC (silent
data corruption) or at best system crash (equally
undesirable) - Low end system may be OK with a reboot not
acceptable for a server - SDC is the most vulnerable wont know unless
computation is repeated at another time
93Redundancy - Spatial
Identical Compute elements
- Spatial
- Duplication compare
- Triplicate vote
- Checkpointing and rollback is a common recovery
technique
checker
Miscompare !
Initiate rollback
Continue if correct Checkpointing regularly
94Redundancy - Temporal
Compute element
- Temporal
- Assume error is not persistent
- Recompute everything (twice)
- First result compare to second
- Checkpointing and rollback is common recovery
technique
Compute twice
Temporary storage
1st result stored temporarily
2nd result forward for comparison
checker
Miscompare !
Initiate rollback
Continue if correct Checkpointing regularly
95Redundancy - Information
parity
- Error detection by extra information
- E.g. parity
- Parity is preserved through computation
- Error correction by extra information
- E.g. Hamming code
data
Parity should match
operation
result
ECC check bits
One bit in a row defective
data
ECC logic
data
Defective bit, correct by ECC
96Hierarchy of Protection
- Start with most unreliable (and highest quantity)
elements - DRAM (highest number of bits)
- Parity, ECC
- Hard disk (mechanical)
- RAID (Redundant Arrays of Identical Disks)
- IO Channels
- CRC, ECC
- Cache
- Parity, ECC
- Register files
- Parity
- Execution units
- DMR
97Section 12.3.5
Defect and Error Tolerance
98Defect Tolerance
- Assumptions
- Defect rate is low
- Most defects cause single cell/row/column
failures - Repair defect elements using redundant resources
99Redundancy Repair An Example
- Design in spare rows or columns or blocks
- Test to identify where bad cell or row/column is
- Algorithm to figure out how best to use spare
elements to repair all defects - Fuse change decoders to address spare elements
instead of faulty elements
spares
100Error Tolerance
- Increasing new computing applications are with
multimedia data - Compression is generally used for these data
types - E.g. MP3, JPEG, MPEG
- Lossy in nature
- Human senses are not keen enough to spot the
difference - Errors may be tolerated for these kinds of
applications - Accepting more faulty chips will increase
effective yield while lowering cost
re Error Tolerance, Mel Breuer, 2005
101Section 12.4
FPGA Testing
102Field Programmable Gate Arrays
- 2-dimensional array
- Programmable logic blocks (PLBs)
- Programmable routing network
- Programmable I/O cells
- Recent FPGAs incorporate specialized cores
- Memory cores RAMs, FIFOs, etc.
- Digital signal processors (DSPs)
- Embedded processors
103Basic PLB Architecture
- Look-up Table (LUT)
- implements combinational logic truth table
- Memory elements
- Flip-flop/latch
- Some LUTs also implement small RAMs
- Carry and control logic
104Programmable Interconnect Network
- Wire segments of varying length
- xN N PLBs in length
- N 1, 2, 4, 6, 8 are most common
- xH half the array in length
- xL length of full array
- Programmable Interconnect Points (PIPs)
- Transmission gate connects 2 wire segments
- Controlled by configuration memory bit
- Four basic types of PIPs
105Programmable Interconnect Points
- Break-point PIP
- Connect or isolate 2 wire segments
- Cross-point PIP
- 2 nets straight through
- 1 net turns corner and/or fans out
- Compound cross-point PIP
- Collection of 6 break-point PIPs
- Can route 2 isolated signal nets
- Multiplexer PIP
- Directional and buffered
- Main routing resource in recent FPGAs
- Select 1-of-N inputs for output
106Ranges of Programmable Resources
FPGA Resource FPGA Resource Small FPGA Large FPGA
Logic PLBs per FPGA 100 22,000
Logic LUTs and flip-flops per PLB 1 4
Routing Wire segments per PLB 50 400
Routing PIPs per PLB 80 1,000
Specialized Cores Bits per memory core 128 18,432
Specialized Cores Memory cores per FPGA 16 500
Specialized Cores DSP cores 0 500
Other Input/output cells 62 1,200
Other Configuration memory bits 32,000 50,000,000
Large FPGAs easily exceed 500 million transistors
107FPGA Testing Problem
- Must test all modes of operation
- Many test configurations must be downloaded
- Long test time
- Large and complex devices
- Large FPGAs exceed 500 million transistors
- Many different types of functions to test
- PLBs
- Routing resources
- Specialized cores (RAMs, FIFOs, DSPs, etc.)
- Frequently changing architectures
108FPGA Testing Approaches
- With respect to system application
- Application independent testing
- Test all resources in FPGA regardless of system
function to be implemented - Application dependent testing
- Test only those resources that will be used by a
given system function - Testing techniques
- External testing
- Test patterns applied and output responses
monitored through I/O pins with external
equipment - Built-In Self-Test (BIST)
109BIST for FPGAs
- Basic idea reprogram FPGA to test itself
- No area overhead or performance penalties for
system applications - Applicable to all levels of testing
- From device-level through system-level testing
- Cost
- Memory to store BIST configurations
- Goal minimize number of configurations
- Download time to execute BIST configurations
- Goal minimize downloads and/or download time
110BIST for PLBs
- Program PLBs as
- Test Pattern Generators (TPGs)
- Multiple TPGs prevent faulty PLBs under test from
escaping detection when there is a fault in a TPG
PLB - Identically configured logic blocks under test
(BUTs) - Output Response Analyzers (ORAs)
- Comparison-based
- Row or column orientation
- Two test sessions required
- TPGs/ORAs and BUTs reverse rolls
Test session 1
Test session 2
ORA
111BIST for Routing Resources
- Program groups of wire segments and PIPs as wires
under test - Program some PLBs as TPGs and ORAs
- Similar to BIST for PLBs
- Two BIST approaches
- Comparison-based
- Similar to BIST for PLBs
- Parity-based
- TPG produces test patterns with parity
- ORA performs parity check
112Concluding Remarks
- FPGAs are more SoC-like with specialized cores
- RAMs, DSPs, etc. can be tested with approach
similar to BIST for PLBs - SoCs are incorporating FPGA cores
- These cores can be tested with BIST for FPGA
techniques - Complex Programmable Logic Devices (CPLDs) are
similar to FPGAs - Can be tested with approaches similar to those
used for FPGAs
113Section 12.5
MEMS Testing
114Introduction to MEMS
- What is MEMS?
- MEMS Micro Electro Mechanical System
- Extremely small (in range of um) devices
utilizing both electrical and mechanical
properties.
MEMS gear chain and a mite for size comparision
(Sandia MEMS)
115Video Clips from Sandia MEMS
- A micro-resonator is used as an actuator to drive
MEMS gears with mites crawling on top.
116Commercially Available MEMS Devices
Digital Micromirror Device (DMD) (Texas
Instruments, Inc.)
MEMS ink jet print head (nozzle) (HP, Inc.)
117Commercially Available MEMS Devices
ADXL50 accelerometer (Analog Devices, Inc.)
LambdaRouter optical switch (Lucent, Inc.)
118Why MEMS?
- Why MEMS?
- Lower cost due to batch fabrication
- Light weight
- Smaller size
- Lower energy consumption
- Higher performance
- MEMS applications
- Automobile industry
- Health care
- Aerospace
- Consumer Products
- RF telecommunications
- Other areas
119World MEMS Market Prediction
- Worldwide MEMS market is increasing rapidly.
- Worldwide MEMS sales gt 5 billion in 2004,
- gt 8 billion in 2007.
120Basic Concepts for Capacitive MEMS Devices
- In static mode where
- dielectric constant of air
- S overlap area between M and F1/F2
- d0 static capacitance gap between M and F1/ F2
121Basic Concepts for Capacitive MEMS Devices
- Assume Vertical stimulus force causes movable
mass M to move upward with displacement x, where
x ltlt d0
122Basic Concepts for Capacitive MEMS Devices
- To sense displacement x, modulation voltages Vmp
and Vmn are applied to F1 and F2 respectively
Where V0 is modulation voltage amplitude, w is
freq of modulation, t is time for operation, sqr
is square waveform
123Basic Concepts for Capacitive MEMS Devices
- According to charge conservation law,
where VM is voltage sensed by movable plate M
- By sensing Vm, we find displacement x.
124MEMS BIST Research
- Statement of problem MEMS testing is becoming an
urgent need. - MEMS is finding applications in safety-critical
areas, such as automobile, aerospace, medical
instruments, etc. - MEMS will be integrated into SoC, so it needs to
be thoroughly tested to ensure reliability. - The commercialization of MEMS technology needs a
thorough and efficient testing solution in order
to ensure reliability and reduce the test cost.
125MEMS Testing A Challenging Topic
- MEMS testing is chanllenging
- multi-field coupling
- diversity in device structure working principle
- analog signals involved
- vulnerable to various defect souces (stiction,
etch variances, particle contamination, etc.) - Just like VLSI testing, built-in self-test (BIST)
is believed to be a promising solution for MEMS
testing. - Research goals develop a robust and efficient
BIST solution for capacitive MEMS devices.
126MEMS Sensitivity BIST Scheme
- How to generate test stimulus for MEMS?
- Apply voltage Vd to F1 and nominal voltage Vnom
to M - Electrostatic force Fd will be experienced where
127MEMS Sensitivity BIST Scheme
- The force will result in a displacement x.
- The sensitivity BIST scheme requires another
similar structure to sense the displacement x. - Sensed VM is compared with a known value for
comparison.
128Symmetry BIST for Capacitive MEMS
129Symmetry BIST for Capacitive MEMS (cont.)
- For a fault-free Device
- C1C2, then VM0
- If C1 ?C2(faulty) VM?0
- Fixed capacitance plates are partitioned into two
equal portions (S1, S2 in top and S3, S4 in the
bottom). - During symmetric BIST, always check VM0?
- Local Defect causing left-right asymmetry is
detected. - Partitioning by movable plate see N. Deb and
R. D. (Shawn) Blanton, - Built-in Self Test of CMOS-MEMS
Accelerometers, ITC02, p.1075.
130Capacitance Partition for Dual-mode BIST
- In Fig (b)
- S1, S2 top sensing plates,
- S3, S4 bottom sensing plates,
- D1,D2 driving plates
- In Fig (b) cont.
- M movable plate
- C1, C2 top sensing cap.
- C3, C4 bottom sensing cap.
131Capacitance Partition for Dual-mode BIST
- Fixed plate at each side of the movable plate is
divided into three portions 1 for electrostatic
driving, other 2 equal portions for capacitance
sensing. - The movable capacitance plate is not partitioned.
- Sensitivity and symmetry BIST can be implemented.
132Voltage Biasing Scheme of Dual-Mode BIST
- D1 D2 also participate the normal operation.
- Analog switches are used for mode-switching.
- Any defect causing sensitivity change or
left-right asymmetry will be detected.
133MEMS Comb Accelerometer
- Device prototype ADXL50 by Analog Devices Inc.
- The serrated comb finger groups are extremely
- vulnerable to defects, thus BISR is highly
desirable.
134Working Principles of Accelerometer
- Uses differential capacitance sensing technique
- Vmp (Vmn) Complimentary
- modulation voltages
135Working Principles of Accelerometer (cont.)
- VM is proportional to displacement x, which is in
turn directly proportional to acceleration a. - By sensing VM, the value of a can be measured.
136BIST for Comb Accelerometer
- Device prototype
- ADXL50
- Since fixed plates are
- separated comb
- fingers, the partition
- can be easily realized.
- M1-M8 movable plates
- D1-D8 fixed driving
- plates
- S1-S8 fixed sensing
- plates
Voltage Biasing Normal Operation SensitivityBIST Symmetry BIST
Vd - D1,D3,D5,D7 D1,D3,D5,D7
Vnom - D2,D4,D6,D8 M1,M4,M5,M8 D2,D4,D6,D8 M1,M4,M5,M8 S2,S4,S6,S8
Vmp S1,S3,S5,S7 D1,D3,D5,D7 S1,S3,S5,S7 S1,S5
Vmn S2,S4,S6,S8 D2,D4,D6,D8 S2,S4,S6,S8 S3,S7
137ANSYS Fault Simulation of Comb Accelerometer
Stiction defect simulation result
- Stiction (on right central movable finger)
sensitivity BIST is more efficient. - Finger height is matched (only in right portion)
symmetry BIST is more efficient.
Defect location Frequency (kHz) SensitivityBIST(mV) Symmetry BIST(mV)
Defect-free 11.85 967.6 0
0 28.80 107.8 2.5
10 32.75 82.2 2.1
20 39.23 56.4 1.6
30 46.80 39.2 1.2
Finger height mismatch simulation
Height mismatch ?H (um) Frequency (kHz) SensitivityBIST(mV) Symmetry BIST(mV)
Defect-free 11.85 967.6 0
0.1 11.85 951.8 42.9
0.2 11.85 937.1 83.9
0.3 11.85 921.8 128.2
0.4 11.85 904.4 181.3
For definitions of simulated defects, see
paper at ITC02, p.1075.
138Conclusions of Dual-Mode BIST
- By partitioning fixed instead of movable
capacitance plate, the BIST technique can be
extended to bulk-micromachining and other MEMS
technologies. - Each sensitivity and symmetry BIST has its own
fault coverage. A combination of both ensures
better coverage. - Sensitivity BIST is necessary during in-field
usage even after calibration. - Some unstable defects may change status.
- New defects may be developed in in-field usage.
- Stiction is also possible during in-field usage.
139Section 12.6
High-Speed I/O Testing
140I/O Test Requirements and Architectures
- I/O test requirements are largely driven by the
interoperability, system performance, and
functional performance goals. - I/O test requirements are closely related to the
link architectures. - For data rates lt 1 Gbps, global common clock (CC)
and source-synchronized (SS) are popular. - Above 1 Gbps, serial architectures are dominant
- Timing, voltage, and bit error rate (BER) are
common testing parameters.
141I/O Architecture (I) Common Clock (CC)
- Synchronized global (common) clock
- Common clocks for Tx data driving and Rx data
sampling - Clock skew on board limits its use to lt a few 100
Mbps data rate - Needs to test
- Data to clock delay at Tx
- Setup/hold time at Rx
142I/O Architecture (I) Common Clock (CC)
- Synchronized global (common) clock
- Common clocks for Tx data driving and Rx data
sampling - Clock skew on board limits its use to lt a few 100
Mbps data rate - Needs to test
- Data to clock delay at Tx
- Setup/hold time at Rx
143I/O Architecture (II) Source Synchronous (SS)
- Tx sends data along with strobe (another clock)
- Rx uses sent strobe to sample the data
- No clock or strobe skew issue
144I/O Architecture (II) Source Synchronous (Contd)
- Some designs use strobe/strobe to improve timing
accuracy. - Needs to test
- Data valid before and after strobe at Tx end
- Setup and hold times at Rx end
145I/O Architecture (II) Source Synchronous (Contd)
- Limited by data to data skew due to uneven
channels - Board layout
- E-M issues e.g., coupling, noises
- Variation in drive among channels
- Achieve up to 1000 Mbps data rates for wide bus
- Can improve data rate with splitting into many
narrower bus
146AC IO loopback self-test
Similar circuit as the receiving end! Testing
hardware already exists! -- test for both
drive/receive -- low overhead
Loop time Tco (or Tvb) Tsetup OR TvaThold
147Defect-based IO test
- A wider spread of data valid time indicate faults
148I/O Architecture (III) Serialier/Deserializer
(SERDES)
- Bit clock is embedded in the serial data and gets
recovered at Rx via clock recovery circuit. - Link layer is composed of encoder and decoder.
- Physical layer (PHY) is composed of Tx, channel,
and Rx.
149I/O Architecture (III) A SERDES PHY Implementation
- Bit clock is recovered via a phase interpolator
(PI) clock recovery. - PI clock recovery tracks low frequency jitter
from reference clock, Tx, and channel. - Differential PLLs reduce the jitter from
reference clock. - Used widely in PCI Express (2.5 Gen I, 5.0 Gbps
Gen II) and FB DIMM (3.2, 4.0, and 4.8 Gbps Gen
I). - Jitter is the major limiting factor/performance
metric.
150Jitter Components and Terminology
151Eye-Diagram, Jitter, and Noise Testing
- For output testing, jitter and noise probability
density functions (PDFs), and eye-opening should
be upper bounded. - For input tolerance testing, jitter and noise
PDFs, and eye-opening should be lower bounded.
152Summary
- Link architecture determines the relevant test
parameters and methods. - For synchronized CC and SS architectures,
critical test parameters include - Data valid to clock/strobe
- Setup/hold time
- For a SERDES architecture, critical test
parameters include - Jitter, includes deterministic jitter (DJ),
random jitter (RJ), and total jitter (TJ) or
timing eye-opening - Noise, and voltage eye-opening
- Bit error rate (BER)
153Chapter 12.7
154Outline of the Section
- Introduce the basic concepts related to RF
- Discuss the various challenges associated with RF
test - Describe different core RF building blocks
- Elaborate various test specifications for RF
devices - System-level testing and associated
specifications - Conclude with present and future trends
155Introduction to RF
- RF stands for Radio Frequency
- Usually very high frequencies where signals can
be transmitted wirelessly - Range of frequencies ? 300MHz 3GHz
- RF is used synonymously with wireless
- Significant growth during the last decade in the
consumer segment - Increased consumer applications
156Applications of RF
- Earlier, consumer applications of wireless
technology were limited - Military, space communications, air traffic
control - Currently, consumer applications are on the rise
- Cell phone, laptop, PDA, satellite radio
- Radiofrequency identification (RFID)
157Challenges with RF testing
- Tests are performed in two steps
- Characterization test
- Production test
- Various challenges make production test hard and
expensive - RF devices need extra attention during testing
(challenge 1) - Impedance matching _at_ input and output ports to
ensure optimal power transfer - Shielding from external wireless signals during
testing
158Characterization test
- Characterization validates the first set of
silicon - Uses highly accurate instruments to
- Verify the functionality of the design
- Ascertain that the specifications are met
- Ensure high repeatability of the measurement
system - Production test needs to perform all of the above
with - Cheaper instrumentation ? a least-cost commercial
tester (challenge 2) - In much smaller duration of the time used for
characterization (challenge 3)
159Repeatability and accuracy issues
- Test procedure should classify
- Good devices as good and
- Bad devices as bad
- This is constrained by measurement noise
- Reduces the accuracy of measurement during
production testing (challenge 4) - Introduces large variability in the same
measurement repeated many times (challenge 5) - These can be mitigated by using
- Accurate test application procedure
- High resolution measurements
160Summary of challenges
- Challenge 1 is very specific to RF devices ?
needs careful measurement setup - Challenge 2 and 3 are also specific to RF
- RF testers are very expensive (gt 1M) compared to
the analog and digital counterparts - RF tests are usually longer compared to analog
tests - Challenge 4 and 5 are general for all electronic
devices - However, these are more prominent in RF due to
the large amount of noise involved
161Typical RF system
162RF specifications
- Linearity specifications
- Gain, conversion gain, output power
- Non-linearity specifications
- Third-order intercept (TOI), adjacent channel
power ratio (ACPR) - Noise specifications
- Noise figure (NF), signal-to-noise ratio (SNR),
sensitivity, dynamic range - System specifications
- Error-vector magnitude (EVM), bit error rate (BER)
163A note on decibel
- Decibel is a very commonly used unit in wireless
domain - Notation for decibel ? dB
- Any number N can be converted to decibel by
- NdB 20 log10(N)
- A similar unit is mili-decibel (notation ? dBm)
- Used to denote power with reference to 1 mW
- P watts of power is converted to dBm by
- PdBm 10 log10(P/1 mW)
- Thus 1W 1000mW 30 dBm 10?W 0.01mW -20 dBm
164Gain
- Measures the small-signal gain of the
device/system - Input is a single-tone stimulus within the
operating frequency - Amplitude is within linear range of operation
- Gain (Output amplitude / Input amplitude),
usually expressed in dB
A2
A1
f1
f1
Gain A2/A1 20 log10(A2/A1), in
dB
165Conversion Gain
- Measures the small-signal gain of mixers
- Mixers translate the input frequency at a
different output frequency - Input is a single-tone stimulus within the input
operating frequency, amplitude within linear
range of operation - Gain (Output amplitude _at_ f2 / Input amplitude
_at_f1), usually expressed in dB
A2
Different frequencies
A1
f2
f1
Gain A2/A1 20 log10(A2/A1), in
dB
166TOI
- Measure of nonlinearity for a device/system
- Two-tone input, within operating range
- Frequencies are closely spaced, difference is
usually lt1 of the device bandwidth - Amplitude is larger than linear range of
operation - TOI Pout (Pout PIMD)/2
Fundamental tones
Intermodulation tones
Pout
DUT
PIMD
f1
f2
f1
f2
2f1 - f2
2f2 - f1
167A note on TOI
Third-order nonlinear term
- Usually, RF devices exhibit third-order
nonlinearity - y(t) A0 A1x(t) A3x(t)3
- TOI is denoted in dBm
- It indicates the output power level where the
fundamental tones and intermodulation tones
attain same power
168Basics of SNR
- SNR is an important factor for any signal
- SNR for a known signal can be easily computed
- SNR denotes the level of purity
- For this sinusoid, SNR 50 dB
- (-23) (-73) 50 dB
- This notion can be extended to any known signal
-23 dBm
Amplitude
-73 dBm
Frequency
169Noise figure
- Noise figure measures the degradation of SNR of a
signal when it passes through the DUT - Noise figure SNRin/SNRout
- This indicates the amount of noise added by the
DUT - NF is usually measured in dB (its a ratio !!)
- NF is measured using NF meter
170System-level