Title: EEE287 California State University Sacramento VLSI Design IC Manufacturing and Test
1EEE-287California State UniversitySacramentoVL
SI Design -IC Manufacturing and Test
2Manufacturing Flow Overview(Typical flow -
variations exist)
- Purchase Blank Wafers
- Wafer Processing
- E-Test
- Wafer Sort
- Assembly
- Burn In (BI)
- Class Test
- Inspect, Mark Pack
3Purchase Blank Wafers
- Current Technology is 8 (diameter)
- 6 still in volume production
- 12 in early production ramp (not all companies)
- Number of die goes up with square of radius (pr²)
- 6 to 8 wafer is a 78 area increase.
- Center die to edge die ratio also improves
- Cost of material (gasses, etc.) little to no
increase. - Difficulty is in consistency across large wafers.
- Photoresist, CVD (deposition) and thermal
consistency. - 12 wafers are reaching limits of human handling.
- Wafers purchased, due to low level of proprietary
content.
4Wafer Processing
- Defines all silicon structures
- Transistors, Resistors, Capacitors in Si and SiO2
- Known in fabrication plants as front end
processing - Defines metalization for interconnect
- Known in fabrication plants as back end
processing - Key Attributes
- Leff
- Tox
- Metal Quality (line width, metal stringers)
5E-Test
- Fabs need a way of monitoring processing
consistency regardless of product being made. - Numerous simple test structures placed in scribe
line (N channel FET, resistor, etc.) - Allows measurement of critical electrical
parameters on each wafer with same test hardware
and software (Vt, IDsat, BV, rho, etc.) - Scrap limits are set to guarantee consistency.
- Trends in fab readily detectable.
6Wafer Sort
- Utilizes a probecard with needles to contact bond
pads. - ATE (Automatic Test Equipment) testers utilized.
- Similar in concept to bench equipment driven by
HPIB. - Power supplies, volt/ammeters, vector
drive/compare logic - Test vectors are logical 0s and 1s applied to
device inputs and compared to device outputs. - Essentially, a truth table for the device.
- Used to screen out bad die before wrapping an
expensive package around them. - Not all test are done at wafer sort.
- AC values not accurate (L and C of package affect
timing). - May be done at hot temperature using a Hot
Chuck - Bad die receive an ink dot, usually at an
Off-line Inking station.
7Assembly
- Mount and saw.
- Saw cuts through wafer but not through plastic
film. - Saw blade cuts away approximately 2 mils of
silicon - Die Attach to leadframe or substrate using
silver-filled adhesive or solder. - Provides mechanical, electrical and/or thermal
connection to leadframe. - Wirebond.
- Gold wire attached by
- ultrasonic thermal bonding.
- ball bond on die pad,
- wedge bond on leadframe
- Solder balls on pads
- replace wirebond for
- flip-chip assy
8Assembly (cont.)
- Mold
- Plastic (epoxy cresol novolac polymer) is
injection molded to encapsulate die and wires. - Wire sweep is largest threat at this stage.
- Tends to be limiting factor for max wire length.
- Ceramic packages receive a metal or ceramic lid
instead of plastic mold injection. - Plate, Trim and Form
- Leadframe is plated with gold or tin to reduce
corrosion and increase solderability. - Leadframe is cut away from carrier.
- Leads are formed into final shape.
9Package Evolution
- As transistor density has allowed integration of
more functions (requiring more I/Os), package
pincount has increased. - DIP (Dual In-line Package)
- PLCC (Plastic Leaded Chip Carrier)
- QFP (Quad Flat Pack)
- BGA (Ball Grid Array)
- Many variations on these packages exist.
- PGA (Pin Grid Array), TSOP (Thin Small Outline
Package), etc.
10Burn-In
- Activates latent failures due to manufacturing
defects - Oxide or silicon crystalline defects
- Diffusion of unintended impurities (e.g. sodium)
- Metal electromigration or bridging.
- High temperature and voltage provide activation
energy to accelerate defects. - Voltage Vcc x 1.25 or more, Temp 125C typ.
- Defect degradation is chemical effect, chemical
processes occur faster with more energy. - Toggle coverage is important to provide voltage
stress across all junctions.
11Burn-In (cont.)
- Takes advantage of the bathtub curve
- Majority of failures in first 10-20 years of life
occur in first 50 hours of use or less (infant
mortality). - Equivalent activation energy to 50 hours of
normal life can be applied in 6 hours or less of
Burn-In (BI). - Exact acceleration is dependent on defect type
and process technology. - Limitations to temperature and voltage
- Degradation (glass transition) temperature of
mold epoxy - Breakdown voltage of transistors.
- BI is a stress, not a test
- Testing is needed after BI to detect activated
failures
12Class Test (aka Final Test)
- Eliminates assembly defects and defects activated
at BI. - Used to separate devices into performance classes
- Example Test at 300MHz. If fail, test at 266MHz.
- Tested at room, hot and cold.
- AC Parametric screen includes package effects
- Capacitive slowdown, Inductive supply bounce
- ATE (Automatic Test Equipment) testers utilized.
- Applies test vectors to test the devices
function - Voltmeters and ammeters are used to test DC
parameters
13Class Test (cont.)
- Hot test is usually most critical since speed is
key differentiator (devices slow down at hot
temp). - Device handler input trays and test site are at
test temp since device does not have time to
self-heat during short test time. - Test temp85C max ambient (Vcc x Idd active x
ThetaJA) - ThetaJA is package thermal coefficient in oC/W
- Why do performance testing at all since designs
are simulated? - Simulations are not 100 accurate
- They are models, not reality
- Weather models are only accurate 1 -2 days in
advance, since the entire weather system is too
complex to model. Circuit models are only as
accurate as the accuracy of the inputs and the
inclusion of second and third-order effects. - Manufacturing has variability (Leff, Vt,
metalization, etc.)
14Inspect, Mark Pack
- Inspection
- Devices are inspected by laser inspection
equipment for package and lead coplanarity. - Devices are marked with Mfg. information
- Lot number, speed grade, manufacturer.
- Devices are mounted for shipping.
- Tape and reel is most common.
- Reels then receive 24 hour bake at 125C and are
sealed in a hermetic bag. - Plastic mold compound absorbs moisture which, if
not baked out, will vaporize during IR reflow or
solder wave steps of PC board assembly.
(popcorning)
15Economics of Si Manufacturing
- Profit Revenue - Expenses
- Revenue product_price volume
- Expenses manufacturing_cost NRE (eng costs)
COS (cost of sales). - Product_price driven by market
- You cannot directly control it
- Product Engineer is responsible for
- Low manufacturing cost / high yield
- High volume manufacturing capability
16Manufacturing Cost and HVM
- Incremental Processing Costs at each step
- Yield losses occur at each step
- Wafers rejected in fabrication line, including
etest - Dice rejected at wafer sort
- Dice and packaged dice (units) rejected at
assembly - Units rejected at final test
- Units rejected at inspect, mark pack.
- Take your yield losses early in the process!
17Mfg. Steps and Typical Costs
- Processed Wafers (8 in) - 2000 /wafer
- (Raw wafers 100)
- Wafer Sort - 50-100 /wafer
- Assembly - 0.05-10 /device
- Burn-In - 0.10-1 /device
- Class Testing - 0.10-5 /device
- Note These numbers are for instructional
purposes only and do not reflect the costs of any
particular product or manufacturer.
18Cost
- Every process step costs money!
- Every process step reduces the number of good
devices! - The value of any added process step (e.g. another
layer of metal) must outweigh its negative impact
on product cost.
19Assembly Yield Loss
- Initial visual inspection rejects
- Die attach problems
- Wire bonding problems
- Injection molding problems
- Package delamination
20Major Sort and Class loss factors
- Point defects
- Predominate at sort
- Parametric yield loss
- Predominates at class
- Gross parametric variation will be caught at
E-test.
21Point Defects
- Characterized by fatal defect density
- Not all defects cause failures
- Affects random die on wafer
- Caused by dust particles and other isolated
defects - Can be modeled with with a Poisson distribution
- Y e(-AD)
- where Y yield (1.0 max)
- A area of die D defects / unit area
- Note Die area is often expressed in mils. (1mil
0.001in) - Yield is exponentially dependent on die area!
22Why die size is critical
- Number of die increases with the decrease of die
size. - Yield increases exponentially with the decrease
of die size. - Therefore, smaller die more die per wafer and a
higher percentage of them being good die. - Two factors working in the same direction to
produce more good die for the same processing
cost!
23Other Yield Models
- The Poisson Model assumes the defect density is
constant across each wafer and from wafer to
wafer. It applies well to devices with small die
size. - Murphy Model Y (1 - e-(AD)) / (AD)2
- Assumes that defect density varies and is
Gaussian with the lowest value at the center of
the wafer - Seeds Model Y e-(AD)1/2
- Assumes that the defect density varies and is
clustered
24Defect Density
- Calculate defect density for a fab process using
area and yield information from multiple products
already being manufactured. - Use this defect density to predict the yield of
future products according to their area. - Defect density can also be used to compare
- Defect density variation between shifts, pieces
of equipment, fabs, etc. - Effectiveness of process improvement changes.
25Parametric Yield Loss
- Caused by process shift away from nominal
- Affects entire wafer
- Leff - Effective length of gate
- Changes performance of devices
- Can cause drain-source leakage
- Called Len for n-channel, Lep for p-channel
devices - Vt - Threshold voltage
- Change trip point of devices
- Performance affected
26Parametric Yield Loss (cont.)
- Gate Oxide thickness
- Capacitance changes
- Performance affected
- Metal problems
- Stringers
- thin metal
- Via connectivity
- All parameters are measured with process
monitors. - In-line monitors and at etest.
27Parametric Yield Loss (cont.)
- Some parametric yield problems still produce
usable parts. - Parametric shifts that cause performance
degradation can be sold as lower speed parts (at
a lower price).
28Other Economic Considerations
- Capacity - The lowest die costs are achieved when
the expensive mfg. equipment is fully utilized. - 2M tester, obsolete in 4 years 500k / year
- 10k units/year 50/unit, 1M units/year
0.50/unit - Die size increases, low yield, long test times,
etc. may result in the inability to make enough
parts to meet the demand with current equipment. - This may require an additional investment of
millions of dollars, in the case of a tester, or
2 billion, in the case of a fab, to meet the
demand. - e.g. 4sec - 5sec test time 25 increase in
testers x 40 testers (_at_3M each) 30M! - These costs end up raising the price of the
product.
29Other cost factors
- The cost of the equipment is not the only factor
which determines processing cost - Other cost factors
- Maintenance costs
- Consumables (electricity, chemicals, etc.)
- Operator labor costs
- NRE for developing the tests
- Other process-specific costs
30Design for Testability
- Design for Testability (DFT) plays a significant
role in reducing test time and improving test
coverage. - DFT starts with the definition of the product,
since test modes are part of the design - Is sometimes called Design for Manufacturability,
although DFM usually refers to device layout
restrictions (bond pad sizes, minimum metal
spacings, etc.).
31Controllability / Observability
- The two main concepts in DFT are
- Controllability How easy it is to control
(toggle) a particular node from primary inputs - Inputs have the highest controllability
- More logic between an input and a node lower
controllability - Observability How easy is it to observe the
behavior of a particular node at primary outputs - Outputs have the highest observability
- More logic between a node and an output lower
observability - Most DFT modes are implemented to increase
Controllability or Observability.
32Common Test Modes
- Common test modes are
- PLL Bypass / PLL Monitor
- All 1, all 0, all Z modes
- Icc standby (Iddq) mode
- Electrical ID readout
- Process monitor
- The listed test modes are neither required nor
comprehensive. - Most of these modes will be used only during
device or board test (not end users, not during
normal operation). - Test modes are similar to breaking device into
smaller pieces to test.
33Test Mode Logic
- In order to enter test modes, additional device
logic is required - Often activated by driving a test pin low
- An internal state machine senses this and enters
test mode (exact mode chosen depends on values on
other pins, contents of a control register, etc.) - Test logic then takes control of the chip and
puts it into desired test mode - May interfere with device operation, e.g. all
tristate - Or only interfere with a few pins (e.g. PLL
monitor)
34Test Mode Logic Diagram
- Simplified test mode logic circuitry
- Req pins normally input to core logic.
- When test is low, req
- pins select which test
- mode is active.
- In case shown, test mode
- is disabling (tri-state)
- all outputs.
enable
test logic
test
addr0
addr1
addr2
35JTAG (IEEE 1149)
- JTAG is a standardized test logic interface.
- Stands for Joint Test Action Group that
developed it. - Requires 4 pins (TCK (clock), TMS (mode select),
TDI (data in), TDO (data out) - Allows for versatile test mode implementation
- User can shift in commands and shift data in/out
- Test logic can be used concurrently with normal
function - Test modes may include a BIST (Built-In
Self-Test) mode. - Same interface and state machine function for all
devices using this standard
36PLL Bypass / Monitor modes
- PLL Bypass
- For some tests, we want to have direct control of
a clock normally generated by a PLL (e.g. slow
speed testing, Si debug) - Insertion of a mux in clock path provides this.
- PLL Monitor
- PLL monitor mode inserts mux in output paths of
non-critical outputs to bring out the internal
clock and PLL lock signal directly to primary
outputs.
37All 1, 0, Z test modes
- Drives all outputs and bi-directional pin values
to logical 0, 1 or tri-state (pullup/down
disabled). - Used to test Vol, Voh and leakage of buffers.
- These tests could be performed without special
test modes by searching vector patterns for
particular required value on each output. - This would be engineering intensive and would
greatly increase test time. - These modes increase controllability
38Icc standby (Iddq) mode
- Used to test current draw of device in
lowest-power state possible. - Tri-states all outputs, disables internal
pullup/pulldown, disables sense amps, PLLs, etc. - Aberrant current measurements indicate
manufacturing faults (latent functional or timing
faults). - May (or may not) find timing failures or latent
failures not caught by functional tests
(resistive shorts) - This will reduce devices which fail during burnin
or at class (speed) test. - Excellent test to perform along with functional
tests.
39Process monitor (Procmon)
- Used to determine speed of core transistors
independent of core circuitry. - Uses numerous inverters in series to amplify
effect of speed changes. - 30pS speed change for one gate becomes a 3nS
speed change for 100 gates in series. - Measurement-to-error ratio is improved by 100x
- Often used by third-party ASIC vendors to
guarantee speed performance.
40Other DFT modes
- Counter test modes
- Break big counters into multiple small ones
- Run them in parallel
- Bring terminal count pulse to output
- Observation test modes
- Bring out hard to observe signals on output pins
- Direct Access Test (DAT) modes
- Fault grade functional blocks with known good
vectors applied directly to them - Provide access to local inputs/outputs from
device pins. - Commonly used for embedded memories (RAM)
- Memory testing utilizes extensive unique test
methods
41DFTs Growing Importance
- DFT is growing in importance due to
- Faster time-to-market requirements
- Leaves less time to develop test vector suites
- Higher quality goals
- Requires more comprehensive testing of devices
- Increased gate count and decreased IO count
- More complex designs continue to reduce
controllability and observability of internal
nodes
42Exhaustive Testing
- Test every possible combination of inputs
- For hex inverter, requires 64 vectors (26)
- For 32 bit adder, requires 265 vectors
- At 1GHz 1170 years!
- For sequential circuits, the problem is worse
(vectors 2(nm) where ninputs, mflip-flops). - Clocks not counted since they are not logical
inputs - Grow exponentially worse with device complexity.
43How do you quantify non-exhaustive testing?
- Since exhaustive testing is impractical, develop
a vector set that seems comprehensive. - e.g. For 32 bit adder, add 1000 pairs of 32 bit
numbers. - How can you judge how well these vectors will
find defects? - How well will it find dead transistors, signals
shorted to Vcc or ground or other signals, open
connections, etc?
44Fault Grading
- Fault grading is the process of developing test
vectors and evaluating their effectiveness in
detecting manufacturing defects. - A measure of the goodness of the vector set
- The stuck at (s_at_) fault model is the most
popular model for evaluating vector sets. - Most defects can be modeled as a node s_at_1 or s_at_0
45Stuck _at_ detection
- To detect a s_at_1 fault
- Propagate a 0 to the fault location
- Propagate a difference in local output to primary
output - A s_at_0 is detected by propagating a 1 to the fault
location.
46Fault Grade Process
- Load device netlist into simulator
- Insert (seed) a logic fault in the circuit (short
node to power or ground) - Run logic simulation of vector set on faulty
circuit and good circuit in parallel. - If any of the primary outputs differ (1 vs. 0)
the fault is detected. - Remove seed from list and rerun with next seed.
- Number of fault detected divided by number seeded
is the fault grade for those vectors. - An 85 fault grade does not mean 15 of
real-world defects escape. - It means that 15 of single-node stuck-at faults
would be missed. - Other tests, such as Iddq, find other type of
faults missed by FG. - A defect is more likely to hit large-area
structures like IO buffers. - True outgoing defect rate must be measured and
correlated to FG.
47Drawbacks to Functional Testing
- Compute intensive (100k faults is common)
- Exponentially diminishing returns with each
vector - May get 50 FG with first vector set
- Doing any basic cycle will use most of devices
major functional blocks (Input buffers, Output
buffers, Control logic, Data paths, etc.) - Second vector set may add 15
- Third vector set may add 3
- By 20th vector set , may be adding less than 0.1
per vector - Methodology runs out of gas at 60-80 FG range.
- Large devices may well require over 80 FG to
meet DPM goals - May require person-years of senior engineer time
to write targeted vectors. - Methodology still requires additional DFT modes
- Counter dividers, RAM direct access testmode,
etc.
48Structural Stuck-at Testing
- Does not run device bus cycles
- These have already been done at silicon debug and
system validation testing - Tests that device was built correctly
- Tests the structure of the device, not the
function - Proves that device is good because
- Design has previously been proven correct.
- Structural testing demonstrates that all the
gates were made correctly and are connected
correctly. - Most common structural test method is scan testing
49Scan Test Concept
- Connect all flip-flops into serial chains by
converting each flop into a mux-flop - Serial scan mode selectable by scan enable
signal
50Scan test methodology
- To Test Sequential Logic (i.e. flip-flops)
- Place device into scan mode, feed unique data
into one end of chain and compare against data
coming out - To Test Combinatorial Logic
- Place device into scan mode
- Shift data into scan chains to preset entire
device - Drive Primary Inputs to known states
- Place device into normal mode and clock once
- Allows data to flow through combinatorial logic
and be captured in the next flop - Place device into scan mode.
- Shift data out and compare against known-good
results (next set of data is being shifted in at
same time) - Repeat Unload/Load - Capture - Unload/Load
sequence until desired fault coverage is attained.
51Scan Advantages / Disadvantages
- Advantages of Scan
- Provides very high level of observability/controll
ability - Provides high fault coverage (90 percent
achievable) and burn-in toggle coverage - Highly automated process
- Facilitates other techniques such as fault
isolation - Disadvantages of Scan
- Costs die size for gates/routing
- Offset by time-to-market, better quality, higher
BI yield, etc. - No extra die size needed if there is unused
whitespace - Adds delay to circuit speed paths
- Constraints on usable circuitry during netlist
synthesis - May not find paths that are functional, but slow
52Built-In Self Test
- Built-In Self Test (BIST) adds DFT circuits and
utilizes some of the existing circuits to test
the device. - For logic devices, scan circuitry, along with
pattern generation and pattern checking circuits
are added to allow the device to test itself. - For memory circuits, DFT circuitry is added to
cycle through addresses and perform read and
write functions.
53VLSI Processing Overview
- Typical Steps for a CMOS process
- Issues encountered in manufacturing
54Simplified Process Overview
- NoteThis list is greatly simplified and omits
many critical steps. It is for general conceptual
teaching purposes only. - Start with P- epitaxial layer (0.3mils) on P
substrate (30mils) - Nwell and Pwell creation
- Field Oxide growth
- Gate Oxide Growth
- Create PolySi Gates
- Create Source/Drain regions
- Open Contacts
- Create Via1, Metal1, Via2, Metal2, Via3, Metal3,
etc. - Create Passivation layer and Open Bond Pads
55Typical Steps for Wafer Processing
- Repetitive Lithographic Process
- Create layer to be patterned (e.g. oxide, metal,
polySi, polyimide) - Spin on photoresist
- Negative resist hardness in light. Faster, but
inferior line control. - Positive resist softens in light due to breaking
of polymer bonds. - Image the structures using a mask and develop
- Etch away unwanted material
- Clean (and possible planarize)
- Repeat for next structure.
562
1
3
4
Etch
5
Clean
57Typical Steps for Wafer Processing
- Used for
- Defining diffusion and ion implant areas to add
dopants to Si - Opening holes in oxide for contact and gate
regions - Shaping connectivity paths of metal and
polysilicon - Typical process can require 12-25 masks
58Si Cross Section
59Processing Techniques
- Oxide Growth
- Plasma or Wet Etching
- Diffusion
- Ion Implantation
- CVD - Chemical Vapor Deposition
- Evaporation
- Sputtering
- Planarization
60Oxidation
- Uses
- Gate insulation
- Can generate a high quality oxide
- Tends to be a slow process so thickness can be
tightly controlled - Diffusion mask
- Circuit passivation
- Created by exposing surface to O2 or H2O and high
temperature - Analogous to rusting of iron
61Etch
- Allows selective removal of material
- Wet acid etching is isotropic (all directions)
- Plasma etching is anisotropic.(vertical wall)
- In plasma etching, high energy plasma sputters
material from surface. - Etches are either timed or evaluated by end point
indicators.
62Etch
Wet Etch
Plasma Etch
63Diffusion
- One of two ways to introduce dopants in a
controlled way. - Relies on concentration gradient to induce flux.
- Semiconductor diffusion carried out at 900-1100C
- Typically use Boron to create P type and
Phosphorus or Arsenic to create N type. - Arsenic diffuses faster than Phosphorus.
- Constant source vs. Limited Source
64Diffusion
65Diffusion Profile
66Ion Implantation
- Second way to introduce dopants in a controlled
way - Ions of the dopant of interest are accelerated in
an electric field - These ions are then focused on the Si wafer.
- The depth these ions reach is dependent on their
energy and their angle to the lattice. - 50keV - 1MeV is typical
- Distribution is gaussian with a wider spread for
deeper implants
67Ion Implantation
68Implant Profile
69Ion Implantation
- Advantage
- Highly controlled vs. diffusion
- Easily masked
- Can form shallow junctions
- Doped regions can be buried (low-R regions)
- Problems
- Heavy damage to Silicon
- Highly peaked distribution
- Expensive
- Solution to damage and peaked distribution
- High temperature anneal (also activates
dopants) - Can be used for accurate Vt adjust
70CVD
- Chemical Vapor Deposition
- Deposits material on top of wafer
- Gas phase reaction e.g.
- SiH4 (Silane) O2 - SiO2 2H2
- SiH4 (at 650C) - Si 2H2
- Used for polySi (gate) or amorphous SiO2 (ILD)
- Typically does not create single-crystal silicon.
- Creates lower quality oxide than oxidation.
- Not used for gate oxide
71Evaporation
- Traditionally used to metalize wafers with
Aluminum - Aluminum is heated in a vacuum until it
vaporizes. - It then condenses on the wafer
- Inexpensive process but
- Step coverage can be a problem
- Not a very clean process
- Must break vacuum to change materials
72Evaporation
73Sputtering
- Uses a high-energy ion to sputter material from
a target to the Si substrate. - Used for substances with high vaporization
temperatures or alloys in which the elements have
greatly different melting points (e.g. TiN or
TiW) - More uniform and cleaner than evap but more
expensive. - Used for Aluminum for consistency and no need to
break vacuum between materials - Metal layers are typically a "sandwich of
multiple materials for better adhesion,
anti-reflectiveness, etc.
74Sputtering
75Planarization
- At a number of points in the manufacturing
process the wafer is ground flat or planarized. - This creates a flat field for
- imaging fine structures.
- evenly depositing material (e.g. metal)
76Si Cross Section
77Issues Classification
- Functionality
- Certain device functions do not work at any
speed. - Performance
- Certain device functions do not work at rated
speed. - Reliability
- Functionality or performance of device degrade
over time.
78Common Processing Problems
- Mask Registration
- Channel Length Variation
- Diffusion Profile (Bloating)
- Interconnect (Metal/Poly) Quality
- Dopant Concentration
- Gate Oxide Thickness
- Gate Oxide Quality (crystalline structure)
79Registration
- Each Layer is Constructed Using 1 or More Masks
- The Registration or Alignment Of Masks For
Different Layers Is Difficult - A misalignment of
- Can cause functionality, reliability and/or
performance problems.
80Example of Registration Error
Source Atlas of IC Technologies by W. Maly
81 Channel Length Variation
- Transistor Channel Length is a critical parameter
for performance - Ldrawn is the polySi gate length drawn at mask
design - Lexposed is the actual polySi gate length
manufactured on the wafer (a.k.a. poly CD) - Lexposed may be larger or smaller than Ldrawn
- Leff is device channel length after out-diffusion
(bloating) - Leff is always smaller than Lexposed.
- Lelectrical is the channel length after the
application of bias voltage on the transistor and
the resultant modulation of the depletion region - Lelec is always smaller than Leff
- Lelec channel modulation has a greater effects
on short-channel devices
82Diffusion Dimension Variation
- Difficult to fabricate exact Widths, Lengths and
Depths of features - Methods used for introducing dopants leads to
fuzzy boundaries - Further compounded by out-diffusion during future
thermal processing steps (bloating) - Can cause functionality, reliability and/or
performance problems.
83Example of Bloating
gate
Bloating of drain under gate
Photos from the Textbook Atlas of IC
Technologies by W. Maly
84Interconnect Quality
- Variations in metal width and thickness cause
variations in resistance. - Metal/Via/Metal connection problems increase
resistance. - Variations in inter-layer dielectric cause
variations in capacitance. - Topological considerations
- Stringers
- Step Coverage
85Topological Considerations
- Different Layers Of Material Placed Upon Each
Other - Each Layer Adds Topological Features
- Bumps
- Valleys
- Difficult To Maintain Constant Thickness
- Can cause functionality, reliability and/or
performance problems.
86Topological Example
Too Thin
Too Thick
ILD
Metal2
Metal1
87Stringers
- Conformal coating over vertical feature creates
extra thick coating. - Etching to remove nominal thickness may leave
material on these vertical walls. - This creates shorts between metal or poly lines
that route over these vertical features. - Shorts can be low or high resistance.
- Solutions
- Tune etch for each design
- Improve consistency of coating thickness
- Planarize each level
88Stringer Drawing
Metal2
ILD
Metal2
Metal1
Stringer
ILD
89Dopant Concentration
- Variations cause
- Shifts in Vt
- Shifts in sheet rho (resistivity)
90Gate Oxide Thickness / Quality
- Thickness
- Typical value 200A (i.e. 200 x10-10 meters)
- Variation can cause shifts in Vt
- Quality
- Interface charges
- Dangling bonds
- Trapped charges
- Affect Vt and/or produce leaky, deteriorating
gate junctions
91CMOS Function and Performance
- Review of MOSFET behavior
- VT
- CMOS performance
- Electromigration
- Hot Electrons
92VDS
VGS
Substrate
Inversion Layer
Gate
Drain
Source
N
N
P
Depletion Region
93Threshold Voltage (VT)
- The gate voltage at which strong inversion takes
place. - When VGS exceeds VT, significant current flows.
- VT is a function of
- Insulation thickness
- Channel Doping
- Gate insulation material
94Threshold Voltage
- It is important for VT to be the proper value.
- If it is too small, noise will cause device to
start conducting incorrectly and higher leakage
will result. - If it is too large, the device will not start
conducting until the input signal is a higher
voltage and will delay the output from the gate.
95Threshold Voltage
- Function Of
- Gate Capacitance
- Doping Concentrations
- May be modified with Ion Implantation
- Temperature
96CMOS Performance
- CMOS switching speed dependent on
- Drive capability of device
- Characteristics of load
- CMOS loads typically capacitive with series
resistance. - Capacitance comes from interconnect and from
gates of succeeding devices. - Resistance comes from interconnect R
- Drive capability comes from Rds(on) and Id(sat)
of transistors
97DC Performance
- For Given Voltages
- As W Increases Ids Increases
- As L Decreases Ids Increases
- Varies with Vt and somewhat with Vds
- Thinner tox increases Ids
- Affected by mobility
- Electrons have higher mobility than holes
98Resistance of a Material
- All Materials Used In CMOS Fabrication Have A
Resistive Impedance - The Amount of Resistance Depends On
- The Type Of Material
- The Shape Of The Material
- The Temperature Of The Material
99Resistance Estimation
Current
w
l
t
r Resistivity, a constant of given material
100Equivalence of Resistance
4w
4l
t
101Resistance Per Square
- Thickness of given material is defined when the
process is defined - Resistivity of each material is known
- Therefore L and W are the only variables
- Resistance quoted in W/Sq
- Material can therefore be measured in Squares
102Example Of Resistance Calculation
- Poly Silicon 20 W/Sq
- There are (50/10) 5 Squares Of Poly
- R 20(5) 100 W
W10 mm
L 50 mm
103Typical Resistance of Common Layers
- Poly 20 W/Sq
- Metal1 .07
- Metal2 .07
- Metal3 .04
- N and P Diff 25
- N-Well and Substrate 2000
- Resistors can be made from any of these components
104Decreasing Sizes Mean Higher Resistance
- As transistor sizes shrink
- Widths of metal interconnect decrease
- in proportion with transistor sizes
- Metal thickness decreases
- to allow for complete etch without stringers
- As devices get more complex with more
transistors, average length of metal traces
increases - Average interconnect resistance increasing
105Channel Resistance
0 VGS
-
Channel
L
106Estimating Channel Resistance
- Channel resistance can be estimated by
resistivity times of squares.
Channel Resistance
Resistivity of Channel Depth Of Channel
Oxide Capacitance per area
107Resistance In Vias
- Vias have smaller dimensions than minimum metal,
therefore minimum size vias have higher
resistance than a minimum metal wire - To compensate, multiple vias may be used.
108Capacitance Per Area
- Capacitance of Material To Substrate Can Be
Estimated By - Very Rough Estimate
- Does Not Include Fringing Effects
- Actual Capacitance Is Slightly Higher
109Typical Capacitances Over Field Oxide
- Layer attoF/mm2 (aF 1x10-18)
- Metal1 30
- Metal2 20
- Metal3 10
- Poly 50
- (C between layers is not accounted for here)
110Drain and Source Capacitances
Cgd
Cgb
Cgs
Cdb
Csb
Depletion Region
111Total Gate Capacitance
- Total Capacitance Is Sum Of All Components of
Capacitance - Cgb tends to dominate
- Remember Cg will depend on operating region FET
is in
112Electromigration
- When high density current flows through metal
interconnect, the electrons collide with the
metal atoms. - These collisions can cause the metal to migrate
and eventually create an open circuit. - This typically happens at imperfections or
discontinuities in the metal where current
density is the highest.
113Electromigration
114Hot Electrons
- With large enough electric fields, electrons
become hot (high kinetic energy). - Hot electrons impact the drain and dislodge holes
that show up as substrate current. - This is called impact ionization
- In addition, electrons can penetrate the gate.
- This can cause a Vt shift (reliability problem)
- Problem gets worse with shorter gate lengths
- Same Vcc over shorter distance higher field
strength - One reason for lower Vcc w/ smaller transistors
- Dopant profiles can be modified to reduce field
gradient
115Device Scaling
- Device scaling (shrinking) is a complex process
- IDSat, Vt, Resistance and Capacitance are all
changing - Some effects improve performance, some degrade
it. - Shorter Leff reduces Rds(on), increasing IDSat
- Same resistivity, shorter length higher I until
pinchoff - Also requires less die size per transistor
- Lower C on gate capacitance of loads
- Lower C, higher R on shorter, but thinner metal
interconnect - Balances out somewhat, but tends to increase RC
- Metal line loads becoming higher of total load
- Problem if simulators do not model interconnect
loads well
116Device scaling (cont.)
- Dopant concentrations change Vt and Rdson
- May be done to reduce hot e- effects
- Vcc may need to be reduced
- Reduced Vcc reduces power density and field
strength. - Lower Vcc decreases power exponentially (P C
Vdd2 freq) - Lower Vcc decreases hot e- effects.
- Lower Vcc also avoids punch-through, decreases
sub-threshold leakage. - Vcc reduction will affect performance
- Lower Vcc decreases Idsat and requires Vt
reduction - Vt reduction may require Tox reduction,
increasing Cgate - Not all features scale
- e.g. Vias may not be able to shrink and still
etch cleanly
117Why is Leff the dominant parameter in process
descriptions?
- Leff is the smallest dimension
- i.e. Most difficult to manufacture accurately
- Can not scale other structures w/o scaling Leff
- Smaller Leff provides better performance
- Increases IDsat
- Decreases Cgate
- Smaller transistors smaller die more die per
wafer and better yield percentage - Reduced Leff produces more good die on each wafer
with better performance!
118Parametric Performance
- Devices are designed to meet all performance
criteria. - Actual Performance may be reduced by global (e.g.
Leff) or point defect (e.g. metal particle)
effects. - Reduced performance due to global effects depends
on gaussian distribution of critical dimensions. - Reduced performance due to point defects depends
on random distribution of faults which produce
leakage currents, increased resistance /
capacitance or reduced Idsat.
119Leff, Voltage, Temperature Effects
- Device performance depends on moving charge
(electrons) in and out of load capacitance. - Larger Leff Lower Idsat slower switching
- Lower Vcc Lower Idsat slower switching
- Higher temperature Lower Idsat slower
switching - Slowest high Leff, low Vcc, high temperature
- Fastest low Leff, high Vcc, low temperature
- Devices must be simulated and tested at both
worst case conditions - Slow-corner testing is why overclocking sometimes
works - Too fast can be a problem (see hold time foil)
120Parametric Performance Analysis
- Overall device performance often falls into two
categories - Maximum clock frequency (Fmax)
- IO timing performance
- Both are based on Setup / Hold / Output_Valid
- Setup Time The amount of time a value must be
present and stable at an input before the clock
transition - Hold Time The amount of time a value must
remain present and stable at an input after the
clock transition. - Output_Valid (a.k.a. Tco) is the amount of time
from a clock edge until the output value changes.
It moves with Vcc, temp, Leff. - Max Valid Time The amount of time until the
output of a device has achieved its new value
after the clock transition - Min Valid Time The amount of time that an
output will retain its previous value after the
clock transition.
121 Setup / Hold / Min Max Valid
(Variation over Vcc and temperature)
O1
Clk
Tsu
Th
In
O2
In
O1
O3
O2
O3
O4
O4
MinV
MaxValid
Clk
122Source of Specifications
- The industry-standard PCI bus has the following
specifications - Tcycle (min) 30nS (i.e. 33MHz Fmax)
- Tsetup 7nS
- Thold 0nS
- TmaxV 11nS
- TminV 2ns
- Tprop(max) 10nS
- Tclkskew(max) 2nS
- Why?
123PCI Bus Timing
0
1
0
D1
Q1
D2
Q2
Clk(external)
Clk(ext)
Clk1
Clk2
Skew
D1
Q1
MaxV
D2
Tprop
Tsu
Q2
124Source of Specifications (cont.)
- Setup Time Equation
- Tcycle Tskew TmaxV Tprop Tsu
- Tprop may include delay through combinatorial
logic - Tprop on PC board traces 5 / nS.
- Fmax 1 / Tcycle for all paths
- e.g. 30nS Tcycle 33MHz max frequency
- Most circuits will have multiple paths using same
clock - If one path Tcycle 30nS, another path Tcycle
28nS and a third path Tcycle 31nS, Fmax will be
less than 33Mhz.
125Source of Specifications (cont.)
- What is Min Valid Time for?
- Provides hold time to the next input
- Hold Time Equation
- TminVTprop TclkskewThold for proper operation
- Tprop may be very close to 0nS (worst case)
- Thold is often 0nS or possibly negative!
- Note that clock skew makes both setup and hold
time equations more difficult to satisfy.
126MinV - Hold Time
0
1
1
Tclockskew 2nS
D1
Q1
D2
Q2
Tprop 0.1nS
0
1
Thold 0nS
Clk
Clk1
Clk1
Clk2
Clk2
D1
D1
Q1
Q1
TminV1nS
TminV4ns
D2
D2
Thold for 0
Q2
Q2
??
4nS min valid produces hold time
1nS min valid violates D2 hold time
127Setup and Hold Must be Met!
- Device Setup and Hold times must be met for
reliable operation!!! - Insufficient setup or hold time produce
metastability - Delayed maxV time, oscillatory outputs and/or
wrong output - Min Vcc, Hot is worst case for meeting setup
- Setup times get worse
- Max Output Valid times get worse (longer)
- Tprop through combinatorial logic gets worse
(longer) - Max Vcc, Cold is worst case for meeting hold
- Hold times get worse
- Min Output Valid times get worse (shorter)
- Tprop through combinatorial logic gets worse
(shorter) - Both scenarios must be simulated / tested
128DC Parametric Specs
- Icc standby - Supply Current draw of device in
one or more low-power states (e.g. suspend). - Icc active - Supply Current draw of device while
running its most compute intensive bus cycles. - IO leakage - Current that leaks out of inputs /
outputs when pin is in tristate condition. - Vol / Voh - Output low / high voltage when the
output is sinking / sourcing specified current - Vil / Vih - Voltage required for input to
recognize value as a logical one or zero. - Can all fail due to point defects or global
effects, point defects will usually get worse
over time.
129Input / Output Voltage
- Vil is maximum voltage guaranteed to be
recognized as 0 - Vol is maximum voltage outputs are allowed to
drive and still be seen as a 0 (uA or mA output
current also specd) - Vih is minimum voltage guaranteed to be
recognized as 1 - Voh is minimum voltage outputs are allowed to
drive and still be seen as a 1 (uA or mA output
current also specd) - Values are not the same to allow for noise.
Vcc
3.3V or 5V
Voh
2.4V
Vix/Vox values are industry-standard
TTL- compatible voltages for 3.3V or 5V
2.0V
Vih
Vil
0.8V
Vol
0.4V
Gnd
0V
130High-Volume Production Test
- Production test is used to eliminate
- Functional failures (Correct 1s and 0s go in and
out) - Parametric failures (AC and DC specifications)
- Some latent failures (e.g. aberrant input
leakage) - It is not used to validate design correctness.
- System validation confirmed the designs
correctness. - Test validates that each part works the same as a
known-good device. - Production test makes go / no-go decision.
- Uses hardware comparators to determine if an
output is at the correct voltage level at the
time it is sampled - No indication of margin to spec.
131Typical Test Program Flow
- Opens / Shorts (pull pin negative, check for
diode drop) - Basic Functional testing (slow clock, loose IO
timings and loose DC values). - Fault Grade vector testing (same conditions as
above, may be scan vectors or large set of
functional vectors) - Full Frequency Function (clock at full speed, all
else loose) - AC testing (tight IO timings, Vil/Vih/Vol/Voh
loose) - Vil / Vih testing (tight input voltage, all else
loose) - Vol / Voh testing (tight output voltage, all else
loose) - Icc dynamic testing
- Icc standby / Iddq testing
- I/O leakage testing
132Combining Tests
- Some tests in the flow may be combined
- e.g. Functionality testing at full frequency with
tight AC and tight Vil/Vih - Saves test time
- There are some down sides to combining tests
- Some tests wont run at full speed (e.g. DAT
modes) - Low yield analysis becomes very difficult
- How many parts failed for each type of test?
- Tests may interact negatively with each other
- e.g. High Iol/Ioh may cause ground bounce which
conflicts with tight AC strobing
133Typical Tester Architecture
- See other handout
- Overall tester architecture
- Vector memory contains
- Data, Drive (yes/no), Timing, Format, mask!
- e.g. 1_1_10010110_10_1
134DC current spec testing
- DC current specs are tested with A/D converters
in power supplies or PEC - Icc active tested during pattern execution
- Icc standby / Iddq tested after a pattern is run
to place device in suspend or Iddq mode - IO leakage tested after a pattern is run to place
device in All Z mode. - Proper selection of current clamp / measurement
range is important - Lower current range more accurate measurement
135Current Measurement Ranges
- Current measurements are mode by measuring IR
drop across series resistor - Resistor value is programmable
- Higher resistance value more resolution
- Too high resistance overanges the voltmeter
10 ohm 100mA max I, 100uA resolution 100ohm
10mA max I, 10uA resolution 1k ohm 1mA max I,
1uA resolution
5V
A to D
A to D converter (voltmeter) is 10 bit (i.e. 1024
steps) 0V 0000000000, 1V1111111111
(1Vmax)
Vcc
(DUT)
136Resolution vs. Accuracy
- Resolution is the value of the Least Significant
Bit of your measurement device (granularity) - /-1 pound from typical bathroom scale, /-1 sec
from wristwatch - Determined by number of bits or indicator marks
- Accuracy is how close the measured value is to
the actual value - Varies by manufacturer of measurement equipment
- Determined by quality of components and design
- Resolution typically cheaper than Accuracy
- 8 - 10 bit converter improves resolution
- Accuracy requires precision Rs, precision AtoD
converters, etc. - Resolution should be finer than accuracy
- /-10mA resolution on /-1mA accuracy meter
wastes accuracy
137Pin Electronics Cards
- See other foil for Pin Electronic Card (PEC)
architecture. - Fail strobe AND data mismatch AND not masked.
- Fail signal tells tester to stop any further
testing and send signal to handler to put device
into fail bin/tray.
138Tester Programmable Loads
- Tester loads are usually programmable
- Iol
- Ioh
- Vfloat
Vcc
Iol
P-channel
DUT pin
V
Vfloat
N-channel
Ioh
139Vector tests
- Device functionality and speed are tested by
applying vectors (0s and 1s) to the device
inputs and outputs - Inputs are driven by the PECs
- Outputs are compared at the PECs
- Outputs are compared when PECs are strobed
- If outputs are masked, strobes are ignored
Clk (in)
D (in)
Q (out)
Logical 1 on Q is compared against value
in vector memory only when strobe signal is pulsed
PEC Strobe
140Functional / Parametric Testing
- Same vectors with different PEC parameters are
used to perform functional and parametric testing.
0nS
100nS
0nS
100nS
Functional Test
Parametric Test
3.3V
3.3V
Pass
1.5Voh
1.5Voh
1.4Vol
1.4Vol