JEDEC

STANDARD

Methods for Calculating Failure Rates

in Units of FITs

JESD85

JULY 2001

JEDEC SOLID STATE TECHNOLOGY ASSOCIATION

 NOTICE

JEDEC standards and publications contain material that has been prepared, reviewed, and
approved through the JEDEC Board of Directors level and subsequently reviewed and approved
by the EIA General Counsel.

JEDEC standards and publications are designed to serve the public interest through eliminating
misunderstandings between manufacturers and purchasers, facilitating interchangeability and
improvement of products, and assisting the purchaser in selecting and obtaining with minimum
delay the proper product for use by those other than JEDEC members, whether the standard is to
be used either domestically or internationally.

JEDEC standards and publications are adopted without regard to whether or not their adoption
may involve patents or articles, materials, or processes. By such action JEDEC does not assume
any liability to any patent owner, nor does it assume any obligation whatever to parties adopting
the JEDEC standards or publications.

The information included in JEDEC standards and publications represents a sound approach to
product specification and application, principally from the solid state device manufacturer
viewpoint. Within the JEDEC organization there are procedures whereby an JEDEC standard or
publication may be further processed and ultimately become an ANSI/EIA standard.

No claims to be in conformance with this standard may be made unless all requirements stated in
the standard are met.

Inquiries, comments, and suggestions relative to the content of this JEDEC standard or
publication should be addressed to JEDEC Solid State Technology Association, 2500 Wilson
Boulevard, Arlington, VA 22201-3834, (703)907-7559 or www.jedec.org

Published by
JEDEC Solid State Technology Association 2001
2500 Wilson Boulevard
Arlington, VA 22201-3834

This document may be downloaded free of charge; however JEDEC retains the
copyright on this material. By downloading this file the individual agrees not to
charge for or resell the resulting material.

PRICE: Please refer to the current

Catalog of JEDEC Engineering Standards and Publications or call Global Engineering
Documents, USA and Canada 1-800-854-7179, International (303) 397-7956

Printed in the U.S.A.

All rights reserved
 PLEASE!

DON’T VIOLATE
THE
LAW!

This document is copyrighted by the JEDEC Solid State Technology Association

and may not be reproduced without permission.

Organizations may obtain permission to reproduce a limited number of copies

through entering into a license agreement. For information, contact:

JEDEC Solid State Technology Association

2500 Wilson Boulevard
Arlington, Virginia 22201-3834
or call (703) 907-7559
 JEDEC Standard No.85

Introduction

The calculation of failure rates is an important metric in assessing the reliability performance of a
product or process. This data can be used as a benchmark for future performance or an
assessment of past performance, which might signal a need for product or process improvement.
Reliability data are expressed in numerous units of measurement. This document uses the units
of FITs (Failures-In-Time), where one FIT is equal to one failure occurring in 109 device-hours.
These methods are assembled here to provide a reference to the way failure rates are calculated in
FITs.
2
Most of these methods only apply to constant failure rates. They assume that a χ distribution is
a reasonable approximation of the failure distribution over time. The examples given use failures
that exhibit an Arrhenius behavior. Other failure models can also be applied to these methods,
but care must be used for those models that do not simulate a constant failure rate.

Please note that the methods described in this document represent simplified ways of calculating
failure rates using various forms of data. More detailed methods exist in other publications, such
as EIA/JESD63 Standard Method for Calculating the Electromigration Model Parameters for
Current Density and Temperature and JESD37 Standard for Lognormal Analysis of Uncensored
Data and of Singly Right-Censored Data Using the Persson and Rootzen Method.

-i-
JEDEC Standard No. 85

-ii-
 JEDEC Standard No. JESD85
Page 1

METHODS FOR CALCULATING FAILURE RATES IN UNITS OF FITs

(From JEDEC Board Ballot JCB-01-02, formulated under the cognizance of the JC-14.3
Subcommittee on Silicon Devices Reliability Qualification and Monitoring.)

1 Scope

The methods described in this document apply to failure modes and mechanisms whose failures
exhibit a constant failure rate, e.g., an Arrhenius behavior characterized by an activation energy
for failure. If data on the distributions of failure with time exist, these activation energies can be
assumed from prior knowledge or failure analysis signatures. If the default activation energies
are not known or can't be determined, an activation energy can be used as a default. Refer to
JEP122 Activation Energies for Failure Mechanisms to obtain initial estimates.

The purpose of this method is to provide a reference to the way failure rates are calculated in
FITs.

2 Terms and definitions

probability density function of the time-to-failure; mortality function [f(t)]: The probability
of failure at a given instant of time t. F(t)dt is the probability of failure in the interval t to t + dt.

cumulative distribution function of the time-to-failure; cumulative mortality function

[F(t)]: The probability that a device will have failed by time t or the fraction of units that have
failed by time t. It is the integral of f(t).

cumulative reliability function [R(t)]: The probability that a device will function at time t or
the fraction of units surviving to time t. R(t) = 1 – F(t)

instantaneous (or hazard) failure rate [h(t)]: The rate at which devices are failing referenced
to the survivors. h(t) = f(t) / R(t)

cumulative hazard function [H(t)]: The integral of h(t).

FIT (failure in time): The number of failures per 109 device-hours. Typically used to express the
failure rate. Similarly, it can be defined as 1 PPM per 1000 hours of operation or one failure per
1000 devices run for one million hours of operation.

population failure distributions: The frequency of occurrence of failures over segments of

time. Typically seen failure distributions are Normal, Lognormal (wearout), Weibull and
Exponential.
JEDEC Standard No. JESD85
Page 2

2 Terms and definitions (cont’d)

bathtub curve: A plot of hazard rate versus time that exhibits three phases of life: infant
mortality (initially decreasing failure rate with time), intrinsic or useful life (relatively constant
failure rate) and wearout (increasing failure rate with time).

acceleration factor (AF): A multiplier relating failure times during an accelerated life test
versus those during useful life application conditions, assuming the same cumulative percent
failures and sigma for the two different stress conditions.

activation energy (EA): The excess free energy over the ground state that must be acquired by
an atomic or molecular system in order that a particular process can occur. Examples are the
energy needed by the molecule to take part in a chemical reaction, by an electron to reach the
conduction band in a semiconductor, or by a lattice defect to move to a neighboring site.

apparent activation energy (EA’): Activation energy that is calculated using the principles of
the physical relationship between stress and failure rate but is not directly related to a basic
change in physical processes. It may be based on too many possible physical “effects” that, when
stressed as a unit, produce a cumulative effect. It is similar to the concept of activation energy
but measures the probability of not exceeding some measurable attribute. A plot of the
reciprocal or absolute temperature (1/T(K)) versus the log of percent failed is linear.

3 Methods for calculating failure rates

The methods detailed below are arranged in an order such that an increasing amount of detail
regarding the failure data is available for each successive model

3.1 Case I: Single activation energy procedure for constant failure rate distributions

Assumptions:

• NO TEST INTERVALS (single read point of 2000 hours)

• UNKNOWN FAILURE MECHANISMS
• NOMINAL VOLTAGES DURING STRESS TEST

This method assumes knowledge of the number of failures in a known sample size from a life
test with a single stress temperature and single end point test. The failures have not been
segregated or analyzed to determine the applicable failure mechanism, so an apparent activation
energy is assumed for all failures.
 JEDEC Standard No. JESD85
Page 3

3.1 Case I: Single activation energy procedure for constant failure rate distributions
(cont’d)

3.1.1 Summarize the data

• f = # of failures = 15
• ss = sample size = 500
• T STRESS = stress temperature = 125 ºC
• T USE = use temperature = 55 ºC
• EA (avg) = apparent activation energy = 0.7 eV
• Very low power consumption during life test (negligible self heating)
• Apparent activation energy may or may not be an average.

3.1.2 Calculate the acceleration factor

Using the Arrhenius equation (1), calculate the acceleration factor for the failures at the given
stress and use temperatures.

 E    1   1   
AF = exp  A  ⋅    −   (1)
 k    TUSE   TSTRESS   

 0.7eV    1   1  
= exp  −5  ⋅    −    
 8.6 ⋅ 10 eV / K    (55 + 273)° K   (125 + 273)° K   

= 78.6

Usually devices draw power during life test. Junction temperature can be derived from the
ambient temperature, power dissipation and the thermal impedance, and should be used for
accurate estimation of the acceleration factor. The self-heating in a device is package-dependent.
The thermal dissipation property of a package (Theta ja, θja) determines the amount of self-
heating for the package. θja is a function of the air velocity around the package body. A good rule
of thumb is that θja is reduced by as much as 30% for moving air compared to still air. As an
example, if we were to assume that θja were 60 °C/watt, the power dissipation were 0.1 watt at
125 °C, and 0.12 watts at 55 °C, then there would be 7.2 °C (60*0.12) difference between
junction temperature and ambient at 55 °C and 6 °C (60*0.1) difference at 125 °C. Thus, the
acceleration factor would be calculated based on 131 °C as stress temperature vs. 62.2 °C as the
actual use temperature. The net effect of the self-heating would reduce acceleration factor to
62.5, compared to 78.6 calculated earlier without a self -heating consideration.

Some mechanisms such as oxide and inter-layer dielectrics may also be accelerated due to the
applied electric fields. Appropriate models may be used to calculate acceleration due to voltage
applied. Refer to JEP122 for the most commonly used models.
JEDEC Standard No. JESD85
Page 4

3.1 Case I: Single activation energy procedure for constant failure rate distributions
(cont’d)

3.1.3 Calculate the point estimate of the failure rate

f ⋅ 10 9
λPOINT =
t ⋅ ss ⋅ AF

15 ⋅ 10 9
=
2000h ⋅ 500 ⋅ 78.6

= 190.84 ~191 FITs

3.1.4 Calculate the upper confidence bound of the failure rate

χ 2 % CL, 2 f +2 ⋅ 109
λCL =
2 ⋅ t ⋅ ss ⋅ AF

33.4 ⋅ 10 9
λ60% CL = χ2 for 15 failures is 33.4 for 60% confidence)
(χ
2 ⋅ 2000h ⋅ 500 ⋅ 78.6

= 212.46 ~ 212 FITs

42.6 ⋅ 10 9
λ90% CL = χ2 for 15 failures is 42.6 for 90% confidence)
(χ
2 ⋅ 2000h ⋅ 500 ⋅ 78.6

= 270.99 ~271 FITs

3.2 Case II: multiple activation energy procedure for constant failure rate distributions

Assumption:

• NO TEST INTERVALS (single read point of 2000 hours)

• KNOWN FAILURE MECHANISMS
• NOMINAL VOLTAGES DURING STRESS TEST

For this method, the data set is the same as in section 5.1, but now it is known that the devices
failed due to several different failure mechanisms to which can be assigned the appropriate
activation energies.
 JEDEC Standard No. JESD85
Page 5

3.2 Case II: Multiple activation energy procedure for constant failure rate distributions
(cont’d)

3.2.1 Summarize the additional data available

• FM#1 from 3 failures at EA = 0.5 eV

• FM#2 from 5 failures at EA = 0.7 eV
• FM#3 from 7 failures at EA = 1.0 eV

Total # of failures = 15

3.2.2 Calculate the acceleration factors

Calculate the acceleration factors for each failure mechanism (known activation energy) at the
given stress and use temperatures using the Arrhenius equation (1)

• AF(FM#1) 0.5 eV = 22.6

• AF(FM#2) 0.7 eV = 78.6
• AF(FM#3) 1.0 eV = 510

3.2.3 Calculate the failure rates

Calculate the point estimate of the failure rate for each failure mechanism and add to obtain the
point estimate of the total failure rate.

f ⋅ 10 9
λPOINT =
t ⋅ ss ⋅ AF

3 ⋅ 10 9
λ FM #1 = (0.5 eV)
2000h ⋅ 500 ⋅ 22.6

= 132.74 ~133 FITs

5 ⋅ 10 9
λ FM # 2 = (0.7 eV)
2000h ⋅ 500 ⋅ 78.6

= 63.6 ~64 FITs

7 ⋅ 10 9
λ FM #3 = (1.0 eV)
2000h ⋅ 500 ⋅ 510

= 13.7 ~14 FITs

JEDEC Standard No. JESD85
Page 6

3.2 Case II: Multiple activation energy procedure for constant failure rate distributions
(cont’d)

3.2.3 Calculate the failure rates (cont’d)

n
λTOTAL, POINT = ∑ λFM #
1

= 132.74 + 63.6 + 13.7

= 210.4 ~210 FITs

3.2.4 Calculate the upper confidence bound of the total failure rate

 χ%2 CL , 2 f +2 
λTOTAL,CL = λTOTAL, POINT ⋅ 
 2f 

210 ⋅ 33.4
λTOTAL, 60% CL =
30

= 233.8 ~234 FITs

210 ⋅ 42.6
λTOTAL ,90% CL =
30

= 298.2 ~298 FITs

 JEDEC Standard No. JESD85
Page 7

3 Methods for calculating failure rates (cont’d)

3.3 Case III: Multiple activation energy procedure for constant failure rate distributions
with known test interval

For this method, the data set is the same as in sections 5.1 and 5.2, but the stress was now
performed using different test intervals to determine when each of the failures occurred. This
type of detail enables a calculation of the average failure rate within a given time interval. The
early life, or nonconstant failure rate, segment can be calculated using a Weibull distribution
(m < 1). Make an assumption about where the constant or steady state failure rate time interval
begins. This time interval should be with respect to the use conditions, not the test time, since
different failure mechanisms exist in the failures. Note that, if the failure mechanisms were
unknown, a steady state time interval for the constant failure rate in terms of the stress time could
be used. The time period before the beginning of the time interval is assumed to represent a
nonconstant failure rate (e.g., infant mortality-type failures) and can be represented by a Weibull
distribution (m < 1). For this example, we assume our application displays a breakpoint between
early life and inherent life at 10,000 hours use time.

0hr < tUSE,WEIBULL < 10,000 h nonconstant failure rate (infant mortality)

tUSE,EXPONENTIAL > 10,000 h constant failure rate (operational life)

It is assumed that this particular life test does not stress the devices to the wearout portion of their
lifetime. If the data shows a nonconstant (increasing) failure rate before the end of the stress
time, then the failure rate for that segment can be calculated using the method described below
for a Weibull distribution (m > 1).

Assumption:

• KNOWN TEST INTERVALS

• KNOWN FAILURE MECHANISMS
• NOMINAL VOLTAGES DURING STRESS TEST

3.3.1 Summarize the additional data available

Stress Time
48 h 168 h 500 h 1000 h 2000 h

Acceleration Sample Size
500 497 495 493 490
Factor # Failed
3 2 2 3 5
22.6 FM # 1 0.5 eV 0 0 1 0 2
78.6 FM # 2 0.7 eV 1 0 0 2 2
510 FM # 3 1.0 eV 2 2 1 1 1
JEDEC Standard No. JESD85
Page 8

3.3 Case III: Multiple activation energy procedure for constant failure rate distributions
with known test interval (cont’d)

3.3.2 Separate the data into early life and inherent life periods

The field equivalent hours of 10,000 (depending on the application, different times may also be
used) is used to find the equivalent time for each mechanism as follows:

0.5 eV = 10,000/22.6 = 177 hours

0.7 eV = 10,000/78.6 = 50.9 hours
1.0 eV = 10,000/510 = 7.8 hours

This means that the first two read points of 48 and 168 hours for the 0.5 eV belongs in the early
life period. The first read point of 48 hours read point for the 0.7 eV group also belongs in the
early life period while all the rest of the read points will belong to the inherent life group (the
constant or steady state failure rate). The early life data points are highlighted in the table in
3.3.1. This data is rearranged separating the early life portion from the inherent life for ease of
illustration.

3.3.3 Calculate the early life failure rate

For the use condition time interval between 0 and 10,000 hours, assume a Weibull distribution
with a declining failure rate to reflect the infant mortality portion of the lifetime bathtub curve
(m < 1). Calculate a failure rate for a given stress time interval inside the nonconstant failure rate
interval.

22.6 FM # 1 0.5 eV Time
48 168

Sample Size
500 497
# Failed
0 0

78.6 FM # 2 0.7 eV Time
48

Sample Size
500
# Failed
1

Since there is only one failure in the early life period, the Weibull distribution can not be applied
to this data set. Also, calculating the failure rates based on more then two read points minimize
the error in estimate. However, with very few failures and large sample sizes, we may use the
exponential model to estimate the early life failure rate.
 JEDEC Standard No. JESD85
Page 9

3.3 Case III: Multiple activation energy procedure for constant failure rate distributions
with known test interval (cont’d)

3.3.3 Calculate the early life failure rate (cont’d)

0.6 ⋅ 10 9
λ FM #1 = (0.5 eV) (for zero failure, substitute .6 for 50% confidence)
168h ⋅ 500 ⋅ 22.6

= 316.05 ~316 FITs

1 ⋅ 10 9
λ FM # 2 = (0.7 eV)
48h ⋅ 500 ⋅ 78.6

= 530.11 ~530 FITs

Total = 316.05 + 530.11 = 846.16 ~ 846 FITs

3.3.3.1 Use of Weibull distribution

Consider a test where the activation energy for the observed failure mechanism and test
conditions result in an acceleration factor of 100. The test started with a sample of 500 pieces
and one failure occurred at 48 hours. The remaining sample was continued to 96 hours wherein
one additional failure occurred.

The cumulative failure values are thus:

F(48) =[1-(499/500)] = 0.002

F(96) = {1-[(499/500)*(498/499)]}={1-(498/500)}=2/500 = 0.004

3.3.3.2 Apply test parameters to Weibull distribution

The basic Weibull parameters are "m," the shape parameter and "c," the characteristic life. m
is defined as follows:

m = ln[-ln{1-F(t)}] / [ln(t) - ln(c)] (2)

/
= ln[-ln{(1-0.00200)}] [ln(48) -ln(c)] = -6.21361 /[ 3.8712 -ln(c)]

= ln[-ln{(1-0.0040)}] / [ln(96) -ln(c)] = -5.51946 / [4.5643 - ln(c)]

JEDEC Standard No. JESD85
Page 10

3.3.3 Calculate the early life failure rate (cont’d)

3.3.3.3 Solving for ln(c)

ln(c) = 10.076

and c= 23762 hours.

3.3.3.4 Then solving equation [2] for m

m = 1.001446

3.3.3.5 The Weibull equation for instantaneous failure rate

h(t) = (m/c) * (t/c)m-1

For t = 1 hour,

h(t) = 41,535 FITs

With very few failures, and m closer to 1, the Weibull distribution reduces to the exponential
form and is a better estimate for failure rates.

2 ⋅ 10 9
λ =
96h ⋅ 500 ⋅ 100

= 416.67 ~417 FITs

3.3.4 Calculate the inherent life failure rate

3.3.4.1 For 0.5 eV mechanism

22.6 FM # 1 0.5 eV Time
1085 3797 11,300 22,600 45,200

Sample Size
500 497 495 493 490
# Failed
0 0 1 0 2

Since 11,300 hours is greater then 10,000 hours for the early life period with one failure and that
there was no failure at the 3797 read point, we do not know if this failure occurred within the
early life period or beyond. Therefore, the failure rate would be averaged beyond the early life
period of 3797. If we assume a constant failure rate between 3797 and 11,300 hours, the
likelihood of the one failure occurring after the early life breakpoint of 10,000 hours is:
 JEDEC Standard No. JESD85
Page 11

3.3.4 Calculate the inherent life failure rate (cont’d)

3.3.4.1 For 0.5 eV mechanism (cont’d)

11,300 − 10,000
a= = 017
.
11,300 − 3797
and f = 2+.17= 2.17

f ⋅ 109
λPOINT =
ss ⋅ tUSE , INTERVAL

3 ⋅ 109 2.17 ⋅ 10 9
λ0.5eV = λ0.5eV =
500 ⋅ ( 45200 − 11300) to 500 ⋅ (45200 − 10000)

= 123.29 ~123 FITs

3.3.4.2 For 0.7 eV mechanism

78.6 FM # 2 0.7 eV Time
3773 13,205 39,300 78,600 157,200

Sample Size
500 497 495 493 490
# Failed
0 0 0 2 2

In the case with 0.7 eV, there are no failures beyond early life and the first read point in inherent
life, thus the failure rate (4 fails) will be averaged between 13,205 and 157,200 hours.

4 ⋅ 10 9
λ0.7 eV =
500 ⋅ (157200 − 13205)

= 55.56 ~56 FITs

JEDEC Standard No. JESD85
Page 12

3.3.4 Calculate the inherent life failure rate (cont’d)

3.3.4.3 For 1.0 eV mechanism

510 FM # 3 1.0 eV Time
24,480 85,680 255,00 510,000 1,020000

Sample Size
500 497 495 493 490
# Failed
2 2 1 1 1

In the case with 1.0 eV, there are 2 failures that could have occurred during early life or inherent
life. Since the inherent life period is assumed to have a constant failure rate and we have more
than one data point in this region to use, the first two failures (shaded in the table above) can be
ignored. Thus the failure rate (5 fails) will be averaged between 24,480 and 1,020,000 hours.

5 ⋅ 10 9
λ1.0eV = = 10.05 ~10 FITs
500 ⋅ (1020000 − 24480)
The total failure rate in the constant failure rate time interval is the sum of all individual failure
rates:
n
λTOTAL, POINT = ∑ λFM #
1

= 123.29 0.5ev + 55.6 0.7ev + 10.05 1.0ev = 188.94 ~189 FITs

The upper confidence bounds can be obtained using the same procedure shown in the previous
methods. A comparison of the results of the first three cases using the same data set can be
found in Annex 3.

3.4 Case IV: Different test conditions (truncated data)

This method is used when dealing with censored (OR Type II) data, which is data that is obtained
after a truncation of a test either after a given number of failures or after a certain period of time.
Other types of censoring include interval (data recorded only during certain intervals) and
suspended (functioning parts removed during life test) censoring. The latter requires the use of
the Kaplan-Meier method of data treatment by collecting the relative (vs. absolute) failures. This
will be illustrated in the following example.

Here, there are three more sets of data with different rates and occurrences of failures, and
stressed at different temperatures so as to produce different accelerations. Thus, different
termination times are produced when calculated for use times. For simplicity, we assume in the
example below the same activation energy (0.7 eV) for all the failures.
 JEDEC Standard No. JESD85
Page 13

3 Methods for calculating failure rates (cont’d)

3.4 Case IV: Different test conditions (truncated data) (cont’d)

Assumptions:

• KNOWN TEST INTERVALS

• UNKNOWN FAILURE MECHANISMS
• NOMINAL VOLTAGES DURING STRESS TEST
• MULTIPLE STRESS TEMPERATURES

3.4.1 Summarize the additional data available

There are four sampling of 500 devices each. The default activation energy for all failures is
assumed to be 0.7 eV. Failures after each read-point for each stress temperature are shown
below. An "x" indicates that a read-point was not taken for that sampling.

sampling TSTRESS 48 h 168 h 500 h 1000 h 2000 h

O
1 100 C 2 1 2 2 x
O
2 125 C 3 2 2 3 5
O
3 125 C 2 2 3 x x
4 150 OC 4 3 2 3 2

3.4.2 Convert stress times to equivalent use times

Using the Arrhenius equation (0.7 eV for activation energy), convert the stress times to
equivalent use times (use temperature of 55 °C) as summarized in the table below. The numbers
in parentheses ( ) below are the ordering of the use hours from shortest to longest.

TSTRESS AF 48 h 168 h 500 h 1000 h 2000 h

o
100 C 20.0 960 (1) 3360 (2) 10000 (4) 20000 (7) 40000 (9)
125 oC 78.6 3773 (3) 13205 (6) 39300 (8) 78600 (11) 157,200 (13)
o
150 C 263 12624 (5) 44184 (10) 131,500 (12) 263,000 (14) 526,000 (15)
JEDEC Standard No. JESD85
Page 14

3.4 Case IV: Different test conditions (truncated data) (cont’d)

3.4.3 Order the failure occurrences by equivalent use time

If we make the optimistic assumption that the failures occurred at the end of the test interval, we
obtain the following failure distribution extrapolated to useful life conditions. The "-" indicates
no reading taken at that equivalent use time.

a sampling(s)=> 1 2 3 4 f = total ΣF=cum ΣF%= n=pop.

use hours: fails@a fails % Fail
1 960 2 - - - 2 2 0.10 2000
2 3360 1 - - - 1 3 0.15 1998
3 3773 - 3 2 - 5 8 0.40 1997
4 10000 2 - - - 2 10 0.50 1992
5 12624 - - - 4 4 14 0.70 1990
6 13205 - 2 2 - 4 18 0.91 1986
7 20000 2 - - - 2 20 1.01 1982
8 39300 x 2 3 - 5 18 1.21 1487
(20+5-7) (1982-500+7-2)
9 40000 x - x - 0 11 1.11 989
(18+0-7) (1487-500+7-5)
10 44184 x - x 3 3 14 1.42 989
11 78600 x 3 x - 3 17 1.57 986
12 131,500 x - x 2 2 19 1.93 983
13 157,200 x 5 x - 5 24 2.45 981
14 263,000 x x x 3 3 12 2.44 491
(24+3-15) (981-500+15-5)
15 526,000 x x x 2 2 14 2.87 488

An “x” indicates that there were no more components from a specific sample in the test.

NOTE For sampling # 1, the last read point was 20,000; for sampling # 2, the last read point
was 157,200; for sampling # 3, the last read point was 39,300 and for sampling # 4, the last
read point was 526,000 (indicated by shading in the appropriate field).
Refer to Annex 1 for sample calculation of cum failures.

F = total number of failures tested to that specific use hour point.

ΣF = cumulative number of failures seen up to that specific use hour point.
ΣF% = ΣF/n = cumulative % failures up to that specific use hour point.
N = the number of functional components at the start of an interval (population).

[ ]
na = na −1 − ns − ∑ f s − f a −1 (3)
Where: a = row number and s = sampling column
 JEDEC Standard No. JESD85
Page 15

3.4 Case IV: Different test conditions (truncated data) (cont’d)

3.4.4 Perform conventional analysis

3.4.4.1 Calculate early life failure rate

Assuming, as in 5.3, that the early life failure rate time interval is in the range between 0 and
10,000 hours of use condition, there are 10 cumulative failures from the four sampling:

∑ f ⋅ 10 9

λELFR = tINTERVAL
ss ⋅ t INTERVAL

10 ⋅ 109
=
2000 ⋅ 10000

= 500 FITs

The upper confidence bounds can be obtained using the same procedure shown in the previous
methods.

3.4.4.2 Calculate intrinsic failure rate

The average failure rate above 10,000 hours (or intrinsic failure rate) under use conditions
has 43 cumulative failures and 10 of these occurred at 10,000 hours or before:

∑ f ⋅ 10 9

λINTRINSIC = tINTERVAL

∑ ss ⋅ (t )
a

a STRESS ,a − t ELFR
SAMPLING =1

33 ⋅ 109
=
497 ⋅ (157200 − 10000) + 498 ⋅ ( 39300 − 10000) + 500 ⋅ ( 526000 − 10000) + 495 ⋅ ( 20000 − 10000)

= 92.92 ~93 FITs

JEDEC Standard No. JESD85
Page 16

3.4.4 Perform conventional analysis (cont’d)

3.4.4.3 Recalculate cumulative distribution function

Using the Nelson or Stolze methods and preparing the data using the Kaplan-Meier
method (decreasing sample size during the test, see appendix-II for example) we recalculate the
cum % failing (CDF) with the adjusted sample size as indicated in the table below:

H survival p R F% # F = # Failed
use hours: f=total ss f/ss p =1-f/ss ∏ (p)
1-∏ CDF100*(1-R) CDF *ss Integer
1 960 2 2000 0.00100 0.999000 0.999000 0.10 2.0 2
2 3360 1 1998 0.00050 0.999499 0.998500 0.15 3.0 3
3 3773 5 1997 0.00250 0.997496 0.996000 0.40 8.0 8
4 10000 2 1992 0.00100 0.998996 0.995000 0.50 10.0 10
5 12624 4 1990 0.00201 0.997990 0.993000 0.70 13.9 14
6 13205 4 1986 0.00201 0.997986 0.991000 0.90 17.9 18
7 20000 2 1982 0.00101 0.998991 0.990000 1.00 19.8 20
8 39300 5 1487 0.00336 0.996638 0.986671 1.33 19.8 20
9 40000 0 989 0.00000 1.000000 0.986671 1.33 13.2 13
10 44184 3 989 0.00303 0.996967 0.983678 1.63 16.1 16
11 78600 3 986 0.00304 0.996957 0.980685 1.93 19.0 19
12 131,500 2 983 0.00203 0.997965 0.978690 2.13 20.9 21
13 157,200 5 981 0.00510 0.994903 0.973702 2.63 25.8 26
14 263,000 3 491 0.00611 0.993890 0.967752 3.22 15.8 16
15 526,000 2 488 0.00410 0.995902 0.963786 3.62 17.7 18

3.4.4.4 Conduct regression analysis

The cum F% vs. tUSE values can then be fitted onto a log-log or log-normal plot using linear
regression. The resultant regression analysis will give the slope and intercept parameters. Care
should be taken in applying these methods as very large values of T50 may result from a poor fit
to the data. The data table is further simplified by eliminating the 40,000 read-point as there were
no failures. A useful tool is to find the saturation point of failure in time (linear plot of %f vs
time, in this case fsat~4%) and then normalizing the cum failures to this value. This type of
analysis identifies the defective sub-population from the main distribution.
 JEDEC Standard No. JESD85
Page 17

3.4.4 Perform conventional analysis (cont’d)

3.4.4.4 Conduct regression analysis (cont’d)
Data Defective Sub-Population Standard Analysis
Use hours (t) Ln (t) Cum%f x 100 Z(def subpop) Cum % fail Z(cum%fail)
[ordinate] fsat [abscissa] [abscissa]
960 6.87 3% -2.0 0.10% -3.09
3360 8.12 4% -1.8 0.15% -2.97
3773 8.24 10% -1.3 0.40% -2.65
10000 9.21 13% -1.2 0.50% -2.58
12624 9.44 18% -0.9 0.70% -2.46
13205 9.49 23% -0.8 0.90% -2.37
20000 9.90 25% -0.7 1.00% -2.33
39300 10.58 33% -0.4 1.33% -2.22
44184 10.60 41% -0.2 1.63% -2.14
78600 10.70 48% 0.0 1.93% -2.07
131500 11.27 53% 0.1 2.13% -2.03
157200 11.97 66% 0.4 2.63% -1.94
263000 12.48 81% 0.9 3.22% -1.85
526000 13.17 91% 1.3 3.62% -1.80

3.4.4.5 Effect on plot using standard inverse CDF

The following plots illustrate the significance of graphical analysis and the defective sub-
population. If we perform a standard analysis by plotting ln t vs. cum f inv CDF:

lo g n o rmal p lo t Data A

2 0 .0 0

y = 4 .3 1 0 6 x + 2 0 .1 4 2
ln (field time in hr)

R 2 = 0 .9 4 4 1 5 .0 0

1 0 .0 0
la rg e s ig m a s u g g e s ts
m u ltip le p o p u la tio n s 5 .0 0

0 .0 0
-4 .0 0 -3 .0 0 -2 .0 0 -1 .0 0 0 .0 0
Inv CDF (c um e %)

In the log-normal plot of cum defective versus log time, we obtain a straight line with reasonable
fit as indicated by the R2 of 0.944 from the regression analysis. However, the large sigma of 4.31
indicates that there are multiple mechanisms present. The same data, when normalized to a
saturated cum failure rate of 4% (saturation point from linear plot of CDF vs time) will also yield
a straight line and a more reasonable values of sigma and T50.
JEDEC Standard No. JESD85
Page 18

3.4.4 Perform conventional analysis (cont’d)

3.4.4.6 Plot of defective subpopulation on inverse CDF

If we use only the defective subpopulation in plotting ln t vs. def sub pop inv CDF:

lognormal Defective Subpopulation

16.00
y = 1.8321x + 11.006 14.00
ln (time in field hr)

2
R = 0.9729 12.00
10.00
Note the sigma is now
8.00
a reasonable 1.83, w here
6.00 value of 1-2 means def ects,
4.00 w hile ~0.5 means w earout
2.00
0.00
-3.0 -2.0 -1.0 0.0 1.0 2.0
Inv CDF (cume %)

In this graph, the R2 value (0.973) indicates a better fit then before (0.944) with more reasonable
values of sigma (1.83) and T50 [Exp(11.006)]. The T50 values are 6x104 (for defective sub-
population) compares to 5.6x108 (normal plot). Also note that a sigma of 1 to 2 usually indicates
defective sub-population while a sigma of less then 1 indicates a wear-out phenomenon.

4 References

1. JEP122, Failure Mechanisms and Models for Silicon Devices

2. JESD63, Standard Method for Calculating the Electromigration Model Parameters for
Current Density and Temperature

3. James R. King, (1981), Probability Charts for Decision Making, 2nd Printing, TEAM,
Tamworth, New Hampshire

4. Paul A. Tobias and Dave Trindade, (1986), Applied Reliability, Van Nostrand Reinhold, New
York
 JEDEC Standard No. JESD85
Page 19

Annex A

NOTE 1 Line 14-8 in above table reduced the cum failures to 18 (cell J14). The figure below illustrates
the formula used to calculate the cum failure. There were 20 cum failure at prior read point to which 5
new failures are added (E14+F14). Since group 1 has ended the stress test, the cum failures for this group
(D7+D8+D10+D13) are removed from the cum count (=20-7+5=18).

Similarly, the figure below illustrates the total sample size (cell O14) reduced by sample size
from group-1 (493) and by the failure encountered in immediate prior read point (2).
JEDEC Standard No. JESD85
Page 20

Annex B

Kaplan-Meier Approximation (estimator) (4, pp124-126)

A procedure to calculate survival probability estimates at various times. It allows calculations of

reliability statistics with multi-censored data. Basically, the probability of a component surviving
during an interval of time is estimated from the observed fallout of units starting the interval. The
estimates at each time interval are multiplied together to provide a survival or probability
estimate from time zero. The cumulative distribution function can be estimated at each failure
time by subtracting the survival values from 1.0. Following example illustrates this method:

1 Small Sample size

Time Sample Failure Removed Survival Survival Estimates Survival

Size for probability Estimates
analysis
200 8 1 1 7/8 7/8 0.875
300 6 1 5/6 (7/8)*(5/6) 0.729
400 5 1 1 4/5 (7/8)*(5/6)*(4/5) 0.583
500 3 1 2/3 (7/8)*(5/6)*(4/5)*(2/3) 0.389
600 2 1 1/2 (7/8)*(5/6)*(4/5)*(2/3)*(1/2) 0.194

Time Survival CDF Cum Failures

Estimates 1-Survival CDF
200 0.875 0.125 1.25%
300 0.729 0.271 2.71%
400 0.583 0.417 4.17%
500 0.389 0.611 6.11%
600 0.194 0.806 8.06%

2 Example using data from mechanism # 3 and for 1.0 eV

Time # Failed # Removed Sample Size Survivor Cum p CDF

Of survival
24480 2 1 500 0.996000 0.996000 0.400000
85680 2 0 497 0.995976 0.991992 0.800805
25500 1 1 495 0.997980 0.989988 1.001207
510000 1 2 493 0.997972 0.987980 1.202016
1020000 1 4 490 0.997959 0.985964 1.403645
 Annex C Comparison of calculations
Minimum Information Known mechanism Known mechanisms and time interval
F =15 f = 3,5, 7 Stress 48 168 500 1000 2000
Ss=500 Time

Ts = 125 °C Ea = 0.5, 0.7, 1.0 Acceleration Sample 500 497 495 493 490
Tu = 55 °C t=2000 hours Size

Ea = 0.7 ev (Assumed) Factor # Failed 3 2 2 3 5
t=2000 hours

22.6 FM # 1 0.5 ev 0 0 1 0 2
78.6 FM # 2 0.7 ev 1 0 0 2 2
510 FM # 3 1.0 ev 2 2 1 1 1
AF = 78.6 AF = 22.6, 78.6, 510
3 ⋅ 109 2.17 ⋅ 10 9
λFM #1 = (0.5 ev) λ0.5 eV =
2000hr ⋅ 500 ⋅ 22.6 500 ⋅ (45,200 − 10,000)

= 133 FITs = 123 FITs

15 ⋅ 109 5 ⋅109 4 ⋅109

= λFM # 2 = λ0.7 eV =
2000hr ⋅ 500 ⋅ 78.6 2000hr ⋅ 500 ⋅ 78.6
(0.7 ev) 500 ⋅ (157,200 − 13205)
= 191 FITs = 63.6 FITs = 55.6 FITs

7 ⋅109 5 ⋅ 109
λFM #3 = λ1.0 eV =
2000hr ⋅ 500 ⋅ 510
(1.0 ev) 500 ⋅ (1,020,000 − 24,480)

= 10.05 FITs
= 13.7 FITs
n
n λTOTAL, POINT = ∑ λFM #
λTOTAL,POINT = ∑ λFM # 1

JEDEC Standard No. JESD85

1
= 123 0.5ev + 55.6 0.7ev + 10.05 1.0ev
= 133 + 63.6 + 13.7
= 191 FITs = 210 FITs = 188.65 FITs

Page 21
JEDEC Standard No. JESD85
Page 22