Digital Equipment Corporation CMOS semiconductor, Alpha 21064, NVAX microprocessor Technical journal
PDF
문서
광고
광고
Semiconductor Technologies
Digital Technical Journal
Digital Equipment Corporation
Volume 4
Number 2
Spring 1992
Editorial
.J:ttll' C. BlakL·. Ediwr
Helen 1..
Pa t n: r,on..\S>ociate Editor
M. St et son . ,\s">eiatt· Editor
K:nhleen
Circulation
C:lllll'rinc .VI . l' hill ips, Administrator
Shen·y
L.
Gonzalez
P.-oduction
Mildred R . Rosenzweig, Production Editor
Ma rga ret 1..
Burdine. 'J)'pographer
Illustrator
Pt:ter R. Woodbury.
Advisory Board
Samuel H. fuller, Chairman
Hichard \V Beane
Robert M. Glorioso
Richard]. Hull ingsw orth
"'emeth
Ai:ln G.
Victor A. V1·ssotsky
Gam
The
B.
Winters
/Jigilaf Ti!clllliatfjullmal i' pu blish e d quartnll' 111'
D ig ital
J-i6 :\- l ain Streel \ l l l l l .V BI>H � la yna rd .
.\l:l"achu"·tt' fll7'i i-2'>�1. �ub,cription' to thc.fu/1/'llltl arcS iO.OO
l<>r lin1r i"u" and must he p re paid in {I.S. funds. lnilersit\· and
college prufc"or' and l'll.D. students in th<: electrical engineering
and computer scic.:nce tields n.:-cci,·t: complill ll' llt ar y suh:o-.LTiption:-,
upon request. Ord er s. inquiries. ami addrc" cltang" should he
Equipment Corporation.
sent to tltc Oigilal
. ·
.
-
.
h·cblticalfoltl'l[{t{ :It the puhlishnl-lw addres,_
Inquiries can also be sent
ekctronic:llly
toiJ'Ijlu:<:t!I..I>H:.Co.\1.
Single eopic,; and back issues are :ll·ailahk
"'" $1h.OO each
from
Digital Press of Digital Equipment C orporati on , I 1\urlington
Woocb Dri\'c, Burlington, \L\
OlH:\O-·i�lJ7.
Digital L'mployL' l' S nw\· send suhscription ordtTs on the L''Wi' to
IU>VAX::.JOIIRNAI.or hy interoflin· m:til to maibtop \II.<
ll·:l/ll(>H.
Orders should include badge JlltmhLT. sit e· locnioneodc. ami
addr"s. !\II elllploy<TS
Co!l!lllcnr.., on
he Sl'nt
mu st ad1·i,e of ch: tnge> of address.
rhe cont ent of an�
paper arc \\'t:konH.:d and
may
t o tiH: editor a t the pub! i'hcd-h) or net work addr ·ss.
<:opnight
1')')2 Digital Equipment Corporation. Copying
''
\Yithout l'cc i� permitted pro \· i d v d th;ll ....uch copi...:s :ItT
made f(n·
and an: not
Ahstr:Kting 1\ ith credit
usc in cducttinn;.d inslitutinn� h�· Ltcult� JllC!lllwr_..,
di,u·ihutcd ti1rcommnci:tl aL h' ant :t gc .
of tligital EquipmC'nt Corpor:Jtion·, :�uthoJ·.,Jtip is permitted.
.\II riglll' rc"n·ecl.
l'ltc·
inl(>rmation in
thc.Jutrrllrtl i' 'uhjeu to dJ:utgc without
notice :1nd !'<ll!ould nor
ht· nHI,trucd
:L' .t conllnitnu:nt h�·J>igit;tl
Equi pme n t Corpor:tl ion. Digil:tl l- :quipn1l'llt < .orp()r:tt it 111
no n.· ... r on :-, i h i l it �
OX'JH-')01"
I�S.
for ;u1y
Document:Jlioll �umhcr
The
CrJIIer Des(� II
j f!I'IICI'SSCS d!'SCI'ifJ!'d ill Ibis [.,·:me (//'l' 1/S!'d to flllifd
fJigitol :\ I.Jig/.J-jJ('J1i Jl'llltt!JCe ,\'1:1.\ wu f A/jJ/)(1 .! f()(ri ('/.JifJS. 'I he
'f/w (..\10
.•.
fJhofrmlicrop,rafJ/J IIIII JIII' COI'l'l'jtJCI/ses 011
ifl'il'e/' {[ 11!1 /IIIII '('/' jmds
1/.w b��b-sfJeed clue/..•
u{ 1/.Je ..1/jJ/Ill .! /()()-i 1/JiCI'II jii'IICI'SSfJ/;
oj)('l'll les fll (/duck mfc• uf.!OO .till::.
II •/I feb
7be JI!IIJ/Imlicm.r<.ruf'l�l' is f�J' Peter (.{lfillel!u c111d Uria11 f:'duwds
I'�' !Jigitat:�
\('ill iCII//(//1('/(1/' rllllfl)'.� is
I he u ;I ·er design i.\ f�J' IIike (illI
La fi r IJ'((/IJI:J'. 1/udSflll,
.II. I.
u{!Jua 111 ic (,t JIII/II wiicalit ills. Inc.
ern'�"' tll.ll nl.t\
I >igit:tl.
t'OR m \
Ill'<. ,\VS
the· Digital logo.
. and\ \IS.
l>i)\it:tl t:quipmcllt. ( :orporati<llt·
\'AX.\:\"
I' I�< . I· '- :�ntl
I I'!U:\I.i arc
\'A"c!U,tlT,
Computcr S) '1<'111,,
tr:�Jcm:trh.'
>I l.d:111d Stantord .Juni< >r I ni\ er,it 1.
'!'It .I
(>0(1!1.
\>\ \
i' at r:J<kll'l:trh. ot' �l:ll'dcnl < :ontput<'l' �) 'lL'll".
\Ill'!'> i' :t tr:Jdcnt:Jrh. of \til'S
•
.
EY-1.'>21 i'.-1>1'
liJI IIm ing arct r:llklll:trhs o/
Ut:C..
;t�;-.unJc ...
appc:tr in tl u.. ·.frJ/11"1/0/.
h .1 trademarJ
oft It ·
lnc
llo:�rd
of l'ruo,tn·,
uf the l ni'('f'it \ of< ::liifornia at lkrh.L'IL'\'.
llllllh. prmluetion \\.l,dlllll' h\ the
(oraphil: '>LT\ icn. 1m
lk,ign r;roupof,\tl:llltil:
I
Contents
10
Foreword
R. ]. Hollingsworth
Semiconductor Technologies
12 Microprocessor Performance and
Process Complexity in CMOS Technologies
Bjorn Zetterlund, james A. Farrell, and Thomas F Fox
25 Numerical Device and Process Simulation Tools
in Transistor Design
Marden H. Seavey, john V faricelli, Nadim A. Khalil,
Gerd Nanz, Llanda M. Richardson, Christian 0. Schiebl,
Hamid R. Soleimani, and Martin Thurner
39 CMOS-4 Technology for Fast Logic and
Dense On-chip Memory
Andre I. Nasr, Gregory). Grula, Antonio C. Berti, and
Richard D. jones
51
CMOS-4 Back-end Process Development for a
VLSI 0. 75 pm Triple-level Interconnection Technology
Marion M. Garver, joseph M. Bulger, Thomas E. Clark,
jamshed H. Dubash, Lorain M. Ross, and Daniel). Welch
73 Implementation of Defect Reduction Strategies
into VLSI Manufacturing
Mary Beth Nasr and Ellen]. Mager
83 A Yield Enhancement Methodology for
Custom VLSI Martufacturing
Randall S. Collica, X. joseph Dietrich, Rudolf Lambracht Jr. ,
and David G. Lau
100
114
Transistor Hot Carrier Reliability Assurance
in CMOS Technologies
Daniel B. jackson, David A. Bell, Brian S. Doyle,
Bruce). Fishbein, and David B. Krakauer
E/ectromigration Reliability of VLS/lnterconnect
]. joseph Clement, Eugenia M. Atakov, and james R. Lloyd
I
Editor's Introduction
jones review the front-end process (formation of
device and local interconnect) for the CMOS-4
0.75 /Lm technology and the steps taken to meet
-
design requirements. They also describe the effects
on submicron devices related to the scal ing of fea­
ture sizes and examine some solutions.
Goals for the back-end process (form ation of
global metal interconnect) were also driven by the
logic design requirements for higher circui t density.
In addition, back-end development goals included
the continued use o f equipment developed for
Jane C. Blake
Editor
the 1 .0 -�Lm CMOS-3 technology. Marion Garver, joe
Bu lger, Tom Clark, Jamshed Dubash, Lorain Ross,
and Dan Welch relate how tools were modified for
CMOS-4 and describe new blanket tungsten and pla­
The design of semiconductor chips has been the
narization processes for submicron devices.
topic of several past Digital Technical journal
To produce a specified yield of CMOS devices,
issues. With the introduction of Alpha 21064, the
defect reduction and yield enhancements, l ike other
world's fastest microprocessor, this issue focuses
processes, must be i nitiated concurrently with
for the first time on the development of semi­
the design stage. Mary Beth Nasr and E l len Mager
conductor technologies that make possible the high­
review the principles of microcontamination con­
performance of D igital's VLSI chips. Engineers in
trol and outline defect reduction techniques to
Advanced Semiconductor Development present
increase product yield i n the areas of p -gate leakage
in-depth views into CMOS-4 technologies, which pro­
and metal 2 short circuits. The paper that fol lows
duce m icroprocessors with up to 1.7 million transis­
addresses the methodology of yield enhancement,
tors and operating frequencies as high as 200 MHZ.
including processing, process equipment, manu­
The significant increase in p erformance achieved
facturing, and design and test. Randy Col l ica, joe
with each generation of CMOS technology is in part
D ietrich, Rudy Lambracht, and D ave Lau describe
the resul t of a synergistic relationship between
the use of test chip data, yield models, and the
m icroprocessor design and process engineers. In
selected approach to yield analysis and forecasting.
their paper on process technology contributions
An advanced method that helps designers predict
to microprocessor performance, Bjorn Zetterl u n d ,
J i m Farrell, a n d Frank Fox describe t h e sca l i ng the­
circu it hot carrier l ifetime and thus maximize tran­
sistor performance at the required rel i abil ity level
ory that has led to a doubling of gate density and
is the topic of a paper by Dan jackson, David Bell ,
an increase of 30 percent in gate speed in fou r suc­
Brian Doyle, Bruce Fishbein, a n d David Krakauer.
cessive CMOS generations. They discuss process
The authors describe a physically based method
features implemented in CMOS-4, and close with a dis­
for determining the acceptabil it y of bot carrier­
cussion of models that predict process variations.
Models and tools, essential i n providing design­
induced degradation in t ransistor characteristics.
ers early insight into the characteristics of the tran­
ity of the interconnects. I n their paper on electro­
migration reliability, joe Clement , £ugenia Atakov,
and Jim Lloyd provide a helpful overview of the
sistors to be fabricated, are the focus of a paper by
Marden Seavey, john Faricell i , Nadim Khalil, Gerd
Nanz, Llanda Richardson, Christian Schiebl, Hamid
Soleimani, and Martin Thurner. The authors describe
Another predictor of chip l ifetime is the reliabil­
potential erosion in metal interconnect clue to
electron conduction. They then present a scaling
several physical models that accurately simulate
model developed to characterize the rel iabi l i ty of
transistor behavior, and present numerical mathe­
CMOs-4 chip interconnects.
The editors thank Rich Hol l ingsworth and Arlene
Delvy of the Advanced Semiconductor Develop­
ment Group for their guidance and unfail i ng sup­
port in developing this issue.
matical methods used to enhance existing simu la­
tors. An overview of Digital's and others' efforts to
integrate simu lation tools concludes the paper.
The need for both high-density logic gates and
on-chip cache memory in m icroprocessors pre­
sents special challenges to process engineers.
Andre Nasr, Greg Grul a , Antonio Berti, and Rich
2
I
Biographies
Eugenia M. Atakov
As a principal engineer in the Advanced Semiconductor
Development (ASD) Group, Eugenia Atakov is responsible for evaluating on-chip
interconnect rel iabil i ty. Since joining Digital in 1988, she has contributed to sev­
eral CMOS VLSI projects, including the Alpha 21064 chip. Previously, Eugenia
worked at the Moscow Institute of Semiconductor Electronics. She received a
B.S.E.E. and an M.S.E.E. from the Moscow Energy Institute, and a Ph.D. from the
Moscow Institute of Steel and Al loys. She holds t wo patents and has published
17 papers on semiconductor device and interconnect reliability.
David A. Bell
David Bell attended Bates College and received a B.S. degree in
physics and mathematics in 1980. In 1984, he completed his Ph . D. in physics at
the Massachusetts Institute of Technology. David worked at Texas Instruments
on submicron CMOS and BiCMOS development prior to joining Digital's Advanced
Semiconductor Development Group in 1989. H is current work includes sub­
micron CMOS transistor design, characterization, and modeling. David is a mem­
ber of IEEE and P h i Beta Kappa.
Antonio C. Berti
Antonio Berti is currently contributing to the work of the
Fab 4 engineering organization as a senior engineer. I k has developed and imple­
mented CMOS-4 processes in manufacturing, including metal deposition pro­
cesses for cobalt, aluminum interconnections, TiN adhesion layers, ancl TiN local
interconnects. Antonio a lso developed a coba lt silicide process to decrease
defects and implemented TiN deposition hardware and processes to reduce par­
ticles. He graduated from Rensselaer Polytechnic Institute in 1988 with a B.S. in
materials engineering and joined D igital in 1989.
Joseph M. Bulger
joseph Bulger is a principal engineer with the Fab 4 manu­
facturing engineering organization. His m ain areas of involvement are thin film
manufacturing, chemical vapor deposition of tungsten, and metal l ization. Joe
joined Digital in 1988 after 10 years with Analog Devices, Inc . , working on thin
film development and manufacturing. He received his B.S. degree in electrical
engineering from Northeastern University in 1982.
Thomas E. Clark
A principal engineer, Tom Clark is currently
a
member of
the CMOs-6 technology team within the Advanced Semiconductor Development
Group. Tom has developed blanket tungsten plug modu les for CMOS-4, CMOS-),
and CMOS-6 technologies. He also performed feasibil ity studies of Co and Ti
sal icidation processes for contact, source, drain, and gate applications in the
CMOS-4 technology. Tom holds a Ph.D. in inorganic chemistry from the Lniver­
sity of Massachusetts and has published numerous papers on the subject of <:VD
tungsten metallization. He has received one patent.
3
BiograjJ!Jies
}. Joseph Clement
A principal engineer on the i nterconnect physics project
in the Advanced Semiconductor Development Group, joe Clement works pri­
marily on metal reliabil ity and performance issues for VI.Sl. Before joining Digital
in 1987, he held positions at Fairchild Semiconductor and Sandia National Labo­
ratories. Joe received a Ph.D. i n electronic materials and devices from Princeton
University and a B.S.E.E. from the University of Notre Dame. He has pu blished
four papers on i n terconnect reliability and serves as a m e n t or for SRC research in
elcctrom igration at the Massachusetts Institute of Technology.
Randall S. Collka
in
Randal l Collica, a principal yield engineer, joim d Digit al
·
198 7 His primary areas of interest are SRAM fau l t analysis, yield modeling, fault
.
tolerance, and experimental design and analysis. For the CMOS-4 technology, he
established the use of yielcl moclels for redundancy on advanced processor chips
conta i n i ng on-board cache RANis. These yield models are used for l ine control
and productivity optimization for VJ.SI products. Prior to this work, Randa ll was
engaged in failure a n al y sis, test chip design, ancl process experimentation. He
holds a L\.S.F..E. from Northern Arizona University and is a member of IEEE.
X. joseph Dietrich II
Joe Dietrich is a senior engineer whose work involves
yield enhancement for C.VIOS technologies. His responsibilities for the CMOS--1
process included work on the design for manufacturability and y ie l d modeling.
He is currently performing the same yield analyses for the CMOS-5 process. Joe
has also developed photoli thography processes lor earlier generations of C.\'IOS
technologies. Joe joined Digital in 1983 after graduating from Rensselaer Poly­
technic Institute. He was awarded a B.S. degree in physics in 1983.
Brian S. Doyle
Born in Dubl in, Irdancl, Hrian Doy le graduated from Trinity
Col lege, Dubl in, in 1975, and received an ,'vi.Sc. and a Ph . D. from the 1:niversity of
London. He held a postdoctorate position at the University of London from 1980
to 1982. For the n<.:xt six years, Brian was employed by Bu l l, France, and worked
o n cryogenic MOS devices and hot carrier effects. He joined Digital in 19H8 and is
a principal engineer in the Gate Insulator ami Transistor Reliabil i t y c;roup. Brian
is
a
member of IEEE and is on the tcchnic:.ll com mittee for the 1992 International
Electron Devices M ee ting.
Jamshed H. Dubash
Fab
Jamsh<.:d Oubash is a senior process engineer with the
4 manufacturing organizat ion. For the CMOS-4 technology, he developed
corrosion-fre<.: metal e tch processes and a sloped via process for metal step
coverage. Prior to this work, he
c
o n t r i bu t e d to the etch process development
of CVIOS-3 generation sem iconductors. Jamshed is currently responsible for
the process development of the integration of SOG planarization processes
into manufacturing. He joined D igital after receiving his B.S. degree from the
Rochester Institute of Technology i n 1988.
I
john V. Faricelli
John Faricell i is a principal manufacturing engineer in the
Advanced Semiconductor Development Group. Since joining Digital in 1983,
he has worked on projects related to device physics and model ing, including
cryogenic operation of MOS devices, hot electron effects, device isolation, and
CMOS latchup. John has contributed to several device and process simu lation
programs, including MINIMOS, P ISCES-I!, and PRO.\HS. He holds a B.S.E.E. from
Rensselaer Polytechnic Institute, and M.S.E. E. and Ph.D.E.E. degrees from Cornell
University. He is a member of Tau Beta Pi and Eta Kappa Nu .
James A. Farrell
Jim Farrel l is a principal engineer in the Semiconductor
Engineering Group (SEG). He was most recently engaged in the development
of high-performance switching circuits, with particular emphasis on low­
temperature operation. Prior to this work, Jim was a static RAM designer in SEG,
with responsibility for developing the CMOS-3 process with the Advanced Semi­
conductor Development Group and for designing several memory chips. He
joined Digital in 1985, after receiving B.S.E.E. and M.E.E.E. degrees from Rensse­
laer Polytechnic Institute in 1984 and 1985, respectively.
Bruce J. Fishbein
Bn.�t:c Fishbein attended Cornell University, where he
received a H.S. degree in m aterials science and engineering in 1983. lie also
received M.S. (1985) and Ph. D. (1988) degrees in electrical engineering from Stan­
ford University. Since joining the Advanced Semiconductor Development Group
at Digital in 1988, Bruce has worked on gate insulator and transistor rel iability
for CMOS microprocessors. Specific areas of interest include time-dependent
dielectric breakdown, plasma-induced d ielectric damage, thermal instabil ities
in gate dielectrics, and the use of high d ielectric constant materials.
Thomas F. Fox
Frank Fox, a consulting engineer in the Semiconductor Engi­
neering Group, co-led the implementation of the NVAX microprocessor and con­
s u l ted with the Advanced Semiconductor Deve lopment Group on the design of
the C\10S-4 technology. He joined Digital in 1984 and worked on the implemen­
tation of the CVA..'\. microprocessor. Frank received a B.E. degree from University
Col lege Cork, National University of[reland (1974), and a Ph.D degree from Trin­
ity College, Dubl in University (1978), both in electrical engineering. He has pub­
l ished papers on u lt rasonic instrumentation and VL'>I design; he has t wo patents.
Marion M. Garver
Marion Garver was the project leader and supervisor of
the CMOS-4 back-end process development group and is currently performing
the same function for the CMOS-6 technology. Marion joined Digital in 1980 soon
after the opening of the H udson site. She initially worked in bipolar manufactur­
ing fabrication and joined the Advanced Semiconductor Development Group i n
1982 to develop some o f the original plasma etching systems for 'MOS and CMOS
applications. Marion holds a B.A. degree (197 3) from Oberlin Col lege and was
previously employed by Alpha Industries and Ohio Medical Products.
5
Biographies
Gregory J. Grula
A consultant engineer in the Advanced Semiconductor
Development Group , Greg Grula has worked on process development and inte­
gration for CMOS-1 , CMOS-2, and CMOS-4 technologies. Currently he is the super­
visor of the C'l'lOS-5 front-end process technology and has p erformed significant
work in the isolation of transistors for CMOS-5. Greg joined Digital in 1978 after
five years with RCA Corporat ion. He received a B.S.E.E. from Wilkes University
and belongs to the Electrochemical Society and IEEE. Greg was awarded one u.s.
patent and has three patents pending.
Daniel B. jackson
Daniel Jackson received a B.S. E. E. (1978) from Worcester
Polytechnic Institute and was awarded M.S. (1980) and Ph.D. (1985) degrees from
the University of Illinois for work on semiconductor device physics. He then
joined Digital, working on advanced MOS devices and process development.
Daniel has managed the Gate Insulator and Transistor Reliability Group in
Advanced Semiconductor Development since 1989. He is interested i n advanced
process development and in MOS device physics and transistor reliability.
A member of IEEE, Daniel has served on the technical com mittee for the IEDM.
Richard D. jones
A principal engineer in the Advanced Semiconductor
Development Group, Richard Jones developed etch processes such as local inter­
connect etch, polysilicon etch, and metal etch for the CMOS-4 technology. He is
currently developing po lysilicon etch and spacer etch processes for 010S-6.
Richard joined Digital in 1984. He received a Ph. D. (1986) and an M S (1982) in
physics from Rensselaer Polytechnic Institute and is a member of the American
Physical Society and the American Vacuum Society. He has published two papers
on the properties of semiconductor materials.
Nadim A. Khalil
Principal engineer Nadim Khalil is a member of D igital's
Advanced Semiconductor Development Submicron Simulation Group. Cur­
rently developing optimization-based tools for process and device design,
Nadim was involved in graphics and numerical software development and ana­
lytical modeling for CMOS technologies. He joined D igital in 1985, after receiving
a B.S.E.E. (with distinction) from the American University of Beirut and an M.S.E.E.
from Louisiana State University. Nadim is presently pursu ing a Ph . D. degree a t
t h e Technical University o f Vienna.
David B. Krakauer
Since joining Digital's Advanced Semiconductor Devel­
opment Group in 1989, David Krakauer has worked on topics in the area of tran­
sistor reliability, including hot carriers and electrostatic discharge (ESD). I n
particular, he has been i nvolved in t h e design and characterization of ESD protec­
tion for D igital's advanced semiconductor processes. David graduated from the
Massachusetts Institute of Technology in 1989 with S. B. and S.M. degrees in elec­
trical engineering.
6
I
As a principal yield engineer, Rudolf Lambracht per­
Rudolf Lambracht Jr.
forms analysis to identify yield- l imiting process steps and root causes for yield
improvement. He is also i nterested in tools for failure analysis. Rudy worked on
the CMOS-4 technology and is currently involved i n the CM05-5 technology. I n his
previous position as senior ion implant engineer, he was responsible for the
implant process for MO S CMOS, and advanced bipolar processes. Rudy joined
,
D igital in 1978. He holds an A.S.E.E. degree from Waterbury State Col lege and is
coauthor of a paper concerning ion implantation.
As a senior manufacturing engineer for the Advanced Semicon­
David G. Lau
ductor Development Group, Dave Lau was responsible for driving yield issues
during CMOS-4 development and analyzing yields for the CMOS-4 process in con­
junction with the manufacturing yield group. Currently, he is developing yield
strategies for CMOS-6. In a prior position, he developed and implemented device
rel iabi l i ty characterization techniques and designed a CMOS process for l iquid
n itrogen temperature operation. Dave holds B.S. and M.S. degrees from MIT. He
has coauthored three papers for IEEE International Electron Devices Meeti ngs.
james R. Lloyd
Principal engineer James Lloyd of the Design and Rel iabil ity
Assurance Group special izes i n metal rel iabil ity problems, primarily electro­
migration and stress voiding. He is also studying physical mechanisms related to
thin film reliability in integrated circu its at the Max Planck Institut fur Metall­
forschung i n Stut tgart, Germany. Before joining D igital i n 1988, Jim worked
for IBM. He holds fou r patents, has written more than 50 papers for refereed
journals, and is an invited speaker i n h is field. Jim received a Ph.D. (1978) from
Stevens Institute of Technology i n materials and metal lurgical engineering.
Ellen J. Mager
E l len Mager is a member of the Fab 4 process engineering
organization within Semiconductor Manufacturing. An engineering supervisor,
E l len is currently involved with the ongoing defect reduction of 010S-4 mate­
rial, as wel l as preparing for the transfer of the next technology level (CMOS-5).
E l len joined Digital in 1987 after working as a process engineer for N:Hional
Semiconductor Corporation. She received her B.S. degree in chemical engineer­
ing (with high distinction) from Colorado State U niversity.
Gerd Nanz
Gerd Nanz received his D iplommathematiker degree from
the Technica l University of Munich, Germany, in 1985. From 1985 to 1986, he
remained at the U n iversity as an assistant teacher in the D�:partment of C ivil
Engineering, where he was involved in the development of self-adaptive finite
element programs. Gerd then joined Professor Selberherr's group at the Tech­
nical University of Vienna, Austria, and completed his Ph . D. th<.:sis on numerical
methods in device simulat ion in 1989. Since 1990, he has worked on device and
process simulation a t Digital's Campus-based Engineering Center in Vienna.
7
Biographies
Andre I. Nasr
A con su ltan t engi n e er with the Advanced Semiconductor
Develop men t Group, Andre Nasr is the p ro je c t leader in the 0.35 -�-t-m CMOS tech­
no lo gy fo r fro n t- e n d process de ve lop ment and on-chip memory implementa­
tion. He was the proj ec t leader for the CMOS-4 dev ice group , responsible for the
0.75-�-t-m proc ess and device arc hitecture. An dre in tro du c ed and implemented
the drain e ng in eer ing c on c ept in Digital's CMOS proce sses. He also pioneered the
high-performance BiCMOS process development. Andre has on e paten t issued
and two p endin g. He holds an M.S. in physics from t he Univt:r sity of Massachusetts.
Mary Beth Nasr
,\1.ary Beth Nasr is a member of the Advanced Semiconductor
Development (1roup . She ,is cu rren tly the eng inee ring supervisor of the Metro l­
o gy an d Mic rocon tamin ation Group, wi t h direc t re spons ib il it y for defect inspec­
tion tools. Prio r to jo in in g D igi tal i n 1987, Mary Beth was a m icroc o nt am in ati on
con tro l en gin eer for Mitsubishi Semiconductor Corporation in Ja p an for one and
one-half years. I n 1985 Mary Beth received a B.S. in chemical engine eri n g from
Columbia University and a B.S. in mathematics from P rovi den ce Co llege .
Llanda M. Richardson
Consultant engineer Llanda Richardson manages the
Submicron Physic s and Chemistry Group within the Adv an ce d Semiconductor
Deve lop m en t Group. She established the Device and Process Simulation Group
an d is re sp o n sible for ASD's participation in the Campus-based E n g in eer in g Cen ­
ter in Vien n a, Au stria. Llanda is a member o f the CMOS Development Commit tee
an d Digital's alte rn ate rep res en tativ e on the SRC Bo ard of Directors. She jo ine d
Digital in 1979 after receivin g a Ph.D. in phy sic s from the U n ive rs ity of Vermon t.
Lorain M. Ross
Lorain Ross joined Digital in 1984 af te r two years with NCR
Corpo ra ti on . She has worked primarily to ward the p ho to l i th ogr ap hy process
enhancement of several generations of b ip o lar an d CMOS device s. L o ra in partici­
p ate d as a member of the original task force to introduce SPC m e tho ds for con­
trollin g c ritic al par am ete rs of semiconductor manufacturing at the Hudson site.
As a princ ipal proc ess engineer for the Fab 4 o rg an iz a tion , she is re spo n sible for
the int egrati on of newly developed pho to pro c e s ses into manufacturing. Lo rain
ho lds a B.S. de gree in chemical engineering from Yale Un iversity.
Christian 0. Schiebl
Christian Sc h i eb l received a Diplomingenieur degree in
p hy si c s fro m the Tec hn ic al Un ive rsity o f Vie n n a, Au stria, in 1986. The sub jec t of
his master's thesis was qu an titative e lec t ron probe m ic roan aly sis (EPMA) u sin g
c harac t eristic �1-lines. He stu die d at the In stitu t fur Ange wand te u n d Tec lmisc he
Physik at the Tec hn ic al Unive rs i ty of Vienna and received a Ph.D. in 1989. H is dis­
sertation was on c harac te ristic fluore sc en ce correction in qu an tita t ive EPMA.
Christ ian then jo in ed the Campu s-b ased En ginee rin g Center in Vien na, whe re
he currently wo r k s in the fie ld of semiconductor process and devic e simu latio n .
8
I
Marden Seavey, a consultant engineer, came to D igital in
Marden H. Seavey
1981 from Raytheon Company. He was the supervisor of the device and process
simu lation projects within the Advanced Semiconductor Development Group
for five years. Marden pioneered the adaptation of the MINIMOS program to the
ASD CMOS pr o cess and device design. The mobility model that he developed also
made it possible to apply MINIMOS in worst-case circuit design. M;u·den received
a Ph . D . in a ppl ied sol id-state physics from Harvard University. He retired from
Digital in june 1992.
Hamid R. Soleimani
Senior manufacturing engineer Hamid Soleimani joined
Digital in 1987 and is a member of the Submicron Simulation Group within
Digital's Advanced Semiconductor Development Group. He developed the tran­
sient diffusion model that has been incorporated in the SUPREM-III process simu­
lation program and is used tor CMOS technology development work. Hamid
holds B.S.E.E. (1984) and M .S.E.E. (1986) degrees from Louisiana State University,
Baton Rouge, Louisiana. He is a member of the IEEE CAD and E lectron Device
Societies.
Martin Thurner is a principal engineer in the Submicron
Martin Thurner
Physics and Chemistry Group within the Advanced Semiconductor Develop­
ment Group. He is responsible for the three- dimensional device simulations
investigating MOS d evice width effects. Martin has improved the efficiency and
accuracy of the mathematical and physical models used in the MINIMOS pro­
gram. Prior to joining Digital in 1988, he worked at the Federal Research and Test
Institute of Austria. Martin received Diplomingenieur and Ph. D. degrees from the
Technical llniversity of Vienna, Austria.
Daniel). Welch
As a principal engineer in the Fab 4 process engineering thin
films group, Dan Welch is responsible for the development and optimization of
interlevel d ielectric and planarization processes. He was project leader for the
development and implementation of the SOG planarization process for CMOS-4.
Dan joined Digita l in 1991 after affiliation with Intel Corporation as a senior pro­
cess engineer for the high-volume manu facture of advanced microprocessors.
He received a B.S. i n chemical engineering in 1983 from Clarkson University.
Bjorn Zetterlund
Bjorn Zetterlund, a manufacturing consultant engineer,
came to Digital in 1985 after 12 years with Raytheon Company. He is the device
engineering supervisor tor CMOs-5 and is responsible for CM05-5 t ransistor
design and the CMOs-5 technology file . Prior to this work, Bjorn was responsible
for transistor design in CMOS-3 technology. He received B.S., M . E., and P h . D.
degrees in electrical engineering from Rensselaer Polytechnic Institute. Bjorn is
a member of IEEE, Tau Beta Pi, and Eta Kappa Nu.
9
I
Foreword
and
suhs�qu e n t l y
m a n u facture
s em i c onductor
tech nologies.
Digital's sem iconducto r operations group set a
goa l in the early 1980s to achi e ve leadership in t h e
d e vel opm e n t and manufacture o f the world's h igh ­
est performance m i croprocessors. To meet t h i s , a
nu mber of st rategic p o si tions were taken:
• Develop d i stinct ge n e ra t io ns o f < :.VI OS that wou ld
produce a wide range of VLSI d ev i ces, not o n l y
m i croprocessors.
R . J. Hollingsworth
•
Managet; lldtmuced
Semiconductor Development
Be at the lead i ng edge in densi t y ; be ahead of the
i n d u s t ry i n h igh-speed swi t c h i ng devices and
syste m - leve l features.
• Mak e
C MOS tech n ol ogy d ecisions by optim izing a
wide ra n ge of req u i rements necessary to meet
th e goa l : the world's fastest m icroprocl'ssors. The
Dig i t a l has developed and manufactures the world's
approach wo u ld be a rigorous e ng i nee rin g opti­
fastest p ro d u c t i o n com plex i ns t ruction set (CISC)
and reduced i nstruction set
(RISC)
m ization from c o m p u t er arch itecture th rou gh
m icroproces­
chip manu facturing processes.
s o rs. The speed of t hese m i c roproc es so r s is d u e ,
i n large p a r t , to a co m p l e me n t ary metal oxide semi­
conductor (CMOS) technology hav ing faster tran­
•
Develop
CMOS
w i t h a s i ngle, m u lt i- d iscipl i n ed
tech n i c a l t ea m d e dicated to the project from
sisto rs , as dense on-ch i p w irin g , and i n novative
i n i t i a l co nc epti o n through m a n u factu r i ng qualifi­
pe rformance-en h a n c i ng materials and structures.
cation-a four- to five-year e n deavo r.
Th is advanced CMOS t ec h n o lo gy a l lows D ig i t a l the
unique capabi l i t y to design and produce m i c ro­
•
processors that operate t w ice as fast as c o m m o n
poinr fo r CMOS deve l o p m e n t . Tie together the
lead i ng-edge devices pro d u c e d hy t h e wo rld 's pre­
e ffo rts of fu l l - t i me chip arch i tecture , design. test,
mier semiconductor manufacturers.
re l i a b i l i ty, packaging, and m a n u facturing p l' o ple
I n 1980, Digital recognized the strategic role that
fo r tl1l' full project d urat i o n , i . e . , fou r to five
m icroprocesso r c h i ps played as core elements in
years.
reshaping a nd adv anc i ng the computer i nd ustry A
key o bserv at i o n was t h a t t h e u n re l e n t i ng advances
in very l a rge- s c a le i n t egrated
(VJ..SI)
circuits wou ld
Use the m icroprocessor product, targeted to
wor ld- lea d i n g p e rformance, as the specific focal
•
Develop CMOS techno logy in c o n j unction with
t h e m i c roprocessor arch i tecture a n d design­
an essl' n t i a l i ngre d ie n t i n d e l iver i n g l e a d i ng VLSI
co n t i n u e t o a l low vast amou nts of lo gic and mem­
o ry to be e c o n o m i ca l ly produced in a s i n gl e s i l i­
c h ips to the m arket first . This " c o n c u rren t "
con c h i p , t he reby y ielding d ramatic i m p ro vements
approach h a s b e e n a m a i n st ay i n Digi t a l 's CMOS
i n performance, cost, ami re l i ab i l i t y VLSI devices
development since the early .l �l:lOs.
w e re d e m o ns t ra t i n g y ea rl y i m p ro veme n t s of 10 to
15 p e rc e n t in gate swit c h i ng speed and 25 to 35 per­
cent i n densi t y. M icroco n t a m i n a t i o n a n d process
•
Part i c ipate i n , contribute to, and draw u p o n the
best semico nductor research i n the wo rld .
control m e t h o d s coupled w i t h i n c re a s i ng wafer
Many lea d i ng semiconductor compan ies fo l l ow
sizl' a l lowed larger chips to be fabricated at lower
these st r at e g i es . Digi t a l , however, is u n ique in prac­
cost. I t wa s c lear that t he abi l i t y to i n t e gra te m o re
ticing a l l of them.
and morl' fu n c t i o n i n to a piece of s ilicon was funda­
To h e l p guide the tech nical d i rection, a CMOS
m e n t a l l y changing the compu ter i n d u s t ry. The era
te ch no lo g y road map was created in the early 1980s.
of entire compu t i ng s ys t e m s on a s i ngle chip was
It defined the key pac i ng eleml'nt S that cle l i n e a te
ra p i d l y approaching. Ch ips were n o t just compo­
each d istinct CMOS gen e r a t i o n : m i n i m u m featu re
nents in a system, t hey were bec o m i ng t he s ys t e m .
size, swi t c h i ng speed. manufact urable chip and
To fu l ly exploit t h i s a n d lead Dig i tal i n to w h a t
w a fer size, ami other at tri bu tes necessary to del iver
C . Gordon Bel l termed a ·' s e m i - c ompu te r " company
leadi ng-edge microprocessors. This roadmap se t
in the 1990s, the decision was made to d evel o p
the goa l s for the organi zation and a l lowecl e as y
10
I
reference to competitive trends. In its simplest
transistors operating at clock rates up to 200 MHZ.
form, the roadmap defined each CMOS generation
A variety of materials and structures have been
by making logic gates switch 30 percent faster
crafted into CMOS-4, a llowing advanced system­
while occupying half the sil icon area and by inte­
level capabilities not available in other state-of-the­
grating these elements on chips that were growing
art CMOS processes.
40 percent in area compared with the previous gen­
eration. The roadmap today defines eight genera­
tions of CMOS; four have been introduced to
manufacturing (CMOS-1 through CMOS-4), and two
arc now under development (CMOS-5 and CMOS-6).
The papers in this issue of the Digital Technical
journal are focused on Digital's fourth-generation
CMOS (CMOS-4) that is used to build a wide variety
of VLSI chips, most notably the NVAX and Alpha
21064 microprocessors. CMOS-4, presently being
manufactured in Digital's semiconductor facilities
in Hudson, MA, and South Queensferry, Scotland,
del ivers m icroprocessors with up to 1 .7 m i l l ion
This issue spans the breadth of technical areas
necessary to make CMOS-4 a successful element
in establishing Digital's preeminent position in
microprocessor technology. The d iscussions herein
include how manufacturing process and chip
design trade-offs are made; how a large number
of complex manufacturing steps are integrated;
how leading-edge speed , density, and materials are
achieved; and what modeling, simulation, and mea­
surements are critical to ensure reliability and pro­
duceability. The papers are a sample of the range
of technological achievements in Digital's semi­
conductor operations.
11
Bjorn Zetterlund
James A. Farrell
Thomas F. Fox
Microprocessor Performance
and Process Complexity in
CMOS Technologies
Digital's CMOS technology is characterized �)' a scaling methodology that doubles
the gate density and improves the gate speed by approximately
30 percent with
eacb new ge n e ra tio n. Dec reasing feature size from one generation of CMOS tech­
nology to tiJe next isfundamental to improving the pe1jormance of VLS! chips. Each
of Digital's successiue C'v!OS generations has added new technology features to
improve pe1jormance furtbeJ: Digital's latest, qu alifi ed CMOS technology incorpo­
rates features such as low voltage operation, low-resistance topside substrate con­
tacts, low-resistance transistor gate material, local interconnects in Sl?AMs, three
levels of metal interconnect, andfusesfor redundancy
t o prov ide leadersh ip i n product performance a n d
CMOS Microprocessors­
General Considerations
fu n c t i o na l ity, a s most rece n t l y evid enced by the
Several factors dete rmine the perfo rmance of the
The goal of Digit a l 's semiconductor organ ization i s
Alpha 2 1064 and
l\iV�L'X microprocessors. L2 I n ternal
fastest microprocessor that can be bu i l t i n a given
developme n t of complementary metal-oxide semi­
CMOS technology. The p e rformance of a mic ro­
conductor (CMOS) processes has been crucial to the
processor is inversely prop o r t i o n a l to the product
success of these ch ips because i t a l lowed u s to
of clock cyc les per i nstruction
design t he process to meet very specific needs. The
mach ine cyck t i me. From one generation of m i c ro­
identihcation and fu lfil lment of these needs has been
a m u l t igcnerational, ongoing task that closely l i nks
the chip design effort with the process development.
Each new generation of
CMOS technology is scaled
to double the gate density and i m p rove the gate
speed by
30 percent.
ln add i t i o n to the generation­
to-generation improvements t hat are clerived from
sea l ing, a C .VIOS process that is designed specific a l l y
(CPJ).I and the
processors to the nt:xt, improvements i n both
C PJ
and mac h i ne cycle t i me are requ i red in order to
meet th<.:: pe rformance goal.
CPI depends on the
m i c roarchitcctutT as we l l as the m ix of i n struc­
t i o ns executed. Tht: m i n i m u m mach i n e cycle t i me
depends o n l y on circu i t and microarc h i tectural
issues.
Improvements in
CPI are achieved in parr by
for h igh-performance microprocessor app licat ions
pipel i n i ng and in part by adding cache memory
requires a n u mber of fea t u res beyond t h ose nor­
the d i e . As shown i n Figure
to
1 , p i pe l i n i ng is defined
CMOS processes. As a new
as the s i m u l taneous exec u t i o n of two or more
process is being developed, proposed new features
i nstructions; for example, the fi rst part of a two­
mal ly implemented in
arc critical l y evaluated in order to arrive at the opti­
part i nstru c t i o n is cxt:cuted a t the same t i me as the
mum t rade-off between chip performance and pro­
second part of the previous instru c t i o n . S i m u l tane­
cess complexity.
ons execu t i o n reduces the n u mber of CPI; the resu lt
This pape r describes Digital's
CMOS
processes
i s h igher performance at the expense of more
from t he perspective of those process fcatures that
circui try.4 A d d i ng o n - ch ip
co n t ribute to the pe rformanct: and func t i o n a l i t y
reduces the
o f h igh-speed microprocessors. It begins w i t h a
access i nstructions o r data resident i n i ts caches,
short discussion on microprocessor archi tecture,
rat her than wait for this information to bt: t rans­
which strongly influe nces the d irect i o n of process
m i t ted
development.
(RA.\1) chips.
12
cache
memory
also
CPI. The microprocessor can qu ickly
from
Vol. 4 No. 2
on- board
random-access
spri11g IYY2
memory
Digital 7ecbtlicaljourual
Microprocessor Pe1jormcmce and Process Complexity in CMOS Technologies
Generation-to-Generation Scaling
NO PI PELINING
0
-I
I
2
TIME
3
4
1- CLOCK CYCLE
6
5
The pace of improvement i n microprocessor per­
8
7
forma nce inclica tecl by the REX520/NVAX compari­
son
I
is
ahead
perfo r mance
F
S --FI R ST
_
_
_
_
_
_
INSTRUCTION
S ------- F
SECOND
INSTRUCTION
of the
has
industry.!•
Microprocessor
been doubl ing approximately
every two years, and there is no evidence that this
pace of change is slacken i ng. The shrinking of
fea ture sizes with each new generation of CMOS
NO PI PELINING: Second instruction does not start until the
process technology and the increase in yieldable
first instruction finishes. One instruction is executed every
four cycles, so CPI = 4.
die sizes have enabled this rapid im provement in
the performance of very large- scale integra t i o n
(VLSI) chips.
At the center of the reduction i n the feature sizes
PIPELINING
0
-I
I
TIME
4
3
2
1- CLOCK CYCLE
5
6
of CMOS processes is the m i n iaturization of the
8
7
known
I
as
scal ing theory, has been developed to
guide this process . 7. H. 9
F
S --F � R ST
_
_
_
_
_
_
I N STRUCTION
In the fu ndamental form of sca l i ng, ca l led con­
stant field s c a l ing, the transistor's physical parame­
s ------- F
SECOND
INSTRUCTION
ters
and
the
power
supply
vol t age
are
kept
proportional to the feature sizes to maintain the
s ------- F
THIRD
I N STRUCTION
magn itude and the con tours of the e l ect ric fields
within the transistor. Al l the dimensions of the tran­
P I P ELINING: Second instruction starts halfway through the
sistor, e . g . , length, width, gate d ielectric thickness,
execution of the first instruction. Third instruction starts
halfway through the execution of the second. One instruction
is executed every two cycles, so CPI = 2.
and source/drain junction depths, and the power
supply and threshold voltages are reduced by the
Note: I t takes more circuitry to implement a pipeline since instructions are
executed simultaneously.
s c a li n g factor, [ 1/K (where K is grea ter than 1 )] ,
while the el oping concentrat ions are increased by
KEY:
S
F
MOS
transistor. Over the past 15 years, a set of ru les,
K. These rules are also extended, with some excep­
START
FINISH
t i o ns, to guide the mini aturizat i o n of the intercon­
nect. In practice, sca l ing theory is not fol lowed
Figure 1
exactly, for reas ons of both perfo r mance and stan­
Relationship between CPJ and
dardiza t i o n
Pipelining
As the m in i m u m fea ture size decreases, transistor
density i ncreases. Thus, for a given d ie size, a
0.75 -micrometer (J.Lm) m i n i m u m fea t u re size tech­
nology
(CMOS-4)
that
are
d iscussed
below.
Digital's
implementation of sca l i ng through four CMOS gen­
can support four times as many
transistors as a 1.5 -J.Lm techno logy
(Ci'viOS-2).
In aclcli­
erations is shown in Table
2.
Improved Peiformance through Scaling
The reduced feature sizes m ade possible by scal­
i ng have a major impact on node capacitance and,
to
hence, the speed of the chip . The m i n i m u m cycle
increases i n t he s i ze of the largest d ie that can be
time of a microprocessor is i nversely proportional
tion,
advances
in
process
technology
lead
to the capacitances of the gates, sou rces, drains,
bu i l t with an acceptable manufacturing yield.
Microprocessor designers take advantage of the
and interconnect . The gate capacitance is inversely
extra transistors to increase the degree of pipe­
proportional to the t hickness of the gate d ielectric,
CPI,
and transistors with thinn er gate d ielectrics have
which boosts the performance of the mach ine.
higher drive current. Since the m i n i m u m cycle time
Table 1 shows the d ifference between t wo genera­
is a stronger fu nction of transistor d rive current
l ining and the size of caches. This reduces the
tions of scaled
CMOS
processes. The halving of the
feature size has been augmented hy a larger d ie size
and microarchitectural changes that i ncn:ase the
SPECmark' performance by a factor of 4 . 8.
Digital Technical journal
Vol. 4 No. 2
Spring i'J'J2
than
gate
di electric
capacitance,
the
trade-off
shou ld be made in favor of a thinner gate d ielectric.
In Digita l 's fa m i ly of CMOS processes, gate capaci­
tances have scaled with m i n i m u m fe ature size.
L3
Semiconductor Technologies
Ta ble 1
Compa rison of Single Chip VAX M i c roprocessors
Process
M inimum
Feature
Size
Chip
Tape
Out
Date
CMOS-2
1 . 5 p.m
REX520
Sep 87
8.5
1 1 .95
28
460X 460
320,000
CMOS-4
0.75 p.m
NVAX
Nov 90
40.5
5.85
12
636 X 574
1 ,300,000
Performance
SPECma rks*
Cycle
Time
Cycles per (Nanolnstructiont seconds)
Chip
Size
(Mils)
No. of
Transistors
Notes:
'These are combined i nteger and floating point SPECmarks run on a VAX 6000 Model 4 1 0 (REX520) and a VAX 6000 Model 610 (NVAX).
1 C PI depends not only on the CPU chip, but also on the memory subsystem and the particular program being executed. The CPI values
quoted here are a composite for the ten benchmark programs in the SPECmarks suite.
Table 2
Compa rison of Fea t u re Sizes in CMOS Ge nerations
Gate Dielectric T h i c kness (A}
CMOS-1
CMOS-2
CMOS-3
CMOS-4
300
225
1 50
1 05
4/2
2/2
4/2
3/1 .5
2/1
M i n i m um Featu re/S pace (p.m}
Active area
Po lysi l i con/polycide
Metal 1 with contact o r via
Metal 2 with via
Metal 3 with via
M i n i mum Featu re Size (p.m}
Metal 1 contact
Metal 2 contact
5/2
1 . 5/1 .5
3/1 .5
3.75/1 .5
1 /1
2/1
2.5/1
4/6
2
1 .5
1
2
Metal 3 contact
2
8
1 .875/0.75
3/4.5
0.75
0.75
1
1 .5
M i n i m um Spaci n g (p.m)
Metal 1 contact/pol ys i l icon
(in active a rea}
P +/N + active area
1 .5/0.75
0.75/0.75
1 .5/0.75
1 .5
6
4
3
1
0.75
3
4
From the 1 .5 -p.m minimum feature size of CMOS-2
sis tors form P+ IN diodes to the n-wells; the n-well
technology to the 0.75 -p.m size of CMOS-4, the area
of the gates was scaled by a factor of fou r and the
is held at vdd (power supply voltage).
gate dielectric thickness was ha lved . The resu lt is a
function of its shape and size: there is both an
twofold reduction in gate capacitance (C = E0A it).
area component and a perimeter component.
The typical gate dielectric thickness in the 0.75-p.m
CMOS-4 process is 105 angstroms (A). Manufactura­
The capacitance of a reverse-biased d iode is a
c[ Q!al
=
CJJ'l';l X Area + cperimC!lT X Length_of_Perim eter
bility and rel iability considerations have been the
major factors determining the minimum gate
Since the area scales with the square of the mini­
mum feature size and the perimeter scales d irectly
dielectric thickness used for each generation.
with the feature size, C101" 1 scales by somewhat
As shown in Figure 2, the sources and drains of
n-channel metal-oxide semiconductor (NMOS) tran­
more than the minimum feature size: the exact
amou nt depends on the shape of the source or
sistors form N +/P diodes to the substrate. Since
drain. For the NVAX m icroprocessor, which was
the p -type doped substrate is held at a potential of
designed in CMOS-4, the area and perimeter compo­
�s (ground) and (during normal operation) the
sources and drains are a lways at �s or higher, these
diodes are always reverse biased and act as voltage­
nents contribute about equally to C101. .1 . In future
technology generations, the p erimeter component
will tend to dominate.
dependent capacitors. The sources and drains of
In Digital's CMOS processes, metal i nterconnect
p-channel metal-oxide semiconductor (PMOS) tran-
widths and spaces are scaled with the minimum fea-
14
Vol. 4 No. 2
Spring 1992
Digital Tecbnicaljour11al
Microprocessor Performance and Process Complexity in CMOS Technologies
ture size, b u t metal thicknesses a n d d ielectric th ick­
nesses are held constant to avo i d t h ree u n desirable
effects. Sca l i ng the metal t h ickness wou ld i ncrease
the sheet resistance (lead ing to l a rger power supply
voltage drops and
RC
t i me constant de lays) and
decrease the curre n t - carrying cap abil ity of the con­
ductor l i n es (from a n electro m igration view point) .
Sca l ing the in terconnect d ielectric th ickness wo u l d
i n c rease the capacitance per u n i t area . Because the
thicknesses of conductor l i nes and d ielectric layers
are not scaled, the aspect ratios of the spaces
between the condu ctors and the con tacts o r v ias
between i n terconnect l ayers i n c rease. This m a kes
fab rication more d iffic u l t .
A s w i t h sources and d rains, i n terconnect capaci­
tance has both at1' area componen t, which sca les
quadratical ly, and a peri metl'r compone nt, w h i c h
s c a l e s l i nearly. Consequently, the total capacitanc e
of i n tercon nect scales by somewhat more than the
m i n i m u m fea t u re size_ Because neither the i nter­
connect thick ness nor the d ielectric thickness is
scaled, the capaci tance benveen adjacent conduc­
tor J i nes increases. The res u l t is an i n c reased suscep­
t i b i l i t y to cross - t a l k between adjacent bus signals.
Fig u re 2
Diagram of NMOS an d PMOS
which is s hown in Figure :). In the NVAX micro­
Trausistors Shuwi11g Gate,
processor, greater- than-mi n i m u m spaces were used
Source, and Drain for CMOS-4
on some critical buses to reduce cross - t a l k .
CMOS-3
CMOS-4
SU BSTRATE
Note: C13 a d C 4 are the total capacitances lor the center line lor CMOS-3 and CMOS-4. respectively.
Dimensions are typical.
Figure 3
Cross Section of Three /Vlinimum -spaced Meta l
Digital Tecbuical jou.rual
Vul. 4 No. 2
.\j1ring /1)<)2
I Lines Drawn
to Scalefor Ct\IJOS-3 and Ct\IJOS-4
15
Sem iconductor Technologies
The use of industry-standard power supply volt­
ing a CMOS process. A process for low- to mode rate­
ages n::sults i n a significant violation of constant
volume, high-perfo rmance microprocessors differs
field sc a l i ng rules. Nevertheless, power supply volt­
from a process optim ized fo r fast turnaround gate
age is general ly held constant across two or more
arrays or high-volume RAMs. In a high-volume prod­
process generations [5.0 volts (V) i n CMOS-I and
CMOS-2 and
3.3 V
uct, a great deal of effort is devoted to reducing the
in CMOS-3 and CMOS- 4 ] i n order
total nu mber of process steps. For exam ple, com­
to mai ntain volt age compat i b i l i t y with industry­
pensating blanket implants are often used to set the
standard chips such as RAMs. However, a nonscaled
thresholds of the transistors to decrease the num­
power supply voltage presents formidable chal­
ber
lenges for the design of reliable transistors.
approach couples the parameters fo r the NMOS and
of
photo l i t hographic
masking
steps.
This
PMOS transistors, making parameter adjustments
Developing CMOSfor Microprocessors
more d i ffi cult and requiring more development
The particular implementation of transistors, inter­
effort.
connect, and special circ u i t elements in a c�os
process depends on the application. For Digit al's
To produce h igh-performance m icroprocessors,
Digital has developed many unique features fo r its
high-speed m icroprocessors, performance, as mea­
CMOS
sured in SPECmarks, is crucia l . '\ I n a d d i t ion to opti­
ogy features that have been developed for each pro­
mizing the transistors for maximum drive current,
cess generation to meet increasingly demanding
technologies. Ta ble 3
I ists
the new technol­
performance in this appl ication can be improved
performance requirements. We begin our d iscus­
by
sion of the implementation of these features and
adding process
feature s to
provide
denser
interconnect
the requ irements for rel i able circ u i t operation by
with high current capabi l ity, and precision resistors
addressing the issue of power d i ssipation in high­
for impedance matching. As discussed above, per­
speed microprocessors.
on-chip cache static RAM (SRAM),
formance can als o be improved by increasing the
die size. By contrast, a major part of the effort in
designing a process for dynamic RAMs
(DRAM)
is
Power Supply Voltage
I t is well known that CMOS power dissipation is
C X �,/ X J, where C is the switched
vdd is the power supply voltage, and f
clock frequency. A reduction in 1-drt is an excel­
d i rected toward developing a very s m a l l , high­
dominated by
capacitance memory element.
capacitance,
I n addition to performance, the planned pro­
duction volume is an i mportant factor in detin-
Table 3
is the
lent way to counteract the increase in power due to
Features Added by Generation for CMOS-1 to CMOS-4
CMOS- 11
CMOS-2
CMOS-3
CMOS-4
M asks
12
13
20
21
Minimum
D i m ension
2.0 11-m
1 . 5 11-m
1 .0 11-m
0.75 11-m
New
Featu res
1 X photol ithog raphy
Lightly doped dra i n
junction/spacer
5 X g-line
photolithography
Local
i nterconnect
Epitaxial layer
Tu ngsten
s i l i cide gate
n, p-polys i l icon
Tungsten plu g
for M1 C, M2C
Borophospho-
Deep P+ ring
n-we l l
sil icate g lass
planarization
Photoresist
etch-back
plana rization
Cobalt s i l icide
TiN barrier
Spi n-on-g lass
etch-back
planarization
AI : 1 %Cu
Metal 3
Fuse
16
Vt>l. 4 No. 2
Spring /'}92
Digital Techllical jounllll
Microprocesso r Performance and Process Complexity in CMOS Teclmologies
the h igher frequencies and larger switched capaci­
Hot Carrier Degradation
tance (which results from the increase in d ie size).
A CMOS transistor that is subjected to excessive volt­
When developing the CMOS-3 process, we chose tO
ages becomes damaged over time by hot carriers.
reduce the power supply vol tage from the industry
Hot carriers are h ighly energetic current carriers
norm of 5 v to the Joi n t Electronic Device Engi­
neering Council Ot:DEC) 3.3 V standard. 1 "· 1 1
that resul t from the high electric fields in the tran­
With the CMOS-4 process specified for 3.3-V
susceptible to hot carrier degradation than the
power supply voltage, the N YA.."\ and the Alpha
2 1 064 microprocessors consume
16 watts (W) at
100 megahertz (MHZ) and 27 W a t 200 MHz, respec­
sistor. To date, the N MOS transistor has been more
PMOS transistor.
Hot carrier dam age gradua l ly
reduces the satu ration current fos!IT of the N MOS
transistor as the damage i ncreases over time. On
tivdy. If the supply volt age were 5 .0 V, the power
would scale by (5 V) 2/(3.3 V) 2 = 2.3. This increase in
chips with a nominal 3.3-V power supply, some
power dissipation woul d have greatly increased the
transients as h igh as 4.3 v. Table 4 gives details of
complexity and cost of the chip packages.
the origins of these h igh voltage transients.
transistors are subjected to sou rce-drain voltage
Significant changes tO the N,\' IOS and PMOS transis­
Hot carrier rules fo r the CMOS-4 process are illus­
tors were necessary to optimize t he process fo r
trated in Figure 4, which shows the three legal
operation at 3.3 V. The most visible parameter
regions of device operation on a plot of
change was a lowering of t he target threshold vo lt­
source voltage) versus "':ts (drain-to-source voltage).
ages for the NMOS ancl PMOS transistors by about
I 0.4 1 v to 0.5 V and
- 0. 5 V, respectively. To explain
�s (gate- to­
Devices may operate in any, or all , of three regions:
(1) u ncondi t i o n a l ly safe region, (2) region subject
why this is necessary, we must consider the depen­
to turn-on transient rule, and (3) extended safe
dence of both the nodal transition time (which is a
region for "off" devices. Devices can spend up to
good measure of circuit performance) ancl the tran­
sistor current:-; on 1�111. The time required to transi­
UNCONDITIONALLY
SAFE REGION
tion a signal node bet ween the power supply rails is
proportional to the charge (Q) on the node ancl
inversely p roport ional to the drain-to-source cur­
(111) of the driving transistor. Since Q a t.-;1" and ,
first order, Ids a "':,,/. the time required to transi­
/-r-L---,------,v
rent
to
tion a node is inversely proportional to "':td· How­
/
4 . 5 V 1----Vl
(J)
>
ever, when second-order effects are considered, the
3.3-V technol ogy is of about the same performance
as the corresponding 5 -V technology. The second­
order effects include the benefits from lowe ring the
threshold voltages of the transistors in the 3.3-V pro­
REGION SUBJECT
TO TURN-ON
TRANS I E NT RULE
EXTENDED
SAFE REGION
FOR OFF DEVICES
/
0.3 v
3.6 v
0.0
4.3 v
/l
4.5 v
cess and the compromises that would have to be
made to the transistors in the 5-V process to make
Figure 4
them reliable.
Ta ble 4
CMOS-4 Hot Carrier Rules
Orig ins of H igh Vo ltage Tra nsients
(Volts)
vd
s
Subtotal
(Volts)
Comments
Power supply
3.30
3.30
N o m i nal voltage
Power supply tolerance
0.1 65
3.465
5% tolerance, includes r i pple
On-c h i p power supply
0.1 75
3.64
From NVAX SPICE simu lations. Half of peak-to-peak
ringing d ue to package
noise (Vdd-internal with respect to lf,5_internal)
i n d u ctance
Booting above Vdd d ue
to capacitive cou pling
0.66
Total
4.30
Digiln/ Teclmical journal
Vol. 4 No. 2
4.30
Spring 1992
Capacitive coupling to suscepti ble nodes i s l i m ited to
< 20% by desig ners
17
Semiconductor Technologies
100
percent of the t i me in the safe regi on, no more
than
i n CMOS-3 helped t o counteract the effect o f scal­
5 percent in the region subject to turn-on tran­
sient rule, and no more than
10 percent in
ing on J ,vcragc·
the "off"
Scaling the interconnect and dealing with electro­
devices safety region.
m igration issues are some of the most formidable
A wide variety of NVAX circu its wen: simu l ated
challenges that must be faced as feature sizes con­
to determine what constraints shou ld be placed
tinue to decrease in the next decade.
upon circuit design style i n order to ensure that
the
CMOS-4
Substrate Contact
hot carrier rules were not violate d . A
set of general circ u i t design constraints was devel­
As mentioned earlier, the substrate m us t be con­
oped, and a compu ter-aided design (CAD) tool
nected
was written to ensure that all the circuits on the
NVA.X chip observed these constraints. The hot
carrier CAD checks were run prior to fabrication,
and circu its t h a t violated the constraints were
'
�-
to the
of the
chip
through
a low­
i mpedance path to prevent any rise in voltage. I f
t h e substrate voltage r i s e is severe, NMOS source/
drain diodes wil l conduct, and if sufficient charge
is injected , the chip may latch-up. Latch-up is a
redesigned .
destructive
Electromigration Considerations
The process, circu its, and
If the average current density ( '"·mge ) through
designed to prevent latch-up from occurring.
mech an ism
involving
the
parasitic
bipolar transistors fo rmed by the CMOS process.
V:s
substrate contact are
an aluminum conductor l i n e is too high, the con­
The usual industry substrate connection method
ductor line is suscep tible to metal migration. This
depends on a path through bond wires and the
phenomenon occurs over t i me as the e lectron
package to connect between in ternal V55 and the
deposits
substrate. To ensure a good substrate contact,
downstrea m . Eventu a l ly a short circuit or an open
Digital's CMOS technologies incorporate a deep P +
current fo rms
voids
at
one
site
and
circu it develops, which resu lts in a circu it fa ilure.
implant (DPI) around the edge o f the d i e to connect
Chip designers guard against electrom igration fa il­
the
for every conductor l in e
ure by ensuring thar J,v� ra
g�
on t h e chi p is lower t h a n the maximum allowed
V55 metal
on the d ie surface to the low-resistance
substrate . The implant creates a low-resistance
path through the ?-epi taxial layer i n which the
NMOS and PMOS transistors are formed.
value.
For a conductor l ine that is switched every cycle,
The DPI is a low- induct ance path when com­
the relationship between average current density,
pared to the standard method that connects the
mi croprocessor cycle time
sectional area is given by
Cross-Sectional Area) .
C�vc Je), l�trt•
and cross­
!"nag� = ( C X �'" ) /C.Zrcle X
It is interesting to note the changes to
lav�ra ge
substrate through the package. The additional area
enclosed by the path through the boncl wires and
package impl ies greater inductance, which i s unde­
sirable fo r h igh-frequency signals. The DPI path
V . and substrate has low inductance
.\.\
for a conductor I ine as a chip is shru nk from
between
one generation to the next. The node capacitance,
because i t is made directly on-chip.
C,
decreases by s l ightly more than the sca l i ng
scal ing factor since the chip can now run faster;
Technology-limited Gate Dielectric
Thickness
and
As stated above, maximizing the current that a tran­
factor;
�"'
the
remains constant;
cross-sectional
� rck
reduces by the
area decreases
by the
�'"
;"' era�;�
increases by
lf.f1Vnage
now exceeds
transistor gate length, gate dielectric thickness, and
the circu i t must
threshold voltage i m proves the d rive current. The
scaling factor. Consequently,
sistor can supply at a given
is of uppe rmost
s l ightly less than rhe scal ing factor as the wiuth of
importance for circu i t performance. Sca l ing the
the conductor l i ne shrinks.
the maximum
al lowed value,
be redesigned . If there i s enough space , .favera ge can
transistor gate length is constrained by the mini­
be reduced hy widening the conductor l i ne so
m u m polys i l i con l i n e-width feature; the minimum
that the cross -sectional area is increased. From this
threshold voltage is set by the kakage current
brief analysis, it is clear that it becomes more
al lowed when the transistor i s turned off. However,
difficu l t to observe the electromigrat ion l i mits as
s c a l ing does
the technology scales-even when the metal thick­
between feature size and gate d ielectric thickness;
ness i s not scaled. As can be seen from the .fwcra gc
scal ing only determines the change from one gener­
equation, reducing
18
�td from 5.0 V
in CMOS-2 to :).3 V
not
establish a fixed
n:lationship
ation to the next.
Vol. 4 No. 2
!>Pring 1992
Digital Techuical journal
Microprocessor Performance and Process Complexity in CMOS Technologies
Figure 5 shows how the saturation currents for
required development of a new sil icided gate pro­
both NMOS and PMOS transistors in CMOS-4 depend
cess. The new process, known as sal icide for self­
on the gate d ielectric thickness. The dielectric
a ligned silicide, forms the sil icide on the gate and
thickness range plotted spans appl ications from
on the sou rce/drain regions after a l l the required
microprocessors to SRAMs. Both curves are only
transistor implants have been completed. This
slightly sublinear; thinning the gate d ielectric
reduces the sheet resistance of the source/drain
provides almost a one-to-one return in transistor
regions by more than an order of magnitude and
saturation current. In high-performance micro­
al lows them to be considered for use in local sig­
processor appl ications, rel iabil ity and manufac­
nal rou ting. The reduced sheet resistance, how­
turability considerations determine the extent to
ever, does l it tle to i mprove the current drive of the
which the gate d ielectric thickness can be reduced.
transistor; for a typical CMOS-4 NMOS transistor,
D igital 's C\10S technologies have consistently used
MINIMOS simu lations show that the use of sil icided
source/drain regions improves the saturation cur­
rent by only 0.6 percent. 1 2
thinner gate dielectrics than industry norms.
4
Precision Resistor
Although transistors are the dominant element in
logic design, a resistor is sometimes needed, for
NMOS TRANSISTOR
o<(
lf)
example, to match the impedance of an output
driver with the impedance of a board-level trans­
:":
mission l ine that it drives. MOS transistors make
3
poor controlled impedance drivers because they
X
0
f-
change impedance as a function of drain voltage.
@J
One method of controll ing the impedance of a
0.
rn
�
...._
�
�
driver is to use a diffusion resistor as the dominant
element , as shown i n Figure 6. The MOS transistors
2
are sized such that their on-state resistance is much
lower than that of the resistor. Therefore , if the
transmission l ine impedance is 50 ohms, the resis­
tor plus transistor impedance can be sized to match
that value with little i nfluence from the variations
7
10
11
in transistor impedance.
Resistors are constructed from nonsilicidecl dif­
fusion to meet tolerance requirements that would
not be possible with a sil icicled version. Because a
Figure 5
Normalized Drive Current as
a Function of Gate Dielectric
Thicknessfor CMOS-4
silicidecl resistor has lower sheet resistance, it is
much longer and narrower than a nonsil icided
BOARD
CHIP
Silicided Source/drain and Gate
The basic gate material for an MOS transistor is
highly doped polysil icon. The sheet resistivity of
this material in the CMOS-1 process was 40 ohms
per square. For CMOS-2, the RC time constant delay
associated with this sheet resistance would have
J,
created nonuniform turn-on of wide, fast-switching
output transistors. A tungsten silicide layer was
added to the polysil icon to form a polycide. The
DRIVER
TRANSMISSION LINE
RECEIVER
sandwich structu re reduced the sheet resistance
of the gate material to 3 ohms per square.
Changes to the transistor process for CMOS-3
technology, which were continued into CMOs-4,
Digital Techtticaljournal
Vol. 4 No. 2
Spring1992
Figure 6
Precision Resistor Use in
Impedance Jll/atching
19
Sem iconductor Teclu10logies
resistor of the same value. A n arrow resistor i s
more suscept i ble t o varia t i ons in field d ielectric
six- transistor static cell layouts show an array of
fo ur cells with the PMOS load transistors at the top
encroach m e n t . wh ich lowers the tolerance of the
and bottom of t he arrays and the NMOS pass transis­
resistor. The tol c ranu: i s lowered furthn by p ro­
tors, which provide access to the bit l i nes, in the
cess variati o n : it is more d i ftic u l t to contro l the
center of the array. The c e l l without local intercon­
sheet resist a nce of the sil icided d iffusion than that
nect is 120 p.m2 in CMOS-4, compared to 98 p.m2 for
of the nonsil icidcd d i ffu sion.
the cel l with local i n terconnect. The 18 p e rcent
The prec i s i o n resistor is also an important e l e­
p ro t ec­
has ft.:an1res that i n c rease yield . There a re only
an l/0 d river with the ESD prot ection components.
i n terconnect (M l ) a nd d iffusion or polysi l icon fo r
ment of ou r e lect ro s t a t i c d ischarge
(l:SD)
i m p rovement i n area is important, but the cell also
t ion s t r a tegy. Figure 7 shows a simplt.: schematic of
2 . 5 contacts
between the first level of a l u m i n u m
Clamp 1 is the main path fo r shu n t i ng current d u r­
the local interconnect cel l , compared to 8 in the
res i s t o r (R) 1 and clamp 1 ensure t h a t the clamp
local interconnect cel l is a t m i n i m u m pitch (where
ing an
ESD even t .
The posit ion and construction of
impedance is lmvcr than that of a s h u n t i ng p a t h
through H 1 a nd the driver. R 1 is a l s o t h e impedance
n o n- local i n terconnect cel l . None of the M 1 in the
p itch equals .\ 1 1 w i d th plus s p ace), whi le a l l the
\1 1 i n the other cell is at m i n i m u m pitch. Final l y,
.
matching resistor fo r the output d river. R 2 prov ides
the M 2 pitch is smal l e r in the local i n terco nnect
a n ad d i ti onal level of protection for the gates of the
cel l , b u t at 3.19 p.m, it i s s t i l l greater than the mini­
input driver in con j u nction with the smal ler clamp
mum of 2.63 p.m a l l owed by the techno logy. Al l of
2 placed ncar th<: input d rive r. Both Rl and R2 are
these factors add u p to a significantly more yield­
fa bricated using the precis i o n resistor mask l ayer.
able cell; this is important since 1 5 p ercent of the
area and abo u t two - t h i rds of the transistors on the
Alpha 2 1064 chip are SR.AJ.\1 cel ls.
Thick Metal 3 Interconnect
High-speed operation of the dense circuitry in large
m icroprocessors results in a level of power d issipa­
tion not encoun tered in gate arrays o r RA.I\h The
interconnect of a m i c roprocessor has to carry tens
of amp eres of instan taneous cu rrent i n to and o u t
OUTPUT
DRIVER
INPUT
DRIVER
of the chip i n a d d i t i o n t o ro u t ing signals. Further­
m o re ,
since
high-perfo rmance
m icroprocessor
clock frequencies are of the order of h u n d reds of
Fig ure 7
megahertz, the on-chip clocks must be d istributed
Precision Resistor Use in Electrostatic
Discharge Protection
with very low RC time delay constants. These
req u i rements lead to a
bet ween
Local Interconnect in the SRAM
As stated earl ier, one of the microarch itectural
methods fo r increasing p e rfo rmance is to bu i l.d as
large a cache memory as possible o n the d i e . D ue
the
number of d ifferences
interconnects
used
fo r
m icro­
processors, gate arrays, and HA\'ls. RAMs at the 1-p.m
feature size are u s u a l l y designed with two levels
o f aluminum-based cond uctors (M 1 an d M2) that
are very similar i n th ickness (a p prox i m a tely I p.m),
RA\l-specific process
m i n i m u m wid t h , and s paci n g . Gate arrays genera l l y
enhancements an: genera l ly not implemented in a
add a t h i rd level of a l u m i n u m -based interconnect
to fa brication complexit y,
process t a i lored to m icroprocessors. However, the
(M3) to i m p rove signal rou t i ng and gate util ization
CV!OS-1 t echnol og y does include one feature, called
in the a rray. This level has very s i m i lar characteris­
loca l i n terconnect, that significantly decn:ases the
t ics to M 1 and M 2 . Since the transistor density in a
cel l size. In the C\ lOS-'i i m plementation, local inter­
connect is
a titan i u m -n itride (Ti�) l ayer that pro·
vides d i rect contact between the polys i l icon and
diffusion l aye r s . 1 1
Figure H shows a compa rison of layouts of mem­
ory cel l s with and without local i nterconnect. The
20
gate array is low and not a l l the gates are used in a
design, the power d issipat ion is usua l ly mod erate
by m icroprocessor standards.
The first two layers of interconnect on a high­
performa nce microprocessor are very similar t o t he
co rres p o nd in g l ayers on an SRAJ\.1 or gate array.
Vol. ..; No. .!
SjJrinJ!, I'J'J.!
Digital Technical journal
Microprocessor Performance and Process Complexity in CMOS Technologies
KEY:
·-···-1
rL·-··
· -"
•
D
D
•
N-WELL
POLYCIDE
METAL 1
METAL 2
•
D
P+
!81
MlC
•
M2C
LOCAL INTE RCONNECT
N+
(a) Without Local interconnect
Figure 8
Digital Technical jounwl
Vol. 4 No. 2
(b) With Local Interconnect
Six- transistor SRAM Cell Layout
.�pri11g l')'Jl
21
Sem iconductor Technologies
However, because of the high power supply cur­
chip designers must account for these manufactur­
rents and low skew clock d istribution required for
ing process variations when the chips are being
high-speed operation, M3 in CMOS-3 and CMOS-4
designed.
p rocesses is approximately 2. 5 t imes thicker. To
The lot-to-lot variation in characteristics for
avoid an impact on yield, the pitch of M 3 is chosen
CMOS-4 transistors is significantly larger than for
to be approximately three times larger than that
bipolar transistors. Figure 9 shows how the satu­
for i\12. To reduce capacitance to Ml and M2, the
d ielectric u nder M3 is thicker than that between
M l and M 2 .
ration current (I0_1A1) varies in manufacturing for
CMOS-4. The five points on the plot of PNIOS IIJ.\AT
versus NMOS 1/JSAT represent the process extremes
and are often referred to as the process corners.
Fuses
As the number of memory cel ls increases on a
microprocessor, the impact of those cells on the
yield of the die increases. CMOS-3 and CMOS-4 tech­
nologies implement redundancy by a standard tech­
nique of laser-fusible l inks to remove bad cel ls and
incorporate new ones into the array. Digital's pro­
cess d ifkrs from others in the implementation of
the fuses, however. Standard RAM processes use
polys i l icon fuses for their smal l size and ease of
ablation. Because of the thickness of d ielectric l ay­
ers that would need to be etched to uncover the
fuse, the CMOS-3 and CMOS-4 technologies could
Each process corner has a two-letter label: FF, TT,
SS, FS , and SF. The first letter of the pair is used to
refer to the PMOS device : F indicates a fast (i. e . , high
current) device; S i ndicates a slow (i . e . , low cur­
rent) device; and T indicates a typical (i.e . , mami­
facturing target current) device. The second letter
of the pair refers to the NMOS device. Thus, the FF
point in Figure 9 represents the fastest PMOS device
paired with the fastest NMOS device; the 55 point
represents the slowest PMOS device p aired with the
slowest NMOS device; and the TT point represents
a typical PMOS device paired with a typical NMOS
device.
not use polysil icon fuses. Sufficient control of the
dielectric etch rate and the selectivity of the d ielec­
1 .5
tric etch to polysil icon cannot be achieved in a
manufacturing environment for the thin poly­
sil icon fuse layer to be left intact. Instead, the
80-nanometer (nm) 111m of TiN that forms the bot­
tom layer of M3 is used. The upper layer of alu­
minum copper (AICu) is selectively etched through
a special mask to leave a 3-p,m wide strip of TiN that
can be ablated by industry-standard lasers. 1·•
Fusible l inks can be placed on a chip to form a n
identification register. A laser c a n then be used to
program a unique code into this register. The con­
�
-.3'
�
::2
a..
Cl
w
N
::J
<(
::2
a:
0
z
1 .0
F � FF
0.5
�cSF
SS
1 .0
0.5
NORMALIZED NMOS ldsat
0
tents of the identification register on the l\l\IAX chip
can be read by system software so that individual
die can be un iquely identified not only during
system manufacture, assembly, and test, but also in
the field.
Manufacturing Process Variations
and Chip Design Strategy
Figure 9
1 .5
Comparison of Saturation Currents
of NMOS fosAT and PMOS !05A7 for
CMOS-4 Process Corners
The fF SPICE'" models are used to pred ict the
speed of the fastest chips, the m aximum power dis­
As mentioned earlier, the performance of a CMOS
sipation, the transient current demands on the
process is a function of the current driving capabil­
power supply, the maximum vol tage drops in the
ity
on-chip power and ground rou ting, the worst-case
Oc�) of the PMOS and NMOS transistors, as wel l as
the capacitances of the gates, sources, drains, and
current density (checked to ensure that the electro­
interconnect that the PMOS and NMOS transistors
m igration l imits are not v io lated) in power supply
must charge and discharge. Al l of these parameters
and signal l i nes, and the m aximum rate at which
vary in manufacturing. In order to ensure that chips
the chip's signal pins wi l l transit ion. Power supply
will fu nction correctly and at the planned speed,
current transients and signal-pin transition rates
22
Vol. 4 1Yo. 2
Spring 1992
Digital Technical journal
Microprocessor Pe1jormance and Process Complexity in CMOS Technologies
have important impl ications for the electrical
design of chip packages. Both place limits on how
much inductance can be tolerated in the leads. The
a CMOS inverter changes from FS to TI to SF. C ircu it
designers use DC simu lations l ike these to deter­
m ine the safe bounds for the sizes of transistors
maximum power dissipation has obvious thermal
in a variety of common circu it structures. These
implications for package and heat sink design.
bounds are incorporated into t he design methodol­
The ss SPICE models are used to predict the speed
ogy for the project, and CAD tools are used to
at which the slowest chips wiU run. The TI models
search the circuit schematic database for structures
are used to check that the circu its on the chip will
that violate the methodology.
run at the desired speed when the manufacturing
When creating schematics, circu it designers use
process is at the center of its range. The bulk of the
technology-specific rules of thumb to estimate
circuit design work for the NV�'{ and Alpha 2 1064
chips was done using the TI models.
The FS and SF SPICE models are used to determine
the noise margin for circu its (DC circuit analysis)
rather than to predict the speeds at which circuits
will run (AC analysis). The FS model has a semi-fast
P:\10S transistor paired with a semi-slow NMOS tran­
sistor; the SF model is just the opposite. The param­
eters that determine IDSAT for PMOS and NMOS
transistors are correlated. These correlations are
taken into account in the FS and SF models. The cor­
rect operation of some CMOS circuits is particu larly
sensi t ive to the ratio of the PMOS to NMOS currents.
By using the FS and SF models to simu late these cir­
cuits, designers can verify that the circuits wil l
function correctly in spite o f variations i n the man­
ufacturing process. A simple example is given in
Figure 10, which shows how the switching point of
3.3
the interconnect capacitance on signal l ines, etc.
Layout for the schematics is then generated, exact
capacitances are extracted from the layout using
CAD tools, the capacitance estimates are replaced
with the extracted values, and the circu its are
resimulated to ensure that they stil l meet the
specifications. The capacitance extraction tool can
be rerun for a different process corner (dielectric
thicknesses, etc.) by changing its parameter file.
Although not discussed here, environmental
effects such as operating temperature and power
supply variations must also be taken in account.
If CMOS chips are to be manufacturable, design­
ers must account for process variations during t he
design phase by fol lowing procedures such as those
just outl ined. In order to get to market quickly,
the NV�'I: microprocessor was being designed while
the CMOS-4 process was being developed. Process
Voo
=
3.3 V
V;n
Vaut
Vss
>� 1 .65
0
3.3
Figure 10
Digital Technical journal
--c>--
Vol. 4 No. 2
CMOS-4 /nverter Switching at Process Corners
Springl992
23
Semiconductor Technologies
simu lators and rest ch ips were used to gem:ratc the
7.
C�'IOS-·1 worst -case and t y p ica l electrical models
R . Dennard , f Gaensslen, H . Yu . V Rideout,
E. Bassous, and A. LeBlanc, '' Design of Ion­
for transisto rs and i n t e rconnect d uring the early
implanted MOSFETs with Very Small Physical
phase of process deve l opment. The accuracy of
D i men s ions ; ·
t imely completion of t h e :'�: VAX design: it was never
nccessarv to nxlcsign circu its due to process or
8.
9.
fr o m scal ing, i n c rc.a s c d d ie area, and new technol­
n:duced instruct ion s c. t (!USC) ami complex instru c­
t ion set (CIS<:) m icroprocessors reported in the
10.
11.
R. Badeau c:r a l . , " 100�1Hz Macrop i p e l ined
C!S C CMOS M i c roprocessor," IEEE Interna­
12.
13.
instru c t ions execu ted . Note: strictly speak­
Hennessy and
D.
Patterson,
Mateo: Morgan Kaufmann, 1990).
5.
SPECmark is a quantitative measure of perfor­
mance determined by r u n n i ng a suite of
10 benchmark programs. For example, a VAX
A.
Schiiltz,
and
H.
Potzl,
A. Nasr, G. Gru l a , A. Berti, and R. Jones.
vol . 4, no. 2 (Spring 1992, this issue): 39- 50
14.
M . Coffey and R. H o l l i ngswort h , " I ntegrated
C ircu i t Hav ing Laser-Al terable Metal l i zation
Layer," U.S. Patent 4,849, 363, Issued J u l y 1989.
Co mp u ter
ArciJitecture: A Quantitative Approa ch (San
Selberherr.
On-chip Memory," D ig i tal Jxl:mical}oumal
i
i ng, CPI is also a function of the system
memory su bsystem , etc.) and the compiler.
S.
"CMOS-4 Technology for Fast Logic and Oense
grams is d ivided by the total number of
c onfigurat i o n ( i . e . , the size and speed of the
Joint Electro n i c Device Engineering Cou n c i l
tor Analyzer."' IEEE J l·cmsactions on Electron
Devices, ED-27 (1980): 1540- 1550
and
clock cycles requ ired to execute these pro­
R. Allmon et a l . , "System, Process and Design
" MJ.'>: IIVlOS-A Two-Di mensional .VlOS Transis­
To compu te clock cycles per instruction
benchmarks arc Jirst run . Then the number of
G. Baccaran i , M . Wordeman, and R . Dennard,
(JEOE C) J EDEC Standard No. 8 Standard fo r
Reduced Operating Voltages and ln tetfa ce
Levels for Integrated Circuits (\Vashington
D.C. : Electronics Incl ustries Associat i o n ,
March 1984)
tional Solid-State Circuits Conference D igest
of Techninil Papers (February 1992): 106- 107.
programs
Det,ices,
M i croprocessor," /L"Ef International Solid­
State Circuits Conference Digest of Technical
Papers (February 1'-)90): 48-49.
I J Dohherpuhl e t a l . , "A 200MHz 64b Dual­
appl ica t i o n
F:"lectron
Impl ications of a Reduced Su pply Vo l t age
References and Notes
tional Solid-State Circu its Co nfe rence Digest
of Tec!Jnical Papers (February 1992): 104- 105.
on
(April 1984) : 4'52 -462.
industry to d a t e .
Issue CMOS M icroprocessor," IEEE Interna­
Tra nsac tio n s
Transactions on Electron !Jeuices. vol. ED-31
implemented in CMOS-4, are the h ighest p e rforming
].
VLSI Technology:
" General ized Scal i ng Theory and Its App l i­
every two years. The Al pha 21064 and N VA '1:
.- chips,
4.
J.Llll MOSFET
E. Wal ker, and
cation to a X M i c ron MOSFFT Design .'" IEEE
ogy features have al lowed increased p erformance
typical
F. Gaensslen,
vol . ED-26 (Ap r i l l979): 32';-333.
specifically for h igh-perform:�nce m icroprocessors.
Generat ion- to-generat ion improvements derived
(CI'l),
·· I
for High-p erfo r m ance Logic Appl ications;·
IEEE
D igital 's CMOS processes have been developed
3.
Dennard ,
Part ! 1 - Device Designs and Characteristics
Conclusion
2.
R.
P Cook,
model changes d ur ing the course of the proje c t .
I.
IEEE Joumal r�( Solid-State
Circuits, vol . SC-9, (1974 ): 256- 268.
t hese models was critictl to the successful and
15.
SPICE is a general-purpose c i rcui t s i m u lator
program developed by Lawrence Nagel and
E l l i s Cohen of the Depart ment of E lectrical
Engineering ancl Computer Sciences, Univer­
sity o f Ca l i ti >rnia, Berkeley
1 1 / 780 system has a SPECmark value of 1 .0.
6.
R. Allmon et a l . , " CMOS I m plementation of a
32b Computer,'" IEEE International Solid-State
Circuits Conference Digest of Tech ni cal
Papers, vo l . 32 (February 1989): 80- 8 1 .
24
Vol. 4 .Vo. 2
SjJri!lg 1')')2
D igital Techuica{ jouruaf
Marden H. Seavey
john V. Faricelli
Nadim A. Khalil
Gerd Nanz
Llanda M. Richardson
Christian 0. Schiebl
Hamid R. Soleimani
Martin Thurner
Numerical Device and
Process Simulation Tools
in Transistor Design
Numerical device and process simulation programs are fundamental tools in the
design and characterization of silicon transistors. These tools employ numerical
mathematical methods to simulate the operation of the elemental transistor struc­
tures that are the building blocks of CMOS VLSI circuitry. When desiRning these
basic structures, CMOS process and device design teams require efficient, high­
performance simulators that use accurate physical models. Digital has developed
thermal annealing, mobilit)J, and avalanche models, and has improved the numer­
ical methods used in its process and device simulation programs. Also, supporting
software was developed to help integrate the various simulation tools.
With the i ncreasing nu mber of transistor functions
provide accurate pred ictions. D igital's Advanced
implemented on each chip in complementary
Semiconductor Development (ASD) Submicron
metal-oxide semiconductor (CMOS) very large-scale
Simu lation Group modifies and extends the capabil­
i ntegration (VLSI) circuitry, computer simulation
ities of simu lators that have al ready been develop ed.
tools have become essential a t all levels of the
These modifications and extensions have resulted in
design process. This is particularly true at the level
accelerated performance, h igher accuracy, and thus
of metal-oxide semiconductor (MOS) device fabri­
improved application of the simu lators to Digital's
cation and elemental transistor design. The con­
CMOS technologies. The process simulators SUPREM3
tinuing reduction in the geometrical scale of the
from Stanford University and PROMIS from t he
transistor structures means that cardu l control and
Technical University of Vienna (TUV), Austria, and
design of these structures is necessary to maintain
the device simu lators MINIMOS from the TIN and
the required switching properties and current drive
PISCES from Stanford are the main simu lators that
Digital modified and appl ied . In add ition, the ASD
capabi li ties.
Process and device simu lators contribute to this
Group has developed software to support the simu­
design and control by employing nu merical mathe­
lators. This software consists of programs to inter­
matical methods to simu late the ion flux that
face
occurs i n the device fabrication process and the
current flow that occurs du ring transistor opera­
simulators, and programs to automatical ly perform
simulator cal ibration and sensitivity analysis.
the
process simu lators with
the
device
tion . Process simulation requ ires physical models
The modified simulators and supporting soft­
of ion implantation, diffusion of ions, and thermal
ware are vital elements in the development of
oxidation, for example. In devic<.: simulation, the
Digital's CMOS technologies. These tools provide
basic semiconductor equations of drift and d iffu­
invaluable i nsight into transistor fabrication and
sion are solved using microscopic physical models,
operation. Using these tools in the design of lot
such as models of the mobi lity of electrons and
spl its considerably decreases fabrication time and
holes and models of electron-hole pair generation,
thus reduces cost. To predict circuit performance
also referred to as avalanche generation.
l i m its, designers use ca l ibrated simu lation results
For practical application, the simu lators must
as input to circ u it simu lators. The abil i ty to pre­
be capable of h igh-speed performance and must
dict these l i mits has made possible concurrent
Digital Technical journal
Vol. 4 .\'o. 1
5>jJriJJg 1 992
25
Semiconductor Technologies
technology and circui t design of Digital's CMOS-2,
CMOS-3, CMOS-4, and CMOS-5 technologies.
This paper describes the nature of Digital's
process and device simu lation tools. Examples of
the important physical models, nu merical mathe­
matical methods, and supporting software for the
simul ators are discussed . The paper closes with a
summary of how these tools benefit Digital's semi­
conductor process development teams.
feet l attice.) Since ion implantation is performed at
room temperature, the collection of implantation­
induced defects is retained in a stable state in the
lattice. However, at h igh temperatures the defects
become highly mobile and influence the migration
of impurity atoms.
The migration of impurities u nder anneal ing
is governed by the d iffusion phenomena, which
are mediated by point defects. Ion-implantation­
induced point defects can cause anomalous dif­
Physical Models
The key to accurate simu lations of transistor char­
acteristics is i n the physical models employed by
the programs. This section describes examples of
models developed by the ASD Group. First, a rapid
thermal annealing model for process simu lation is
presented. Next fol low discussions on ion implan­
tation through nonplanar surfaces in two d imen­
sions and on the mobility and avalanche models
used in the MINIMOS program . The section con­
cludes with information about the use of simula­
tors to predict transistor capacitance values.
fusion of the dopant. This effect is highly transient
i n nature beca use either the ion-implantation­
induced defects d isappear to the surface or to the
depth of the silicon (referred to as the silicon
" bu l k ") , or the defects may recombine while inter­
acting with the dopant. An accurate understanding
of point defect behavior is particu larly important
for smal l geometry transistors requiring ul tra­
shallow j unctions with high eloping levels.
Although research in point defect physics in sili­
con is extensive, point defect behavior is still not
well u nderstood. One contributing factor is the
lack of a reliable technique to study point defects
quantitatively, i .e . , knowledge of point defects is
Rapid Thermal A nnealing Model
obtained only through ind irect observations of
To alter the electrical properties of the semiconduc­
their effect on dopant d iffusion. There remains
tor substrate material, i . e . , the sil icon wafer, atoms
significant controversy surrounding the experi­
from groups I l l and v of the periodic table are
mental observations and the corresponding inter­
used for doping. A known amount of these atoms,
pretations. At the macroscopic level, there are
also cal led impurities or dopant, must be placed
models that solve the diffusion equation using an
in the sil icon lattice. Ion implantation is the main
average cl iffusivity for dopant. However, the con­
technique for incorporating the impurity atoms.
ventional diffusion models do not take into account
However, the implanted atoms do not move into
the details of defect -dopant in teractions resul ting
the proper sites upon
from high-dose ion implantation. Recently, several
implantation.
A high­
temperature treatment, known as thermal anneal­
models, both physically based and empirical, have
ing, is required to achieve this.
been proposed to simulate transient-enhanced dif­
There are two types of thermal annealing used in
fusion under high-close ion implantation. The phys­
semiconductor processing: conventional furnace
ical ly based models require an accurate account
of ion-implantation-induced defect concentration
annealing (CFA) and rapid thermal anneal ing (RTA).
CfA is a long-term (minutes to hours) annea ling
step carried out at moderate temperatures (below
1000 degrees Celsius). RTA is a recent technique
conducted at higher temperatu res (usua l ly above
1050 degrees Celsius) for extremely short times
(seconds).
and often use time-consuming Monte Carlo meth­
ods. Empirical models, on the other hand , are fast,
b u t must be based on detailed physics and wel l ­
cal ibrated parameters.
The ASD Group developed a new empirical
model cal led implantation-enhanced transient dif­
Ion implantation is a defect-producing process
fusion (IETD). 1 This model has been used success­
that creates lattice disorder and point defects, such
as silicon vacancies and interstitiab, in an other­
ful ly in the CMOS technology development i n
wise perfect lattice. (A silicon vacancy occurs
H udson, MA . T h e IETD model i s phenomenological
when a sil icon atom is missing from a perfect sili­
and i s based on a dual vacancy-interstitial mecha­
Digital 's semiconductor manufacturing facility i n
con lattice, whereas a sil icon interstitial occurs
nism, with model parameters determined empiri­
when an extra silicon atom is squeezed into a per-
cal ly. The model uses a relationship that li nks the
26
Vof. 4 No. 2
Spring 1992
Digital Technical journal
Numerical Device and Process Simulation Tools in Transistor Design
amou nt of ion-implantation-induced defects to ion
implantation conditi ons, such as dose and energy.
The overa ll transient diffusion process, which
depends on the annealing temperature, takes from
several seconds to a few minutes to complete. Con­
sequently, a relatively short time interval is available
to l imit the role of point defects and to control the
transient diffusion, particularly when fabricating
shallow junctions below 0.2 micron (!J.m) for devices
lc.:ss than 0.5 !J.m in size. The IETD model provides
good predic tabil ity of transient diffusion, based on
point defect behavior. This feature allows designers
to study the effect of various processing conditions
on transient diffusion and thus to optimize the pro­
cess. Figure 1 shows a comparison of secondary ion
mass spectroscopy (SIMS) data with arsenic diffu­
sion, both with and withou t the use of the IETD
moue! in the SUPREM3 process simu lator. Clearly,
using the I ETD model gives more accurate results.
ffi 1 021
f­
w
�
i=
z
w
u
u
iii
:::J
u
a:
1 019
�
a:
1 0 18
w
0..
z
0
f­
z
w
u
6
u
'
'
-
- ...
ARSENIC I M P LANT:
2
5 · 1 0 1 5 ATOMS/CM
AT 80 KEV ENE RGY
1 o 1 7 L.._
0
KEY
"
... -
'
'
'
'
_
_,_
_._
_
" �
_,_
_
_
_
_
__,
_
_
_
__c_
_
_
0.1
0.2
DEPTH INTO S I LICON (MICRONS)
0.3
SIMS DATA WITH ARSENIC D I F FUSION
SUPREM3 WITHOUT IETD MODEL
Comparison ofSJMS Data with
Resultsfrom Annealing Models
Implantation through Nonplanar Surfaces
A problem frequently encountered in semiconduc­
tor technology is the implantation of a dopant
through thin dielectric layers on semiconductors
or, in general, the implantation in multilayer struc­
tures. Calcu lating the exact depth distribution of
implanted ions in a multi layer structure requires
the solu tion of Boltzmann transport equations or
Monte Carlo simulation. This process, however, is
Digital Technical jour11al
Vol. 4 No. 2
-0.2
(j)
z
0
a:
u
SUPR EM3 WITH IETD MODEL
Figure 1
very complicated and requires a great deal of com­
pu ting time.
To circumvent these disadvantages, several ana­
lytical models have been developed that give rea­
sonably good results under certain conditions. One
of these models, the numerical range-scal ing
model, is applicable in the limit of both very thin
films and very thick films and leads to the most real­
istic profiles of all known analytical models 2·'"
Therefore, this model is implemented in the two­
dimensional (2-D) process simulator, PROMIS. Using
this numerical range-scal ing model makes it easy to
extend the 2-D model for arbitrarily shaped, non­
planar mu ltilayer structures. In addition, the tilt
angle of the implantation can be varied. Thus, it is
possible to see the dependence of the doping pro­
file on the tilt angle. The change in the analytical dis­
tribution function caused by changes in channeling
by different i mplantation angles has not yet been
considered.
Figure 2 shows an appl ication of the newly
implemented ion implantation feature in PROMIS .
The simu lation starts from a rectangular semicon­
ductor region. The first process step is a local wet
oxidation for 30 minutes at 1 100 degrees Celsius.
Boron ions are implanted into this structure with
an energy of 50 kilo electron volts (keV) and at a
dose of 5.0 X 10 1 1 /cm 2 The iso-concentration lines
of the 2-D as-implanted boron profile can be seen
in Figure 2 . Each contour line represents an order
of magnitude change in boron concentration.
Spring 1992
�
I
f0..
w
0
OXIDE
0.0
0.2
0.4
0.6
SILICON
0
�
2
LENGTH (MICRONS)
>
10
19
CM-3
1 01 4 C M - 3
3
4
Nole: Each conlour line represenls an order of magnilude change in
boron concenlralion.
Figure 2
Two-dimensional Boron Implant into
Silicon, Simulated Using PROMJS
27
Semiconductor Technologies
Although the numerical range-scaling model is
the most accurate way to analytical ly describe
implanted dopant distributions for multilayer struc­
tures, t he model has some shortcomings and
rest rictions. For example, the analytical distribu­
tion function does not depend on the tilt angle and
the sil icon orientation. This shortcoming could
probably be solved by introducing a term that
depends on these two parameters and cou ld be
fitted using measured profiles. Another shortcom­
ing, as compared to a Monte Carlo simulation, is
that the lateral distribution function does not con­
sider interfaces. This deficiency cou ld be especially
important fo r structures with very steep interfaces,
such as trenches. However, this neglect of the inter­
faces in the case of the lateral distribution function
is only significant if the mean atomic numbers of
the materials (on both sides of the inte rface) differ
considerably. I n the case of Si and sil icon d ioxide
(SiO), for example, it is not a major shortcoming.
The model described above is implemented in
PROMIS in such a way that there are, in principle, no
restrictions in the simu lation geometry This means
that the implantation module can handle arbitrary
mul tilayer structures. The implementation of this
ion implantation model is thus a significant step
towards extending PROM!S to a fu lly multilayer sim­
ul ation tool.
Lllfobility Model
Carrier mobilities in semiconductor material are
determined by a large variety of physical mecha­
nisms. Electrons and holes are scattered by thermal
lattice vibrations, ionized impurities, neu tral impu­
rities, vacancies, interstitials, dislocations, surfaces,
and the carriers themselves. The saturation of the
drift velocity caused by interactions with lattice
vibrations resu lts in a further mobility reduction.
For MOS transistors, however, the effect of the sur­
face, i.e. , the silicon-sil icon dioxide (Si-Si02 ) inter­
face, is of overriding importance. The sheet of
conducting charge (either electrons or holes),
cal led the inversion layer, is forced by the applied
electric fields to flow close to this interface and
interact with it.
The su rface scattering is a complex and poorly
understood process consisting of a combination of
roughness, interface ch arge , and su rface phonon­
scattering mechanisms. Nevertheless, there is a uni­
versal empirica l model, first demonstrated in 1979
and 1980, t hat can be cal ibrated against measured
resu lts.' " An effective mobility exists that depends
28
only on an effective field perpendicular to the sili­
con su rface and is independent of the doping level
near the surface.
The ASD Group modified the microscopic mobil­
ity model in the MlNIMOS simu lator to reflect this
u niversal mobility modeP The m icroscopic model
contains three adjustable dimensionless parame­
ters cal led MR, MT, and MX, which are general ly
close to unity. These parameters scale the magni­
tude and the two field dependencies of the micro­
scopic mobility. The model was originally
cal ibrated against CMOS-2 electrical data in 1986.
Note particu larly that recent CMOS-4 data compare
wel l to simu lated data with only minor adjustments
made to these three parameters. Figure .) is a com­
parison of MINIMOS simu lation fo r a device having a
long-channel (51-micron) length with the linear
region drain current graphed as a function of
the gate voltage (lj;). In this section and the fol low­
ing two sections, the width of a l l the simulated
and measured MOS devices is 50.5 11-m . The scaled
mobility parameters fo r this excellent fit are an MR
of 1 .02, an MT of 1 .10, and an MX of 1 .00, as com­
pared to the default 1986 values of 1 .00 for each
parameter. Figure 4 shows the fit for a CMOS-4
short-channel device, i.e., 0.62 11-m in length (an
effective length of 0.5 11-m), using the same mobil ity
parameters as fo r the long-channel device. Adjust­
ments were made to a l low for the effects of inter­
face charge and contact resistance. The agreement
20
Cii
w
a:
w
�
<{
0
a:
0
15
� 10
t­
z
w
a:
a:
::J
0
z
;:;:
5
a:
0
2
0
KEY:
0
-
GATE VOLTAGE (VOLTS)
3
4
MEASU RED CMOS-4 DEVICE DATA
M I N I MOS SIMU LATION
Figure 3
Comparison ofMINIMOS Simulation
to Long-channel Cl10S-4 Device Data
Vol. 4 No. 2
Spring 1992
Digital Tecbnicaljourual
Numerical Device and Process Simulation Tools in Transistor Design
1 .2
UJ
a:
UJ
a._
::2
<{
(j)
extremely small. As a fa l lback, it is common prac­
tice to use the substrate current as a monitor of
hot electron damage, since both currents arise
from the same high field conditions in the device.
Using an accurate substrate current model can
assist device designers in optimizing the CMOS tran­
sistor for reliability.
In late 1988, at the time of early CMOS-4 devel­
opment, the avalanche generation model in the
1 .0
::; 0.8
-'
�
f-
z
UJ
a:
a:
::>
u
z
0.6
0.4
<( 0.2
a:
0
0
KEY:
1
2
GATE VOLTAGE (VOLTS)
3
4
0 MEASURED CMOS-4 DEVICE DATA
- MIN IMOS SIMULATION
Figure 4
Comparison ofMINIMOS Simulation
to Short-channel CMOS-4 Device Data
of the simulated resu lts with the measured data is
again excellent and demonstrates the validity of the
MIN!MOS mobility model.
Avalanche Model
An important part of submicron MOS device design
is device reliabi l ity. Aggressively scaled devices
contain high electric fields. These fields are
unavoidable, even with the reduced power supply
of 3.3 volts for the CMOs-4 device. In add ition, cir­
cuit effects, such as ringing, can cause voltages on
devices to rise well above the magnitude of the
power supply. Electrons in these high electric fields
accelerate and begin to acquire energy faster than
they can dissipate it to the underlying sil icon lat­
tice. A small fraction of the electrons, namely ener­
getic or "hot" electrons, gain enough energy to
generate an electron-hole pair in a process cal led
impact ionization. The newly created electrons are
swept to the drain of the transistor, and the holes
are col lected as a current called substrate current.
Other energetic electrons surmount the barrier
at the Si-Si0 interface and inject themselves into
2
the oxide, appearing as gate current, which flows
out of the gate terminal. Carrier injection into the
oxide can damage the oxide and cause a shift in the
device threshold voltage or degrade the ability of
the device to conduct current. This injection is the
physical cause of device degradation.
Unfortunately, gate current is difficult to mea­
sure directly, because the currents involved are
Digital Techt�ical ]our11al
Vol. 4 No. 2
Spring 1992
MINIMOS simulator was reexamined. The design
team found that the peak substrate current for a
given drain voltage ( v,i ) occurred at gate voltages
lower than predicted by simulation. They resolved
this problem by modifying the microscopic model
for impact ionization in the MINIMOS simu lator to
include a depth-dependent term, similar to the one
used by Slotboom et aJ.8.9 The modification sharply
reduced impact ionization near the Si-Si02 inter­
face. Impact ionization model parameters were
derived from CMOS-3 and CMOS-4 measured data.
Figure 5 shows measured and simu lated CMOS-4
substrate current data for drawn gate lengths of
0.62 and 2.00 JLm at a �� equal to 3.3 volts. Figure 6
shows the substrate current as a fu nction of the 'i
for the 0.62-micron gate length device for three
drain voltages around the design center of 3.3 volts.
The agreement shown using one set of model
parameters is quite remarkable, given the limited
4
r;;
UJ
a:
UJ
�
�
a:
u
3
�
f­
z
UJ
a:
a:
a
�
-'
L ; 0.62 M I C RONS
0
2
0
0
1
Vd ; 3.3 VOLTS
::>
Cll
0
GATE VOLTAGE (VOLTS)
KEY:
3
4
0 MEASURED CMOS-4 DEVICE DATA
- M I N I MOS SIMULATION
L GATE LENGTH
Vd DRAIN VOLTAGE
Figure 5
Comparison of MINJMOS Simulation
to CMOS-4 Substrate Current Data
for Two Gate Lengths
29
Semiconductor Technologies
25
w
a:
w
(j)
a._
:::E
<(
0
a:
()
�
1z
w
a:
a:
80
o 0 o
20
0
w
()
0
15
z
::;)
;,:
-'
::;)
(!l
KEY :
0<( Cfl
a.. o
<( <(
0
0
10
()
;::
0
0
5
GATE VOLTAGE (VOLTS)
3
0 MEASURED CMOS-4 DEVICE DATA
- M I N IMOS SIMULATION
Vd DRAIN VOLTAGE
Figure 6
() 0:
� ;1:
40
2
20
cr:
::;)
o
Cfl
·
O
1:::E
w
!:::..
uJ
0
Comparison of MINIMOS Simulation
to CMOS-4 Substrate Current Data
for a Short-channel Device and
Three Drain Voltages
60
1<(
(.9
1
0
KEY :
2
3
GATE VOLTAGE (VOLTS)
4
0 MEAS U R E D CMOS-4 DEVICE DATA
- M I N IMOS S I M ULATION
Vd DRAIN VOLTAGE
Figure 7
physics and empirical nature of the model.
Although the modified model has proven satisfac­
tory for CMOS-4 devices, we are examining other,
more physically based, models of substrate and
gate current for our future generations of CMOS
devices.
Comparison ofMINIMOS Simulation
Data to CMOS- 4 Gate-to-sou1·ce
Capacitance Data for Four Drain
Voltages
80
Vd = O VOLTS
Capacitance Simulation
The speed of MOS circuits partially depends on the
amount of device capacitance that must be charged
du ring circuit state transitions. Device simulators
can tell device designers how large the device capac­
itances wil l be and what effect changes in the man­
ufacturing process will have on capacitance size.
The gate capacitance of the MOS transistor is one of
the two key types of device capacitances; the other
consists of the so-called "diode" capacitances of the
source and drain j unctions. The gate capacitance is
split into three parts: gate-to-source (Cg5), gate-to­
drain (Cgd), and gate-to-bul k (�b). Figures 7 and 8
show simulated and measured C s and C.�rt values
g
versus gate and drain voltages for a device with a
width of 50.50 fLm and a length of 0.62 fLm. c3b is not
shown because it is a negligibly small quantity for
voltages above the transistor threshold voltage.
Achieving the good agreement between simu la­
tion and measurement shown in Figures 7 and 8 is
a difficul t u ndertaking. Accurate capacitance mea-
30
2
0
3
GATE VOLT AGE (VOLTS)
KEY:
4
0 MEAS U R E D CMOS-4 DEVICE DATA
- M I N IMOS SIMULATION
Vd DRAIN VOLTAGE
Figure 8
Comparison ofMINIMOS Simulation
Data to CMOS-4 Gate-to-drain
Capacitance Data for Four Drain
Voltages
surements on such short -channel devices require
carefu l experimental techniques to reduce the
effects of parasitic capacitance and to ensu re that
Vol. 4 No. 2
Spring 1992
Digital Technical]our11 a l
Numerical Device and Process Simulation Tools in Transistor Design
the DC power supply can sink sufficient current.
For example, with the aid of simulation, it was
shown that small parasitic resistances, on the order
of 10 ohms, can significantly shift the capacitance
cu rves of the device. Such resistances must be min­
imized to yield accurate resu lts and to verify that
simulation and measurement can be correlated.
On the simulation side, uncertainties in the con­
centration of dopant in the channel region and
the source-drain regions have a significant impact
on the capacitance characteristics. The ASD Group
used this fact to advantage in a "reverse engineer­
ing" of the doping profiles. They compared the mea­
sured data with the simu lated capacitance using
doping profiles generated by the process simula­
tors. Then, they rev ised the profiles until the simu­
lated capacitances matched the measured ones.
The ASD team applied this reverse engineering
method to both channel doping into depth and lat­
eral source-drain doping. In the former case, the
deep depletion capacitance-to-voltage measure­
ments of doping were verified using simulation. In
the latter, the doping was extracted by exam ining
the relationship between the cgd and the � when
there is no �· Using this method, it was possible to
extract doping profiles that were u nattainable by
other analytic means, such as SIMS.
The CMOS-4 devices exhibit a slight polysilicon
depletion effect, which decreases the measured
capacitance of the MOS transistor gate. Work per­
formed in parallel by the ASD Group and at the TUV
has resulted in an implementation of a polysilicon
depletion model in the MINIMOS simulator. 10 The
model is unique in that it analytically solves the
Poisson equation i n the polysilicon gate and thus
can be imp lemented in MINIMOS as a simple modifi­
cation of the gate boundary condition. Figure 9
shows a comparison of the model with simulation.
The measured total gate capacitance of a large-area
MOS diode (50.5 by 50.5 J.Lm) as a fu nction of the l;g is
presented, along with MINIMOS simulation results.
For reference, Figure 9 also shows the simulated
gate capacitance without the polysilicon depletion
effect, computed using the PISCES device simu­
lation program. The PISCES results were used as
an independent check on the MINIMOS simulator
results. The curves indicate that the modeled
resu lts are a good approximation of the measured
data . At CMOS-4 dimensions, the polysilicon deple­
tion effect is rather smaiJ, i.e., on the order of 5 per­
cent. The effect becomes more pronounced at
thinner gate oxides.
Digital Technicaljottrnal
Vol. 4 No. 2
Spring 1992
rJi
0
<1:
a:
<1:
lL
0
6
(.)
n::
w
(.)
z
4
<1:
f0
<1:
a.
<1:
(.)
2
w
f<1:
�
KEY:
Q L--�----�----L--�----�----�--��-­
-2
-4
4
-6
0
2
6
GATE VOLTAGE (VOLTS)
0 MEASURED CMOS-4 DEVICE DATA
MINIM OS SIM ULATION
PISCES S I M ULATION ( I DEAL)
Figure 9
Comparison ofMINIMOS and PISCES
Simulation Data with Capacitance­
voltage Data in the Presence of the
Polysilicon Depletion Effect
Numerical Mathematical Methods
The efficient use of CPU time and memory in the
device and process simulators is directly depen­
dent on the efficient implementation of state-of­
the-art numerical mathematical algorithms and
physical models. Because of the l imitations on rep­
resenting numbers in computers, these algorithms
and models must i ncorporate many detailed
enhancements in order to maintain stability of the
simulation programs in the numerical and physical
sense. This is especially true in the case of simula­
tion i n three space d imensions. This section first
describes the improvements to the MINIMOS device
simu lator achieved during the last few years. An
application of the improved MINIMOS to a parasitic
device in shal low trench isolation is then pre­
sented, fol lowed by a discussion of automatic space
grid design.
Performance Improvements Achieved
for the MINIMOS Device Simulator
A three-dimensional (3-D) simulation program
requires the solution of very large matrices and is
thus time-consuming. Therefore, an accurate use of
resources is obligatory. This goal was achieved for
the MINIMOS device simulator by means of software
techniques and hardware options. u , J2
31
Semiconductor Tedmologies
Various algorithms and n umerical methods were
thoroughly tested with practical examples, because
The VA.'<
FORTRAN
high-perfo rmance option
(HPO) compiler was also used to improve the per­
not all algorithms presented in the l itera t u re are
formance of the MINIMOS 3-D simu l a tion program.
appl icable. Most of the proposed algorithms, and
The ASD Group investigated both parallel ization
most of the i mprovements related to ove ra l l perfor­
and vecto rization on a m u l t ipr o c ess or VAX .system
nunce gain, are found among the l inear solvers.
with vector hardware. The p e rfo rmanc e improve­
These solvers are appl ied to the sol u t ion of the
ment resu l t i ng from parallel ization (using t wo
large sparse matrix systems that arise from the
CPUs) and vcctorization (using t wo vector u n its)
discretization of the partial differential equations.
was as much as a factor of ftve. During the last
The best-known l inear solver algorithm is cal led
few years, software enhancements, includ ing the
Gaussian elimination; this algorithm is the most sta­
removal of hot spots, th e improve m e n t of algo­
ble method of solving a system of l inear equations.
rithms, and the creation of common code , have
U nfortunately, Gaussian el imination is generally
resulted i n a thirtyfold p e rformance improvement
not applicable to the 3-D simu lations because enor­
of the progra m. Thus, the combined software and
mous amounts of memory
and
CPU
time are
hardware
enhancements
3-D
improved
required. Instead, iterat ive solvers have been tested
m a nce of the MINIMOS
for stabi l ity, using several techniques, which are
approximately a factor of 150.
the
perfo r­
simulation program by
required because of the variation over a range of
Many of these improvements also enhance per­
many orders of magnitude in the coefficients and i n
formance on scalar processors. The MlNIMOS pro­
t h e dependent variables.
gram is run on a variety of VAX and MIPS processors
After an accurate comparison of the convergence
at several sites in the United States and in Europe.
behavior of the different l inear i terative solvers, a
Engineers can submit MINIMOS (or other) simula­
final set of algorithms was ident ified, one which is
tion programs from their VAX.cluster systems run­
a d equa te for the present. These algorithms make
the 3-D simulation program more efficient by taking
ning under the Vi'vi S operating system to various fast
processors on the local network. This im p ro ves t he
into account the different properties of the Poisson
overall turnaround t i me of comp u tationally inten­
equations (solut ions for electric potential) and the
sive j o bs ancl provides a pain less way fo r engineers
continuity equations (solu tions for current flows).
to better use available computational resources.
Strong emphasis has been placed on reducing the
amount of dupl icate code in the simu lation p ro­
Application to a Parasitic Device
gram. MINIMOS i s a hierarchical ly ordered program
The improved performance of the
which first solves an approxim ate 2-D problem in
u lation program made i t possible to analyze the
of
shal low
trench
3-D
M I NI MOS sim­
isolated
d ev i c e s.
three steps and then proceeds to solve the actua l
behavior
3 - D problem with t h e 3-D algori thms. This method
Electrical data on these devices in early develop­
computer
ment work showed a " bump" in the subthreshold
resources. However, sometimes the incompatibility
drain cu rrent versus gate voltage characteristic
guarantees an
efficient
usc
of the
of the 2-D and 3-D codes means that separately func­
under cert ain condit ions. The ASD Group simu lated
ment work has been done to create com mon code
this bump using the 3-D MINI�v iOS program and
demonstrated that the origin of the bump was in
in a l l important areas, including physical models
a parasitic c u rrent along the t rench sidewa l l . The
t ioning code must be used. Additional develop­
(e.g. , mobil ity models) and the numerical algo­
rithms. Future upgrades and modifteat ions for 2-D
3-D
2-D version of the :VIINIJ\1!05 program was unable
to explain the phenomenon. Only when the 3-D
codes can be done simultaneously. This
version was appl ied to the problem (taking into
improvement also contributes significan t l y to the
acco u n t a sl ight overlap of the gate polysil icon into
and
performance improvement.
the trench) was it apparent that a parasitic device at
An investigation using the VAX Performance and
the edge of the trench was turning on at a lower
Coverage Analyzer (PCA) was carried out to identify
gate voltage t han was the main device, thus causing
so-cal l ed "hot spots," which are code sequences
the bump.
that consume h igh amounts of CPU time when exe­
Figure 10 shows a comparison of the M l IMOS
cuted. E l i m inating or improving the performance
simu lations with the elect rical data. The bump is
of these code segments signifi cant ly reduces simu­
lation time.
32
resolved by the 3-D MI Nn10S simu lation but is not
present in the case of the 2-D s i m u lation. The para-
Jiul. 4 Nu. 2
.'ojJrittg 1992
Digital Technical jourual
Numerical Device and Process Simulation Tools in Transistor Design
sitic device turns on before the main device, but its
effective width is much less than that of the main
device. Thus at higher gate voltages, the main
device, which can be simulated with the 2-D ver­
sion of the MINIMOS program, is dominant. The sim­
ulations have been adjusted for threshold and back
bias. Further simulations ind icated that increased
back bias on the device greatly enhances the bump;
the parasitic device at the corner of the trench
is partially shielded from the effects of the back
bias. The improved 3-D MINIMOS program is cur­
rently used in the design of the CM05-5 technology,
fo r continued electrical analysis of shal low trench
isolation.
1 0-
(i)
w
a:
w
[l_
z
w
a:
a:
::>
4
1 0- 6
:2
:-:5.
f-
1 0-
�
.:: �
8
�" ��
��
�(S ��
u 1 0-1 0
z
<(
a:
0
sions, equidistribu tion of the discretization error,
the degree of coupling of the differential equations
to be solved, and final ly, the control of the ratio of
adjacent mesh distances. Attention to these details
results in the achievement of theoretical conver­
gence rates and reasonable CPU times.
Figures 1 1 and 12 show respectively an implanted
boron profile and the error indicator for the masked
implantation of boron. The mask ends at the origin,
and the error criterion is based on the d ifference
between the second mixed derivatives of the solu­
tion function. Figure 12 clearly shows that the error
is concentrated at the curvature of the profile
around the mask edge.
1 0-
��
\::5 �
12
0
KEY:
0.5
1.0
GATE VOLTAGE (VOLTS)
1 .5
� \.\
'll �
8 !!=_
"
0 MEASU RED CMOS-4 DEVICE DATA
- 3-D M I N IMOS SIMULATION
- - 2-D MINIMOS SIMULATION
Figure 10
Comparison ofMINIMOS Simulation
with Datafrom a Shallow Trench
Isolated Device with Parasitic
Sidewall Current
Automatic Grid Design
In process and device simulation, the numerical
mathematics requires the construction of a dis­
cretization scheme, an underlying space grid and,
for time-dependent problems, a time grid, which
should be controlled automatical ly. This basic grid­
ding determines not only the accuracy of the solu­
tion but also the r u n time of the program, and
in some cases even the success or failure of the
simulation.
The design of an automatic space grid requires
both mathematically based and physically based
strategies. 1'·14 Mathematical criteria for self-adaptive
gridcling involve the remainders in series expan-
Digital Teclmical]ournal
Vol. 4 No. 2
Sp-ri11g 1992
Figure 11
Masked Jon Implantation ofBoron
Showing the Space Grid
Almost all physical criteria are heuristic, i . e . , they
use quantities l ike the current densities or the net
generatio n/recombination of carriers to design the
grid properly in regions of physical interest. For
example, numerical experiments indicate that in
device simulation, mesh refinement is essential in
regions where the net generation/recombination of
carriers becomes higher than 1 . 0 X 10 22/cm·'� ·s.
Conservation laws play an important role in the
definition of these strategies. For instance, coarse
grids at curved pn-junctions introduce a significant
integration error, thus leading to insufficient charge
conservation, which is essential for accurate capac­
itance calculations. ' ' For the automatic calculation
33
Scmicond uctor Technologies
Figure 12
Error lndicatorfnr the Masked Ion
Implantation ofBoron
of time steps, heuristic criteria are used al so. These
criteria take into consideration the change in
boundary conditions and try to estimate the time
discretization error by analyzing the solution for
two successive time steps. More mathematica l ly
based criteria, such as utilizing the change of the
potential energy, are under investigation.
Grid design strongly affects the run t i mes of a
simulation program. Usual ly the numerical effort of
a two-dimensional simulation program depends
quadratica l ly on the number of grid points. Thus a
50 percent reduction in the number of grid points
reduces the run time by a factor of four. Recent
work at the Campus-based Engineering Center
(CEC) in Vienna on gridding strategies in the PROMIS
simulator has resu lted in improving performance
by a factor of 2.5 for the same accuracy. Automatic
grid design is absolutely necessary to optimize the
rat io between the number of grid points and the
desired accuracy fo r each specific boundary condi­
tion. For time-dependent simu lations, in particular,
moving grids is the only way to guarantee the same
accuracy fo r all time steps. Work is continuing in
this area at both the ASD Group in Hudson and
users must combine different tools to per form
required simu lations. The ASD Group has devel­
oped interface programs to assist in this effort. On a
larger scale, research is going on in industry ami
in academia to develop the so-cal led technology
computer-aided design (technology CAD or TCAD)
frameworks for integrating the various simulation
tools. t6• 17. 1 H The ASD Group has been involved in this
eft<)rt by actively participating in the United States­
based TCAD group of the CAD Framework Initiative
(CFI) and by interacting with the researchers a t the
TIN, who are developing the Viennese Integrated
System fo r Technology CAD Applications (VISTA)
framework discussed later in this sect ion. These
efforts hold much promise for the fu ture, b u t the
ASD Group has required intermediate-term solu­
tions to speed up and automate the various proce­
dures used by its transistor designers. This sect ion
presents two such solutions fol lowed by a brief
description of the VISTA system currently under
development.
TCAD Command Language
The simulation needs and demands of Digital's
CMOS technology development have increased
beyond isolated simulation ru ns to i nclude more
complex tasks such as sensitivity analysis, optimiza­
tion, and macromodeling. To support these higher­
level needs, the ASD Group developed the TCAD
Command Language (TCL). TCL is a specialized pro­
gramming language for TCAD task management and
execu tion, an instance of a command extension
language tail ored to TCAD users' needs. The follow­
ing l ist highlights features of TCL that are necessary
in the ASD TCAD environment:
•
General programming constructs and control
mechan isms
•
Specialized subroutines suitable for optimiza­
tion and sensitivity analysis, with arguments
divided into input parameters, and input and
output variables
•
Tool control and manipulation that support envi­
ronment customization, journaling, distributed
execution, and parameterized tool invocation
the CEC.
Technology Computer-aided Design
Historically, process and device simu lation tools
have been developed individually as standalone
tools. No single tool covers the whole range of
processing steps and device simu lation needs that
can arise in CMOS transistor design. Consequently,
34
•
Special analysis and optimization commands
bund led i n a callable run- time library
TCL has been used in a variety of design and anal­
ysis tasks that include exploration of the design
space, parameter sensitivity, design through opti­
mization, model parameter extraction and charac-
Vol. 4 No
.
.!
Spring 1992
Digital Technical journal
Numerical Device ana Process Simulation Tools in Transistor Design
•
terization, and statistical analysis. For example,
Figure 13 shows TCL code that illustrates the appli­
cation of the TCL OPTIMIZE command for simulator
•
calibration. Using this code, the linear region mobil­
Grid handling (e. g . , creation of new grid, inter­
polation, and merging of grids)
Tools (e.g . , plotting and rotation)
ity parameters used for the MINIMOS program and
Additionally, powerful macros can be defined as
discussed in the Mobility Model section are auto­
procedures in command files. The expansion of 1-D
matically adjusted so that the MINIMOS calculated
doping profiles to 2-D by elliptic rotations is an
current values fit the experimental data. This au to­
example of such a macro. To illustrate the expan­
mates a routine and time-consuming task formerly
sion, this section describes the combination of the
done manual ly.
2-D source and drain profiles (arsenic and phos­
phorus) of a MOS transistor simu lated by PROMIS,
Vienna Interactive Data Editor
and the 1 -D channel profile (boron) simulated
The Vienna Interactive Data Editor (VIDE) provides
by SUPREM3. First, the 1-D boron profile from
a set of tools for data manipulation and data trans­
the SUPREM3 simulator is expanded to 2-D and
formation. The program, i .e . , toolbox, consists
stretched to the appropriate geometry. Then, the
of three parts: input, manip u l ation, and output.
profile is interpolated on the grid defined by
Separating the parts in this way makes i t easier to
PROMIS. The boron profile is the acceptor profile,
whereas the sum of the arsenic and the phospho­
rus profiles gives the donor concentration. The
result, consisting of the net 2-D doping, is shown in
Figure 14.
The applications of VIDE are not restricted to pro­
cess and device simulation. The toolbox allows the
manipulation of data of arbitrary origin. The pro­
implement additional data formats from a variety of
programs. For this toolbox, there are five classes of
operations that apply to either one, two, or three
dimensions:
•
•
Geometry handling (e . g . , stretching, scal ing,
and shifti ng)
Quantity handling (e. g . , stretching, scal ing,
gram has been successfu l ly used to analyze the
shifting, and least squares fit)
accuracy of simulation results for the calculation of
• Quantity arithmetic (e.g . , m u ltiplication, square
the absolute and relative error, as well as for least
root, and logarithm)
squares fits to measurement data. Furthermore,
BEGIN
1
Fi rst
D E F I NE
the
block
BLOCK
def i n i t ion.
TUNE
PA R A M E T E R S
MR,
I NPUTS
L,
VGS
OUTPUTS
I DS
MX,
MT
BEGIN
READ
COMMAN D S
TEMPLATE
$MR-VALUE$
/BY=MR
/ IN=TEMPLATE
R E P LA C E
$MX-VALUE$
/ BY=MX
/ IN=TEMPLATE
R E P LA C E
$MT-VALUE$
/ BY=MT
/ IN=TEMPLATE
R E P LA C E
$VGS-VA LUE$
/ BY=VGS
/ IN=TEMPLATE
R E P LA C E
$L-VALUE$
/BY=L
/ I N=TEMPLATE
FWRITE
M I N I RUN . MM I
MINIMOS
READ
!
!
T1 . MMI
REPLACE
TEMPLATE
M I N I RUN . MM I
M I N I RUN . MMO
IDS
/ OUT=MI N I RUN . MMO
/ DOPI NG=T1 . 2 DOP
/ I NTO= I D S
END
Ini t i a l
p a r a me t e r s
Expe r i menta l
OPT I M I Z E
TUNE
data
va l u e s
i n
in
I N I T . PAR,
F i n a l
va l u e s
T1 . DAT .
/ DATA=T 1 . OAT
I I N I T=T1 . PAR
in
T1
. S AV
/ SAVE=T1 . SAV
END
Figure 13
TCAD Command Language Code Illustrating the Application of the OP11MIZE Command
Digital Technical journal
Vol. 4 No. 2
Spring 1992
35
Semiconductor Technologies
process flow simulations is supported. The TCAD
shell automatically performs impl icit paralleliza­
tion and job control; the shel l is also capable
of quick standard visualization. Interfaces to
advanced visualization tools, such as DEC AVS , pro­
vide state-of-the-art graphics. An on-line documen­
tation system with automatic documentation
generation from the source code always guarantees
up-to-elate information for users and programmers.
Figure 14
Net Doping Profile Combining the
2-D Source and Drain Profilesfrom
PROMIS with the 1-D Channel
Profilefrom the SUPREM3
Simulator
VIDE serves as a transformation program between
different data formats of simu lators.
VISTA Framework
VISTA, developed by Professor Selberherr's group at
the TUV, is the first working TCAD framework based
on a standardized data format. 1'U0 A brief descrip­
tion of the VISTA system fol lows.
Tbe Back End A multilanguage programming
i nterface, e.g. , FORTRAN, c, or XI.ISP, permits access
to the simulation database. Such an interface has a
well-defined data format, which is a basic require­
ment for tool-to-tool communication. The Program
Interchange Format (PIF) data format in the VlSTA
environment provides a standardized way to trans­
port simulation data, while at the same time
remains open to future demands and extensions
through its h ighly flexible structure.
Tbe Front End The point-and-click interface of
the TCAD shell, together with a visual program­
ming interface, al lows easy interaction for inex­
perienced users. The user interaction surface can
be customized to accommodate user ancl program
requirements. Interactive development of complex
36
The Tool Aspect Within the VISTA package, a
generic toolbox enables common data manipu la­
tions such as interpolation, gradient calcu lation,
and arithmetical operations. The toolbox covers
most of the standard data manipulations occurring
in process and device simu lation and thus al lows
program developers to focus on the main issues of
their tasks. A tool abstraction concept helps with
the automatic generation of front and back inter­
faces for new simulation tools. Strict rules ensure a
consistent extension of the VISTA system to new
and complex simulation tools.
General Aspects To face future challenges, a TCAD
environment requires a consistent architecture
together with high-level concepts. Modern soft­
ware development techniques, such as automatic
code generation and documentation, layered struc­
tures, ancl a high abstraction level of the underlying
concepts, are the basis on which VlSTA is built.
Present Status of VISTA witbin D igital's Develop­
ment Work The most recent version of V ISTA has
been installed at the CEC in Vienna. Because of the
close proximity of the CEC and the TIN, feedback
on the concepts and the implementation details
will directly influence the future progress of VlSTA.
This test version permits the application of the
common data interface, i.e., the PIF application
layer, and the simple coupling of the VlSTA-PROMIS
and VISTA-MINIMOS simu lators, both of which are
based on PROMIS and MINIMOS but have been fur­
ther developed for use within a TCAD framework.
It is expected that during 199 3 , VlSTA wil l replace
parts of the intermediate solutions for Digital's
needs. Integration of the existing tools such as VIDE
into VISTA is in progress.
Conclusions
The use of process and device simulation tools pro­
vides Digital's semiconductor process development
teams with the following benefits:
Vol. 4 Nu. .!
.SjJring 1992
Digital Technical journal
Numerical Device and Process Simulation Tools in Transistor Design
•
Decreases the n u mber of experiments required
to optimize the fabrication process. Experimen­
tal lots may take many months to process,
whereas simulators can give results in minu tes
or hours. Simulation can never totally replace
experimen tation, however, because simulation
is only a model of what we know about process
and device physics, not reality itself. Smaller
dimensions and new manufacturing processes
requ ire constant revision of our physical models
and simulation tool capabilities.
•
•
Allows the design teams to better estimate the
spread of device performance, i . e . , the so-called
"worst case" conditions, before the process is
wel l established. Thus, circuit designers can
begin design earlier and with more confidence.
Gives i nsight into the i nternal behavior of pro­
cesses and devices to back up engineering ju dg­
ment. For example, a layout design rule had been
violated in an obscure way in the design of an
1/0 driver circuit. Redesign wou ld have cost both
time and space on the chip. Device simulation
verified the device engineer's opinion that the
violation would not cause any problem i n actual
practice.
6.
S. Sun and ). Plummer, " Electron Mobility in
Inversion and Accumulation Layers on Ther­
mally Oxidized Silicon Surfaces," IEEE Trans­
actions, ED-27 (1980): 1497 - 1508.
7.
S. Sel berherr, W Hansch, M. Seavey, and
). Slotboom, "The Evolution of the M INIMOS
Mobil ity Model," Electronics and Communi­
cation (1990): 44, 161 .
8.
]. Faricel li, "Improved Substrate Current Pre­
diction Using Dept h Dependent Impact
Ionization Parameters;' MINIMOS User's Sym­
posium (Ju ne 1989).
9.
). Slotboom, G. Streutker, G. Davids, and
P. Hartog, "Surface Impact Ionization in Sili­
con Devices," Proceedings of the IEEE Inter­
national Electron Devices Meeting (IEDM)
(1987): 494-497.
10.
the Physical Modeling of the Gate-depletion
Effect," IEEE Transactions (Forthcoming,
1992).
11.
H . Soleimani, " Modeling of High-dose III­
induced Dopant Transient Diffusion and
Dopant Transient Activation," journal of the
375- 381 .
12.
Electrochemical Society (Forthcoming, 1992).
2.
3.
H. Ryssel , ). Lorenz, and K. Hoffmann, " Mod­
els for Implantation i n to Multilayer Targets,"
Applied Pbysics, vol. 41 (1986): 201 - 207.
5.
H. Ishiwara, S. Furukawa, ). Yamada, and
M. Kawamura, Jon Implantation in Semi­
conductors, edited by S. Namba (New York:
Plenum, 1975): 423.
A. Sabnis and ). Clemens, " Characterization
of the Electron Mobility in the Inverted < 100>
Si-surface," Proceedings of the IEEE Inter­
national Electron Devices Meeting (IEDM)
(1979): 18-21 .
Digital Tee/mica/ ]our11al
lkll. 4 No. 2
Spring 1992
K. Traar, M. Stiftinger, 0. Heinreichsberger,
and S. Sel berherr, "3D Simu lation of Semi­
conductor Devices on Supercomputers;' Pro­
ceedings of the ACM International Conference
on Supercomputing (1991 ).
13.
G. Nanz, W Kausel, and S. Selberherr, "Self­
adaptive Space and Time Grids in Device
Simulation," International journal for
Numerical Methods in Engineering, vol . 3 1 ,
n o . 7 (1991): 1357- 1374.
14.
G. Nanz, "Self-adaptive Space Grids in Process
Simulation," Microelectronics journal (Forth­
coming, 1992).
15.
G. Nanz and C. Schiebl, " Remarks on Numeri­
cal Methods in Capacitance Calculations,"
H . Ryssel, W Kriiger, and ]. Lorenz, " Com­
parison of Monte Carlo Simu lations and
Analytical Models for the Calcu lation of
Implantation Profiles i n Multilayer Targets,"
Nuclear Instruments and Methods in Pbysics
Research B, vol. 19/20 (1987): 40-44.
4.
M. Thurner, P Lindorfer, and S . Selberherr,
"Numerical Treatment of Nonrectangular
Field Oxide for 3D MOSFET Simulation," Simu­
lation of Semiconductor Devices and Pro­
cesses (SISDEP) Proceedings (Bologna, 1988):
References
1.
P Habas and ). Faricelli, " I nvestigation of
IEEE Transactions on Computer-aided Design
(Forthcoming, 1992).
16.
TCAD Architecture Committee, "Technology
CAD Framework Architectu re," edited by
D. Boning (Research Triangle Park, NC: Semi­
conductor Research Corporation, May 1990).
37
Semiconductor Technologies
17
18.
19.
38
CFI TCAD Framework Group, Semiconductor
CAD Shell," Proceedings of the International
Wafer Representation Working Group , "The
SWR Programming In terface, Version 0.9c,"
edited by A. Wong and G. Chin (Austin, TX:
CAD Framework Initiative, Inc . , February
1992).
Conference on Simulation of Semiconductor
Devices and Processes, vol. 4, edited by
S. Selberherr et a l . , "The Viennese TCAD Sys­
tem," Proceedings of the International Work­
W Fichtner and D. Aemmer (Konstanz:
Hartung-Gorre, 1991) .
20.
F. Fasching, H. Read, C. Fischer, S. Selberherr,
H. Stippel, and W Tuppa, "A PIF Implementa­
tion for TCAD Pu rposes," Proceedings of the
shop on VLSI Process and Device Modeling
(1991 ).
International Confe1·ence on Simulation of
Semiconductor Devices and Processes, vol . 4,
H . Pimingstorfer, S. Halama, S. Sel berherr,
K . Wimmer, and P Verhas, "A Technol ogy
edited by W Fichtner and D. Aemmer (Kon­
stanz: Hartung-Gorre, 1991).
Vol. 4 No. 2
Spring 1992
Digital Technical Journal
Andre I. Nasr
Gregory]. Grula
Antonio C. Berti
Richard D. jones
CMOS-4 Technology for Fast Logic
and Dense On-chip Memory
Digital's fourth-generation CMOS technology has produced the industry's highest
performance microprocessors. The NVAX and Alpha
21064 chips are based on
0. 75-p.m, 3.3-V CMOS technology capable ofproducing operatingfrequencies ofup to
100 MHZ and 200 MHz respectively The higb-pe1jormance CMOS transistors consist
of a 105-A gate oxide, symmetric n + and p + dopedpolysiliconfor surface channel
conduction, low threshold voltage, and good turn-off characteristics. The transistor
has an on-wafer electrical gate length of 0.5 J.Lm, a shallow medium doped drain
junction for hot electron immunity, a CoSi2 salicided gate, and source and drain
regionsfor low interconnect sheet resistance. A TiN/CoSi2 local interconnect scheme
was used to strap the drain and gate regions to form a six-transistor memory cell
with an area equivalent to 100 J.Lm2.
H igh-performance
complementary metal-oxide
(CMOS) microprocessor design
began in D igital's Hudson, Massachusetts, site in the
m id-I980s. D igital's strategy calls for the scal ing of
feature sizes with each new generation of CMOS
technology, coupled with larger d ie size to achieve
higher circuit density and h igher system perfor­
mance. CMOS CPU operating frequency has doubled
every two years since I98 6, from I O megahertz
(MHz) for the CVA.X chip fabricated with the CMOS-I
technology, to 100 MHz for complex i nstruction
set computer (CISC)-based architecture as demon­
strated by the NVAX chip . 1 I t reached 200 MHz in
1991 for the Alpha 2106 4 reduced instruction set
computer (RlSC) architecture.2 On-chip caches
increased from I kilobyte (KB) for the CVA.X chip
in I98 6 to a total of I 6KB for the Alpha 2 10 6 4 chip
semiconductor
fabricated with CMOS-4 technology in I991 .
The transistor count implemented on a single
microprocessor chip beginning with CMOS-I tech­
nology swel led from 100,000 in I98 6 to 1 .7 m i l l ion
devices in 1991 . D uring the same period of time,
the polysilicon l ine that forms the transistor
channel length was scaled from 2 .0 microns (p.m )
to 0.75 p.m, and the gate oxide thickness was
decreased from 300 angstroms (A) to 105 A for the
CMOS-4 process. Concurrently, the power supply
was scaled from 5 volts (V) to 3.3 V to reduce power
High-performance m icroprocessors have posed
a number of technological challenges in the design
of transistors and on-chip caches. A nu mber of
process and device innovations have been intro­
duced . For example, symmetric submicron n + and
p+ doped polysilicon transistors must provide
low threshold voltage and good turn-off character­
istics. Also, graded drain junctions must balance the
driving current of the high-performance transistor
with adequate hot carrier resistance. Self-al igned
cobalt silicide (CoSi 2) polysil icon gate, and source
and drain regions must provide low interconnect
sheet resistance for improved circuit density and
performance.
On-chip, high-density, static random-access mem­
ory (SRAM) is required for high-performance CMOS
microprocessors.
Previous
technology
scal ing
restricted the size of the on-chip cache due to the
rel a t ively l arge cell area. A variety of techniques
were proposed to improve the RAM density, includ­
ing the four-transistor (4T) cel l with a second-layer
polysil icon resistor and the six-transistor ( 6T) cel l
with buried contact (BC) or local interconnect
(LI) . The most attractive technique was the local
interconnect scheme that was used to fabricate the
100-p.m2 cell , the cell being used successfully i n the
Alpha 2106 4 and NVA.X m icroprocessors.
This paper describes the front-end process flow
d issipation and to provide extra protection against
for D igital's fourth-generation 0.75 -p.m CMOS tech­
gate oxide wear-out mechanisms.
nology. It emphasizes the methods used to form
Digital Technical jo-urnal
Vol. 4 No. 2
Spring 1992
39
Semiconductor Technologies
the low-resistance CoSi2 and the titanium-nitride­
based local interconnect used for the dense on-chip
cache. The paper also discusses the CMOS transistor
design and device characteristics. It concludes with
a description of the SRAM cell .
Process Description
The front end of the CMOS-4 process is d ivided
into six process modules: well formation, device
isolation, gate formation, medium doped drain
ju nction formation, CoSi2 formation, and the local
interconnect process. CMOs-4 technology was built
upon the previous CMOS generations. Additional
steps were incorporated into the process flow to
meet new circuit design and layout requirements.
Process steps were modified to accommodate scal­
ing of device dimensions. A schematic process flow
that depicts the various process modules is shown
in Figure 1. The CMOS-4 layout design rules are
given in brief in Table 1 .
•
\
(e) Spacer Oxide Deposition,
Densification, and Etch
(a) Well Drive
\
(b) Initial Oxide/nitride
Active A rea Mask
(/) Source/drain Drive
•
\
'
(c) First Gate Oxide: Field Oxidation
•
�"'>
\ ""�
(g) Cobalt Silicide Formation
D
D
\
'
(d) Second Gate Oxide: Polysilicon
Deposition, Polysilicon Mask
and Etch, Wet Reoxidation
(h) Titanium Nitride Deposition,
Pattern, and Etch
(i) Local Interconnect
Figure
40
1
CMOS-4 Front-end Process Flow
Vol. 4 No. 2
Spring 1992
D igital Teclmicaljournal
CMOS-4 Technologyfor Fast Logic and Dense On-chip Memory
Table 1
Layout Design Rules for CMOS-4
Process
Dimension
Design Rule
M i n i m um active a rea width
1 . 5 J-lffi
p + to n + spacing
3.0 J-lffi
n +/n + or p + /p + spacing
0.75 J-lffi
Polys i l icon wid th/space
0.75/0.75
Metal contact to polysi l i con
or active area
0.75 J-lffi
t-tm
Well Formation
Digital's CMOS technology uses an n-well formation
process. Starting silicon wafers, or substrates, are
doped p-type. The starting wafe r is composed of
a thin, high-resistivity epitaxial layer on a low­
resistivity substrate (p on p+). A thick oxide is ther­
mally grown on the wafer. Well regions are opened
in photoresist using a photomasking process. The
oxide is removed in the well regions. The n-wel l is
formed by phosphorous ion implantation (approxi­
mately JO l� atoms per square centimeter [cm2))
into the open regio ns. A high-temperature diffu­
sion step is performed to drive the n-well implan t
to a specified depth in the epitaxial layer. For the
CMOS-4 process, the d iffusion step was carefu l ly
adjusted to account for the closer spacing between
p and n transistors and the thin epitaxial layer thick­
ness (see Figure l a).
Epitaxial and well process requ irements for
Digital's CMOS-I and CMOS-4 technologies are given
in Table 2. The time required for we ll diffusion was
shortened to 2.5 hours from 9 hours. Epitaxial
thickness was reduced to 6.5 fLm to improve latch­
up immunity. After the well diffusion step, all oxide
is removed from the wafer before formation of the
device isolation regions.
Device Isolation
Isolation regions between devices are formed using
a conventional, semi recessed local oxidation of sili­
con scheme referred to as LOCOS isolation.' After
the pad oxide is grown and the silicon nitride i s
Table 2
deposited, device areas (or active areas) are imaged
i n photoresist. Sil icon nitride is removed from the
isolation regions (or field regions) and retained on
the active areas. In the substrate field regions, it is
necessary to increase the concentration of p-type
dopant to prevent unwanted current flow between
devices. Boron is selectively implanted (approxi­
mately lQl'> atoms per cmZ) into the substrate field
regions and is blocked from entering the well
regions by the photoresist l ayer. Note also that the
nitride not covered with photoresist must be thick
enough to block the implant (see Figure l b). Next
a thick thermal oxide approximately 0 .45 fLm is
grown in the field regions. The n itride, however,
prevents oxidation of the active area.
The LOCOS isolation has been tailored to CMOS-4
dimensions. Narrow-width transistors, routinely
used in dense SRAM layou ts, are particularly sensi­
tive to isolation process conditions. Pad oxide and
sil icon nitride thicknesses have been selected to
minimize excessive lateral oxidation of the active
area, which could cause a reduction in the physical
transistor width and thus lower the saturation cur­
rent. The oxidation temperature and time must be
optimized to grow the desired field oxide thickness
without negative impact on other device parame­
ters. For CMOs-4, the lateral oxide encroachment
was reduced to 0.25 fLm (per side). To minimize
undesirable lateral diffusion of the field dopant, a
relatively low temperature (950 degrees Celsius) is
used for field oxidation (see Figure l c) .
After field oxidation, t h e nitride is chemically
removed from the active areas, and the pad oxide is
chemically stripped. The field oxide regions are
semirecessed. Approximately 50 percent of the
field oxide is below the sil icon surface due to sili­
con consumption during the oxide growth .
Gate Formation
To achieve the desired electrical and reliability
behavior, microcontamination must be controlled
during gate region formation. The gate oxide mate­
rial must be free of any defects and contain minimal
amounts of ionic and met al l ic impurities. Rigorous
Epitaxial a n d Well D iffusion Process for Two CMOS Technolog ies
Technology
Epitaxial Thickness
CMOS-1
C M OS-4
Digital Technical journal
Temperature
Well Diffusion
1 1 30 degrees Celsius
1 1 00 deg rees Celsius
Vol. 4 No. 2
Spring 1992
Time
9 hours
2.5 hours
41
Semiconductor Technologies
chemical cleaning of the wafer is required before
gate oxide growth. Gettering (collecting) of mobile
ions is performed during gate oxidation by the use
of chlorine species. The wafer is transferred from
gate oxide to polysilicon deposition immediately to
minimize exposure of the dielectric film to airborne
contamination.
Channel Doping A 450-A thermal oxide is grown
on the active areas to condition or remove impuri­
ties from the silicon surface. Often referred to as
a "sacrificial " oxide, the thermal oxide is etched
from the sil icon surface before the real gate oxide
is grown . Before the growth of the gate oxide,
channel dopant for p- and n-type transistors is
implanted through the sacrificial oxide. Photo­
masking steps are done to selectively implant
boron into n-channel regions and phosphorus into
p-channel regions. For n-channel regions, two sepa­
rate boron implants are done: a shallow implant
(approximately 1012 atoms per cm 2 at 20 kilo elec­
tron volts [keY]) for threshold voltage adjustment
and a deep implant (approximately 1012 atoms per
cm 2 , 1 10 keY) to guard against various punch­
through mechanisms.
Being a surface channel device, the p-channel
region receives only a phosphorous threshold
adjust implant (approximately 1012 atoms per cm2
at 100 keY) since the well concentration is suffi­
cient for proper subthreshold operation.
Next the sacri­
A
ficial oxide is removed, and 105- thin gate oxide is
grown at 900 degrees Celsius. Then polycrystal l ine
silicon of 3500 A thickness is deposited, patterned,
and etched to define the CMOS n-channel and
p-channel transistor gate electrodes (see Figure l d) .
The requirement for appropriate symmetric device
design for the n- and p-channel devices is discussed
in greater detail in the Sym metric Device Require­
ment section of this paper.
Gate Oxide and Gate Electrode
Medium Doped Drain junction Formation
Next, a thin, 170-A silicon dioxide layer is grown on
the polysilicon and the active area regions. This
l ayer is fol lowed by a photoresist step that defines
the n-channel device. Phosphorus is implanted
(approximately w n atoms per cm 2 at 25 keY) to
form a shal low, 0.1-f.tm, medium doped drain (MOD)
junction for hot carrier protection4 and good
turn-off characteristics. The p-channel junctio n is
formed next. A photomasking step defines the
42
p-channel region, fol lowed by a shal low boron
difluoride (BF2) implant (approximately 1015 atoms
per cm 2 at 50 keY)
A 2000-A sil icon dioxide layer
is deposited and densified at 850 degrees Celsius.
The oxide is reactive ion etched. This particular
etch is anisotropic and leaves a vertical oxide side­
wal l 2000-A wide along the 3500-A thick poly­
silicon lines. The MOD ju nction is located u nder
the spacer oxide wall and extends 2000 A toward
the drain contact region (see Figure l e) . The oxide
spacer protects the MDD j u nction from receiving
the heavy dose source and drain implants, and it
protects the gate oxide during the CoSi2 formation.
Spacer Formation
After the
spacer formation, the photoresist step that defines
the n-channel transistor is repeated. An arsenic
implant (5 X 1015 atoms per cm2 , 100 keY) forms the
n+ source/drain junction. A final high-temperature
drive is performed to anneal the implant damage
and to drive the junctions for the n-channel and
p-channel devices to a final depth of 0.22 f.tm (see
Figure l f) .
SUPREM-based simulations of the CMOS-4 chan­
nel, MDD, and source and drain doping profiles are
shown in Figure 2a for the n-channel device and in
Figure 2b for the p-channel device. The channel
surface concentration for both devices is approxi­
mately 1 X 1017 atoms per cm 2 ; the MOD junction
depth is 0.1 f.tm; and the n + and p+ j unction depths
are both 0.22 f.tm.
Source and Drain junction Formation
Cobalt Silicide Module
The CMOS-4 process technology has relied on the
self-aligned CoSi (salicide) to improve device per­
2
formance by reducing the parasitic RC delay time
associated with the gate and active area inter­
connects. CoSi2 also provides good ohmic charac­
teristics for metal contacts to gates and active areas,
and acts as an etch stop during metal contact for­
mation 6 The term "salicide" refers to the forma­
tion of sil icide on the gate and the source and drain
region without the use of a masking layer. Because
the silicide forms only in areas where the deposited
metal can react with the exposed silicon surface,
no silicide forms over silicon dioxide areas.
Prior to CoSi formation, a wet chemical clean
2
removes any surface contamination, and a hydro­
fluoric acid dip removes any residual silicon diox­
ide (approximately 100 A) . The wafers are then
Vol. 4 No. 2
Spring 1992
Digital Technical journal
CMOS-4 Technology for Fast Logic and Dense On-chip Memory
"' 1 021
:::;;
u
a: 1 020
w
a.
�
�
z
1 0 19
1 0 18
0 1 0 17
�
a:
f-­
z
�
z
8
1 0 15
1 o'• c__
2.0
1 .5
1 .0
0.5
_
_
_
_
_
_
....
_
._
_
_
__._
---'
'---'-
o
DEPTH (IJm)
KEY:
0
D.
is self- limiting (once a l l the uoreacted cobalt is
removed, the reaction stops as the acid etches
cobalt and nothing else). A one-minute 700 degrees
Celsius anneal in n itrogen is then performed to
form 700 A of CoSi2 with sheet resistance at approx­
imately 5 ohms per square. (See Figure lg.)
Figure 3 shows a cross-sectional photomicro­
graph of CoSi2 film formed simu ltaneously over
the polysil icon gate and the active area region. The
CoSi 2 film thickness is approximately 700 A . The
spacer oxide separates the silicide in the active area
from the sil icide on the polysilicoo gate.
N + SOURCE AND DRAIN
N-CHANNEL
0 MOD
(a) N-channel Device
�
u
a:
w 1 01 9
a.
(/)
:::;;
18
0 10
f--
�
z
0
i=
<(
a:
f-z
w
u
z
0
u
0.2
0.4
0.6
0.8
KEY:
1 .0
1 .2
DEPTH (�m)
1 .4
1 .6
1 .8
2.0
0 P+ SOU RCE AND DRAIN
D.
P-CHANNEL
(b) P-cbannel Device
Figure 3
Figure 2
SUPREM Process Simulations
for CMOS-4 Channel Devices
introduced into a multistation, high-vacuum (down
to 5 X 10-8 torr) spu ttering system where another
100 A of sil icon dioxide is etched. This is fo llowed
by a spu tter deposition of approximately 200 A of
pure cobalt film on the su rface of the wafer.
The initial high-resistivity phase of CoSi is
formed using a rapid thermal annealer. Each wafer
is annealed at approximately 475 degrees Celsius
for 90 seconds in nitrogen gas. Next the wafers are
immersed in a selective etch, which is based on
phosphoric acid to remove all the unreacted cobalt
on the sil icon dioxide without attacking the already
formed CoSi or silicon dioxide. This 30-minute etch
Digital Tecbt�ical journal
Vol. 4 No. 2
Spring 1992
Cross Section of CMOS-4 Process
Salicide CoSi2 over Polysilicon
ana Active Area
Local Interconnect Process
The requirements for local interconnect materiaF
include low-resistivity film with good etch selectiv­
ity to the u nderlying silicide film. Also the film must
have good electrical contact resistance to both the
sil icided gates and silicided source and drain
regions. The material selected was a laminate of
200 A of titanium and 600 A of low-resistivity
titanium nitride (TiN). Titanium reduces the coo­
tact resistance through a chemical reaction with
any surface oxides over the CoSi2 to form a good
ohmic contact. The titanium nitride has a higher
resistivity, but is more chemica l ly stable and does
43
Semiconductor Technologies
not oxidize as readily as titanium during subse­
quent processing steps, such as oxygen plasma
strips. Titanium nitride is also much easier to pat­
tern with photoresist for the subsequent reactive
ion etch step.
The local inter­
connect is deposited using the same high-vacuum
sputtering system used for cobalt deposition. The
titanium and the titanium nitride are sequenti a l ly
deposited in the same cham ber. Nitrogen is reacted
with the titanium on the target surface w t<Jrm tita­
nium nitride. Subsequently, the ti tanium nitride is
sputter deposited onto the wafer. Va rying the nitro­
gen flow can control the chemical and electrical
properties of the titanium nitride film. The deposi­
tion process is adjusted to produce TiN with a resis­
tivity of 50 p,ohm X em.
Local Interconnect Deposition
Interconnect Etch Local interconnect
etch consists of patterning TiN i nterconnects
between source/drain and unrelated polysil icon
(see Figure 1 h). At this layer, there are two main
concerns: (1) complete removal of TiN residue
(stringers), and (2) etch selectivity to the other
exposed materials on the water.
At this point i n the fabrication process, the TiN
is 4.5 times thicker at the side of an oxide spacer
than over a flat region on the wafer. Continued
etching to remove the TiN residue by the spacer
wou ld resu lt in unacceptable overetching in the flat
regions. D ue to the contormality of the TiN to the
spacer, the width of the TiN at the side of the spacer
is the same as the thickness of the TiN in the flat
regions. By controll ing the ratio of the anisotropic
etch component to the isotropic etch component,
a portion of the TiN residue can be removed later­
ally, which significantly reduces the amou nt of
overetch requ ired . This combination of vertical
and lateral etching is accomplished by using a
chlorine/trifluoromethane (CiziCHF�) plasma chem­
istry at low pressure.
During the TiN etch, four other materials are
exposed to the reactive plasma: silicon oxide,
CoSi2 , photoresist, and sil icon. To minimize mate­
rial loss, the TiN etch rate must be optim ized to
significantly exceed the etch rates of these other
materials. This is accompl ished by using response
surface methodology.
The optimized etch process consists of a two­
step etch. The first step has an anisotropic char­
acteristic. The pressure is 25 millitorr (mtorr);
bias voltage is - 220 V ; Cl 2 is 20 standard cubic cen­
ti nH..: ters per minu te (seem); CHF; is 30 seem; and
boron trichloride (BCl�) is 120 seem. The second
step is a more isotropic overetch st<.:p to rcm ov<.:
residual TiN. I t consists of a 25-mtorr pressure and
a 94 v bias voltage. The C 12 is 60 seem; CJ IF3 is
40 seem; and BC13 is 90 s e e m . The temperature of
the cathode is m aintained at 50 degrees Celsius.
The photoresist is stripped using a three-step pro­
cess. The first step is a w<.:t solvent strip that
removes any soluble residues and plasm a-hardened
resist from the wafer. The second step is an O:\.)'gen
plasma strip using a single-wafer stripper. The final
step is another wet solvent strip (see Figure l i) .
Figure 4 i s a photomic rograph of the top view of
the active area and polysilicon regions. It shows the
thin layer of TiN local i nterconnect strap that short­
circuits the two l ayers together.
-
-
Local
44
Figure 4
Photomicrograph Showing the Top
View of the Active Area, Polysilicon,
and liN Local interconnect
Transistor Design Considerations
Figure 5 shows a cross section of a CMOS-4 transis­
tor. It highlights t he poJysilicon gate region, spacer
r<.:gion, CoSi2 r<.:gion, and the metal contact regions
filled with TiN and tungsten films.
Unless adequately designed, submicron CMOS
devices suffer many u ndesirable electrical dfeets.
These effects are rel ated to the sealing of the tran­
sistor channel length and the gate oxide thickness.
Scaling the effective channel length for both n- and
p-chan nel devices requ ires a reduced ju nction
depth and an increased channel surface concentra-
Vol. 4 No. 2
Spring /'}92
Digital Tecbnical]ournal
CMOS-4 Technologyfor Fast Logic and Dense On-chip Memory
The
two-dimensional
device
opttmtzation
resul ted i n the fol lowing graded ju nction process
parameters: a phosphorous dose of 7 X 1013 atoms
per cm2 at 25 keY, d iffused to a 0.1 -fl-m junction
depth with a spacer width of 0.15 to 0.2 fl-m.
The phosphorous surface concentration was set
to approximately 10 1 9 atoms per cm 3 to reduce the
source and drain series resistance, R1 , and to accom­
pl ish the highest possible saturation current, Idsat'
while
maintaining
low substrate
cu rrent
for
improved hot carrier rel iabil ity. The graded phos­
phorous junction was called medium doped drain
Figure 5
Cross Section of CMOS-4
0. 75-�-Lm Transistor
tion aimed at improving the short channel effects.
These solutions may also contribute to an increase
in the built-in electric field, an increase in the
impact ionization, an increase in the substrate cur­
rent, and a decrease in the pu nch- through voltage.
Arsenic doping profiles are usua l ly used to fabri­
cate shal low ju nctions in n-channel devices. They
produce a high electric field that can cause a high
rate of impact ionization between the electrons
injected from the source and the fixed ions in the
depletion region of the drain ju nction. On impact,
(MDD) to refer to the relatively high doping concen­
tration (1 X 10 19 atoms per cm3) . In contrast, the
l ightly doped drain (LDD) junction, with eloping in
the 10 17 to 101 8 atoms per cm3 range, suffers from
large R1 and low driving current capability.
By implementing the M D D process, we were able
to accomplish the fol lowing n-channel device char­
acteristics: (1) m inimum device l ifetime of 20 years
at a drain voltage bias (Vas ) of 4 . 3 V, (2) source and
drain series resistance of0.05 ohm X em, (3) driving
current capability Ittsat of 0.385 m i l l iampere (mA)
per fl-m, and (4) punch-through voltage (BVDSS)
above 7 V. See Table 3.
Table 3
electron-hole pairs are generated . The holes are
swept to the source or substrate region and are
known as the substrate current. The electrons sub­
jected to the high drain electric field can gain
enough energy to inject in the gate oxide region
above the drain junction. These "hot electrons" may
create interface states or may lose enough energy
and be trapped in some defect location. Long- term
effects of the trapping mechanism give rise to trans­
conductance and saturation current degradation
which eventua l ly lead to circuit fai lure.
N-channeljunction Formation
To protect against the rel iabil i ty hazards associated
with arsenic profiles, a graded junction was imple­
mented. Heavy emphasis was placed on the SUPREM
process simu lator and the MINIMOS device simu­
Electrical Parameters for C MOS-4
Transistors
N-channel
M DD X
i
P+ Xi
Xwell
Tax
P-channel
0.1 f.lm
NA
NA
0.1 8 fJ-m
NA
1 . 75 f.lm
1 05 A
1 05 A
Vrx
0.5 v
- 0.5 V
0.5 f.lm
0.5 f.lm
Delta L
0.25 f.lm
0.25 f.lm
Delta W
0.55 f.lm
0.65 f.lm
fdsat
0.385 mA per f.lm
- 0.1 67 mA per f.lm
>7 V
>-7 V
Left
BVDSS
Figure 6 shows the resu lts of two-dimensional
MTNIMOS simulation of the lateral electric field d is­
lator.8 The use of these simulat ion tools a llowed us
tribution in the CMOs-4 n-channel device with MDD
to accurately predict the device behavior u nder cer­
j unction profiles. The electric field is plotted a long
tain electrical conditions, and ensu red that the
the transistor channel region starting from the
device characteristics were optimized for high per­
source toward the drain region. The device has a
formance and superior reliability. Use of these sim­
drawn polysil icon length of 0.75 fl-lll and a gate
u lators also reduced the dependence on wafer
oxide thickness of 105 A. The drain voltage biases,
usage for process optimization.
�Is ' were set to 4.3 V and 3.3 V, and the gate voltage,
Digital Technical journal
Vol. 4
No.
2
Spring 1992
45
Semiconductor Technologies
0
_J
w
u:: - 4
u
a:
f-
�
u:J
- S L-�--�----�
-0.4
KEY:
--8--
····
0
····
Figure 6
-0.2
Vd s
Vcts
=
=
0
0.2
0.4
LENGTH (�m)
0.6
0.8
1 .0
3. 3 V
4 .3 V
N-channel MINIMOS Simulation
of Lateral Fieldfor Medium Doped
Drainjunetions
�s' was set to
2.2 v.
Notice that the peak field for
3.3-V operation is approximately 0.3 X i06 v per em
compared to the worst-case field condition of
0.4 X 10 6 V per em, where l�1s equals 4.3 V.
P-channeljunction Formation
The p-channel device requires special care in the
design of the p + source and drain junction depth to
ensure that boron does not penetrate the thin gate
oxide and enter the n-well region. The p+ junction
dose, j unction depth, and temperature cycle were
optimized to accomplish a low R1 , by using BF2
(1 X 1015 atoms per cm2 and an energy of 50 keY). I n
addition, the diffusion time was minimized, and
good threshold voltage control and punch-through
protection were maintained. The boron distribu­
tion in polysilicon obtained with extensive sec­
ondary ion mass spectroscopy (SIMS) analysis, as
wel l as negative bias and temperature instabil ity
(NBTI) tests, ruled out any boron penetration.
Synzmetric Device Requirement
High-performance submicron CMOS technologies
require the simultaneous optimization of the
n-channel and p -channel devices for high driving
current capability and excel lent short-channel
device characteristics. This is best accomplished
with symmetric design where channel eloping,
junction depth, and threshold voltage of both tran-
46
sistors are equivalent.? Furthermore, the n- channel
has an n-type doped polysilicon formed during the
n + source/drain junction and the p-channel has a
p-type doped polysil icon formed during the p +
source/drain junction. The technique o f symmetri­
cally designing the devices al lows the threshold
voltages to be equal in magnitude, and suitably low
for high driving current capability while maintain­
ing good punch-through characteristics.
Figure 7 shows the resu lts of the MINIMOS simu­
lation of the potential distribution in CMOS-4 tran­
sistors with p-type doped polysil icon gate . The
bias points were set fo r �s equals 0 V and �Is equals
- 3.3 V. The channel length was 0.75 JLm , and the
gate oxide was 105 A . Superior pu nch- through char­
acteristics are observed since the potential contour
l ines do not spread significantly toward the source.
0.2
1 0.4
I
t­
o..
�
0.6
0.8
-0.4 -0.2
KEY:
0
-3.45
-3.02
-2.59
Figure
7
0.2
0.4
0.6
LENGTH (�m)
-
-·-
---
-2. 1 6
-1 .73
--- - - -1 .31
0.8
1 .0
1 .2
1 .4
. . . . .... -0.877
-
-·-
-
-0.449
-0.02
P-channel MINIMOS Sitnulation
ofPotential Distribu tion
Device Characteristics
The extrapolated threshold voltage for an n-channel
device plotted as a fu nction of gate length is shown
in Figure 8. Excellent threshold voltage control
is shown for channel length down to 0.5 JLm. The
n-channel drain current, Ids, is plotted in Figure 9 as
a function of drain voltage, Vds' while �s is varied
from 0 to 5 V with 0.5-V steps. In Figure 10, the drain
current is plotted on a logarithmic scale as a func­
tion of gate voltage to highlight the su bthreshold
Vol. 4 No. 2
Spring 1992
Digital Technicaljounzal
CMOS-4 Technology for Fast Logic and Dense On-chip Memory
0.60
1 0-2
0.58
1 0-5
0.56
0.54
6
0
0.52
�
:J
�
_9 1 0-8
0.50
10- 1 1
0.48
0.46
1 0- 1 4
-0.5
0.44
0
1 .5
2.0
VGs (V)
0.42
0.40
1 .0
0.5
1 .8
1 .4
1 .0
0.6
2.6
2.2
3.0
Lelf(�m)
Figure 8
N-channel 17Jreshold Voltage Plotted
as
a Function of Effective Channel
Length
Figure 10
N-channel Drain Current Plotted
as
a Function of Gate Voltage
Table 3 shows typical CMOS-4 transistor process
and device parameters. The j unction depths, xi'
7.0
and the n-well depth, Xweli' are simu lated with the
SUPREM process simul ator and verified with SIMS
6.0
analysis. Tax is the physical gate oxide thickness; Vrx
5.0
<(
.s
is the extracted threshold voltage; Lcrr is the nomi­
nal final channel length, and delta L and delta w are
4.0
electrically extracted using the Terada method,
!/)
.9
which accounts for the parasitic series resistance.
Idsat is the saturation current measured with the
drain and gate voltage at 3.3 V BVDSS is the pu nch­
through voltage measured with
1 .0
0
1 .0
2.0
3.0
4.0
5.0
6.0
7.0
Silicided Interconnects Characteristics
Table 4 shows the effects of the sal icide process on
Vos (V)
Figure 9
�s set at 0 Y.
the parasitic resistance in four consecutive tech­
N-channel Drain Current Plotted
as a Function of Drain Voltage
slope behavior for �ts of OJ V and 3.6 Y. The sub­
threshold slope was measured to be 86 mY per
decade and is characterized by good drain-induced,
barrier-lowering characteristics. The drawn di men­
sions of the transistor are 12.5 J-Lm wide by 0.75 J-Lm
long.
Similar characteristics are observed for the
nologies. The CMOS-1 process uses no sil icided gate
or drain and therefore is expected to have a high
interconnect sheet and contact resistance. CMOS-2,
on the other hand, uses a low sheet tungsten sili­
cided (WSi 2) polysilicon gate with a sheet resistance
of 3 ohms per square. The CMOS-3 and CMOS-4 tech­
nologies use sal icided low sheet resistance CoSi 2
for both the polysil icon gate and the source/drain
region with a sheet resistance of 5 ohms per square.
p-channel device and are shown in Figures 1 1 , 12,
SRAM Implementation
and 13. The subthreshold current conduction and
A six- transistor (61) cell was selected for its process
punch-through characteristics are very similar to
simplicity and cell stabili ty. To provide a dense ,
those of the n-channel device.
cost-effective SRAM capabil ity, the 6T cell was
Digital Technical journal
Vol. 4 No. 2
Spring 1992
47
Semiconductor Technologies
0.60
0.58
0.56
0.54
�
>->
LJ.J
f:::>
_J
0
(/)
0..
co
<(
0.52
0.50
0.48
0.46
-1
0.44
o-14
/
7----L---=---'---=-'c::----L-�=--_,_-��-�
0.5
0
-0 .5
0.42
0.40
- 1 .0
-1 . 5
-2.0
VGs (V)
0.6
1 .0
1 .4
1 .8
2.2
3.0
2.6
Len (�m)
Figure 11
P-chcmnel Threshold Voltage Plotted
Figure 13
P-channel Drain Current Plotted
as a Function o{Gate Voltage
as u Function ofrffective
Channel Length
region withou t the need for area-consuming met:�l
contact. Unfortunately, this technique is not readily
-3.0
compatible with symmetric n + and p+ doped poly­
silicon structures. In addition, sil icon grooves
might form during polysilicon etch, which could
jeopardize the ju nction in tegri t y and cause leakage
-2.0
or short circu its to the s i l icon substrate.
The preferred method to access the source/drain
region was the use of TiN strap over CoSi2. TiN local
i nterconnect is a condu ctive m a terial. When it is
-1 .0
sputter deposited on the wafer, pattern and etch
can be used to strap the node of one transistor to
the gate or drain of another transistor. Also, the TiN
-1 .0
-2.0
-3.0
-4.0
-5.0
-6.0
-7.0
Vos (V)
local interconnect provides excel lent etch selectiv­
i t y to the u mkrlying CoSi 2 material . The TiN local
interconnect process p roved superior to the buried
contact scheme because the im proved etch selec­
Figure 12
P-channel Drain Current Plotted
as u
Function of Drain Voltage
chosen over the 4T cel l, which requires complex,
two-level polysilicon films.
tivity to CoSi2 prevents junct ion leakage.
In standard layout techniques, the metal 1 con­
tact is spaced 0.75 JLDl from the edge of the
polys i l i con gate and the isol ation. This spacing
resu lts i n a 2.25 -t-tm wicle act ive area, as shown in
During the initia l 6T cel l process development,
Figure 14a. In contrast, the local interconnect tech­
the TiN local i n terconnect scheme was considered
nique does not require a contact region ; therefore
advantageous to the buried contact scheme. I n the
the active area width can be scaled to 1 .5 t-tm, as
buried contact procedure, the gate oxide is pat­
shown in Figure 14 b. The usc of local i n terconnect
terned and etched, and then a polysil icon gate is
has reduced the 61' cel l area from 120 JLlll 2 (no Ll)
sil icon and the active area. This technique a l lows
has improved yield due to a relaxed metal 1 contact
the polysilicon ti.lm to access the source/drain
requ irement and metal spacing.
deposited to provide the contact between poly­
48
to 100 t-tm2 (with LI). In addition, the use of L I
Vol. i No. 2
Spring 1992
Digital Technical jounwl
CMOS- 4 Technology for Fast Logic and Dense On-chip Memory
Table 4
Sheet Resistances for C M OS Techno log ies (Ohms per Squa re)
CMOS-1
CMOS-2
CMOS-3
CMOS-4
N + /P+
N +/P+
N + /P+
N + /P +
Sou rce/d rain sheet resistance
40/75
40/75
5
5
Po lysilicon sheet resistance
40/NA
3
5
5
Local i nterconnect
NA
NA
NA
6
_/ POLYSILICON GATE
-"'
0.75 1Jm
ACTIVE AREA
,
!
"' ' "m
0. 75 IJm
�
2.25 �Jm
0.75 1Jm
active area regions, compared to eight contacts i n
the cel l with no LI. The minimum transistor width
in the ceJI is 1 .5 p.m.
Summary
CMOS-4 technology for the Alpha 21064 and the
NVA.X microprocessors was discussed in detail.
Process and device features for fast logic and
dense on-chip SRAM were presented. The high­
performance transistor requires the simultaneous
optimization of the drain ju nction for hot carrier
resistance and for high driving current capabi l ity.
Low-resistance silicided interconnect uses a robust
CoSi process. On-chip SR.Ai\1 based on a 6T cell
2
with TiN local interconnect provides high-density
and high-yield products.
M ETAL CONTACT
(a) Standard Layout
LOCAL INTERCONNECT
Acknowledgment
The authors wish to acknowledge the contribu­
tions of all the members of the CMOS-4 group and
the FAB-4 manufacturing group.
0.75
-- :
References
1 . R. Badeau et al. , " 100 MHZ Macropipelined CISC
CMOS Microprocessor," IEEE International Solid­
'
· - - - - - - - - - - - - .. - '
State Circuits Conference Digest of Technical
Papers (February 1992): 104- 105.
(b) Local Interconnect Layout
Figure 14
Layout Schematics Comparing
Metal Contact with Local
Interconnect
Figure 15a is a photomicrograph of a CMOS-4 6T
SRAM cel l taken after LI etch and photoresist strip.
It highl ights the active area regions covered with
CoSi 2 and TiN Ll straps. Figure 15b shows a layout of
the SR.A.J\1 cell used in the Alpha 21064 micro­
processor. Transistors Tl, T2, etc . , are highlighted
in Figures 15 and 16 to simpl ify their identification.
The cell area is 6.75 by 14.8 p.m 2 (100 p.m2) . The cel l
uses only three metal 1 contacts ( V00 and Vs.>) to
Digital Tecbnical jountal
Vol. 4 No. 2
Spring 1992
2. D. Dobberpuhl et al., "A 200MHz 64b Dual-Issue
CMOS Microprocessor," IEEE Jntemational Solid­
State Circuits Conference Digest of Technical
Papers (February 1992): 106- 107
3. J. Appels et al. , " Local Oxidation of Sil icon and
Its Application in Semiconductor Device Tech­
nology," Philips Research Reports, vol. 25 (1970) :
1 18- 132.
4. P Tsang, S. Ogura, and W Wal ker, "Fabrication
of High Performance LDDFET's with Oxide
Sidewall-Spacer Technology," IEEE Transactions
on Electmn Devices, vol. ED-30 (1988): 652 -657
5. C . Ho, ]. Plummer, S. Hansen, and R. Du tton,
"VLSI Process Modeling SUPREM-II I ," IEEE Trans-
49
Semiconductor Technologies
BIT LINE
T2
V55
T4
(b) Layout Schematic for the
Local Interconnect Cell
(a) Photomicrograph of the
Local Interconnect Cell
Figure 15
T6
Local Interconnect Cell Used in the Alpha 21064 Chip
actions on Electron Devices, vol . ED-30 (1983):
1438.
6. S. Murarka, D. Fraser, A. Sinha, and S. Hil lenius,
"Self-Aligned Cobalt Disilicide for Gate and
Interconnection and Contacts to Shallow Junc­
tions," IEEE Transactions on Electron Devices,
vol . ED-34 (1987): 2108- 2114.
7. T. Tang, C. Chia-Wei, R. Haken , and T. Holloway,
"VLSI Local I nterconnect Level Using Titanium
Nitri de,"
IEDM
Technical Digest (December
1985): 714-717.
8. S. Sel berherr, A. Schultz, and H. Potzl,
" M INIMOS-A Two-Dimensional MOS Transistor
BIT
LINE
WORD LINE
BIT
LINE
Analyzer," IEEE Transactions on
Devices, vol. ED-27 (1980): 1540- 1550.
Electron
9. K. Cham and D. Wenocur, "Submicrometer Thin
Figure 16
Six- transistor Cell Identifying
the Various Transistors
Gate Oxide P-Channel Transistors with P +
Polysilicon Gates for VLSI Application," IEEE Elec­
tron Device Letters, val. EDL-7, no. 1 (January
1986).
50
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Marion M. Garver
joseph M. Bulger
Thomas E. Clark
jamshed H. Dubash
Lorain M. Ross
Daniel]. Welch
CMOS-4 Back-end Process
Developmentfor a VLSI 0.75- J.Lm
Triple-level Interconnection
Technology
Digital's CMOS-4 on-chip interconnect technolOg)J, developed for and used in pro­
duction of the NVAX and the Alpha 21064 microprocessor chips, is a three-level alu­
minum alloy metallization process, with planarized TEOS-based silicon dioxide
dielectrics, tungsten-jilted contacts and vias, and a minimum feature size of
0. 75 J.Lm. The process development effort was a twofold approach based on the maxi­
mum use of existing manufacturing capability and the introduction of required
new process features. For photolithography, plasma etch, and PVD metallization,
the 1. 0-J.Lm manufacturing equipment set and processes were modified and reopti­
mized for the submicron regime. In addition, two new process features, a blanket
CVD tungsten process and a TEOS-based oxide planarization process, were developed
and implemented in manufacturing to meet the CMOS-4 technology requirements.
Each generation of Digital's complemen tary metal­
oxide semiconductor (CMOS) very l arge-scale inte­
gration (VLSI) microprocessor development has the
goal of providing a 30 percent net incremental per­
formance improvement and a twofold area density
improvement from the previous technology. This
logic design need for higher density and improved
performance places a considerable demand on
u ltra-l arge-scale integration (ULSI) circuitry to pro­
vide processes that permit a scal ing of horizontal
geometries with vertical film thicknesses remaining
constant. 1
The technology goals for fourth-generation CMOS
(CMOS-4) were met by providing a 25 percent algo­
rithmic reduction of horizontal feature size from
1 .0 micron (,urn) to 0.75 ,urn, accompanied by mini­
mal or no reduction in back-end interconnect or
dielectric thicknesses. The process materials and
critical parameters are described in Table 1 . The
small spatial resolution required vertical-walled
vias to access smal ler-pitched metal layers effi­
ciently. Interconnect reliability was maintained
through implementation of a tungsten-filled via­
plug to improve current spreading and to maintain
D igital Techuical]otn-nal
Vol. 4 No. 2
Spring 1992
metal step coverage into contacts 2 The CMOS-4
interconnect structure is shown in Figure 1 .
The CMOS-4 process was developed in a manufac­
turing fabrication clean room originally configured
for the preceding 1 0- ,u m CMOS-3 technology. The
goal of the Advanced Semiconductor Development
(ASD) and Manufacturing Engineering Groups was
to introduce as few process changes and new
pieces of equipment as possible. For two of the pro­
cesses, j oint development efforts at equipment ven­
dor sites were conducted to develop hardware and
assess process feasibility. The equipment was p ur­
chased and instal led in the manufacturing clean
room with final process characterization and inte­
gration performed at Digital's Hudson facility.
This paper discusses how the existing tools
were modified for use in the CMOS-4 process in
the areas of photolithography, plasma etch, and
physical vapor deposition (PVD) metall ization. It
describes the addition of blanket tungsten and
plasma-enhanced tetraethylorthosilicate (PE-TEOS)
oxide processes that were developed in clustered,
multichamber tools and optimized to meet the sub­
micron-level requirements of CMOS-4 technology.
.
51
Semiconductor Technologies
Table 1
CMOS-4 F i l m Types, Thicknesses, and Critical Dime nsions
Process Level
Material'
Final Thickness Ta rget
Dielectric 1
Phosphorus-doped PE-TEOS
75oo A
and boron oxide
A
Critical Dimensions
Metal 1 contact
Ti / Ti N and W p l ug size
Plug recess <3000
Meta1 1
Al:1 % C u / TiN cap
75oo A
Di electric 2
PE-TEOS and SOG
75oo A
Metal 2 contact
Ti /Ti N and W plug
Plug recess <3000 A
0.75 by 0.75 JLffi via size
Metal 2
Al:1 %Cu/Ti N cap
75oo A
1 .88/0.75 JLffi line/space
Dielectric 3
PE-TEOS and SOG
1 8 ,ooo A
Metal 3 contact
Al:1 %Cu
> 0.3 /Lm
Metal 3
Ti N/AI : 1 %Cu/TiN cap
2o,ooo A
Passivat ion
PE-TEOS
75oo A
0.75 by 0.75 JLffi contact
1 .50/0.75 JLffi l i n e/space
3.0 by 3.0 JLffi via size
4.5/3.0 JLffi l i n e/space
9
�
35§:#,�
KEY:
1
2
3
4a
4b
FIELD OXIDE
POL YSILICON
DI ELECTRIC 1
TUNGSTEN 1 PLUG
TO POL YSILICON
TUNGSTEN 1 PLUG
TO SOURCE/DRAIN
5
6
7
8
9
10
11
METAL 1
D I ELECTRIC 2
TUNGSTEN 2
METAL 2
DI ELECTRIC 3
METAL 3
PASSIVATION
Fip,ure 1 Cmss-sectirmal Photornicrograpb and Schematic of CMOS-4 /nterconnect Structure
Modification ofBack-end Processes
tion. The photoresist t hickness used at metal 1 and
The details of de ve l op m e n t
metal
t:fJ( >rts and
their
resol ution fo r photo! i t hograp hy plasma e t c h and
,
,
PVD met a l l ization an: d iscussed i n the fo l l owi n g
scctions.
2 is 2.0 /Lm in o rd er to prevent total photo­
resist erosion from the higher steps during the etch
process. Metal 3 contact and metal 3 use a thi c ke r,
3.5 -JLm p h oto r es i st t o accommodate tapered oxide
etching, extreme topography, and high loss of photo­
Photolithography
resist during etch . In the manufacturing line, m u l­
The CMOS-1 photo l it hography process uses single­
t iple track and stepper combinations are al lowed.
layer ph ot or es i sts
Critical d i mension control is mainta ined by run­
,
reduction steppers with a n
exposing wavek:ngth of 4.)6 n ano m e t e rs (nm) and
n ing fixed exposures and m o n i toring p rocess E,
.4S :--JA and .S4NA, and
once a shift fo r the possible coat-and-expose equ ip­
s p ray/pu d d lc develop . A photoresist th ickness of
1 . 2 /Lm is used at contact l e ve l s to enhance rt:solu-
ach ieved by running a lot p i lo t wafer and using
n ume ri c aperture lenses of
">2
ment scquences. Overlay control of ::!::: .20 !Lm is
Vol. 4 No. 2
Spring 1992
Digital Technical journal
CMOS-4 Back-end Process Developmentfor a VLSI 0.75-f.l-m Triple-level Interconnection Technology
a lignment offsets to center the lot d istribution at
Maximum metal 2 contact
zero.
Minimum metal 2 l ine width = 1 .55 f.l-m
The CMOS-4 photolithography process was trans­
ferred into the existing CMOS-3 production l ine
without any modifications to existing equipment.
However, the introduction of back-end processes
did present resolution problems at the metal 2 and
contact layers and poor overlay performance at a l l
metal layers.
=
0.95 f.l-m
Misalignment tolerance = ( 1 .55 - 0.95) / 2 = 0.30 f.l-m
Similar calculations performed at metal I contact
and metal 3 contact ind icated similar tolerances
were needed. However, signal - to-noise rati os with
the a l ignment systems were extremely low and
overly sensitive to minor process fl uctuations, such
as changes in film thickness, film reflectivity, sur­
Resolution The challenge for many of the upper
levels was to routinely resolve 0.75 -f.l-m geometries.
For contact and via levels, part of the solution was
to use thinner photorc�ist, overexposure, and only
the higher NA steppers. Initial ly, material processed
at these levels periodically exhibited d istorted o r
missing contacts. Experimentation with focus/
exposure matrices showed no further improve­
ment with overexposure. Also, sl ight focus shifts
were enough to considerably d istort the contact
pattern. A 0.5 -f.l-m focus offset provided the neces­
sary latitude to el iminate resolution problems.
Metal 2 Processing Poor pattern definition caused
both bridged photoresist lines over topography and
rounding of the tips of l ines. The photoresist could
not be thinned below 2.0 f.l-m due to the high loss of
photoresist during etch. Elimination of dyed photo­
resist was an option because titanium nitride
(TiN) antiretlective coating (ARC) was used below
the photoresist layer. Undyed photoresist elimi­
nated the bridging problem. A focus sh ift similar to
that used at the contact levels was necessary to best
resolve the tips of the l i nes. Factorial ly designed
experiments indicated that higher exposure and
further defocus wou ld help minimize metal short
circuits. The increased exposure widened the gap
between potentially short-circuited lines, and the
defocus increased the degree of proximity effect,3
leaving the dense l ines significantly smaller than
the isolated l ines. The 0.5-f.l-m focus offset was again
selected as opt imum. It decreased the photo­
lithography contribution to metal short circuits yet
sti l l produced l ine widths that met design gu ide
criteria.
face roughness, extent of planarization , and grain
boundary highl ighting. Initial attempts ind icated
a l ignment, when possible, had an average value
above 0.40 f.l-m for a lot. New a lignment systems
specifical ly designed for metal and planarized
d ielectric were considered. However, the decision
was made to develop a new processing technique
on the existing equipment.
A logical approach to the problem was to elimi­
nate the metal from the a l ignment target areas.
Al l efforts at optimizing target size still rendered
only a marginal ly acceptable process. "Cutout" pro­
cessing at metal 3 was used from the beginning of
CM05-4 development and was gradually introduced
to the other metal layers as wafer lot volume
increased and problems with al ignment increased
cycle times and scrap rates.
The cutout process involves running an extra
masking step prior to metal alignment, which
exposes the underlying a lignment targets. Since the
cutout openings are large, misal ignment tolerance
is on the order of microns and is achievable by al ign­
ing the cu tout mask in a manual global a l ignment
mode fol lowed by a blind step sequence. Wafers are
subsequently either dry or wet etched to remove
the TiN and aluminum (AJ), then returned to photo­
l i t hography for standard metal al ignment process­
ing. Overlay results using the cutout at metal 2 and
metal 3 average approximately 0.20 f.l-m ( :± 3 a).
Metal 1 performance is approximately 0.15 f.l-m
( ± 3 a) by a lign.ing to the sti l l visible active area
marks.
Plasma Etch/Strip Processes
The CMOS-4 back-end oxide etches, metal etches,
and photoresist strip processes were developed on
Alignment Maximum misal ignment tolerance is
d ictated by the rel iabil ity requirement to ensure
100 percent contact coverage. The al ignment toler­
ance calcu lation was based on maximum al lowable
contact size and minimum metal l ine widths. For
example:
Digital Tech11ical]ournal
Vol. 4 No. 2
Spring 1992
the existing manufacturing equ ipment. A straight­
walled contact process for metal 1 con tact (M 1 C)
and metal 2 contact (M2C) needed to be developed
for incorporation with the tungsten plug technol­
ogy. Additional ly, the 0.75 -f.l-m wide contacts with
straight sidewa l l s (greater than 85 degrees) and
53
Semiconductor Technologies
aggressive aspect ratios (height/width), requi red
resist p u l lback) were deter m i n ed to be the most
optim ization of the photoresist strip processes. The
critical parameters. The optimized process resulted
metal 3 cont act (i\{)C) tapered etch process was
in a h igh-throughp u t , u n ifor m , 0.75 -1-Lm straight ­
also rnlcvcloped to meet CMOS-4 electromigration
wal led contact process with a n oxide-to-cobalt s i l i­
rel iabi l ity requirements.
cide selectivity of greater than 2 5 : 1 . The u nderlying
In add i t ion to a straight-wal led contact process
for use with tungsten p lugs, excellent selectivity to
material a t }!2C etch is an a l u m i n u m (Al) al loy.
Because of a very h ig h selectivity of oxide-to-Al
underlying materials was required to compensate
etch rates, an overet c h of up to 100 p e rcent at M2C
for the increased planarity of t he C .\10S- 4 d ielec­
etch d i d n o t i m p act the contact profile or the con­
t r ics. For example, the planarity of the cl i e lectric
tact resistance.
between polys i l icon and metal 1 meant that the con­
tacts to a polys i l icon gate would be etched through
the t h innest d ielectric (approx i mately 0.75 11-m),
whereas the contacts to active area regions would
Optimized Photoresist Strip Process The opti­
M l C/M2C straight -wal led etch process
m i zed
altered the amount of sidewal l polymer formed
need to be etched over a much thicker d ielectric .
in the contacts d u ring the contact etch process_(, In
The worst-case difference is approximately two
addi t ion, the change i n aspect ratio affected the
times the thickness of the d ielectric above the poly­
ability of wet chemicals to remove all the photo­
silicon layer, or approxi mately 1 .5 11-m, as shown
resist and polymer during photoresist strip process­
in Figure 2. This d ifference in d ielectric film t h i ck­
i ng. It was shown empirically that residual poly­
ness meant that the coba lt s i l icide (CoSi 2) film
over the polysil icon contact would be overetched
i m p acted contact resistance. The contact photo­
by approx i m ately 100 percent d u r i ng the M l C etch
mer rem a i n i ng in contacts after photoresist strip
resist strip process was mod ified from a two-step
process. Therefore, the M 1 C etch process needed
dry/wet process to a three-step wet/dry/wet pro­
a very h igh d i fferential etch rate or selectivity
cess. The first wet strip was requ i red to pre-wet
bet ween the etch rate of the oxide as compared to
the polymer/photoresist in the contacts. The b u l k
the etch rate of the C o S i 2 . 1
o f the photoresist was t h e n removed i n a n oxygen
COBALT SILICIDE
ON TOP OF
POLYSILICON
lowed w i t h a final wet strip to remove any residues.
downstream plasma stripper (dry) and was fol­
POLYSILICON
CONTACT
ACTIVE AREA
CONTACT
Beginning with a wet strip cycle also i mproves s ide­
wal l polymer removal for metal etch processes.
Consequently, all the CMOS-4 back-end resist strip
i
processes fol low a wet/dry/wet strip process flow.
2.x
Metal 3 Contact Etch Process
Initial ly, the M:JC
tapered e t c h p rocess did not consistently meet the
FIELD OXIDE
ACTIVE AREA
POLYSILICON/M1
POLYSILICON
DIELECTRIC
Figure 2
Schematic Drawing Showing
the Difference in Step Heightfor
Metal 1 Contact to Polysilicon
and Source/Drain Regions
Stra ight- ll'allcd Contact Process
The
straight­
CMOS-4 e lectromigration m i n i m u m step-coverage
requirement of 0.30 11-m of metal on M3C sidewa lls.
A u nique set of problems existed at M3C etch . The
nonun iformity
in
the
underlying
topography
caused the photoresist coat to be t h i n ned over iso­
lated metal I i nes. This t h i n coat led to early p hoto­
resist
breakthrough
and
subsequent
d ielectric
erosion d uring M3C etch . A thicker photoresist coat
led to steeper s idewalls, which resu lted i n a degra­
dation of metal step coverage.
An exp erimental design was used to optimize the
photoresist and etch processes in u nison 7 The
walled contact process development was accom­
photoresist t h i ckness was i ncreased to 3.4 11-m , and
pl ished
voltage,
a focus offset was implemented during exposure to
tritl uoromethane: ox�·gen ( CHF,:02) gas ratio, and
slope the photoresist profile prior to etch . The opti­
by
experiment i ng
with
bias
p ressu re. " The substrate bias (to control photo­
resist, oxide, and CoSi 2 etch rates) and the ratio of
CHF5 :o2 flows
54
(to control sidewall p rofile and photo-
m i zed p h o toli thography process resulted in a pre­
etch photoresist profile of 80 degrees :±: 3 degrees.
A m u ltiple-step, tapered etch process was devel-
Vol. 4 No. 2
SjJring 1992
Digital Teclmicaljournal
..
CMOS-4 Back-end Process Developmentfor a VLSI 0. 75-pm Triple-level interconnection Technology
oped in which step 1 etched 40 percent of the con­
Figure 3 depicts this progression. Step 6 is a final
tact depth anisotropically to maintain final contact
clean-up ste p . The optimized M3C process demon­
critical dimension control. Steps 2 and 4 are photo­
strates a post-etch contact slope of 65 degrees
resist pul lback steps. These critical etch steps were
:±: 5 degrees, which consistently results in metal step
optimized by running a series of designed experi­
coverage exceeding the 0.30 -/Lm requirements.
ments to characterize the responses of photoresist
Measurements obtained by SEM are compared in
etch rate, uniformity, and lateral-to-vertical erosion
Figure 4.
rates. Etch rates were determined using patterned
oxide test wafers which were cross -sectioned and
Physical Vapor Deposition Metallization
analyzed using the scanning electron microscope
(SEM). Substrate bias and oxygen flow were the
primary parameters contro l l ing lateral-to-vertical
photoresist erosion rates. Steps 3 and 5 transfer the
tapered photoresist profiles into the d ielectric film .
The PVD metal lization processes are p erformed in a
single-wafer, multichamber, high-vacuu m spu tter
deposition system. The chambers include a radio fre­
quency (RF) etch , titanium, cobalt, and aluminum­
copper cathodes. The RF etch provides a sputter
PHOTORESIST LOSS
DURING ANISOTROPIC
STEP 1 ETCH
-----t· ------..\
=-=--=- ::... ::_-_.....
PHOTORESIST
(- 3.4 �m)
PHOTORESIST
�
t
t 0.4
X
M2/M3 DIELECTRIC
KEY:
PRE-ETCH PHOTORESIST PROFILE (-80°)
•
t
X
M2/M3 DIELECTRIC
ORIGINAL PHOTORESIST PROFILE
PHOTORESIST PROFILE AFTER STEP ANISOTROPIC OXIDE ETCH
PHOTORESIST PROFILE AFTER STEP 2 PHOTORESIST PULLBACK STEP
PHOTORESIST PROFILE AFTER STEP 3 ANISOTROPIC OXIDE ETCH
KEY:
81
1
(b) After Photoresist Erosion Etch
(a) After Etching of Oxide
r;;_
�
)
/:
�::1
���i���SIST
PROFILE
e,�---_- ���;����1ST
�= .
MVM3 DmCOR"
KEY:
e, PRE-ETCH PHOTORESIST PROFILE (-80°)
POST-ETCH CONTACT PROFILE (-65°)
82
(c) Final Contact Profile
Figure 3
Digital Technical journal
Schematic Drawing Showing Metal 3 Con tact Tapered Process
Vol. 4 No. 2
!c>pring
1992
55
Semiconductor Technologies
0.90
0.80
0
0
0.70
E'
2
� 0.50
<>::
a:
>
0
(_) 040
()._
w
,0
w
Ien
til
4
o ' I
0
to.
0.30
0
!::.
0
0 0
0
D. D. D.
t;,
I
I
6
2/
tl'
0
!::.
8
0
to.
b
0
0
0
0
6.
a
0
t;,
8
£j
\11
8
!::.
CJ
!::.
�
tJ
8
0
0.20 .
0
0
0.60
w
0
Q
0
0
0
0.10 i
0.00
1
KEY:
0
0
D.
Figure 4
3
0
OLD
5
3
5
7
9
LOT/WA FER NUMBER
11
NEW
13
15
17
19
M3 DEVICE STEP COVERAGE MEAN (�m)
0 M3 HIGH STEP COVERAGE MEAN (�m)
D.
M3 LOW STEP COVERAGE MEAN (�m)
M I N I M U M SPECIFICATION LIMIT (0.30 IJm)
Compuriso11 of/11etal 3 Step Coverage with Standard and Optimized Contact Tapers
cleaning of t h e wafer to r�move any native oxide
To lower the nu mber of defects, Digital process
from the wafer surface before the film is deposited .
engineers worked with the equipment vendor to
The titanium target is used to reactively form the
design and develop a new cathode. The CM05-3 cath­
TiN film that is used fo r local interconnect, tungsten
ode used fo r both the TiN and A.l films was a m ag­
adhesion layers, antiretlect ive coatings, and a fuse
netron configu ra tion, d�signed to enhance the
l i n k . The cobalt target is used fo r s i l icide for mation
deposition rate of the target material by creating
on gate and source/drain regions. The AI: 1 %Cu a l loy
additional bombarding ions. The magnetron has a
is used fo r three levels of on-chip low-resistance
fixed set of magnets oriented to confine the ion iz­
interconnect.
i ng electrons and thus cause the target to erode in
The reactive sputt�ring of the TiN film has pro­
a racetrack p a t tern . This confinement results i n
vided a number of chal lenges in the ar�a of particle
re-deposition on areas o f t h e target that are n o t
control. Early model ing in the 0105-4 development
spu t tered . These areas become a major particle
cycle using the C\lOS-:3 yield model predict�d the
source. The new design is based on a set of rotating
impact of the TiN particle levels o n the n:duced
magnets that move the erosion pat tern over the
m�tal pi tch areas of the C.\!OS-4 proc�ss. The model
entire surface of the target and keep the surface
ind icated that the number of defects p�r 100 mct�rs
free of any material b u i ld-up. Because of this
of interconnect would have to he reduc�d from the
enhanced spu t ter u niform ity, the rotating configu­
level of 20 defects added . To prevent the Ti
film
ration mainta ins a low particle count t h roughout
and associated interconnect from being among the
the life of the target . The conventional magnetron
top yield l i miters, a level of less than 5 d efects had
and rotating magnetron defect density levels are
to be attained.
compared i n Figure S B
56
Vol. 4 No. 2
Spring 1992
Digital Tcch11icaljournal
CMOS-4 Back-end Process Developmentfor a VLSJ 0.75-p..m Triple-level Interconnection Technology
I
(])
r-
z
:::l
>a:
<>:
a:
r-
co
a:
�
>-
rVi
z
w
0
r-
(.,)
w
LL
w
0
0
O DD
25
20
0
I
>-
J
15
0.99
6
0.95
6
0
0
0
0
6
r- 0.90
::::i
co
<>:
en
0 0.70
a:
0,_
w 0.50
10
>
i=
Q
5
0
0
D--El
0
Figure 5
:::l
:2
:::l
(.,)
0
50
KEY:
:5
.P
1 00
1 50
200
250
300
WAFERS PROCESSED
350
400
STANDARD CATHODE
Cathode Configuration as Compared
to Deject Density
Titanium-nitride Film The TiN process required
re-optimization to characterize the rotating cath­
ode and accommodate the CMOs-4 application of
TiN as a tungsten plug adhesive layer. In addition
to lowered particle levels, the process focused on
achieving low contact resistance between the vari­
ous metallurgical interfaces. Screening studies
were performed to characterize adhesion layer
properties for M 1 C and M2C via plugs. While minor
effects on via resistance were observed for some of
the factors, the single most important factor was
the presence or absence of a titanium u nderlayer.
With as l ittle as 15 nm of titanium beneath the TiN,
low via resistance was obtained . Without titanium,
no conditions resulted in acceptably low resis­
tance. These resu lts are shown in Figure 6, which is
a cumu lative probability plot of via resistance for
three adhesion layer processes: ( 1 ) 120 nm of TiN
deposited using conditions acceptable at the M l C
level, (2) 120 n m of TiN deposited using a modified
deposition process, and (3) a film with 40 nm of tita­
nium beneath 80 nm of TiN.
The importance of overall process integration
became apparent when it was determined that the
rotating cathode was damaging transistor charac­
teristics during spu ttering of the local interconnect
TiN. The TiN deposition process was modified to
decrease the substrate bias, and the new process
recipe was characterized and retrofitted into the
Vol. 4 No. 2
0.10
0.05
6
6
0.01 6
0
0
0 ROTARY MAGNET CATHODE
Digital Technical journal
0.30
Spring 1992
2
KEY:
o
,
6
4
6
8
10
M2/M1 VIA RESISTANCE (OHMNIA)
12
1 200
A TITANIUM NITR IDE
A TITANIUM NITRIDE "IMPROVED"
400 A TITANIUM/800 A TITANIUM NITRIDE
1 200
Figure 6
Titanium Nitride Adhesive Layer
Optimization
contact and via levels.9 This modification resulted
in a more manu facturable single recipe to maintain
for all CMOS-4 TiN levels.
Deposition of the Al interconnect film for CMOS-4
required l ittle alteration from the CMOS-3 recipes.
Film thicknesses were reduced by 500 angstroms
(A) for metal 1 and metal 2 to maintain aspect
ratios of no greater than 1 . Metal 3 film thickness
remained the same. The incorporation of the tu ng­
sten plug process into CMOS-4 technology simpli­
fied the metal step-coverage requ irements because
filling the contact with tungsten reduced the effect
aspect ratio significantly. The metal 3 contacts used
a tapered non-tungsten process and therefore
required that the high step-coverage process devel­
oped for CMOS-3 technology be maintained and fur­
ther improved for the CMOs-4 process.
The PVD Al step-coverage process is a three-step
process, optimized for wafer temperature, aspect
ratio, and the u nderlying material. The first step is
a low-temperature deposition of a nucleation layer
that provides a continuous coating over all sur­
faces. Step 2 uses a low-power temperature ramp to
a llow enhanced surface mobi l ity. Controlling the
time during this step increases the average distance
57
Semiconductor Technologies
that material can move along the surface and into
pose a problem for blanket plug techno logy (see
the contacts. A h igh- temperature, high-power third
Figure 7b).
step is used to reach the final film thickness. 10
Development efforts provided a manufacturable,
Although selective tungsten is the simpler of the
two tungsten plug formation schemes, it proved
high-throughput, and high-yielding meta l l i zation
not to be manufacturable i n the CMOS-4 develop­
process. A modified TiN cathode made the existing
ment time frame. Our evaluations demonstrated
equipment set serviceable for another technology
a general inability to reproducibly generate low­
generation. The use of the historical equipment data­
resistance interconnections while simu ltaneously
base minimized development time for the new tech­
controlling selectivity loss. As a consequence, blan­
nology, and increased the level of understanding for
ket tungsten plugs were chosen for use in the
a deposition technology with known benefits. The
CMOs-4 process.
incremental nature of the changes from the CMOS-3
to the CMOS-4 process al lowed an efficient technol­
ogy transfer due t o a reduced number of learning
cycles and the use of a fami l iar equipment set.
Blanket Tungsten Plugs Process
Development
As stated previousl y, new process technologies
were developed to meet the CMOS-4 process crite­
ria . Blanket tungsten plugs are an example of a new
technology As Figure 1 il lustrates, blanket tungsten
plugs are used in the CMOS-4 techno logy to verti­
cally interconnect metal 1 to the sil icon substrate
or t o po lys i l icon (i . e . , contacts), as well as to verti­
cally interconnect metal 1 and metal (i. e . , vias).
Blanket tungsten plugs were selected fo r use in the
CMOS-4 technology because they minim ize spatial
requ irements fo r contacts and vias by al lowing
the use of vertical-wal l openings rather than the
tapered openings used in earlier technologies.
At the outset of the CMOS-4 process development
cycle, two tungsten plug technologies, selective
and blanket, were evaluated. Al though these plug
technologies are very similar in their final structural
fo rm, their formation i nvolves important d iffer­
ences. As shown in Figure 7a, select ive tungsten
plugs are formed by the deposition of tungsten only
on conductive surfaces and not on the surrounding
oxide surfaces. Such growth leads to a fi l l ing of
openings from the bottom u p . In comparison, blan­
ket tungsten plugs, as shown i n Figure 7 b , are
formed by a three-step process:
1 . Sputter deposition of an adhesion layer
2. CVD of a conformal blanket tungsten film
3. Reactive ion etch back of the tungsten and adhe­
sion layer to leave tungsten plugs surrounded
by the adhesion layer
Blanket tungsten plug processing thus fiJ is open­
Blanket Tungsten Plug Requirements
Tungsten plugs and their processing have simple
structural and electrical requ irements. Structural ly,
the plugs must be free of voids and flush with
the surrounding oxide surface after all u ndesired
tungsten and adhesion layer residues h ave been
etched from the top surface of the entire wafer.
The void-free constraint ensures that potentially
damaging process materials are not trapped in
voids. The requirement that the plug be flush with
its surrou nding surface ensures good step coverage
during the subsequent deposition of AJ. The photo­
micrographs in Figure 8 i l l u strate these structural
a t tributes. Electrical ly, the plug resistivity must be
high enough to induce current spreading in the
plug, but the i nterfacial resistance between the
plug and other conducting materials must be low
enough not to adversely affect circuit performance .
In addition, the processing of the plugs must not
induce device damage. The latter point refers to the
res u l ts of early investigations of tungsten plugs,
which indicated that device clamage cou l d occur as
a result of the attack of underlying materials during
the CVD process of tungsten. 1 1 · 12
Final ly, in addition to t h e structural and electrical
objectives, the plug for mation process had to be
optimized from a cost perspective. Cost was of par­
ticu l a r concern because the indust ry-wide blanket
tungsten deposition method of choice a t the t ime
(low-pressure CVD) involved very low depo sition
rates and made very inefficient use of the expens ive
source gas, tungsten hexafl uoride (WF6) .
Blanket Tungsten Plug Formation­
Equipment and Process
An Applied Materials Precision 5000 tungsten sys­
tem was used fo r tungsten plug processing. The
Precision 5000 system performs both tungsten
ings from the sides as wel l as from the bottom up.
deposition and etch back withou t breaking vac­
As a resu l t , variable contact and via depths do not
uum. Processing involves first loading a wafer i n to
58
Vol. 4 No. 2
Spring 1992
Digital TechuicalJournal
CMOS- 4 Back-end Process Developmentfor a VLSI 0. 75-/J-m Triple-level Interconnection Technology
CONTACT PATTERNING
TITANIUM -­
NITRIDE!
TITANIUM
ADHESION LAYER DEPOSITION
i n+
p
TUNGSTEN
TITANIUM
"-.
NITRI DE/ "'-.
TITANIUM
TUNGSTEN DEPOSITION
ETCH BACK
(b) Blanket Tungsten
Plug Formation
(a) Selective Tungsten
Plug Formation
Figure 7
Comparison ofPlug Technology
the deposition chamber. Then a nucleation layer
approximately 50 nm thick is deposited at approxi­
mately 475 degrees Celsius by the silane reduction
of WF6:
SiH4 + WF6 � W + SiF4
(1)
The bulk of the tungsten layer, approximately
800 run, is then deposited using the hydrogen
reduction ofWF6:
(2)
Hydrogen reduction chemistry is used fo r the
bulk of tungsten deposition because it yields good
step coverage, whereas silane reduction does not. l3
However, silane reduction chemistry is used to initi­
ate tungsten growth because hydrogen reduction
chemistry i nvolves an incubation period before film
deposition begins on TiN 14 , a step not required by
silane reduction. 's
Digital Technical journal
Vol. 4 No. 2
Spring 1992
Following tungsten deposition, the wafer is
raised to expose its backside, and a short nitrogen
trifluoride (NF3) plasma etch is then performed in
this same chamber to remove the small amount of
tu ngsten that deposits on the backside edges of
the wafer. The wafer is then transferred to a cham­
ber in which a two-part etch back is performed.
The bulk tungsten film is first etched in a sulfur
hexafluoride/argon (SF6/Ar) plasma according to
equation (3):
(3)
until an optical emission from nitrogen, which has
been l iberated from the underlying adhesion layer,
is detected. At this point, a chlorine/argon (Clz!Ar)
plasma is used to remove any remaining adhesion
layer according to equation (4):
Tii'l/Ti + Cl2 � TiCI4 + N 2
(4)
59
Semiconductor Technologies
TUNGSTEN
PLUG RECESS
TRENCHING OF THE
ADHESION LAYER
TUNGSTEN
PLUG VOID
VOID-FREE FLUSH PLUG
Figure 8
Photomicrographs Showing Several Possible Plug Structural Attributes
Both etching steps may employ a rotating mag­
over the ranges given in Table 2. Four factors were
netic field , which serves to improve the u n iformity
identified to have an impact on the responses o f
of the etching.
i nterest (i.e . , t h e properties listed above).
Blanket Tungsten Deposition
fou r factors was next pe rformed i n the context
A detailed determ ination of the effect of these
The more important propert ies associated with the
tu ngsten deposition process include:
•
Thickness u niformity
•
Deposition rate
•
Film resistivity
•
Step coverage
•
Film st ress
•
Su rface smoothness
•
Table 2
Process Factors a n d Ra nges Used
i n the Blanket Tungsten Deposition
Screening Study
Factor
Spacing
Range
t
200 - 600 m i l s
Susceptor tem peratu re t
400- 475°C
Total pres s u re
6 0 - 80 torr
Partial pressu re, H2t
t
1 6 - 32 torr
1 -1 .9 torr
Tungsten hexafluoride conversion
Partial pressu re, WF6
D ur i ng development of the blanket tungsten
Partial pressu re, carrier gas
1 . 9- 7.2 torr
Backside pu rge, N 2
300- 700 sccm
deposition process, these propert ies were studied
ami "globally" op timized with respect to the plug
Notes:
objectives. A screening study was performed, fol­
'Table 2 is adapted from data p u b lished in reference 1 6 .
t
l n d icates factors found to have t h e g reatest i m pact on
Table 2 shows the resu lts of the screening study of
:J:
lowed by a response surface modeli ng (RSM) study.
these properties. Seven process factors were varied
i:
responses of interest.
scc m � standard cubic centi meters per m i nute
Vol. 1 No. 2
Spring 1992
D igital Technical journal
CMOS-4 Back-end Process Developmentfor a VLSI 0. 75-p.m Triple-level Interconnection Technology
of an RSM study. The process factors and ranges
used for this study are given in Table 3. RSM stud ies
produce mathematical models that relate factors
and responses of interest. In this study, seven differ­
ent models, one for each response, were generated.
The models were then used to search for po ints of
interest either computationally or graphical ly. An
example of a set of contour plots used fo r a graphi­
cal search for tungsten step coverage and sheet­
resistance uniformity (used to mon itor thickness
un iformity) is shown in Figure 9.
Ta ble 3
Process Factors a n d Ra nges Used
i n the Blanket Tungsten Deposition
RSM Study
Factor
Range
Preferred
Settings t
Spacing
200 - 600 m i ls
-400 m i ls
Susceptor temperat u re
430 - 490°C
475°C
Partial pressu re, H2
6 - 30 torr
-1 8 torr
Partial pressu re, WF6
1 - 2 torr
-1 .75 torr
Notes:
*Table 3 is adapted from data p u b l i shed in reference 1 7 .
t
Aiso shown are the preferred deposition cond itions that
satisfy the optimization criteria shown in Table
4.
The specific criteria for the tungsten deposition
optimization search are given in Table 4. The seven
mathematical models/contour plots described
above were used to find a region of the factor space
where all of the optimization criteria cou ld be
met simultaneously. The preferred deposition con­
ditions associated with this region are shown in
Table 3. The models also ind icated a relatively large
process range within which the optimization objec­
tives could be met. This range is important data fo r
a production process.
Contour plots for deposition rate , resistivity,
stress, WF6 conversion, and reflectance (used to
monitor surface smoothness) are not included
i n this paper. However, evaluat ion of these plots
showed that the relevant optimization criteria
cou ld be met throughout the factor space studied.
As a result, these responses did not impose con­
straints on the selection of the optimized process.
The criteria set for deposition rate and WF6 con­
version corresponded to values significantly higher
(by a factor of approximately 3 to 10) than those
typical of blanket tungsten processes at the time of
our stu dy. 1 ' · 1 6· 17 The improvement resu lted mainly
from t he use of higher deposition pressures com-
Digital Tee/mica/ jou1'Twl
lirJI. .j So. 2
Spring 1992
pared to those previously used (80 torr versus
less than 1 torr). Higher pressure also improved
smoothness of the film, as seen in Figure 10. A
smoother film is important because roughness on
the tungsten film can be transferred into the under­
lying oxide during tu ngsten etch back.
The optimization criterion for film stress was
set at a level corresponding to a mechanical ly stable
film (i.e., one that would not peel spontaneously).
Thus, altl1ough the stress values obtained are rela­
tively high, they are below the critical level associ­
ated with delamination. For the tungsten resistivity,
the observed values ranged from approximately
7.7 to 10. 5 p.. o hm per centimeter (em) and were all
acceptable.
Unlike the other optim ization criteria, those for
step coverage and sheet-resistance unifo rmity were
not met throughout the factor space stud ied (see
Figure 9) . Tungsten step coverage can directly
impact void fo rmation, and tu ngsten thickness uni­
fo rmity can impact plug recess control . Figure ll
i l l ustrates how thickness variation across a wafer
can lead to variations in plug recess fol lowing etch
back. Because of the importance attached to meet­
ing these two optimization criteria, the allowed
process window was diminished i n size. Figure 9
shows that a step coverage greater than or equal to
95 percent restricted the WF6 partial pressure to
approximately greater than or equal to 1 .5 torr,
hydrogen (H 2) partial pressure to approximately
less than or equal to 18 torr, and gas-inlet-to-wafer
spacing to less than or equal to 400 mils. A sheet­
resistance uniformity less than or equal to 3 per­
cent served to further restrict the spacing to values
between approximately 400 and 300 mils.
In addition to the tungsten deposition process
properties mentioned, the tu ngsten thickness had
to be optimized. The upper l imit on thickness was
influenced by cost considerations ancl by the fact
that a thinner tungsten deposit has less likelihood
of being trapped in d ielectric troughs. A thinner
tungsten deposit also requires less ove retch to
remove tungsten spacers that may fo rm on the
trough sidewa lls. As shown in Figure 12, the size of
the spacer formed on a nonplanar dielectric fo r
zero overetch with a nonisotropic etch clepencls on
the absolute magnitude of the deposit thickness, r,1 ,
and the worst-case dielectric sidewal l slope, a ,
which together determine the local tu ngsten th ick­
ness range, T11 - T,1 .
The lower limit on thickness was influenced
by the need to fi l l contact openings with tungsten
61
Semiconductor Technologies
2.00
98
\
era:
\
I
I
30
'
I
\
\:,
\
I
1 .75
cr- 24
a:
0
!::UJ
a:
0
!::UJ
a:
:::> 1 .50
(/)
(/)
UJ
a:
CL
18
:::>
(/)
(/)
UJ
a:
a.
u_«> 1 .25
3:
"'
r. 1 2
92
445
475
460
490
\
I
I
\
\
\
\
\
\
I
\
I
1 1 .5
\1
I
I
I
I
I
I 2
I
\
95
I
\
\
\
\
445
TEMPERATURE ('C)
600 \
\
96.5
460
I
\
cr-
1 1 .5
a:
I
I
/ 96.5 /
/I /
I
I
I
I
/
/
/
/
'5
'f,__5_ - - - - -
\
/
/
/
460
'
96.5
"'
r. 1 2
�Q. - - - - _4
98
'
' , 2.5
18
(/)
(/)
UJ
a:
a.
�-­
/
445
24
g
UJ
95
,
490
\\ ��',
\
,'
475
TEMPERATURE ('C)
1------
_ _ _ _
475
490
1 .25
1 .5
175
2
WF6 PRESSURE (TORR)
TEMPERATURE ('C)
600
600
I
\ �
n��·
I
/
/
500
93.5
en
_J
�
96.5
98
UJ
(.)
4:
a.
(/)
400
3
_______
300
_
4 --1 .25
1 .5
1 .75
WF6 PRESS URE (TORR)
KEY:
Figure 9
12
18
24
30
H2 PRESSURE �ORR)
TUNGSTEN STEP COVERAGE, PERCENT INTO TRENCHES 1 . 5 �m WIDE BY 1 .0 �m DIAMETER
TUNGSTEN SHEET-RESISTANCE U N I FORMITY. 1cr EXPRESSED AS A PERCENT OF THE
AVERAGE OF 49 SITE MEASUREMENTS
Contour Plots of Tungsten Step Coverage and Sheet-resistance Uniformity
(Reprintedfrorn tbe Journal of Vacuum Science Technology; see reference 1 7.)
62
Vol. 4 No. 2
Spring 1992
Digital Technica l jourr.al
CMOS-4 Back-end Process Developmentfor a VLSI 0. 75-1-Lm Triple-level Interconnection Technology
Ta ble 4
Tu ngsten Deposition-related
Opti m i zation Criteria for a Blanket
Tu ngsten Plug Appl ication
Parameter
Optimization Criteria
Growth rate
;:;: 300 nm per min
•
Tungsten etch rate
•
Tungsten etch-rate u n ifo rmity
•
Tungsten/titanium nitride etch-rate ratio
•
Isotropy (lateral etch rate/vertical etch rate)
•
Microloadi ng (tungsten pl ug/tungsten bu l k etch­
Resistivity
;:;: -8 J..t O hm-cm
Sheet-resistance
un iformity
� 3%
Te nsile stress
,;; 1 6 X 1 09 dyne per cm
Step coverage t
� 95%
WF6 conversion
;:;: 1 2 %
ratio, etch isotropy (lateral etch rate versus vertical
Reflectance
;:;: 25% vs Si @436 nm
etch rate), and microloading (tungsten plug etch
rate ratio)
(a)
2
rate u n ifor m ity, tungsten/titanium nitride etch-rate
rate/bu l k tu ngsten etch rate) all impact the ability
Notes:
*Table
For b u l k tu ngsten etch, the tungsten etch rate
impacts throughput and cost. The t ungsten etch­
4 i s reprinted from the Journal of Vacuum Science
Technology; see reference 1 7.
t
step coverage for a trench 1 . 5 �-tm deep and
1 .0 �-tm wide.
to produce a residue-free surface, while simu lta ne­
ously m a i n taining flush plugs. A nonuniform etch
rate affects plug recess because it leads to the clear­
i ng of o n e region of the wafer before the rest. This
prior to etch back, and further, to planarize the
tungsten over the opening to m i n i m ize plug recess
fol lowing etch back, as shown in Figure 13. Tung­
sten thickness stud ies showed that a tungsten film
greater than approxi mately 650 nm is required for
the subm icron-level contacts and v ias of CMOS-4
technology.
first-to-clear region becomes the s i te of the worst­
case plug recess on a wafer. Figure 14 shows the
plug recess that occurs in high etch-rate regions on
a wafer when a tungsten film of un iform thickness
has been etched to the proper end point for the
lowest etch-rate regions.
The tungsten etch isotropy affects plug recess by
i ncreasing the etch time required to clear t he final
residues on a nonplanarized dielectric. Figure 1 5
Tungsten Etch Back
i l lustrates that an etch isotropy that equals 1 (i.e . ,
In this sectio n , the tungsten a n d adhesion layer
lateral etch rate equals vertical etch rate) can
etch-back processes are discussed.
lead to uniform clearing of tungsten on n o n planar
Bulk Tungsten Etch Chemistry (SF6 /Ar)
su rfaces, and that an etch isotropy less than 1
The more
tends to produce spacers that m ust be removed by
important etch-back process properties for bulk
overetching. The additional time required to clear
tungsten etch i nclude:
tungsten spacer residues leads to increased plug
(a) Deposited at 80 Torr
Figure 10
Digital Technical journal
(b) Deposited at Less Than 1 Torr
Photomicrographs Showing Surface Smoothness of Timgsten films
Vol. 4 No. 2
-
Spring 1992
63
Semiconductor Teclu10logies
�
v
TUNGSTEN
/
SI LICON
DIOXIDE
v
TITANIUM/
TITAN I U M NITRIDE
TUNGSTEN
ratio in S FcJAr using only the factors in Table 5,
while simu ltaneously maintaining high tungsten
etch rates (greater than 500 nm per minu tt:), were
largely unsuccessfu(,lV The desired improvement
was eventually obtained through the incorporation
of active wafer temperature control. Contro l l ing
the wafer temperature between 20 and 40 degrees
Celsius improved the tungsten/titanium nitride
(a) After Tungsten Deposition
RECESSED PLUG
I I
FLUSH PLUG
(b) After Etch Back
Figure
Plug Recess as a Function
of Thickness Variation
II
etch-rate ratios from approximately 2:1 to between
10: 1 and 50: I, respectively. 20
Adhesion Layer Etch Chemistry (C/2/Ar)
The
more important etch-back process properties for
adhesion layer etch include:
•
TiN etch rate
•
TiN etch-rate u n iformity
•
TiN / T u ngsten ( W ) etch-rate ratio
•
TiN/oxide etch-rate ratio
For adhesion layer etch chemistry in Cl /Ar, the
tit;m i u m nitride etch rate impacts throughput
and cost, while etch-rate uniformity affects the
recess. Final ly, a high tungsten/titanium nitride
level of adhesion layer undercu t t i ng or trenching
etch-rate ratio affects plug recess control indirectly
that occurs around a plug (see Figure Sc). The otlwr
by preventing the l i beration of oxygen from
two etch properties impact plug recess and oxide
the underlying d iele.ctric. Oxygen l i beration has
loss. Consequent ly, a process is derived that etches
bet:n shown t o greatly increase the tungsten micro­
TiN at a high and uniform rate while it simu l­
loading factor. IH
taneously and s lowly etches tungsten and si licon
Development of the b u l k tungsten etch in SFr, /Ar
began with an RSM study of the tungsten etch rate
oxide.
and etch-rate u n iformity.19 Table 5 shows the fac­
in CI/Ar began with an RSM study of the TiN etch
Proct:ss development for the adhesion layer etch
tors and ranges used in the initial stu dy, along with
rate and etch-rate uniformity. 1Y Table 5 shows the
the preferred etch conditions identified . At tempts
factors and rang<:s used in the initial study, along
to improve the tungsten/titanium nitride t:tch-rate
with the preferred
etch cond itions identified.
Cos o: =
�
Tn
SILICON
Iigure 12
64
Photomicrograph and Schematic Drawing of a Tungsten Spacer
Vul. · i
No
2
Spriug 1992
Digital Technicaljourual
CMOS-4 Back-end Process Development for a VLSJ 0. 75-1'-m Triple-level interconnection Technology
NONPLANAR TUNGSTEN SURFACE
PLANAR TUNGSTEN SURFACE
OXIDE
TUN G STE N
1
/
�
OXIDE
TITANIUM NITRIDE!
TITANIU M
TUNGSTEN
./ TITANIUM NITRI DE/
./ TITANIUM
(a) After Tungsten Deposition
NONPLANAR TUNGSTEN PLUG SU RFACE
PLANAR TUNGSTEN PLUG SURFACE
OXIDE
OXIDE
TUNGSTEN
TUNGSTEN
(b) After Tungsten Etch Back
Figu1·e 13
Nonplanarity in a Blanket Tungsten Deposit Transferred into Plug Recess
(Reprintedfrom the Journal of Vacuum Science Technology; see reference 1 7.)
HIGH ETCH-RATE REGION
LOW ETCH-RATE REGION
mities than those for the tungsten etch step can be
tolerated. Wafer temperature control provides addi­
tional latitude against plug sidewall trenching.
TUNGSTEN
S I L I CON
DIOXIDE
/ TI TAN IUM/
TITANIUM
NITRIDE
TUNGSTEN
fl<"
Integration of Tungsten Plugs into
CMOS-4 Technology
After the tungsten deposition and etch-back pro­
cesses were developed, the overa l l plug formation
process was i ntegrated into the CMOS-4 technology.
(a) After Tungsten Deposition
To ensure that electrical requirements were met
and that adequate process latitude existed, the fo l­
lowing factors were considered in the integration
) ] I I
(b) After Etch Back
Figure 14
Plug Recess in High Etch-rate Regions
on a Wafer
studies:
•
Dielectric pl anarization
•
CoSi thickness
2
•
Contact depth
•
Contact overetch
•
Adhesion layer
- Sputter etch preclean
- Deposition temperature
- Substrate bias
Acceptable Ti N etch rates (approxim ately 145 nm
per minu te) and etch-rate uniformity (less than or
equal to ::!::: 5 percent , 1 a) were achieved. Since the
- Th ickness
- Material (TiN or TiN/Ti)
After the integration studies were completed,
adhesion layer is only approximately 120 om th ick,
a relatively robust process was developed. At that
lower etch rates and higher etch-rate nonunifor-
time, it was determined that acceptable electrical
D·igitnf 1ecbnicnf journal
H1/. 4 No. 2
SjJring 1992
65
Semiconductor Technologies
SILICON
(a) After Tu11gsten Deposition
�
LATERAL
ETCH RATE
:
"OTRO Y . '
r
/
I
RESIDUE-FREE
SURFACE
I' _ M E T L
;
-
�
1
I
\
'-\
LATERAL
ETCH RATE
�i��'CAL
ISOTRO P
RATE
<
OXIDE
1
VERTICAL
ETCH
RATE
OXIDE
� �
----------------s-IL-Ic_o_N----------------�
�
�----------------sI LIc
_
o
_-N --------------�
(b) After Etch Back
Figure 15
Comparison of Results of Tungsten Etch Jsotropies
resu lts can he ohtainl:d over relatively large process
metal
ranges.
include low moisture adsorption, low stress for
concentration.
Mechanical
requirements
cracking resistance, low particu late levels, and a
Dielectric Process Development
The CMOS-4 process presented two specific techno­
low pinhole density. Important deposition require­
ments include high deposition rate, good un ifor­
logical chaLlenges requiring dielectric development
mity, and low deposition temperature (not greater
efforts. First, horizontal scal ing without reducing
metal thicknesses resulted in high aspect ratio
h i l lock formation.
spaces. The existing dielectric technology could
than 450 degrees Celsius) for prevention of metal
not fi ll these spaces void-free. Second, the introduc­
Gap Filling
tion of blanket tungsten plugs required the develop­
The conventional silane-based oxides could not
ment of a dielectric planarization process.
meet the CMOS-4 technology gap fill requirements
In addition to meeting gap fi l l and planarization
due to poor conformality characteristics. A silane
requirements, a d ielectric film must also meet a
oxide profile is typical ly described as a " bread l oaf,"
nu mber of electrica l, mechanical, and deposition
that is, the film is thicker on the top of a structure,
requirements. The required dectrical characteris­
but thinner on the sides and bottom, and forms
tics fo r a dielectric film include low current leakage,
cusps along the sidewalP1 Silane oxides nucleate in
high breakdown voltage, high electrical resistance,
low dielectric constant, low mobile ion, and heavy
profile and l imits the aspect ratio (height/width)
66
the gas phase, which causes this reentrant type
Vol. 4 No. 2
.\jHing 19')2
Digital Technical journal
CMOS-4 Back-end Process Development for a VLSJ 0. 75-/J-m Triple-level Interconnection Technology
Table 5
Process Fa ctors a n d Ranges for
the I n itial Tu ngsten Etch-ba ck
RSM Study and the Preferred
Etch Cond itions Found
Factor
Range
Preferred
Settings
stabil ity and less h i l lock formation. Mobi le ion and
heavy metal concentrations are lower with TEOS
oxides, and device testing ind icates lower defect
densities. 24 · 26
In the CMOS-4 process, fill ing gaps between
minimum-spaced metal l i nes was achieved with
Tungsten Etch
profile-al.tering techniques in conjunction with a
85 m i l litorr
SF6 flow rate
20 - 250 m i l litorr
t
20-60 sccm
Ar flow rate
1 5 - 60 seem
60 seem
RF power
200 - 500 watt
475 watt
tooL Deposition of a planarized dielectric was com­
Magnetic field
0-1 00 gauss
20 gauss
pleted in one cassette-to-cassette operation.
Electrode
temperature
60°C
60°C
Gap Fill Process Flow
5 0 - 250 m i l l itorr
85 m i l l i torr
Total pressu re
60 seem
Total pressu re
C l2 flow rate
5 - 35 seem
1 0 seem
Ar flow rate
50-1 20 seem
1 1 5 seem
RF power
1 50 watt
1 50 watt
Magnetic field
0 -1 00 gauss
75 gauss
Electrode
temperat u re
60°C
60°C
Notes:
*Table
PE-TEOS b u l k dielectric . Wafers were moved back
and fo rth between a series of deposition anu etch­
back steps in a load-locked, mu ltichamber cluster
The gap fi l l process is shown in Figure 1 6 . It begins
TiN Etch
5 is reprinted with permission from the IHS
1 9.
P u b l i s h i n g Group; s e e reference
t
sition temperatures, TEOS oxides exhibit thermal
seem = standard cubic centi meters per minute
with a PE-TEOS conformal deposition that is halted
prior to reaching a thickness that would fil l the
smallest gaps. Next, an argon sputter etch is per­
formed on the oxide film deposited in the pre­
v ious step. The oxide removal rate is direction
dependent, with the maxi mum removal occurring
45 degrees from verticaL The original 90- degree
corners are beveled into positively tapered angles.
Then an ozone-TEOS film is deposited to completely
fiJJ the remaining gaps. Ozone-TEOS is formed by a
therm al reaction in which oxygen atoms are pro­
duced by the rapid decomposition of ozone. Ozone­
that can be fi l led without forming a void to approxi­
mately o.s . z z
To fill the aspect ratio of approx imately 1 .2
(0.9-!J-m height/075-/J-m width) for the CMOS- 4
process, a tetraethylorthosil icate (TEOS)-based oxide
was used . The conforma lity ofTEOS oxides is much
better than that of silane oxides. Because the organa­
silicon compounds produced d uring a TEOS-based
CVD process have a significantly higher su rface
mobi l i ty, the reactive molecules diffuse on the sur­
face before reacting, which resul ts in better step
coverage without cusps. 2-' TEOS-based oxides can fill
aspect ratios u p to 1 .0 void-free. When used in con­
junction with profile-altering deposition/etch-back
TEOS film is not desirable as a bulk d ielectric due to
its characteristically low density and tensi le stress.
However, the superior step-coverage characteris­
tics of ozone-TEOS al low it to fi l l very small gaps.
Fol lowing ozone-TEOS deposition,
a
second
profile-altering etch back is performed . The ozone­
TEOS is a sacrificial film that is removed in a CHF ,
chemistry u ntil it remains only in the small gaps
and as a spacer along the sidewa l l s of larger fea­
tures. The combined effect of sputter etching and
ozone-TEOS processing is a void-free surface with
positively sloped sidewa lls. Final ly, a b u l k PE-TEOS
dielectric film is deposited over this smoothed sur­
face to reach the desired dielectric film thickness.
techniques, as in CMOS-4 technology, aspect ratios
Planarization
u p to 1 .8 can be fil led 22 24.25
TEOS oxides are wel l suited for use as interlevel
Planarization is a dielectric smoothing process
that is performed to smooth or reduce the steps
dielectrics. Oxides fo rmed through the plasma dis­
created by u nderlying interconnect features. A pla­
sociation of oxygen (02 ) in the presence of TEOS
(PE-TEOS) are denser than silane oxides and there­
narization process minim izes reflective notching,
reduces the extent of metal overetch, increases the
fore more resilient to moisture adsorption. PE-TEOS
thickness of metal over u nderlying topography, and
films arc typical ly u nder low compressive stress,
redu ces interconnect clefect densities. Enhanced
which results in higher cracking resistance than
planarization is required to successfu l ly form tu ng­
the tensile-stressed silane films. Due to lower depo-
sten plugs with a blanket etch-back process. The
Digital Technical journal
Vol. 4 No. 2
Spring 1992
67
Semiconductor Technologies
TITAN IUM N I TRIDE
ANTI R E FLECTIVE COATING
METAL
(a) Initial PE-TEOS Deposition
PE-TEOS
OZONE-TEOS
TITANIUM N ITRIDE
ANTIRE FLECTIVE COAT I N G
--+---- /
(b) A rgon Sputter Etch and Ozone-TEOS Deposition
TITAN I U M NITR I D E
ANTIREFLECTIVE COATING
PE-TEOS
OZONE-TEOS
(c) Anisotropic Etch and PE-TEOS Cap
Figure 16
Schematic Drawing of and Photomicrograph of Gap-filling PE-TEOS Dielectric Deposition
Process Stages
slope of the su rface must be less than 30 degrees
involved depositing a sacrificial planarizing boron
from the horizontal to ensure removal of tungsten
oxide film and etching the planarized pattern into
stringers.
Two planarization techniques were considered
for use in the C:VIOS-'l process. The first technique
68
the oxide . The second technique was similar,
except a spin-on-glass (SOG) process was used as
the planarizing agent.
Vol. 4 No. 2
Spring 1992
Digital Tech-nical journal
CMOS-4 Back-end Process Development for a VLSI 0. 75-M m Triple-level In terconnection Technology
Initial ly, the boron oxide planarization was cho­
the topography. The isotropic etch back transfers
sen fo r use in the CMOS-4 process. This technique
the planarized surface into the underlying oxide.
was selected based on a perceived manufacturabil­
ity benefit over the SOG process . The boron oxide
The photomicrographs i n Figure 17 illustrate boron
oxide planarization. The boron oxide film is a sacri­
planarization can be done in a single cluster tool
ficial film t h at is completely removed d u ring this
that uses the same serial process as the gap fill pro­
step . The deposition/etch sequence is repeated to
cess. One cassette-to-cassette operation can both
further improve planarity. After all boron oxide
fill the minimum gap and planarize the surface. On
is removed, the etch chemistry is switched to a
the other hand, SOG planarization requ ires two sep­
h igher-rate carbon tetrafluoride (Cf4) anisotropic
arate depositions, a spin coat, a cure, and an etch
operation.
etch process that removes the thick bu l k deposi­
tion to the final desired thickness .
Boron Oxide Planarization
duction for the dielectric between polysil icon
Boron oxide is a cvo film deposited by the plasma
and metal 1 . For the metal 1 to metal 2 d ielectric,
The boron oxide process is used in volume pro­
decomposition of trimethylborate (TMB) in the
the boron oxide process cou lei not meet the wafer
presence of 02 . The film has a low melting point
volume and u n ifo rmity requirements due to prob­
and flows at deposition temperatures as low as
lems with equ ipment rel iabi l ity, thickness variabil­
400 degrees Celsius. Boron oxide can be etched
ity,
back in a CHF3 plasma chemistry and a 1 : 1 boron
oxide :oxide selectivity. Spaces up to 25 Mill can be
process.
and
low
throughpu t .
The
SOG
planariza­
tion scheme was selected as a more cost-effective
fu l ly planarizecl with boron oxide.27
Spin-on-glass Planarization
Boron Oxide Process Flow
The SOG process consists of a series of simple single­
Upon completion of the gap fi l l process flow,
step operations. The overall i ntegrated process
the final bulk cl ie.l ectric deposition step is targeted
characteristics exhibit good throughput and pro­
at approximately three times the desired final
cess control. The photomicrographs in Figure 1 8
thickness. This overdeposition provides some ini­
i l lustrate S O G etch-back planarization.
tial smoothing of the u nderlying interconnect fea­
SOG is a low-viscosity l iquid polymer that is
tures and also fills spaces in the 2- to 5-Mm range to
app lied to the wafer using a spin-coating process
eliminate formation of u nwanted gaps. Next, two
similar to that used for photoresists. The SOG m ate­
sequential boron oxide deposition and etch-back
rial used for CMOS-4 technology is a siloxane (methyl
steps are performed. The boron oxide film flows
group containing pol ymer). Siloxane SOGs ex hibit
as deposited, thereby smoothing and planarizing
improved cracking resistance and lower dielectric
(a) Minimum-spaced Metal Lines
Figure
Digital Teclmical journal
17
(b) Wide-spaced and Isolated Metal Lines
Photomicrographs Showing Boron Oxide Planari:zation
Vol. 4 No. 2
Spring 1992
69
Semiconductor Technologies
(b)
(a) Minim unz-sjJaced Metal fillf'S
Figure 18
Wide-spa ced aud Isolatect J'vletal Lines
Photont icrogmjJhs Showing SOC Etch-back Plcmarization
constants as compared to other ava i l able SO G mate­
Summary
gaps and planarizc topographical surfaces. The
is a t h r e e - lev e l a l u m i n u m a l loy metal l ization pro­
rials. Because it i s a l iquid , soc; can til l very s ma l l
material is solid ified i nto a glass by c u ri ng in a low­
temperature furnace cycle. The c u red SOG film has
Si02 - type mechanical and electrical properties, and
can t herefore be left behind as part or the b u l k
d ielectric. Typically, a partial e tc h back o f t h e mate­
rial i s performed t h a t leaves soc; onlv in th e gap
areas.2H·29
Digit al's
CMOS-4 on-chip intercon nect technology
c ess with p l a n a ri zed TEOS-based s i l icon d ioxide
.
d i electrics. tungsten -til lcd contacls and vias, ancl
a m i n i m u m feature s ize of 0.75 t..L m . The process
development goals required the m a x i m u m use of
the existing m a n u facturing capabi l i ty ancl
the
i nt roduction of new process features. for photo­
li thography, plasma etch, and PV D met a l l ization,
the 1 . 0 -J..L m manu facturing equipment set and pro­
SOG Process Flow
cesses were m o d i fied and reopt i m i zed for the
The i n i t i a l gap ti l l deposition is deposited thick
enough to provide a buffer for the so<.; etch­
back overetch. so< ; is then spu n on to fil l the
larger spaces and planarize the surface. A l o w
­
temperature fu rnace cure is performed to remove
submicron regime. In a d d i t i o n , two new process
features, a blanket CVD tungsten process ancl a
TEOS-based
oxide
planarization
process,
were
deve l oped and imp lemented in manu facturing to
meet the CMOS-4 technology requirements.
t h e s o l ven t s from the SOG and t ransform the m ate­
rial from a l i q u i d i nt o a g l a ss.
a
A partial et c h back of t he soc ; i s performecl usi_ng
C :HF/CT·j 02 chemistry.
The selec t i v i t y of SOG to
References and Note
1.
nology with Tungsten and Planarizatio n ,"
t he u nderlying PE-TFOS is t a rge t e d at l : I a nd is con­
t ro l led by t h e ratio of the
Proceedings
C I I F , and < >2 gas tlows.
of i n terconnect l i nes, but re m ain s in t h e gaps of the
(1988): 21 -2H.
2.
remove soc ; from locations at which contacts w i l l
70
YLSI
International ll:.'EF VLSI Multilevel Inter­
connection Co nfere nce (1990) : 149- 1 5 6.
contacts c a n lead to p roblems w i t h v i a '·poisoning"
ness and encapsulate t he soc ; .
in
I n terconnects," Proceedings of the Seventh
b e formed. E x p o s i ng SOG along t h e sidewa l l s o f
deposited to achieve the desired d ielectric t h ick­
A. Enver and J Clement, " Fin ite Element
Numerical Model ing of Cmrcnts
l a rger-spaced l i nes. The e t ch b a c k is t argeted to
a nd poor contact profiles. F i n a l ly, a PE-TEOS layer is
of th e Pijth International IEEE
Vl..\1 Multilel'el Interconnection Conference
Since the 02 flow also affects the etch uniformity,
t radc-offs between select ivity and uniformity were
n e c e s sa ry. SOG is et ched c om p le t e ly from t he t o p s
C . Kaanta et a l . , " S u b m icron Wiring Tech­
3.
P. C h ien and M . Chen, " Proximity Effects in
Submicron Optical L i thography," SP!E Pro­
ceedings, vo l . 772 (March 1987): 35 -40
Vol. "i No. 2
Sprillg /')')2
D igital Teclmical]ow-nal
CMOS-4 Back-end Process Development for a VLSI 0. 75-/J-m Triple-level Interconnection Technology
4.
5.
A . van Roosmalen, "Review: Dry Etching of
14.
(1984): 429-436
Layers for Tungsten Deposition," i n Tungsten
and Other Refractory Metalsfor VLSI Applica­
tions 11, edited by E. Broadbent (Pittsburgh, PA:
J. Chang, "Selective Reactive Ion Etching of
Sil icon
D ioxide," Solid State
Technology,
Materials Research Society, 1987): 187- 195.
C. Dobson, "Polymer Formation D uring Sil i­
15.
M u l t i level
S. Roth, W Ray, and G. Wissen, "In situ,
16.
burgh, PA, 1990): 167- 178.
from D igital's South Queensferry, Scotland
9.
17.
Vapor Deposited Blanket Tungsten," journal
Reactive Deposition Process Using Design
of Vacuum Science Technology B, vol. 9 (MayI
of Experiments," Advanced Semiconductor
J u ne 1991 ): 1478- 1486.
Manufacturing Conference and Worksbop
10.
18.
Scheme," Proceedings of the Sixth Interna­
minum Metal l i zation, Part I I Planarization,"
11.
tional IEEE VLSI Multilevel Interconnection
Conference (1989): 129- 135.
19.
Etchback," Microelectronic Manufacturing
Tungsten Deposition on Shal low Junction
and Testing, vol . 1 3 (November 1990): 29- 31 .
Leakage,'' in Tungsten and Otber Refrac­
VLSI
Applications
II, edited
by E. Broadbent (Pittsburgh, PA: Materials
20.
Etch Back Process in Non F u l l y Planarized
Topography," Applied Surface Science, vol . 53
C. Yang, ). M u l tani, D. Paine, and ]. Bravman,
(1991 ): 54-61 .
"Junction Leakage of Selectively Deposited
LPCVD Tungsten for Contact F i l l Applica­
tions," P1·oceedings of tbe Fourth Interna­
21 .
TEOS and SiH4 PECVD Oxides," Proceedings of
the Fifth International IEEE VLSI Multilevel
Inte1·connection Conference (1988): 366-373.
C. McConica, S. Chatterjee, and S. Sivaram,
"Step Coverage Prediction during Blanket
LPCVD Tungsten Deposition from Hydro­
22.
(1988): 268- 276.
Digital TechtticalJour·nal
D. Wang, K. Law, and ). Marks, " Planarized
Dielectrics for 0.5 /J-m Applications," ACET in
gen, Silane and Tungsten Hexafluoride," Pm­
ceedings of the Fifth International IEEE
VLSI Multi-level Interconnection Conference
C. Magnel l a , T. Ingwersen, and E . Fleck,
"A Comparison of Planarization Properties of
tional IEEE VLSI Multilevel Interconnection
Conference ( 1987): 200-207.
13.
K. Kol ler, H. Erb, and H. Korner, "Tungsten
Plug Formation by an Optim ized Tungsten
Research Society, 1987): 227-234.
12.
T. Clark and P. Riley, " M u l t i-Chamber, Single­
Wafer Integrated Tungsten CVD and Plasma
G. Georgiou e t a l . , "Influence of Selective
tory Metals for
) . van Laarhoven, H . van Houtum, and L. de
Bruin, "A Novel Blanket Tungsten Etchback
A. Aronson and I. Wagner, "Advanced Alu­
Transactions of the 47th and 48tb Schools
on Thin Film Technology: Advances in Mag­
netron Sputtering (1990): DPS-VII-4.
T. Clark, M. Chang, and C. Leung, " Response
Su rface Modeling of H igh Pressure Chemical
A . Berti, "Characterizing a Titani u m N itride
(1991 ) : 179- 184.
T. Clark, A. Constant, M. Chang, and C. Leung,
by S. Wong (Materials Research Society, Pit ts­
Figure 5 is courtesy of D. Hardie and C . Jones
facility.
Technologies,"
CVD Tungsten, Copper, and Other Refmctory
Metals for ULSI/ VLSI Applications V. edited
national, vol . 1 1 , no. 6 (May 1988): 138- 142.
8.
Interconnection
Transactions on Semiconductor Manu­
facturing, vol . 3 (November 1990): 150- 157.
IEEE
18-20.
Tapered Contact Etch," Semiconductor Inter­
P. Ri ley, T. Clark, E. G leason, and M. Garver,
"I mplementation of Tungsten Metal l i zation in
con Oxide Plasma Etching," Microelectronic
Manufacturing and Testing, (October 1981) :
7.
L. Holschwandner, and
N. Tai, "Thin Layers of TiN and Al as Glue
vol. 27, no. 4 (April 1984): 2 14-219.
6.
V Rana, ]. Taylor,
Sil icon Oxide," Vacuum, vol . 34, no. 3-4
Review (Fall l990): 1 1 .
23.
R. Levin and K. Evans-Lutterodt, "Step Cover­
age of Undoped and Phosphorus-Doped Si02 ,''
Vol. 4 No. 2
Spring 1992
71
Semiconductor Technologies
24.
Journal of Vacuum .\'cie11ce and 1ecl:mology,
Tetraethylorthosil icate," journal of Vacuum
Yo!. B l , no. 1 (19l U): ')4-61 .
Science Technolog)\ vol. 85 (1987 ): 15')'5 - 1563.
.J. Perchard et a l . , " Characterization of a
27
Mu ltiple-Step I n-situ PECVD-TEOS Planariza­
Mul tichamber
and
In-Situ
SPIF
Processing
Oxide ." Proceedings of the Sixth Intema­
tional IEEE VIS! Multilevel Interconnection
Conference (1989): 89-95.
of
Proceedings, vol.
l lfiH ( 19H9): 75 -85.
25.
28.
Molecular Structure," Proceediltgs of the Sixth
International IEEE VLSI ,Hultilevel Inter­
connection Conference ( 1989): 397 -403.
Si02 Using Deposition/Spu t ter Etch Cycles,"
jou rnal of the Electn>ehemical Socie�y,
26.
72
C. Chiang and D. Fraser, "Understanding of
Spin- On-Glass (SOG) Properties from their
G. Schwartz and P Johns, '' Gap-fill with PECVD
vo l. 139, no. 3 (1992): 927
Law, and D. Wang, ·'In-Situ Planar­
ization of Dielectric Su rfaces Using Boron
tion Scheme for Submicron Manufacturing,"
Electronic Materials,
J Marks, K.
29.
S. Gupta, "Spin-on Gl ass for Dielectric Planar­
F Becker et al . , " Low-pressure Deposition
ization,'' Microelectronic Manufacturing and
of High-yuality Si0 Films by Pyrolysis of
2
Testing, vol . 12, no. 5 (April 1989): 10- 12 .
Vol. 4 No. 2
SjJring 1992
Digital Techllical journal
Mary Beth Nasr
Ellen J Mager
Implementation of Defect
Reduction Strategies into
VLSI Manufacturing
CMOS-4 technology combines a high-performance microprocessor with afast, dense
RAiVI. Consistently obtaining a specified die yield on CMOS-4 devices required the
implementation of a series of defect reduction procedures. To achieve high yields,
microcontamination and defect reduction plans needed to be in place well before ini­
tiation ofproduct manufacturing. Levels ofoverall cleanliness had to be specified and
controlled. Process equipment was monitored at the new particle level of 0.375 f.tln
and greater to collect data. Defect density test reticles were designed and wafers were
processed. Electrical results were then incorporated into a yield model and used to
prioritize yield enhancement activities. Experiments were designed to reduce the
defect levels ofprocess areas, such asp-gate leakage and meta/ 2 short circuits.
Fourth-generation complementary metal-oxide
semiconductor (CMOS-4) technology calls for a d ie
area greater than 2 square centimeters (cm2), geo­
metries of 0.75 micron (�-Lm), a gate oxide of 10.5
nanometers (nm), u nique metallurgy, 1 .7 million
transistors, and 23 masking levels. These very large­
scale integration (VLSI) process features required
for chip performance dictate the need for increased
defect reduction and microcontamination control
in the semiconductor production environment.
Production of one fully functional CMOS-4 device
is virtual ly impossible without substantial efforts in
defect reduction and microcontamination control.
Obviously, a single 0.75 -�-Lm particle in the active
area of a die can create a short circuit and cause
the entire dev ice to fail. A 10.5-nm particle at gate
oxide can have the same effect. A high level of
sodium in rinse water can lead to premature gate
oxide breakdown. In fact, there are approximately
250 processing steps that cou ld contribute to the
failure of a chip .
This paper describes the principles of micro­
contamination control and relates their application
to VLSI manufacturing. It next discusses improve­
ments that we implemented for wafer hand ling,
cleaning, and monitoring. It then outlines the over­
all defect reduction techniques to increase product
yield, focusing on efforts in the areas of p-gate leak­
age and metal 2 short circuits. The paper concludes
Digital Teclmical]ourrwl
Vol. ·1 No. 2
-
Spring 1992
with considerations for defect reduction in the next
generation of CMOS technology.
Application of Microcontamination
Control Principles
Implementation of a successfu l set of defect reduc­
tion procedures depends on understanding the
principles of microcontamination control. This
field of study encompasses a wide variety of areas.
Microcontamination control seeks to minimize the
presence of any substance, particle, monolayer, or
ionic contaminant in the wafer production environ­
ment, that could cause a device to fail.
The facility i n which CMOS-4 devices are manu­
factured was constructed for the production of the
1 .0 -�-Lm CMOS-3 technology. However, because of the
high capital costs associated with clean room con­
struction, the faci l ity was originally designed to
meet the fabrication needs for three generations of
CMOS technology: CMOS-3, CMOS-4, and CMOS-5.
The original design specifications for the air quality
of the clean room, the clean room suits, and the
water, chemicals, and gases used in semiconductor
manufacture are discussed in the fol lowing sections.
Clean Room Air
The ambient air qual ity is carefu l ly controlled in a
clean room. Typical ly, indoor ambient air contains
more than one bil l ion particles per cubic foot. ' To
73
Semiconductor Technologies
obtain a high yield on CMOS-4 devices, wafers must
be processed in a well-characterized clean room.
The level of cleanliness within a clean room is
defined by the clean room class, which is deter­
mined by the number of particles greater than
0.5 p,m per cubic foot. The CM05-4 devices are man­
ufactured in a class 10 environment, i . e . , in a clean
room that has fewer than ten 0.5 -p,m particles per
cubic foot. To obtain the class figure, air is mea­
sured with laser-scattering techniques when the
clean room facility is at rest. The air filtration proce­
dure employs high-efficiency particulate air (HEPA)
filters with filter efficiencies greater than 99.9999
percent. Since HEPA technology is well advanced
and u nderstood, most of the particles in clean
rooms are generated when process equipment and
personnel are introduced. Airflow is maintained at
vertical l aminar to prevent particle deposition on
the product wafers.
Clean Room Suits
The material used for the clothing worn by clean
room workers was carefully selected. Humans can
shed up to a million particles of a size greater than
0.5 p,m every minute. To protect the wafers from
microcontamination and maintain a class 10 envi­
ronment, clean room workers must don special
suits. This attire must cover the worker from head
to toe.
When device line widths were greater than
1 .0 p,m, polyester fabrics were used to contain the
particles emitted by clean room workers. The pore
size of the best polyester fabric is 17 p,m, and the filter
efficiency at 0.5 p,m is poor at less than 60 percent.
With submicron line widths, new materials needed
to be evaluated for their ability to contain particles.
A new materia l , an expanded polytetrafluoroethy­
lene, with a pore size of less than 1 .0 p,m and a filter­
ing efficiency of greater than 99.99 percent, was
selected as the garment material of choice. 2
Ultrapure Deionized Water
Each wafer is exposed to hundreds of gal lons of
u ltrapure deionized (DI) water during the 250 pro­
cessing steps involved in producing CMOS-4
devices. The pu rity of the DI water, which is mea­
sured primarily by resistivity meters, is critical to
obtaining a high yield on CMOS-4 devices. Not only
must the number of particles be minimized, but the
levels of cations, anions, total oxidizable carbon
(TOC), sil ica, and bacteria must also be carefully
regulated and monitored. For example, bacteria
74
contain phosphorus and can be a source of uncon­
trolled dopant. Excess levels of TOC can double the
rate of initial thermal oxidation. Silica is known to
decrease the rel iabil ity of thermal ly grown oxides,
and the presence of ionic contaminants can change
semiconductor carrier l ifetimes.3 Table 1 l ists the
specifications for the DI water system used for
CM05-4 production.
Ta ble 1
De i o n i zed Water Specifications
for C MOS-4 Techno logy
1 8.0 megohm per em @ 25°C
Resistivity
Bacteria
Particles (>0.5
0.05 colon ies per m i l l i l i ter
max i m u m
p, m)
200 per l iter max i m u m
Total orga n i c carbon
50 ppb max i m u m
Si l ica
1 0 ppb maximum
All cations and an ions 1 . 0 ppb maximum
Note:
ppb equals parts per billion
Chemicals and Gases
Wet chemistry is used to clean wafers and in the
photolithographic process to develop and strip
photoresist. The particle and impurity levels of the
incoming chemicals used during wafer processing
had to be specified and monitored. Studies have
shown that bare silicon wafers placed in an ammo­
nium hydroxide/hydrogen peroxide (NH40H!Hp2)
solution exhibit a l inear correlation between the
metal content in the peroxide and the metal surface
contamination. The same studies have also shown
that iron and zinc are more important than other
metals.4
Throughout the manufacturing process, gases
are employed during gate oxide growth, metal
depositions, and plasma etches. Both impurity and
particle levels in these gases are critical. Impurity
level changes can alter plasma etch-rate u niformity
or could lead to corrosion. The process of reducing
the impurity and particle levels of gases is better
understood than that of wet chemicals. Implement­
ing new filter technology and employing electro­
polished materials in gas distribution systems h ave
reduced impurity levels.
Improvements to Wafer Cleanliness
Prior to implementation of a structured defect
reduction plan to increase yield on CMOS-4 devices,
three areas known to require better control were
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Implementation of Defect Reduction Strategies into VLSI Manufacturing
improved at a m inimal cost and effort. These were
Since wafer cassettes are known to become
wafer hand ling, wafer cassette/box cleaning proce­
porous and in time to contaminate the wafer sur­
dures, and particle per wafer pass monitoring.
face, a test was i ntroduced to determine when
boats start to degrad e. Resu lts indicated that if cas­
settes took longer than 30 seconds to rinse to resis­
Wafer Handling
Clean room workers use tweezers to handle wafers
tivity in DI water, they should be replaced.
when they read wafer numbers and load wafers
into equipment. A fu l l wafer cassette contains 25
Particle per Wafer Pass Monitoring
150 - m i l limeter (mm) wafers, and each wafer is sep­
Each p iece of process equipment is monitored on a
arated by only 5 m m of clearance. Wafer inspec­
tions of CMOS-3 devices revealed that scratches
contributed 10 to 15 percent of the die loss.
Three corrective actions were implemented to
reduce the number of scratches. Since the use of
tweezers to hand le wafers was the primary cause of
scratches, their use on product wafers was banned
from the production l ine. Vacuum wands were
installed throughout the fabrication area. Vacuum
wands restrict contact to the wafer backside only.
With proper training in the use of wands, workers
can achieve better vertical wafer control.
In add ition to vacuum wands, automatic wafer
transfer systems were insta l led i n the production
area. These mass transfer systems al low a worker to
move an entire 25 -wafer Jot from one type of cas­
sette to another, fo r example, to transfer wafers
from polypropylene to quartz cassettes prior to a
photoresist strip. This procedure elim inates man­
ual rol l transfers of wafers, which are known to gen­
erate particles.
Automatic wafer hand lers were i nstal led on engi­
neering microscopes as were automated wafer
sorters. Use of these devices e l i minated a major
source of scratches on experimental lots, which are
inspected and sorted often. Sneeze guards installed
at the microscopes were an added insurance
against damage to any die from spittle.
Cassette/Box Cleaning Procedure
Several cleaning procedures for the wafer cassettes
and boxes were implemented . A surfactant, which
routine basis to detect particles added at each wafer
pass; these checks are performed once per shift.
The selected wafers are measu red on a high-angle
laser-scattering system designed for u npat terned
wafer particle detection. The wafers are then pro­
cessed t hrough the production equ ipment. After
final processing, the wafers are measured again for
particles. These measurements are subtracted from
the initial readings, and the d ifference is recorded
on a trend chart.
For the 1 .0-J.Lm technology, bare silicon wafers
were measured for particles greater than 0.5 J.Lm, or
half the minimum polysilicon l i n e width. This size
was chosen based on the premise that a conductive
particle of this size could degrade device perfor­
m ance and rel iabil i t y. Using the same reasoning,
and prior to the implementation of CM05- 4 tech­
nology into manufacturing, the particle monitoring
size was decreased to 0.375 J.Lm.
Particle per wafer pass (PWP) monitoring must
closely simulate the wafer processing environ­
ment to which product wafers wil l be exposed. If
a process involves an oxide deposition, then the
PWP process shoul d also. However, i t is important
not to dam age the wafer su rface d u ring the PWP
run. For example, gases should be flowed, if pos­
sible, during a PWP process on an etcher, but a
bias should not be applied. A bias m ight cause sur­
face damage that could result in false particle
counts.
A continuous production P\XfP program must be
maintained on the process equ ipment. Small parti­
helps wet the surface, was added to the c leaning
cles are held to a surface by strong van der Waals
solution of the boat/box washing equ ipment. This
forces. These forces increase over time due to the
improved the cleaning efficiency over the pre­
particles conforming to the surface, thus increasing
viously used method of DJ water only. Wafer cas­
the contact area. Therefore, once particles are
settes were cleaned more often. Cleaning cycles
deposited on a wafer's surface, they are very d iffi­
were added at several front-end operations, includ­
c u l t to remove.; Even if each process step con­
ing wel l oxidation, initial oxidation, and first- and
tributed only five particles to each wafer, the
second-gate oxidation. In addition, weekly cleaning
cumu lative effect of 250 process steps wou ld be
of cassettes dedicated to specific equipment was
more than 1000 particles deposited on a fu l ly pro­
initiated .
cessed wafer.
Digital Technical}oU1-nal
Vol. 4 No. 2
Spring 1992
75
Semiconductor Technologies
Defect Reduction Procedures
to Increase Product Yield
AJ I microcontamination control efforts would be
fruitless without vehicles to assist in yield predic­
tion and defect prioritization. The procedures
developed for CMOS-4 technology are out lined in
this section. Yield modeling and test chips are
described elsewhere in this issue.6 Therefore, the
fol lowing discussion is brief.
General Yield Model
For the purposes of this paper, the Poisson yield
model is used . 1 This model is very simple, but it can
be used to i l lustrate some key points. The yield
model is given by the following equation:
Y = e-AD
where Y equals yield, A equals chip area in square
centimeters, and D equals defect density per square
centimeter.
It is easy to see that if the yield is equal to 50 per­
cent, AD must be 0.69. Now if the chip area is
i ncreased by 50 percent, AD becomes 1 .04, and the
yield is 35 percent, if all else remains equal.
However, each new CMOS technology reduces the
l ine widths and decreases the film thicknesses. The
size of a "kil ler" defect therefore decreases, which
automatically increases the basel ine defect density.
For each successive generation of CMOS devices,
substantial improvements are needed in the reduc­
tion of defect densities.
since they can be processed with less than a two­
week cycle time.
Defect Reduction Priorities
After electrical testing of failed structures, visual
inspections were performed to identify various
defect types. It is very common fo r l arge defect den­
sities to be caused by more than one defect type.
These inspections helped to design the experi­
ments used to reduce the defec t levels found in
p-gate leakage and metal 2 short circuits.
In addition, laser- and holography-based auto­
mated inspection tools were initiated in-line at var­
ious process steps, including active area after strip
inspect (ASI), polysilicon ASI, local interconnect
ASI, tu ngsten plug 1 and plug 2 ASls, and metal I and
metal 2 ASis. These inspections were performed
routinely on all ful l- and short- loop test chips. The
tungsten plug and local interconnect steps were
chosen because they were not part of previous
CMOS generations. The remaining steps were cho­
sen because they were known areas of concern
based on previous electrical results.
Defect reduction priorities are determined by
incorporating electrical resu lts from fu ll- loop
defect density test chips into a detailed yield model.
Based on this information, yield enhancement activ­
ities are prioritized . Two of the areas selected for
defect reduction were p -gate leakage and metal 2
short circu its.
P-gate Leakage Enhancements
Test Chips
The test chips used during the manufacture of
CMOS-4 devices were both full- and short -loop
defect density test vehicles. Full-process test chips
included snake structures to capture intralevel
open circuits, comb patterns for intralevel short cir­
cuits, and capacitors for interlevel short circu its.
The fu l l-process test chips were run routinely to
determine defect reduction priorities, as wel l as to
assist in reducing defect levels.
Short- loop test chips, which are processed
through 20 to 25 process steps, contain the same
snakes, combs, and capacitors as the ful l-process
test chips. Their purpose is to focus attention on
certain layers, such as back-end levels, which are
known to contribute many defects. These short­
loop chips can be used in designed experiments to
compare different processes. Short-loop chips also
monitor shifts in defect density from week to week,
76
Historical data obtained from CMOS-3 processing
highlighted two potential contributors to high
p-gate leakage values: surface damage and metal l ic
contamination. The wel l etch process, which
opens up the gate areas, was performed in a hexoid­
configu red, reactive ion etch (RIE) batch etcher.
Designed experiments, therefore , examined ways
to reduce potential lattice damage caused by the
known physical etch process. The batch reactor
processed 12 wafers at a time, and initial process
development ensured complete oxide removal
from all wafer surfaces. Based on uniformities both
within the wafer and from wafer to wafer, a 30 per­
cent overetch was chosen. The overetch proce­
dure, however, exposed the sil icon surface of the
wafers to the plasma for an extended period of
time, thus exacerbating any surface damage. Initial
attempts to improve the gate leakage varied the
overetch between 30 percent and 0 percent.
Vol. 4
No. 2
!>pring 1992
D igital Technical journal
Implementation of Deject Reduction Strategies into VLSJ Manufacturing
Results showed the p-gate leakage values consis­
tently ran at a lower value with the decreased
overetch , confirming a decrease in the amount of
surface damage. Optimization work continued to
lower the p-gate levels further. This work is out­
l ined in the following sections.
Silicon Lattice Damage
With the improved overetch process, stacking
faults and pits continued to be seen in the well
region . This confirmed the presence of su rface
damage, which required fu rther experimentation.
Spl it lots were processed to examine the impact of
changing power, bias, oxygen flow rate, and pres­
sure on p-gate leakage and silicon lattice damage.
Repeated attempts produced identical resu lts;
none of the changes affected p-gate defect density
or visual su rface damage. At this point, a spl it lot
was designed to study various starting materials.
The origina l starting material was compared to a
polysilicon-backed starting materiaL Polysilicon i s
a known "getterer" o f surface damage, that is, i t
attracts the damage t o the backside o f the wafer; it
was expected that this type of starting material
would show an improvement. The visual inspec­
tion of the polysilicon starting material found
no stacking faults or pits in the well regions of
the wafers. The p-gate defect density values also
showed an improvement. Confirmation material
verified the initial findings, and the new starting
material was placed on the manufacturing line.
In spite of significant improvements to the process,
p-gate leakage values continued to show intermit­
tent failures. The one area not previously investi­
gated concerned the potential presence of metallic
contamination at the wafer surface. Patterned and
unpatterned wafers were sent for total reflectance
X-ray fluorescence (TXRF) analysis. The su rface
analysis detected the presence of metall ic contami­
nants: specifically cobalt, iron, and nickeL Iron and
nickel are common elements in stainless-steel com­
ponents, and it was discovered t hat the gas distri­
bution tubes being used in the hexoid etcher were
constructed of stainless steeL Replacement alu­
minum gas tubes were installed in the etcher, and
additional surface analysis tests were taken. As
expected, the iron and nickel elements were no
longer detected; however, cobalt was present.
Consequently, a complete wet clean of the etch
system was performed, and one final set of etched
Vol. 4 No. 2
Ta ble 2
Results of TXRF Su rface Ana lysis
(Un its of 1 012 atoms per cm2)
Spring 1992
Iron
Cobalt
Ni ckel
2
3
2
0.5
0.4
2
2
0
2
I n itial wafer
Wafer center
45 deg rees from center
225 deg rees from center
Aluminum gas t u bes
Wafer center
45 degrees from center
Metallic Contamination
Digital Technical journal
wafers was analyzed by TXRF. The resu lts con­
firmed the elimination of the cobalt contamination.
The results of TXRF analysis on all metal l ic con­
taminants are given in Table 2. Since wet cleans
were performed routinely on the system, a trend
chart of p-gate leakage was available to show the
dates of all completed wet cleans. The trend chart
showed that coincident with every wet clean was
an improvement in p-gate defect densi ty. The
p-gate performance would begin to degrade when­
ever a metal 1 contact process was run in the same
etcher. The contact etch process opened contacts
to a cobalt silicide l ayer, which confirmed the con­
tact etch process as the source of the variable
cobalt levels. Cobalt cross -contamination of the
gates was occu rring whenever a wel l etch was pro­
cessed immediately fol lowing a metal 1 contact
etch, thus inducing p-gate variabil ity. For this rea­
son, it was decided to dedicate separate etch tools
for the wel l etch process and for the metal 1 con­
tact etch process.
225 deg rees from center
0
0
0.7
0.8
0
0.6
0
0
0
0
0
0
Post wet clean
Wafer center
45 deg rees from center
225 deg rees from center
0
0
0
0
0
0
Single Wafer Etch System Evaluation
At the same time the metal l ic contaminant sources
were isolated, a well etch process on a single wafer
etcher was developed. Single wafer etchers have
improved etch-rate uniformity control over batch
etchers. Also, the single wafer etch process is more
chemical (as opposed to physical) than the etch
process used in batch systems. Optimization of pro­
cess parameters (e .g., gas flows, pressure, power,
and gap) was performed on patterned monitor
wafers. As the final process step, a low-power sur­
face cleanup was added to remove any remaining
77
Semiconductor Technologies
surface contaminants from the wafer surface as
well as from the top layer of damaged silicon.
Cross-sectional micrographs were taken to verify
the integrity of the slopes of the patterned lines.
Once th e process was final ized on monitor
wafers, fu ll-process split lots were run through the
line to compare the hexoid-conl1gured etch pro­
cess with the planar-co nfigured etch process. The
results of one of the spl its are shown in Figure 1 .
A substantial improvement in p-gate defect density
was obtained when using the single wafer etch pro­
cess. Defect density levels on the batch reactor por­
tion averaged 300 defects per cm2, whereas the
single wafer etcher portion averaged 15 defects per
cm2. Confirmation product lots were processed to
ensure the probe yield was not adversely affected
by the etch enhancements, and t he wel l etch pro­
cess was released again on the single wafer etch sys­
tems. The final requirement of this process release
was the continued segregation of the well etch pro­
cess and the metal 1 contact etch process. The
metal 1 contact etch process remained on the batch
etcher.
1 000
0
800
w
(:J
<{ 600
;.::
<{
w
__j
w
f<( 400
(:J
o._
0
200
process-step defect density (D0) le ve l of approxi­
mately 40 detects per 100-meter (m) length.
Equipment-induced Short-circuiting
Mechanisms
P\VP data on two sets of process equipment empha­
sized the need for concentrated particle reduction
efforts. These two systems were metal deposition
and dielectric deposition. Task forces, with mem­
ben; from defect reduction, process engineering,
and equipment engineering, were organized and
chartered with reducing these !'WI' numhers.
Metal Deposition System UfJ,� mdes
A known
yield l i miter of the CMOS-3 process was the pres­
ence of a high number of titanium nitride (TiN) par­
ticles on the wafer su rface. Information obtained
fro m the equipment vendor and confirmed by
Digital's semiconductor manufacturing fabrication
plant in Scotland indicated that a new planar tita­
nium target wou ld provide a cleaner TiN film, thus
decreasing metal short circuits and enhancing
yield. The planar target, a rotating magnetic experi·
mental (R.MX) cathode, was able to decrease the
yield loss attribu ted to TiN particles by 75 percent.
Since several TiN layers were used in the CMOS- i
process, fewer particles on TiN film would defi­
nitely benefit the ov<:ra l l yield.
The PWP data in Figures 2 and 3 show how the
particle levels dropped once the R.tVIX cathode was
installed . Dat:� obtained on one of the initial RMX
spl it lots showed a 50 percent red uction in metal
short circuits. A substantial nu mber of confirma­
tion lots were processed to examine metal short cir­
cuits and electrical test data. O nce the lots were
an:�lyzed, the decision was made to release the Ri\1X
cathode into production and an improvement in
metal 2 short-circuit levels was realized.
Q L-------���-MEAN
Figure 1
BATCH
ETCHER
300 . 7
SINGLE WAFER
ETCHER
1 5.7
P-gate Area Defect Densi(y levels
as a Function ofEtcher Type
Reduction in Metal 2 Short Circuits
The implementation of the CMOS-4 metal process
into manufacturing brought a new set of chal­
lenges. The estimated yield impact of the metal 2
short circuits on the first set of fu l l-process lots
was approximately 40 percent, based on a single78
18:
200
z 1 50
<{
w
�
__j
<{
1 00
(f)
50
(,)
i=
<{
i=
f(f)
MEAN
=
58.429
0
Figure 2
TiN Particles before Installation
of RMX Cathode
Vol. 1 No. .!
5jJring 1992
Digital Teclmicnljournal
Implementation ofDefect Reduction Strategies into VIS! Manufacturing
60
lg
z
<{
w
="
_j
<{
u
f=
(f)
f=
<{
I(f)
30
lg
z 20
<{
w
="
_j 1 0
<{
g
I0
(f)
50
40
30
20
�
-10
I-
10
(f)
0
Figure 3
TiN Particles after Installation
ofRMX Cathode
Dielectric Deposition System Particle Reduction
The d ielectric deposition systems bad the h ighest
particle levels of any equipment in the fabrication
area, as shown in Figure 4. Since dielectric particles
can induce metal short circuits, and metal short cir­
cuits typically i mpact yield more than any other
defect structure, a concerted effort was made to
improve the particle stabil i ty. The four major areas
of change were ( 1 ) the instal lation of a new type
of 0- ring, (2) the initiation of a continuous pump
ballast, (3) pressure adjustments in the load lock,
and (4) modifications to the "clean" recipes. These
changes improved the average PWP coun t from
537. 7 to 2. 6 particles greater than 0.375 fLID, as
shown in Figure 5.
-20
Figure 5
Improved Dielectt"ic Particle Baseline
appeared intermittently (not all lots were affected
to the same extent), and were clustered around the
wafer edge. The metal 2 short circuits consistently
were twice as high as the metal 1 short circuits;
therefore, processes u nique to the metal 2 process
sequence were evaluated first. A lot history investi­
gation performed on a l l lots exhibiting corrosion or
grain-bou ndary stringers found no equipment com­
monalities, confirming possible process-induced
mechanisms. Several processes unique to the metal
2 process flow include the intermetal dielectric, the
metal 2 cutout sequence (used to remove metal
from al ignment targets al lowing metal 2 alignment
without manual intervention), and the tungsten
process sequence.
Corrosion
Figure 6 shows the corrosion on a
30
metal l ine in the test circuit. A common cause of
lg
z 20
<{
w
=" 1 0
_j
<{
� 0
corrosion is the interaction of chlorine with mois­
� -10
back processing procedures on corrosion. The
ture. The tungsten etch-back process was i nves­
tigated because it uses chlorine as an etch gas. A
short-loop monitor wafer experiment was designed
to study the effects of various post-tu ngsten etch­
(f)
1-­
(f)
level of corrosion was determined by means of a
-20
MEAN:
Figure 4
patterned wafer defect detect ion tool. Practices
commonly used in the semiconductor i ndustry to
537.680
el iminate chlorine-induced corrosion are a fluorine
passivation process and/or an i mmediate water
Initial Dielectric Particle Baseline
rinse. The data in Figure 7 shows that, without
Process- induced Short-circuiting
Mecbanisms
an i mmediate water rinse, the corrosion counts
increase with time. After only 90 minutes, the rate
of increase changed dramatical ly. Initially, the
Once the equipment PWP data improved to the lev­
incorporation of a rinse provided more stability,
els shown in Figures 3 and 5, and the metal 2 short­
but the corrosion levels were unacceptable after
circuit levels fel l to approximately 10 defects per
just 11 hours. The addition of a 5 - second sulh1r
100-m length, two systematic defects were u ncov­
hexafluoride passivation process, however, com­
ered: corrosion and grain-boundary stringers. Both
p letely prevented any corrosion.
defect types were noted during the inspection of
To verify this data o n test chips, several lots were
failed sites on the test chip structure. The defects
split. The resul ts from one of the splits are shown
Digital Technical journal
Vol. 4 No. 2
Spring 1992
79
Semiconductor Technologies
30
�
if>
z
w
0
f­
(.)
w
LL
w
0
I::
20
0
::::J
u
a:
0
.:..
gj 1 0
I
(f)
N
-'
<(
f­
w
2
Figure 6
0 �-----Corrosion on a Metal Line
MEAN:
SF6
STANDARD
4.1 60556
1 1 .238818
500
Figure 8
Resultsfrom Split Lot Test ofMeta/ 2
Short Circuits
400
verified that the defects were a wet type of corro­
(/)
fz
::>
0
sion formed on metal 1 in the bond pads and guard
300
rings and were caused by exposure to chlorine
du ring the tungsten etch-back process at tungsten
(.) 200
plug 2.
Grain-boundary Stringers
A short-loop lot was
designated to look for potential contribu tors to the
grain-boundary stringers. It examined three factors:
200
400
600
800
1 000
TIME (MINUTES)
lot are shown in Figure 9. They indicate that the
KEY:
D
6
IMMEDI ATE RINSE
5 SECONDS SF6, 0 GAUSS
10 SECONDS SF6, 0 GAUSS
0
0 1 5 SECONDS S�, O GAUSS
•
.to.
e
+
NO RINSE
5 SECONDS SF6, 50 GAUSS
10 SECONDS SF6, 50 GAUSS
15 SECONDS SF6, 50 GAUSS
Figure 7
the dielectric fi l m , the c u tout process sequence,
and the metal 2 deposition process. Results of the
most signi.fi cant factor affecting grain-boundary
stringers was the cutout process sequence. Since
the grain-boundary stringers were seen o n an i nter­
mit tent basis, the spli t was purposely designed to
exagge rate the impact of the cutout process. This
was ach ieved by reducing the metal overetch and
processing the designated wafers through the
photolithography and strip cycles twice. The metal
Particle Counts ofBare Aluminum
2 short circu its on these wafers were dramatically
Wi�Jers as a Function of Time after
higher than those on the wafers that did not receive
Tungsten Etch Back
a cutout process; they improved from 27 defects
per 100 m to 1 .4 defects per 100 m. An example of a
i n Figure 8.
The metal 2 short -circuit levels
grain-boundary stringer is shown in Figure 10.
improved from 1 1 .2 defects per 100 m to 4.1 defects
Since processing without a cutout process was
per 100 m . Additional spl its performed on product
not an option, a split lot was designed to study vari­
and test lots confirmed that the sulfur hexafluoride
ous versions of the cutout strip process. Although
passivation process was as good as, if not better
than, the origi nal process. The electrical results
80
the electrical resu lts of the split were not conclu­
sive (all defect levels were excel lent), a scanning
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Implementation of Defect Reduction Strategies into VLSI Manufacturing
resist that had been plasma-hardened by the metal
>f-
etcher. By transferring the metal etch process
40
from a plasma etcher to a wet sink, the option of a
Ui
z
w
0
fu 30
w
lJ...
w
0
fs
u 20
a:
downstream plasma strip became available. Several
additional splits were processed to study the com­
bination of a wet chemical cutout etch fol lowed by
a downstream plasma strip. They confirmed that
the elimination of the interaction between the
grain boundaries and the solvent/water combina­
0
�
tion resulted in a reduction in grain-boundary
a:
0
I
(f)
N
_,
stringers. Metal 2 short circu its on one of the lots
improved from 5.1 defects per 100 m to 3.2 defects
10
per 100 m and showed a corresponding improve­
<(
fw
:2
ment in yield.
0 L-----�--�---MEAN:
NO CUTOUT
STANDARD
1 .4387
27.069
Figure 9
From the initial stages of CMOS-4 manufacturing
up to the present, an overall improvement in metal
2 defect levels has been seen. The test chip metal 2
short-circuit D0;
Results of Test on
Grain-boundary Stringers
levels
have
diminished
from
approximately 40 defects per 100 m to approxi­
mately 4 defects per 100 m . Corresponding metal 2
visual defect inspection data has decreased from
approximately 1 .5 defects per cm2 to approxi­
2
mately 0.2 defects per cm
Future Considerations
For the successful production of future generations
of CMOS technology, improvements need to be
made in the areas of general microcontamination,
wafer handling, defect review, and data manage­
ment. The needed improvements are outl ined
below.
Microcontamination
Initia l resul ts from particle studies on PWP measure­
ments from current equipment indicate that 70 per­
cent of the total particles are between 0.25 J.Lm and
0.375 J.Lm. As discussed previously, once deposited
on wafers, these small particles are very difficul t to
Figure 10
Photomicrograph of a
Grain-boundary Stringer
remove. Efforts must be made to isolate and el imi­
nate the source of these particles before the equip­
ment is introduced into manufacturing. To provide
electron microscope (SEM) analysis of the same d ie
early identification, in-situ particle measurement
location on a l l spl i ts showed a significant difference
for process tools needs to be i ncorporated as a part
among splits. All splits incorporating a wet solvent
of the tool specification.
had evidence of grain-boundary stringers, whereas
I n addition, as geometries shrink and gate oxides
the spl i ts stripped only in a downstream p l asma
become thinner, the importance of understanding
had none of the stringers. Therefore the cause of
the effect of nonparticulate contamination such
the grain-boundary stringers was the reaction of
as zinc, sodi u m , and hydrocarbons on device yield
the grain boundaries with the solvent and subse­
is critical . The cost of reducing such contaminants
quent water treatments.
from the wafer environment is astronomical . We
The purpose of the solvent/water portion of the
must learn which contaminants are harmfu l and at
strip process was to assist in the removal of photo-
what level. Surface analysis techniques, such as
D igital Technical journal
Vol. 4 No. 2
Spring 1992
81
Semiconductor Technologies
TXRI' and Auger electron spt.:ctroscopy, need to be
incorporatt.:d i nto tht.: manufacturing t.:nvironment
to providt.: the necessary data.
Wafer Handling
Summary
The success or fai lure of a sem iconductor produc­
tion facil ity l ies i n the ability to control contam­
ination and reduce defects. E fforts are u nder \Vay
to improve Digita l 's ability to obtain high yield s
As die size conti nues to increase, wafer size also
on current and fu ture microprocessors. Semi­
increases. As technology moves toward 200- and
conductor companies must join with inspection
300-mm wafers, manual handling of \Vafers is a l l
and analytical equ ipment manufacturers, as wel l as
but impossible. O n e broken wafer wil l resu lt i n
with research i nstitu tes, to develop the required
the loss o f thousands o f dol lars. Automated han­
tools to support future technologies.
d l ing of wafers must be incorporated across each
process step .
Acknawledgments
Defect Review
mentation of the defect reduction work high­
Many people contribu ted to the successfu l imple­
As the killer defect size continues to decrease,
l ighted i n this paper. We would l ike to thank a l l
visual inspections with optical microscopes wil l
members o f t h e manufacturing process and equip­
lose thdr value. A s many a s 5 0 percent o f t h e elec­
ment groups who worked long hours to qual ify the
trical failed combs on our 0.5 -JLm process are not
CMOS-4 process. Special thanks to Tom Estel le and
observed d uring o p t ical m icrosc o pic review. This
Jan Haanpaa for their contributions in coordinating
has highl ighted the need for two t hings: test chips
the work outl i ned in this paper.
designed for fai lure analysis and easy- to-use in-line
inspection S EMs.
References
active involvement and input of defect reduction
1 . C. Osburn e t a t . , "The Eftects of Contamination
Test chips can no longer be designed without the
personnel. A test chip must not only be wsed to
prioritize defect reduction efforts, but i t must also
help to determine and isolate defect sources. A test
chip that is clifli.cult to inspect provides only ha lf
the inform ation needed to reach the ultimate goal
of high yield .
In-line inspection SEMs are requi red to review
defects found du ring process inspections and to
analyze process problems. They should be capable
of energy d ispersive X- ray spectroscopy (EDXS)
on Semiconductor Manufacturing Yield," The
journal of Environmental Sciences
April 1988): 45-57
(March/
2. W Whyte and P. Bailey, " Particle Dispersion in
Relation to Cloth ing," The journal of Environ­
mental Sciences (March/April 1989): 44.
3. I� Burggraaf, "Ultrapure Water Focus," Semi­
conductor In ternational (May 1987): 129- 132.
failing locations can he easily reviewed.
4. M. Meuris et a t . , " Correlation of Meta l l ic
Impurity Content of ULSI Chemicals and Defect­
Related Breakdown of Gate Oxides," Proceedings
of the / 78th Meeting of the Electrochemical
Society (Washington, D.C. , 1991 ): 4 88 .
Data Jlfanagernent
5. R. Bowling, "An Analysis of Particle Adhesion on
analysis to provide information on the elemen t a l
components o f the defects. In add ition, t h e t ransfer
of electrical test data to an SE.\1 is req u ired so that
The fu ture of defect reduction efforts depends o n
the abi l ity t o manage a n d analyzt.: large quantities of
data. Defect inspection tools perform f u l l wafer
i nspections in less than five m i n u tes. A thorough
understand ing of statistical methods, such as con­
trol charts and sampli ng procedures, is required
to determine the extent of defect review efforts.
Semiconductor Surfaces," journal of the Electro­
chemical Society: Solid-State Science and Tech­
nology, vol. 132, no. 9 (September 1985 ) : 2209.
6. R . Col l ica, X . Dietrich, R. Lambracht, and D. Lau,
"A Yield Enhancement Methodology for Custom
VLSI Manufactu ring," Digital Technical journal,
vol. 4, no. 2 (Spring 1992, this issue): 83-99.
Automatic storage of images o n optical d isks for
subsequent review is another key area. Eftorts are
ongoing with Digital's Campus - based Engineering
Center in Vienna to determine the bene li. t s of using
fractal analysis on airborne particle da t a .
82
Vol. 4 Ao. 2
Spring 19Y2
D igital Technical jour11al
Randall S. Collica
Dietrich
RudolfLambrachtJr.
David G. Lau
X. Joseph
A Yield Enhancement Methodology
for Custom VLSI Manufacturing
Integrated circuit yield enhancement is a complex issue due to the many steps
involved in the manufacturing process and the number of variables governing the
overallyield. The task isfurther compounded by industry technology goalsfor con­
tinually improving performance achieved by decreasing minimum feature size,
increasing chip area, and incorporating more on-chip functionality from genera­
tion to generation. In thefinal analysis, the cost ofproducing chips is directly related
to the yield, hence the necessity for a comprehensive yield improvement strategy.
The industry technology goal for continually
improving complementary metal-oxide semi­
conductor (CMOS) very l arge-scale integration
(VLSI) chip performance has been achieved by
decreasing minimum feature size and incorporating
more on-chip functionality in a larger chip area.
At Digital, four generations of CMOS technology
have been developed. Each generation possesses
successively more complex manufacturing pro­
cesses as wel l as more individual process steps to
fabricate the chips. These complex processes and
additional steps have increased the number of vari­
ables that have the potential for affecting yield.
Digital's Alpha 21064 microprocessor chip 1 has a
peak operating frequency of 200 megahertz (MHz),
contains 1 .7 million transistors, and has chip dimen­
sions 1 .68 X 1 . 39 square centimeters (cm2). The
Alpha microprocessor is the h ighest performance,
highest density, and largest chip currently manufac­
tured in volume by Digital. With these chip com­
plexities, achieving chip yield goals is imperative
for successful cost-effective manufacturing.
In the manufacture of i ntegrated circuits, yield is
defined as the fraction of the total number of die
sites introduced into processing that are completed
as fully functional chips. The cost of producing
a ful ly functional chip is inversely proportional
to the yield . In the CMOS-4 manufacturing l ine,
22 wafers are processed through all the steps as a
single lot. On a lot basis,
Cost
of chip
Cost of lot
Yield of lot · Total number of die sites in lot
Hence, improving the yield directly affects reduc­
ing the cost of production per chip. This motiva-
Digital Technical journal
Vol. 4 No. 2
Spring 1992
tion for yield enhancement appl ies to a l l product
chips for the l ife of the respective chip.
Once the product chip is introduced, demand
for the chip increases up to some level and then
declines to the end of its l ife . D u ring a product's
development, circuit and system design teams
require quantities of chip prototypes for design veri­
fication and debug purposes. The rapid increase i n
demand after product introduction requires a steep
yield learning curve and hence rapid yield enhance­
ment to supply production quantities.
This paper discusses the yield enhancement
methodology used to evaluate processing, process
equipment, manufacturing methods, design, and
testing in relation to yield. It describes the yield test
vehicles, designed experiments, yield models, and
special-purpose analytical tools to identify and pri­
oritize defects and to focus resources on appropri­
ate defect inspection and failure analyses tasks.
Before our discussion of the specific techniques,
we present a brief overview of the methodology
to enhance yield.
Overview of Yield Methodology
The yield enhancement methodology applied to
chip manufacturing at Digital involves the use of
test chip data, product chip data, static random­
access memory (SRAM) data, yield models, special­
purpose analytical tools, and defect inspection to
p erform defect identification and prioritization .
The information gained is relayed to process engi­
neering for rapid yield learning and the design of
experiments for yield improvement. In addition,
yield, development, and design teams use the
83
Semiconductor Technologies
information to estimate the manufacturabi l ity of
Product wafers from process en g i neering are
future products .
also electrical ly tested, and parametric data a nd
product functional ity data are measured on these
Figure I presents the overall methodol ogy and
the relationships of the various stages of yield anal­
wafers. The resu lts of measurements at electrical
ysis. After a chip has been designed its design is
test, the lot process records, and defect i nspection
converted into a chip l ayo u t t o prod uce t he reticle
data undergo parametric and prod u ct y ie l d a nal y ses
ufacturing are electrical ly tested. Parametric data,
of analyses are used by process engineers for exper­
,
set for the chip. Test chip wafers from process man­
to determine the factors affecting y ield . Al l results
imenta l des i gn and yield learning. The use of test
defect stru cture data, and SRA�·I test vehicle fu nc­
ch ips is one of the st ar r i ng poin ts in early process
ti o n al ity data are measured on test chips. The lay­
defect l e arn in g and is described next.
out information for the chip, the measurements
from defect structures on the test chip, and SHAM
functional yield are used to model yield.
Test Chip
The yield model generates a priority l ist (defect
pareto) of the yield- l im i ting defect mechan isms for
the layout of that particu l ar c hi p . If the layout of
the chip is determined to be especial ly sen si tive to
a part i c u lar defect type the chip layout can be
al tered to optimize for h igher yield wit hout chang­
ing the design f11 nctional ity. In addition, in the early
stages of design yield information from a previous
generation technology can he used to fo recast the
yield of chips to be manufactured in a succeeding
generation of technology. The effects of choosi ng
d ifferent redundancy schemes can also be forecast.
Test vehicles and structures on the test chip pro­
vide the process and e l ectrica l information needed
to esti mate chip yield . The CMOS-4 yield test chip is
shown i n Figure 2 . The chip is d ivided into three
,
funct ional are as:
1 . Defect density te s t struc t ures
,
�
3. Scribe lane structures
The defect densi ty structures are used to determ ine
the component defect densities for the i nt eg rated
I
•
-
YIELD
FORECASTING
PROCESS
MANUFACTURING
PROBE
2. 128 k i lobit (Kb) SRAM
CHIP DESIGN
-
�
�
WAFER LEVEL
ELECTRICAL TEST
LAYOUT
I
PRODUCT CHIP
YIELD ANALYSIS
�
PARAMETRIC
YIELD ANALYSIS
�
YIELD MODEL
�
DEFECT PARETO
'
•
l
l
ROOT CAUSE DETERMINATION
EXPERIMENTAL DESIGN
AND ANALYSIS
Figure I
R4
Yield Methodology Flow Diagram
Vol. 4 No. 2
S{Jri11g 1')')2
Digital Techuical Jourual
A Yield Enhancement Methodologyfor Custom VLSJ Manufactlu·ing
SCRIBE LANE STRUCTURES
I
DEFECT
DENSITY
TEST
STRUCTURES
128Kb
SRAM
Figure 2
Yield Test Chipfor the CMOS-4 Process
process, and the scribe lane structures are used to
defect type. The complete set of defect structures
determine the parametric yield . Since the SRAM and
combines to represent the ful l process.
the defect density structures are both on the chip,
the SRAM fail ure modes can be correlated with the
component defect levels measured o n the defect
density structures. In addition, with the compo­
The defect structures test for the fo llowing gen­
eral electrical fau lts:
1 . Intralayer short circu its and open circuits
nent defect densities factored into the yield model,
2. Interlayer short circuits
a defect pareto can be generated that prioritizes the
3. Contact/via chain open circuits
defect mechanisms according to their yield-l imiting
effect on the SRAM.
The defect density structures have the following
qual ities:
1. Simple design, with large area structures for
determining specific component defect densities
l ntralayer short circuits and open circuits are mea­
sured using snake/comb structures. For example,
Figure 3 shows a schematic d iagram of a metal 1
(M l ) snake/comb structure that tests for M l short
circuits and open circuits. Testing the continu ity of
the snake detects open circuits in the Ml line.
Testing for continuity between the snake and
2. Testable in-line
combs detects Ml short circu its.
Figure 4 shows an example of an i nterlayer short­
3. Compatible with the yield model
circuit test using an M 1 to metal 2 (M2) capacitor.
4. Inspectable i n-line
A test for continuity between M l and M2 conductor
plates detects short circu its in the dielectric layer
The CMOS-4 technology integrates approximately
between M l and M 2 .
250 process modu les on a si ngle complex chip . J
I f a process aberration o r defect renders a chip
inoperable, it is very difficult to diagnose at which
process step the defect occurred.
The defect structures partition the integrated
process i n to critical physical process features that
are more easily characterized than an entire com­
Contact/via integrity is tested using chains of
contacts or vias. Figure 5 shows how an M 2 to M l
via chain i s tested for continuity t o check for any
chain open circu its.
Table 1 l ists the yield test chip defect structures
and respective electrical faults or defect mecha­
nisms that they detect.
plex chip. A number of defect structures are
The test structures must be testable in-line, that
designed; each one is designed to detect a specific
is, they must a llow electrical tests to be performed
Digital Technical journal
Vol. 4 No. 2
Spring 1992
85
Semiconductor Technologies
METAL 1 \
METAL 1
! METAL 2
7
I
META L 1
I
I
M
ETAL 1
I
I
Figure 3
(a) Top View
METAL 2
METAL 1
Metal 1 Snake/Comb Structure
after M l , M 2 , or metal 3 (M3) patterning. Conse­
quently, st ructures that do not require process
steps after M l must be designed only in those layers
(b) Side View
up to and including M 1 . For example, M 1 to poly­
sil icon contact chains are connected to pads by M 1
Figure 4
and not by M l to :VI 2 vias in conjunction with M 2 .
Metal 2 over Metal 1 Capacitor
The tormer chain i s testable in-line after M l pattern-
f
METAL 2
I
I
I
I
I
I
f
METAL 1
�--I
I
I
I
_ _ _ I
�--I
I
I
I
_ _ _I
I
I
I
I
I
I
I
I
I
I
I
I
� - - -I
I
. - - -.
I
I
I_ _ _ •
I_
_ _
:
I
I
I
I
I
I
(a) Top View
METAL 2
METAL 1
LJ ---�
L�---
I:J
(b) Side View
Figure
86
5
Metal 2 to Metal 1 Via Chain
Vol.
4 No. 2
.SjJrin8 IYY2
D igital Tecbnicaljounzal
A Yield Enhancement Methodology for Custom VLSI Manufacturing
Table 1
Defect Structures a n d Defect Ty pes
Electrical Fault to Detect
Structu re
Dielectrics
M3/M2/M1 /polysil icon/substrate capacitor
Gate oxide, d i electrics 1 ,2,3 short circuits
M3/M2/M1 /polysili con/wel l capacitor stack
Gate oxide, d ielectrics 1 ,2,3 short c i rcu its
B i rd 's beak struct u re - polysilicon plate
m i n i m u m pitch active area stripes in well
F ield/active area over periphery short c i rcuits
B i rd's beak struct u re - polysi l icon plate over
m i n i m u m pitch active area stripes in su bstrate
F i eld/active area periphery short c i rcu its
Contacts
M1 to n+ contact chain
M1 to n + open circuits
M1 to P+ contact chain
M 1 to p+ open circuits
M 1 to n+ polysi l i con contact chain
M1 to n+ polysi l i con open c i rcu its
M1 to p+ polysi l icon contact chain
M 1 to p+ polysilicon open circu its
M 2 to M1 contact chain
M 2 to M1 open c i rcu its
M3 to M 2 contact chain
M3 to M 2 open circuits
Polysil icon to local i nterconnect to n+ chain
Polysi l icon to local i n terconnect to n + open circu its
I nterconnect
Polysil icon open circuits and short c i rcuits
Polysilicon snakes and combs
M 1 snakes and combs
M1 open c i rcu its and short c i rcu its
M 2 snakes and combs
M 2 open circuits and short circuits
M3 snakes and combs
M 3 open circuits and short circu its
N + snakes and combs
N+ open circu its and short circuits
P+ snakes and combs
P+ open circuits and short circuits
Local i nterconnect to n + combs
Local i nterconnect to n+ short circu its
N-channel gate meander
N-gate meander short c i rcuits
P-channel gate meander
P-gate meander short circuits
Local i nterconnect snakes and combs
Local interconnect open circuits and short circuits
Local interconnect to n+ polysil icon combs
Local i nterco nnect to n+ polysilicon short circuits
ing; the latter has to continue processing to be
tested after M2 patterning.
The structures were designed with minimum
p itch metal lines and minimum surrounds for con­
tacts to give them the greatest sensitivity to defects.
In addition, the structu res covered an area large
enough to detect the minimum defect density
desired . The yield model required that the yield test
chip contain fou r or more instances of each type of
defect structure to determine the clustering param­
eter in the yield model.
The structures are i ntegrated onto the chip to
faci litate manual or automated visual inspection of
failing test sites. For example, i n the process flow,
Ml is patterned several process steps after poly­
silicon is patterned. To keep the polysilicon layer in
view, some polysilicon snake/comb structures
must not be covered with any M l structures.
Should a fai l ing polysilicon comb be found at M l
Digital Technical journal
lk>l. 4 Nu. 2
Spring 1992
electrical test, that failing comb can then be visu­
a l ly inspected to identify the type of defect causing
the comb to fai l .
The scribe J a n e contains t h e minimum s e t of
electrical test and process monitor structures
required to characterize and monitor the process i n
a manufacturing mode. The scribe l a n e exists on all
chips. Table 2 l ists parameters measured from the
scribe structures. If a critical parameter does not
comply with its specification on more than a cer­
tain number of die sites on a wafer, that wafer is
rejected . Hence, scribe lane structures determine
parametric yield loss.
Electrical Failure Specifications
To estimate the defect density from an electrical
test structure monitor, specification limits must
be established that determine a fault. Usually, the
DC parametric testing applies either a voltage or a
87
Semiconductor Technologies
Table 2
Scribe La ne Electrica l Test
Parameters
SRAM Ana�ysisfor Process Fault Signatures
The SRAM is a useful circ u i t vehicle for yield ana ly­
Transistor t h reshold voltages
Transistor sat u ration reg ion cu rrents
Transistor leakage c u rrents
Transistor effective channel lengths
D i ode breakdown vo ltages
Field transistor threshold vo ltages
l'nterconnect sheet resistances
Gate oxide t h i ckness
Contact resistances
Die lectric leakage cu rrents
current value. Either the leakage cu rrent or volt­
age is then recorded or the resistance is computed
on metal interconnect vias, on cont act chains, or on
serpen tine l ines. A fault at a gate capacitor is
recorded if the level of current passed after a volt­
age ramp is 1 microampere (p.A) or more. The fault­
level specifications arc period ica l ly reviewed and
changed as necessary.
sis. On the yield test chip, the probe yield of the
SRAM measures the capabil ity of the fu l l , i ntegrated
process to yield product. The regu lar array of the
SRAM and its bit mapping capability al low some cor­
relation of fa ilure modes to defect mechanisms. The
use of SRA\1s processed with the defect rest struc­
tures al lows the determ ination of process defects
that affect circuit fa ilures. A relatively l a rge SRAM
array (i . e . , 12RKb or larger) typical l y captures most
of the fa u l ts within the memory cdls, which com­
prise up to 90 percent of the memory chip layout.
These cel l s are tested after the basic con tinuity and
leakage requirements have been met du ring the
SRANI wafer level testing. Since the memory cel l lay­
out is a regular array, cert ain defect mechan isms
have specific functional fai l u re patterns within the
circu it. Figure 6 is a typical bit fai l map showing
the types of patterns analyzed with pattern recogni­
tion techniques.2 These patterns were analyzed for
their signature of possible defect mechanism with
a probability of fail ure (POF) matrix. 1
CLUSTERED
BIT FAULT
VERTICAL
DOUBLET
DIAGONAL
DOU BLETS
\.
o
X X
X X
X X
X X
X X
X X
X X
X X
X
X X
X
X X
X
· ·
X�X
__�,+------------------�
���� X X���
X X X X X X X X X X X
X X
X X
X X
X X
·
·
X X
X X
X X
X X
X X
X X
X X
X X
•
·
_
_
X
__ .·
_
X X X
o
: :
__
_
_
�,...------------+-'
"-!-----._
PARTIAL WORD
L I N E FAU LT
X
X X
DOUBLE
/ BIT L I N E FAULT
X
X
X
X
X
: (ill : :
�9
� ·
�-r-: �
:--�
S I NGLE WORD
LINE FAU LT
PARTIAL
BIT L I N E FAULT
®
CD) .
SINGLE BIT
LINE FAULT
HORIZONTAL
DOUBLET
I
SINGLE BIT
FAULT
KEY:
X
SRAM BIT CELL
NON FUNCTIONING BIT CELL
Figure 6
88
Bit 1Hap of SRA.M Failure Patterns
Vnl. 4 No. 2
Spring t<J92
D igital Technical jounwl
A Yield Enhancement Methodologyfor Custom VLSJ Manufacturing
Table 3 gives the POF matrix for a 128Kb SRAM.
�
�
The rows of the matrL'< are the individual defect
types monitored in the process using the special
test structures described above. The columns of the
matrix are the types of SRAM failure patterns gath­
ered from pattern recognition programs. Each cel l
in the matrix is the average probabi l ity that a defect
of that type wil l cause a circuit fau l t type . This anal­
ysis was origin a l ly done using Monte Carlo simul a­
tion techniques on the memory cell using the
CI: UJ
UJ a._
�::J �
;;:
Z u..
0.48
•
0.44
0.40
0 . 36
0.32
z z 0.28
..:
� :2
3
8
�
VLASIC yield simulator.<.> This matrix is currently
0.24
0.20
0. 1 6
.
./
!/ /
/ .
0 . 1 2 �_.___._____._�-��-�----'---'-�
2 0 22
18
16
2
4
6
8
10
12
14
METAL 2 SHORT-CIRCUIT D0
being used to compile defect statistics from the
KEY:
defect-sensitive test structures together with the
SRAM pattern fail data.
An example of SRAM analysis shows the yield and
failure d iagnosis of colu m n failures. Since the SRAM
columns are designed i n the M 2 interconnect l ayer,
e
OBSERVED DATA
0.01 7325X + 0 . 1 36872
95% CONFIDENCE CURVES
Figure 7
the level of M2 to M2 short circ u i ts measured from
the defect test monitors correlates to the level of
column failures. Figure 7 shows the level of single
Comparison of SRAM Single and
Double Column Failures to Metal 2
Short Circuits
and double colu m n failures and the level of lateral
M2 short circuits on a per lot basis. SRAM analysis is
used because the defects accumu lated during the
Single Layer Yield Model
manufacturing process do not have a strong electri­
The single l ayer yield model identifies the defect
cal fau l t signature on a large, complex microproces­
types that have the greatest impact on yield . Proc­
sor circuit.
ess engineering uses this information, known as the
The yields of the SRAM chips are typical ly com­
defect pareto, to improve the yield by designing
pared to those forecast with the yield model. The
experiments to reduce the defect density of the
test structures described earlier are used col lec­
highest priority defects. Design engineering uses
tively in the yield model , which is discussed in the
this information to lay out product chips so that
following section.
they are less sensitive to these defect types.
Ta ble
3
1 28Kb SRAM Probab i l ity of Fa i l u re M atrix
Row
Defect Ty pe
Partia l
Row
Fault Ty pe
Double
Column
Column
Column
Fa ils
Single
Fa ils
Partial
Fails
Fails
Di electric 1 (flat)
0.20000
0.0000
0.00000
0.00000
0.00000
Dielectric 2 (worst-case step)
0.00000
0.0000
0.00000
0.95000
0. 00000
Dielectric 2 (flat)
0.00000
0.0000
0.00000
0.05000
0.00000
Polysilicon short circuit
0.1 5640
0.0242
0.021 60
0 . 1 831 5
0.00000
M1 short circuit
0.59000
0.0000
0.1 741 0
0.20453
0.00000
M2 short circuit
0.00000
0.0000
0.73221
0.22870
0.00000
Polysi licon ope n circ u i t
0.1 9602
0.0681
0.00000
0.00000
0.00000
M1 open c i rcuit
0.3631 0
0.6020
0.00000
0.00000
0 . 00000
M2 open circ u i t
0.00000
0.0000
0.1 21 00
0.21 950
0.63900
N-gate area
0.00000
0.0000
0.00000
0.00000
0.00000
P-gate area
0.00000
0.0000
0.00000
0.00000
0.00000
Digital Technical journal
Vol. 4 No. 2
Spring 1992
Fails
89
Semiconduc tor Technologies
The single layer yield model uses the negative
binomial yield equation (1) to p redict the yield
impact on a given product for each defect type .6·7· s
(1)
where A c is the critical area of the chip, Du i s detect
density,
�;
b
is a measure of spatial distri ut ion of
defects,9 and i is an index used to denote a specific
defect typ e .
The critical area o f t h e chip i s the area that is
most susceptible.: to a particu lar type of defect.
Critical area can be measured in various u nits,
including square centimeters, meters, and numbers
of contacts. The defect cknsity is the nu mber of
Table 4
Critical Areas by Defect Type
Defect Ty pe
Units of Critical Area
l n terlayer open
N u m be r of nonredun dant
contacts
c i rc u its (contact
chai ns)
l nterlayer short
c i rcuits (caps)
Area of overlap between
sequential i nterco nnecting
layers
I ntralayer open
circu its (snakes)
Hun d reds of meters of
m i n i m um width
i nterconnect
l ntralayer short
c i rc u its (combs)
H u n d reds of meters of
m i n i m u m spacing
interconnect
defects per uni t an:a where the un its of area are the
same as the critical area. Defect density is indepen­
dent of prod uct type, and depends only on the
cleanl iness of the fabrication. The spatial d istribu­
tion of defects is independent of product type
(depending on chip size 10) and depends to some
d istribu tion of defects on a wafe r 9 Small values of a
indicate that defects are more l ikely to be cl ustered
in isolated areas on the wafer. High values of a indi­
cate that defects occur across the wafer in a more
degree on the clean liness of the fabrication . 1 1
u n iform fashion. A single test pattern has multiple
Critical Areai Extractor
nu mber of fai l ing test structures on a test pattern
Integrated circuits an: designed using compu ter­
aided design (CAD) technology. The geometries that
eventual ly become the wire , transistors, resistors,
and other circuit elements are stored in a layout art­
work filL. Design R u le Check software calculates
critical areas for different defect types. The algo­
rithms for this usual ly involve Boolean algebraic
copies of the same test structure on it. Each of these
test structures is independently testable so that the
can be counted. From this data, a d istribution can
be created of n u m ber of occurrences as a function
of nu mber of fai ling test structu res. The mean and
variance of this d istribution can be computed and
then a can be calcu lated from the fol lowing equa­
tion (2). 1 1
a; = m /l(v; - m;)
operations and sizing of layout geometries.
One exa mple of a critical area extraction is to
define a temporary layer to be the intersection
of the M l , polysil icon, and M l contact layers. The
number of geometries on this temporary layer is the
total number of M l to polysil icon contacts on the
product chip. The critical area of the M 1 to poly­
silicon open-circuit defect type is related to this
nu mber (there are additional steps to the algorithm
that elim inate counting redundant contacts). Other
algorithms are used to calculate the critical areas
for the other defect types. 12. 1.1
Tahk 4 gives the units of critical area fo r the fou r
basic defect types. These area extractions are made
on both the test pattern and the product chip and
are used in the negative binomial yield equation as
the critical areas.
(2 )
where m; is the average number of failing test struc­
tures per test pattern, v; i s the variance of fa iling
test structure per test pattern , and i i s an index indi­
cating the particu l ar defect typ e.
Defect Densityi Calculation
The binomial yield equation can be solved for
defect density; if the yield i' a;, and critical area ;
are known. Therefore, the defect density can be
calc u lated for the test pattern because the yield ;
and a ; are measured d irectly at electrical test, and
the critical area ; is calcu lated from the area extrac­
tion software.
Product Yieldi Calculation
Defect density; and a; are assumed to be the same
for all prod ucts and rest patterns. This assu m ption
Computation ofAlphai
states that the amount and the spatial d istribution
The negative binomial yield model requires the cal­
of defects are not product dependent. Their values
culation of a parameter that is usually referred to as
are established from the electrical testing of the test
As stated previously, a; is a measure of the spatial
pattern. The critical area ; is product specific and
a.
90
Vol. 4 No. 2
Spri11g
1992
Digital Technical journal
A Yield Enhancement Methodology for Custom VIS/ Manufacturing
determines how defect densityi and a i affect the
yield for that particular product. The critical area i
for the product is extracted from the layout art­
from which the values for yield i have been removed
is given in Table 5. Information from the defect
pareto is relayed to process and design engineering
work file. These three values are then inserted into
to complete the process enhancement and design
the negative binomial yield equation, and the yield i
for manufacturability.
of the product is solved.
Composite Layer Yield Model
Defect Pareto
The modeled yields of each defect type i n the single
After a l l the values for yield i have been calcul ated,
layer yield model are assumed to be independent of
they are l isted in order of increasing yield to create
each other. The product of the modeled yields rep­
the defect pareto. An example of a defect pareto
resents the overal l modeled yield.
Table
5
Ty pical Yield a n d Defect Pa reto
Defect
Do
Units
M2 short-circuited l i nes
1 1 .350
1 00 m
Estimated
Yield
O.xxx
0 . 1 08
O. xxx
.490
O.xxx
M1 /polysil icon contact open c i rcuits
3.397
ppm
M1 short-circuited l i nes
4.949
1 00 m
0.1 50
O.xx x
Local interconnect-polysil icon
short-circuited l i nes
5.376
1 00 m
0.059
O. xxx
M 2 open l i nes
0.941
1 00 m
1 .000
O. xxx
Active area p+ short circuits
4.804
1 00 m
0.098
O.xxx
Active area n+ short ci rcuits
4.700
1 00 m
0.080
O . xxx
M2/M1 contact open circuits
0.266
ppm
0.061
O.xxx
M 1 /n+ contact open circuits
0.1 1 3
ppm
0.056
O.xxx
Polysilicon short-circuited l i nes
1 .739
1 00 m
0.035
O.xxx
Polys i l icon open l i nes
2.61 0
1 00 m
0.027
O.xxx
M 1 /p+ contact open circuits
0.1 1 7
ppm
0.034
O.xxx
Active area n+ open c i rcuits
0.831
1 00 m
0.01 7
O.xxx
Local interconnect-n+ active area
short-circuited l i nes
1 .674
1 00 m
0 . 1 01
O.xxx
Active area p+ open c i rcuits
0.458
1 00 m
1 . 000
O.xxx
Dielectric 2 (capacitor)
0.31 6
cm2
0.045
O.xxx
Local i nterconnect short-circuited l i nes
2.1 1 0
1 00 m
0.1 09
O.xxx
Dielectric 3 (capacitor)
0.260
cm2
0.030
O.xxx
Dielectric 1 (flat)
0.047
cm2
1 .000
O.xxx
Local interconnect open l i nes
0.397
1 00 m
1 .000
O.xxx
M3/M2 contact open c i rcuits
0.201
ppm
0.027
O.xxx
area contact open c i rcu its
0.007
pp m
0.004
O.xxx
M3 short-circuited l i nes
3.596
1 00 m
0 . 1 46
O.xxx
M1 open l i nes
0. 1 38
1 00 m
0.003
O.xxx
Dielectric 2 (worst-case step)
0.661
cm2
0.055
O. xxx
M3 open l i nes
2.338
1 00 m
0.01 3
O.xxx
Local i nterco nnect- polys i l icon-active
Notes:
1 00 m = defects per 1 00 meters of length
ppm = parts per million defective
em' = defects per square centi meter
Digital Technical journal
Vol. 4 No. 2
SjJring 1992
91
Semiconductor Technologies
y ield = product of yield Ji1r all values of i
where i is an index describing each ckfect type.
The continuous product ion of h igh-yield product
the
The probe fail ure bins can be ana lyzed by several
methods. Composite wafer mapping is useful for
Product Chip Yield Analysis
wafers requires u nderstanding
labeled " $ $ ." These bins can be analyzed fo r trends
and commonal ities.
impact of
design, manufacturing, processing, process equ ip­
ment, and testing on yield . Root cause analysis of
a low-yield wafer with 54 chips, each containing
1 .7 million transistors, capable of operating at
200 t\' I Hz, is a difficult task. The fol lowing section
describes the testing methodology applied by the
product yield enhancement engineer.
After the wafers have completed the fabrication
process, all die are electrical ly teste d . The electri­
cal test code starts with simple continuity checks
(open circui t s and short circu its) on s ma l l areas
describing the fai l ure pattern. A composite map
graphically displays a probe bin for the entire lot as
a percent of d ie fail ing a certain type of rest.
Composite maps can qu ickly show if lots have com­
mon causes of fai l u re. For example, after extensive
ana lysis, the cause for a certain fai l ure pattern can
be correlated to a certain process operation.
Analysis of composite maps from other lots quickJy
reveals if they were affected by a similar process
operation. An example of a composite m ap for sili­
cide over growt h is shown in Figure 8.
BINCODES: CS
2
of the die. Tests that require l imited functionality are
3
performed early in the test sequence. Electrical tests
incrementally cover a larger area and more function­
ali ty. First-pin fail identity and parameter value (volt ­
2
age, current, o r test vector) are retained for each die.
3
The code written in the test sequence allows the
4
yield engineer to determine the cause of the failure.
5
For example, high current failures (short circuits)
can often be correlated to metal short circuits
caused by inadequate metal etch, poor planariza­
tion, or partick deposition. Electrical testing that
6
0
13
13
0
0
13
0
13
0
0
7
0
0
13
38
4
13
0
0
0
88
5
0
13
13
8
0
13
0
88
88
0
88
88
75
0
0
1 00 1 00
88
1 00 1 00
75
7
6
75
50
50
0
0
13
9
0
13
0
0
0
0
can identify a specific area within a d ie as the prob­
able fa ilure site aids in analysis of the failure . If pin 1
is short circuited to ground ( � 1), i t is probable that
the cause of the fai lu re is in close proxjmity to
Figure 8
Composite �W"ifer Map for Silicide
over Growth on CMOS-4 Process
pin 1 . Visual inspection and the scanning electron
microscope (SEM) are often employed for this type
Another method to correlate probe results to a
of analysis. If no cause i s fou nd , more i n tensive fail­
process change is to plot probe data on a cumula­
lilT
tive sum control chart. Typ ical ly, probe data is plot­
analysis i s p u rsued.
functional testing in a production environment
limits the amount of data that can be stored. The
ted as a response to the sequence of a lot (wafers
are processed i n lots of 22 each) through fabrica­
testing often stops at t he first test failed. It is there­
tion processes. The effectiveness of employing
fo re important that initial functional tests require
these control charts relies on two factors: the
only minimal fu nct ional i ty. Stored data that iden­
sequencing of lots must be randomized from one
tities the fai l i ng pin and test vector provides the
fabrication
abi l i ty to perform commonal ity studies on manu­
changes must be meticulously recorded. When a
process
to
the
next, and
process
facturing data. If a failure mode can be isolated,
slope inflection point correlates to the date of
analysis is simplified.
a process change, a highly probable cause for
Tests that are similar or specific to an area are
change in probe data can be identified, as shown in
grouped in bins. For example, a die that fails
Figure 9. If two or more systems are used for a par­
because of a short-circu ited pad is col lected in a bin
a
ticu lar process inrerchangeably, a control chart fo r
vector Ji1r floating
each system can be generated. For example, an alu­
point is stored in a bin labeled "" !'BOX " ( functional
minum etcher with low etch rate may consistently
failure in the mod u le where the floating point is
cause more cs probe fai lures than another alu­
processed); fu lly functional die arc stored in bins
minum etcher with high etch rate.
labeled "CS"; a d ie that fails
92
Vol. 4
.Vu.
2
Spring 1')')2
Digital Technicaljourual
A Yield Enhancement Methodology for Custom VLSI Manufacturing
®
60
50
40
@)
30
0
--'
w
;;::
1-z
w
u
a:
w
a._
A
20
10
0
-10
-20
-30
--40
-50
-60
-80 �------
LOTS IN PROCESS SEQUENCE
KEY:
CD
@
@
@)
PLANARIZATION I M PROVEMENT
D I E LECTRIC LAYERS 2 AND 3 CHANGED
DIELECTRIC LAYER 1 THICKNESS INCREASED
METAL 2 PARTICLE LEVEL INCREASED
Figure 9
CMOS-4 Yield Cumulative Sum Control Chart
Record ing lot process history is crucial in per­
The CMOS-4 process was debugged and qual ified
forming probe yield analysis. As a CMOS-4 wafer is
using a 128Kb SRAM. As discussed earlier, the SRAM
processed, hundreds of parameters are recorded.
is designed to offer increased analysis capab i l ity
All of these process parameters must be easily
accessible to software tools for analysis. Equally
over custom-designed CPUs. As the process con­
important, all the process parameters must be
from projected product yields. The yield engineer
under statistical process control. Poor probe yields
needs to understand the simi larities and differences
verts to prod uct, actual prod uct yields may differ
with occasional high-yielding lots indicate a poorly
among the chips. The SRA.M yield determined by the
controlled process. Probe analysis of a process with
pertains only to a single lot. Probe analysis of a
lower level processes (u nder M l ) is very similar
to the Alpha 210 6 4 chip yield because large areas
of the 2106 4 chip are designed the same as the
process u nder statistical control normally identi­
fies limited causes and represents the majority of
SR.AM. The upper layers of the 2106 4 chip can devi­
ate in layout from the SRAM and can respond differ­
the lots.
Commonality studies to investigate process­
ently to variations in the process.
Due to the circuitry of the CMOs-4 die, foreign
related causes for d ifferences between high- and
particle control and monitoring are critical. Nearly
low-yielding lots are often performed. Software
every process step deposits some particles. Particle
tools extract all process-related parameters con­
size and type are important parameters to correlate
cerning the two groups of lots. Each group is ana­
to yield. These parameters are correlated with the
lyzed for a common parameter that is different from
aid of the defect density structures. The final and
the other group. Commonality studies that identify
most significant correlation must be done to prod­
wide variability often reveals several causes and
processing differences between high- ancl low-yield
uct yield. Visual inspection and characterization
lots are often confirmed by experimental analysis.
of particles on fa il ing die compose the first- order
Digital Tech1lical]ottr71al
Vol. 4 No. 2
Spring 1992
93
Semiconductor Tech nologies
analysis. The second order of analysis is to correlate
particle size and d istribution (measured with au to­
and targets for each layer are used to proj ect a yield
estimate for the foreseeable l ift· of the product.
mated laser inspection tools) to yield . This analysis
I f the suhhlock is new, an estimate can often
define� the type of particles and the particle size
be made by u nderstanding the type of logic or cir­
that w i l l impact yield most signi ficant ly. Furt her­
more, it uncovers the sources of the defects that
dom logic, buses, or memory-like (e .g. , in a cache
cuitry being considered. lf the circuitry is pure ran­
most contribute to yield loss. Yield engineers can
controller) , then the critical area estimate of that
then prioritize defect reduction.
subblock wil l assume the artwork properties of
these circuits. This scenario may appear overly
Yield Forecasting
The yield model is often used d uring feasibility
complex, however, when the chip being estimated
is 2 to 3 cm 2 in footprint area and is tightly packed
studies to foreca�t the yield of a planned chip prior
with minimum ground rule artwork, this complex
to its design. Typical ly, the process of predicting
procedure is necessary for a reasonable yield esti­
the yield of a new planned integrated circuit chip
mate. A reasonable estimate is considered to be
starts by examining the basic layout of a chip. As
within ::+::: 20 percent of the actual y ield; if one used
shown i n Figure 10, the structure of a chip is parti­
more simplified approaches, errors u p to 300 per­
tioned into functional subblocks. By understanding
cent could easily occur.
how much a subblock wil l change from its previous
Figure 11 compares the forecasted y ield and the
use, yield engineers can estimate changes to the
actual yield for the 128Kb SRAM and the Alpha 21064
new subblock. Frequent ly, a subblock will not
microprocessor. The vertical axis is the forecasted
change enough to cause its yield to be significantly
chip yield norm al ized to a relative scale. The open
different from the layout used in a pr<.:vious chip.
circle is the 128Kb SRAM chip forecast, and the
These subblocks are available in a l ibrary of artwork
closed square is the actual SRAM yield . The AI ph a
layouts kept in the form of their extracted critical or
model yield is plotted with a closed circle, and the
susceptible areas. 1"
Alpha actu a l yield is plot ted with an open square.
CLOCK
The estimates were made approximately 4 to 6
_I
months prior to the product chip being prototyped
in the fabrication facility The projected estimates
CENTRAL
FLOATI NG-POINT
ACCELERATOR
CPU
CACH E CONTROL
I
I
Figure 10
I
DATA CACHE
I
CAC H E TAG
I
I
SimjJlijiecl ChijJ Layout
(jj
x
<(
(.9
0
�
0
...J
w
;;::
w
>
f=
<(
...J
w
a:
1z
w
()
a:
w
a.
2
In the next step, the critical areas estimated from
each subblock are used to obtain an estimate for the
KEY:
entire chip. Cache memories are usual ly added to
G-€l
the random logic areas by taking critical areas of
memory support circu itry and adding the mul tipl i­
cation of the cell areas to the number of total bits
required in a cache array Once the total critical area
estimates are complete, the defect density goals
94
11-11
�
G-El
3
4
4
6
7
8
DATE
9
10 11
1 2 13 14
1 28Kb SRAM FORECAST
ACTUAL 1 28Kb SRAM
ALPHA MODEL Y IELD
ACTUAL ALPHA YIELD
Figure 11
Vol.
5
No. l
CMOS-4 Yield Forecast Projections
Spring 1992
Digital Tecbnicaljournal
A Yield Enhancement Methodology for Custom VLS/ Manufacturing
are relatively close to the actual chip yields during
the same manufacturing time frame .
Table 6
D0
Defect Density Types
Para meter
Units
Redundancy Yield Models of Caches
and Memory Chips
Dielectric 1 (flat)
Dielectric 2 (worst-case step)
cm2
Relatively large cache RAMs are being used today on
Dielectric 2 (flat)
cm2
microprocessor chips to increase the processor's
bandwidth performance. The bit capacity typically
Polysilicon short circuit
1 00 m
ranges from JOOKb to 1 .5 megabits. This cache
memory size can use up to 50 percent of the entire
M 1 short circuit
1 00 m
M2 short circuit
1 00 m
Polysilicon open circuit
1 00 m
M1 open circuit
1 00 m
footprint area of a microprocessor chip. Because
of these d imensions and bit capacities, on-board
cm2
M2 open circuit
1 00 m
N-gate area
cm2
ber of repairable faults is less than or equal to the
number of available spares, the ch ip can be repaired
P-gate area
cm2
Gate meander
1 00 m
to a ti.I I Iy ti.mctioning device.
Gate bi rd's beak
1 00 m
M2/M1 contact
ppm
spares are sometimes used to provide fau l t toler­
ance for the processor's cache memory. If the num­
Redundancy Yield Model
The red undancy yield
M1 -polysil icon contact
ppm
model consists of two sets of parameters. One
M 1 - n + contact
ppm
set characterizes the process and the other set
M 1 - p+ contact
ppm
Active area short circu its n+
1 00 m
Active area open circu its n+
1 00 m
Active area short circuits p+
1 00 m
Active area open circuits p+
1 00 m
Di electric 3 (cap)
cm2
Dielectric 3 (wo rst-case step)
cm2
describes the product. The set of defect densities
modeled i n an older generation process is shown
in Table 6; this set of parameters characterizes the
random defects i n the process. These defect densi­
ties can be expanded in mathematical for m to
A = Ac - D
(3)
where A is a vector representing the four fau lt types
M3/M2 contact
ppm
of memory pattern bit failures as given be low.
M3 open circuit
1 00 m
M3 short circuit
1 00 m
1 . Single bits
Notes:
2. Single word lines
1 00 m = defects per 1 00 meters of length
ppm = parts per million defective
3. Single bit lines
cm2 = defects per square centimeter
4. Chip kil l
These fault pat terns are the parameters describ­
ties of failure have been analyzed using Monte Carlo
ing the random faults on the memory portion of the
product. �c is t he critical area matrix of the prod­
simulation techniques using the VLASIC yield simu­
modeled layers. T h e critical. area matrL'< relates the
cache memory can be simulated to the number of
total bits required by the design. The random logic
uct, and D is the vector of defect densities at a l l
defect densities modeled i n the process to the
cache memory circuit fau l t pat terns of the product
as given in Table 7 These arrays of numbers repre­
lator as described by Wal ker) By establishing
l i braries of critical areas for different circuits, a
sent the sensitivity of the cache memory circu it
critical areas are thus lu mped into the chip-kill
category as seen in Table 7 In this fashion, the mean
number of fails per chip for each circui t fau l t type
to random defects. The sensitiv i ty to defects is
can be computed. In many circumstances, the
obtained
by
calculating
the
critical
area
as
mean number of fails per chip can be obtained
described in the yield model section of this paper.
from existing memories that are similar in design
The probabi l ity of failure for a given size defect
to the future cache memory Fault statistics can
is the fraction of defects of that size which has been
then be collected , and the failure distribution
can be modeled by using the negative binomial
determined to have caused a fau l t . These probabili-
Digital Tech11icaljournal
I'll/. 1 No. 2
Spring 1992
95
Semiconductor Technologies
Table 7
Critical Area M atrix for Processor C h i p with Cache Me mory
Defect Type
Single
Cells
Dielectric 1
0.01 6560
Dielectric 2 (worst-case step)
0.088800
Fau lt Ty pe
Single
Single
Bit Lines
Word Lines
Chip
Ki l l
0.020250
0.000050
0.002430
Di electric 2 (flat)
0.1 41 500
O.D1 81 67
0.01 6060
0.033670
M1 short circuit
0.052460
0.1 701 50
M2 short circuit
0.060980
Polysil icon open circuit
0.048640
M1 open circuit
0.065060
M2 open circuit
0.065600
Polysilicon short circuit
0.000304
0.0001 49
0.1 57040
0.1 1 4590
0.89041 5
0.000374
0.0001 75
0.1 1 6670
N-gate area
P- gate area
0.01 4060
0.023650
Gate meander
0.037340
0.026550
Gate bird's beak
0.009220
0.002770
M2/M1 contact
0.230700
0.1 08990
M 1 - polysil icon contact
0.296600
0.082260
M 1 - n+ contact
0.426600
0.227200
0.1 42300
M 1 - p+ contact
0.21 8400
Active area short ci rcuits n+
0.002733
0.01 2060
Active area open circuits n+
0.002730
0.01 2060
Active area short circuits p+
0.01 2420
0.000081
0.077647
Active area open c i rcuits p+
O.D1 2420
0.000081
0.077674
Dielectric 3 (cap)
Dielectric 3 (worst-case step)
0.003353
M3/M2 contact
0.021 534
M3 open circuit
0.000250
M3 short circuit
0.000335
dist ribution
w
i th A ami
a
as parameters. 10· 1 6 17 This
past. Figure
12
shows the yield as a function of
is accomplished through use of the p roba b i l i ty
redundant spares and sets of banks where the
mouel given as
total set is eight
P(X = x )
J'(.x + a)
=
x
.
1 [ (ex)
(A laY
( 1 + A lex)"'
(4)
· (.(
A nonl inear least - squares technique is used to fit
tlu: parameter!; to the ob!:icrvcd distribution; t hese
parameters can be used instead of defect densities if
desired . 17 Thi s :.tlternative can give the model more
fl exibi l i ty, depend ing on which data is most appro­
priate to us c f<>r an estimate. Exam ples of the fi t ted
distributions of single
cd l fa i l u res, column f:.t i l u res,
and double column failures are given in Table H.
The estimates of c h ip yield using difterent combi­
and the dcsired number of good
banks is at least si-x out of e i ght . These computa­
tions are aga in perfor med using a m o del descri bed
by Stapper. 10· 1'· 1R The yields were csti mated approxi­
mat ely one year before actual manufacturing data
existed for that product. These types of analyses are
p:.trt of the feasibility stud ies that help the design
engineers determ ine thc opt i m u m product yi el d s .
This model has been expanded and mod iJied to
include c l ustering of fa ults within a chip when a
chip becomes rather large. I R
Design for Manufacturability
nations of redundant spares and turning off banks
The yield equation (1) shows that yield is
of the c:.tchc memory have also been U!:ied i n the
of bot h critical area and defect density. Since criti-
Vol. 4 No. 2
Spring
1992
a
function
Digital Technical journal
A Yield Enhancement Methodology for Custom VLSI Manufacturing
Table 8
Faults
per Chip
Fitted D istri butions of SRAM
Functional Fa i l u res
Observed
Predicted
Resid uals
0
0.626866
0.629694
- 0.002829
1
0.1 99627
0. 200209
- 0.000582
2
0.082090
0. 087091
- 0.005001
3
0.033582
0.041 282
-0.007700
4
0.01 6791
0.020374
-0.003583
5
0.01 3060
O.D1 0293
0.002766
0.0021 82
6
0.007463
0.005281
7
0.003731
0.002739
0.000993
8
0.007463
0.001 432
0.006031
9
0.005597
0.000753
0.004844
10
0.001 866
0.000398
0.001 468
11
0.000000
0.00021 1
- 0 .00021 1
12
0.000000
0.0001 1 3
- 0 .0001 1 3
13
0.001 866
0. 000060
0.001 806
14
0.000000
0. 000032
- 0.000032
15
0.000000
0.00001 7
- 0.00001 7
F inal
Parameter Value
Lower
Standard 95%
Upper
95%
0.709781 0.01 9695
0.667540 0.752021
0. 575939 0.02741 0
0.51 71 50 0.634728
Figure 12
Chip Yield Estimate with Spare
and Bank Redundancy
quently cause defects, the work to date has focused
on reducing the critical area of M2 short circuits.
cal area is extracted from the circu it layout, it fol­
lows that changing the layout can affect the yield.
However, the total number of d ie manufactured on
a wafer is also a function of area, and changing
the layout can also affect the number of d ie that can
fit on a wafer. The design for manufacturabil ity
(DFM) method quantitatively analyzes these two
relationships and max i m izes the number of good
die per wafer.
Figure 13 is a flowchart that shows the role the
single layer yield model plays in process enhance­
ment and design for manufactu rabi l it y. Two loops
have been ott t l ined in the d iagram: the process
enhancement loop and the design for manufactura­
bility loop.
DFM currently uses three software tools: the
yield mode l , the cell counter, and the die counter.
The yield model has been made available to design
engineers for their use to model yield based on lay­
out artwork. This software tool can be used in the
design phase to evaluate the manufacturabil ity of
chip subblocks. The subblocks that produced the
The idea is extendible to other circuit features such
as l ine widths and contacts.
The cell counter routine counts the number of
t imes that a specified portion of a circuit is repeated
in a product layout. Critical area extractions can
require large amounts of CPU time, but the c e l l
counter a l lows t h e extraction t o be performed on
a sma l l portion of the chip qu ickJy. The resu l t can
then be multipl ied by the resu l t of the cell counter
to model the yield . The cel l counter can also be
used during layout to determine the effect on yield
of increasing or decreasing the size of the product.
This technique is especial ly effective with very
repetitive layout, for example, SRA.Ms.
The third software tool included in DFM is the die
counter. This rou t i ne counts the number of d ie or
chips that can fit on a wafer given the d ie size. The
die counter is used during layout along with the
yield model to optimize the number of good chips
per wafer.
Conclusions
greatest nu mber of good die per wafer can then be
Enhancement of integrated circ u i t yield a t times
used in the chip design. Since 1\12 short circuits fre-
requires many and diverse analytical techniques
Digital Technical journal
Vul. 4 No. 2
Spring 1992
97
Semiconductor Technologies
.- --- -----
�
PROCESS
MANUFACTURING
�
YIELD
ENGINEERING
DESIGN LAYOUT
�
'
TEST PATIERN
ELECTRICAL TEST
�
�
PRODUCT LAYOUT
ARTWORK FILE
-- -- --- - --
�
CELL
COUNTER
y
DIE
COUNTER
l
PRODUCT YIELD ; CALCULATION
I
�
COMPOSITE YIELD MODEL
�
I
PROCESS
ENHANCEMENT
LOOP
DEFECT PARETO
Figure 13
�
I
I
I
I
I
I
I
I
I
Do ; CALCULATION
�
�
t
CRITICAL AREA;
EXTRACTOR
COMPUTE YIELD;
AND ALPHA;
�
'
TEST PATIERN
LAYOUT ARTWORK
FILE
I
GOOD DIE PER WAFER
l
DESIGN FOR
MANUFACTURABILITY
LOOP
- ----- ---- --------
Yield Model/Designfor Manufacturability Flmu Diagram
for the detection and elim ination of yield- limiting
2.
P. Gangatirkar,
R. Presson, and L. Rosner,
mechanisms. The techniques described in this
"Test/Characterization Procedures for High
paper d iscuss some of the analytical tools necessary
for the successful manufacture of very large, com­
Density Silicon RAMs," Proceedings of the
International Solid-State Circuits Conference
plex 0105 d igital circuits. These tools have been in
(February 1982).
use at Digital for approximately ten years, and the
tool set continues to evolve with each new genera­
tion of technology. As these techniques are refined
and new ones are developed, the overa l l u nder­
standing of integrated circuit yield and yield loss is
3.
88 -46, October 1988.
4.
5.
98
" Exp erience with
the VLASIC
Research Report No. CMUCAD-90-41 , Septem­
D. Dobberpuhl et a l . , "A 200MHz 64b Dual­
tional Solid-State Circuits Confe1·ence Digest
of Technical Papers (February 1992): 106- 107.
D. Wal ker,
System in Defect Probabil ity Prediction," CMU
References
Issue CVIOS Microprocessor," IEEE Interna­
D. Wal ker, Yield Simulation for Integrated
Circuits (KJuwer Academic Publ ishers, 1987).
increased.
1.
D. Wal ker, "Yield Analysis for Fau l t -Tolerant
Arrays," CMU Research Report No. CMUCAD-
ber 1990.
6.
]. Cunningham, "The Use and Evaluation of
Yield Models in In tegrated C ircuit Manufac-
Vol.
4 No. 2
Spring 1992
Digital Technical]out-nal
A Yield Enhancement Methodologyfor Custom VLSI Manufacturing
turing," IEEE Transactions on Semiconductor
Manufacturing, vol. 3, no. 2 (May 1990).
7
18.
C. Stapper, A. McLaren, and M. Dreckmann,
"Yield Model for Productivity Optimization of
VLSI Memory Chips with Redundancy and Par­
tially Good Product," IBMjournal of Research
and Development, vol. 24, no. 3 (May 1980).
19.
C. Stapper, "Small Area Fault Clusters and
Fault Tolerance in VLSI C ircuits," IBMjournal
of Research and Development, vol. 33, no. 2
(March 1989): 174- 177
T. Okabe, M . Nagata, and S. Shimada, "Analysis
of Yield of I ntegrated C ircuits and a New
Expression for the Yield," Electrical Engineer­
ing injapan, vol. 92 (December 1972).
8.
C. Stapper, "LSI Yield Modeling and Process
Monitoring," IBM journal of Research and
Development, vol. 20 (May 1976).
9.
A. Rogers, Statistical A nalysis of Spatial Dis­
persion (London: Pion Ltd . , 1974).
10.
C . Stapper, "Large Area Fault Clusters and
Fault Tolerance in VLSI C ircuits: A Review,"
IBM journal of Research and Development,
vol. 33, no. 2 (March 1989): 162- 173.
11.
R. Collica, "The Effect o f the Number of
Defect Mechanisms on Fault Clustering and
its Detection Using Yield Model Parameters,"
IEEE Transactions on Semiconductor Manu­
facturing, vol. 5, no. 3 (August 1992).
12.
]. Pineda de Gyvez and ]. ]ess, "Systematic
Extraction of Critical Areas from IC Layouts,"
International Workshop on Deject and Fault
Tolerance in VLSI Systems (October 1989).
13.
A. Ferris-Prabhu, " Modeling the Critical Area
in Yield Forecasts," IEEEjournal ofSolid-State
Circuits, vol. SC-20, no. 4 (August 1985).
14.
C. Stapper, "The Effects of Wafer to Wafer
Defect Density Variations on Integrated C ir­
cuit Defect and Fault Distributions," IBMjour­
nal of Research and Development, vol. 29,
no. 1 Oanuary 1985).
15.
A. Ferris-Prabhu, "Role of Defect Size Distribu­
tion in Yield Modeling," IEEE Transactions on
Electron Devices, vol. ED-32, no. 9 (September
1985).
16.
S. Kikuda et a!., "Optimized Redundancy
Selection Based on Failure-Related Yield
Model for 64-Mb DRAM and Beyond," IEEE
journal ofSolid-State Circuits, vol. 26, no. 1 1
(November 1991).
17
C. Stapper, "On Yield, Fault Distributions,
and Clustering of Particles," IBM journal of
Research and Development, vol. 30, no. 3
(May 1986).
Digital Tecbnicaljoumal
Vol. 4 No. 2
Spring 1992
99
Daniel B. Jackson
David A. Bell
Brian S. DOJ'le
Bruce] Fishbein
David B. Krakauer
Transistor Hot Carrier Reliability
Assurance in CMOS Technologies
Hot carrier-induced degradation ofMOS transistors is an essential consideration in
the deuelopment of CMOS processes. Most empirical approaches that characterize
transistor bot carrier lifetime onlyprovide indications of relative degradation; they
do not make a connection between circuit operation and hot carrier degradation
under e.xperimental stress conditions. Digital's Advanced Senziconductor Develop­
ment Group has devised a physically based method for ensuring that the hot
carrier lifetime of transistors produced by a new process technology is acceptable.
The models used incorporate degradation under three voltage bias conditions and
allowfor the effect ofdominant manufacturing variations on transistor hot carrier
lifetime. The method also takes into account the sensitivity of the circuit design
to transistor hot carrier degradation. This bot carrier reliability assurance gives
developers the ability to predict circuit hot carrier lifetime and thus allows them to
maximize transistorpe1formance.
Hot carrier-induced degradation of meta l-oxide
Consequently, much work has been devoted
semiconductor ( MOS) t ransistors is an essential
to u nderstanding hot carrier degradation , with a
consideration in complementary metal-oxide semi­
COIH.luctor (CMOS) technology development. The
reduction of metal-oxide semiconductor field­
view toward developing accurate pred iction tech­
niques for both transistor and circuit l ifetime, such
as those discussed in this paper. 1 - H Substantial
effect transistor (MOSFET) channel dimensions to
progress has also been made in increasing the hot
micron and submicron sizes has placed increasing
carrier robustness of transistors through drain
importance on the rel iability of the gate oxide
j u nction design and gate oxide optimization; a d is­
and its interface with the u nderlying s i l icon .
cussion of this is beyond the scope of this paper
Achieving optimum transistor performance while
however. Most of the effort to date has been
maintain ing the necessary circuit rel iability is
focused o n n-channel transistors, where, for cur­
fundamental to manufacturing high-performance
rent
CMOS microprocessors.
degradation is more severe than i n p-channel
semiconductor process technologies,
the
The hot carrier-induced transistor degradation
transistors. However, for technologies with an
arises from the high l'nergy acqu ired by channel
effective channel length of less than 0.5 m icron
carriers, either electrons or holes, as they move
(Mm), p-channel hot carrier effects wil l become
from the MOSFET source to drain . ' - H The elec­
tric field through which the carriers move has
increased with succec.:d ing generations of process
increasingly important. This paper thus focuses on
n-channel transistors.
A nu mber of empirical approaches have evolved
technology because.: transistor dimensions have
in the semiconductor i ndustry to characterize tran­
scaled faster than opnating voltages. The high
sistor hot carrier l ifl'time. Most of tiKsl' approaches
energy of the channd (_·arricrs leads to a gradual
provide indications of rdative degradation, al low­
characteristi(.:S
ing the comparison of difkrent transistor Lksigns.
through charge trapping in the gate oxide and gen­
However, these schemes do not make a connection
degradation
of
the
transistor
eration of intnface statl's. This hot carrier-induced
between circ u i t operation and hot carrier degra­
degradation can bc.:comc.: large enough to cause.: cir­
dation u nder experimental stress conditions. The
cuit failure.
lack of a method for pred icting circuit lifetime can
1 00
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Transistor Hot Carrier Reliability Assurance in CMOS Technologies
lead to a conservative approach of compromising
and become energetic or "hot:' Since the mean free
transistor performance in order to improve transis­
path of an electron in silicon is small, approxi­
tor hot carrier lifetime, whether this improvement
mately 60 angstroms, most electrons lose the
is warranted or not.
excess energy they acquire by moving through
fo r determining whether the hot carrier-induced
the high field region via collisions with l a t tice
phonons. However, some fraction of the electrons
degradation in transistor characteristics is accept­
wil l traverse the high field region without su ffer­
This paper presents a physically based method
able. The procedure is d ivided into two parts, an
ing enough coll isions to lose a l l of the energy
experimental measurement of transistor degrada­
gained from the electric field. These e lectrons are
tion, and a determination of the maxim u m permis­
the hot carriers that cause degradation in transistor
sible transistor degradation for continued circuit
characteristics.
operation. For convenience, we will refer to both
transistor l ifetime and circ u i t l ifetime. The circu i t
lifetime i s the length o f time a circuit wil l operate
GATE
in conformance to its stated specifications. The
transistor l ifetime is the time it takes a transistor
u nder stress to reach a chosen degree of degrada­
tion. The Circuit Considerations section shows that
these two lifetimes are not necessarily the same. In
practice, of course, it is the circuit l ifet i me that is
important. A significant contribution of the work
DRAI N
JUNCTION
described in this paper is the ability to predict cir­
HIGH FIELD
REGION
c u i t l ifetime from transistor l ifetime.
The method begins with experimental measu re­
ments of transistor degradation u nder static stress
conditions, which resu lts in three fundamen t a l l y
d ifferent types o f dam age t o t h e transistor. Using
these measurements, a model is developed that
Figure 1
Schematic of a Transistor in
Saturation Showing the Device
Biased into Pinch-off
a l lows the determination of transistor lifetimes
u nder dynamic bias conditions for an assumed max­
Under conditions that generate hot carriers, a rel­
i m u m acceptable degradation. Also, for a given cir­
atively large nu mber of "lucky" electrons (those suf­
cuit, the method specifies the worst-case transistor
fering few coll isions) can gain the 1. 7 electron volts
produced as a resu lt of manufacturing process vari­
in energy necessary for electron hole avalanche gen­
ations in such parameters as gate oxide thickness
eration. The generated holes move out through the
and channel length. The circuit-dependent part of
substrate and can be measured as a substrate cur­
the method relies on two important quantities: the
rent IB' Fewer of the l ucky electrons will gain the
set of worst -case, time-varying biases seen by tran­
3.1 electron volts in energy necessary to overcome
sistors i n the circuit, and the maximum amount of
the silicon/si l icon dioxide (Si/Si02) barrier and pass
transistor degradation that the circuit can tolerate
into the gate dielectric. At sufficiently high voltages,
and stil l remain fu nctional.
the holes created in the avalanche can also have
enough energy to be injected into the gate dielec­
Physical Mechanisms-Measurement
of Transistor Degradation
Hot carrier degradation in a MOS transistor is usu­
a l ly local ized to the region where most of the volt­
age drop between the drain and the source occurs,
i . e . , between the pinch-off point and the drain j unc­
tion, as shown in Figure 1 . The size of the high field
tric of the transistor. Because the avalanche-gener­
ated holes can be measmed as a substrate current,
substrate current is frequently used as a measure of
the driving force for hot carrier degradation. I t gives
a convenient measure of the n u m ber and energy of
high energy carriers in the pinch-off region.
region near the drain junction is typical ly on the
Location of Hot Carrier Damage
order of 0 . 1 J.LID. In this region, the charge carriers i n
The injection of charge into the oxide is, by itself,
t h e inversion layer are accelerated b y t h e high field
not a cause for concern. Measurements u s i ng the
Digital Technical journal
Vol. 4 No. 2
Spring 1992
101
Semiconductor Technologies
floating gate technique, for example, show that this
current is on the order of picoamperes per JLm of
where the constants K and n are empirically deter­
mined from experimental data; the value of n is usu­
gate width for electron injection at high drain volt­
a l ly between 0.3 and 0.7 From curves such as those
ages Vos -� For hole injection, the current is several
orders of magnitude smaller.
shown in Figure 2, it is possible to interpolate or
However, a problem does arise from the s m a l l
obtain the transistor l ifetime at a given voltage. The
fraction o f the injected charge trapped in t h e oxide,
choice of the transistor lifetime criteria, in this case
Nn.,·• and the interface states generated by the hot
carriers, Nw The oxide and interface damage causes
extrapolate to a certain level of degradation and
5 percent change in
IDSAT>
i s discussed more ful ly
in the C ircuit Considerations section.
changes in the l i near region transconductance gm
and/or the threshold voltage
saturated drain current
Iosu
IosA T
as wel l as in the
VT ,
of the transistor. The
of the transistor is the drain current with the
gate and urain voltages equal to the positive power
supply voltage �JD and the source voltage equal to
the negative power supply voltage
Vss·
Oxiue traps and interface states that are uni­
formly d istributed along the channel can be identi­
fied from the poststress current versus voltage
characteristics. 10 A uniform N causes a change in
"
the subthreshold current characteristics, and a uni­
form Nux causes a shift in the
VT .
H owever, hot car­
rier stress dam age is not unifo rm . It is local i zed
to the region around the drain junction shown in
Figure 1 . The degradation resul ting from this
nonuniform damage is l imited primarily to gate
voltages Vc;s above V:r , i.e . , the drain current I0 is
reduced when 1-G·s is greater than VT> irrespective of
whether N . or !\�s is created . ' '
0.,
This interpretation is supported
by
two­
dimensional simulations. 12 The similarity in effects
of interface states and oxide traps has resulted in a
poor u nderstanding of the types of damage that
occur in hot carrier stn.:ssing and , in turn, has led to
diffic u l ties in predicting transistor l ifetimes u nder
dynamic stress conditions.
Figure 2 shows typical stress results at five different
drain voltages. The gate voltage in each case corre­
sponds to the peak substrate current. Although
many criteria are used to monitor stress damage,
I0 ,
the criterion used to
obtain the results presented in this section is the
change in the saturated drain c u rrent when both
V0s and "Gs equal VfJo • IDSA T · From a circuit stand­
point, IDSAT has been identified as one of the most
meaningfu l hot carrier monitors. Figure 2 i l lus­
trates that, at a l l stress voltages, the degradation as a
fu nction of time t obeys the equation
� 1 0.0
z
0
�
0
<X:
ffi
lU
0
...
1 .0
<(
(f)
.9
0 ·1 1L--
-1'-0
-
1':-:02
-1'-:
03:-- -1'-:04
_..J.:J._ :-L--
_
-
TIME (SECONDS)
1
:-.1.:
os
KEY:
DRAIN VOLTAGE (VOLTS)
-
e
-
·--0---
____,.____
-- -{3 - -
--e--
Figure 2
5.25
5.5
6.0
6.5
7.0
Degradation as a Function of Time
for Devices Stressed at Different
Drain Voltages
In the stressing of n-channel metal-oxide semi­
conductor transistors, there are three major regions
of damage in the gate vol t age range.4 - 8 The three
regions are distinguished by the charge injected
into the oxide du ring hot carrier stress. Figure 3
shows the gate current Ic as a function of gate volt­
age for a fixed drain voltage. The figure is divided
into three regions: the low gate voltage region,
region I , in which holes are the predominant com­
ponent of the gate current; the medium gate volt­
age region, region II, where both electrons and
holes are injected in approximately equal n umbers;
and the h igh gate voltage region, region III, where
electrons are the main current species. In region I,
(1)
102
i='
z
lU
u
cr
Three Types of Stress Damage
Kinetics of Damage Evolution
e . g . , VT , g111, and l i near
1 00.0
the predominant damage mechanism is the genera-
Vol. 4 No. 2
Spring 1992
Digital Tecbt1icaljour11af
Transistor Hot Carrier Reliability Assurance in CMOS Technologies
tion of electron traps in the bulk oxide by the
injected holes, Nox, h· In region II, the dominant dam­
age mechanism is the generation of interface states
N55. In region III, the dominant damage mechanism
is the generation of electron traps in the bulk oxide
by the injected electrons, Nox,e· In circu it operation,
all three types of damage occur. The relative impor­
tance of each depends not only on the voltage of
operation, but also on the relative rate of degra­
dation in each region as shown i n the section
Dynamic Hot Carrier Effects.
(obtained from data similar to that in Figure 2) on
Under these gate voltage conditions, transistor
lifetime can be predicted simply by establishing the
values of A �nd m in equation (2) and extrapolating
to the known substrate current at the operating
drain voltage.
I8.
- (c)
0.2
0.0008
g
Cii
0
2:,
-8
0 0006 �
LiJ
n.
::<
z
REGION
-10
ly-
REGION
REGION
II
Cii
LiJ
a:
0
i=
�
0
�
ffi
LiJ
:'S
0.1
0.0004
0
0
....J
� -12
0
::<
-14
8
....J
�
::::>
()
LiJ
�
I
(/)
LiJ
a:
:'S
S}
1z
LiJ
0.0002
�
a:
I­
(/)
co
::::>
(/)
-1 6
0
4
6
8
VGs (VOLTS)
VGs (VOLTS)
Figure 3
KEY :
t:>
0
Gate Current as a Function of
Gate Voltage for a Device Biased
at High Drain Voltages
The most widely recog­
nized damage mechanism occurs in the medi u m
gate voltage region a n d i s caused b y interface state
generation. Curve (b) in Figure 4 shows the amount
of degradation suffered as a function of gate voltage
for a series of devices stressed at a fixed drain volt­
age. The amoun t of degradation peaks at the same
gate voltage as the peak in substrate current,
depicted in curve (a). A direct relationship exists
between the degradation and the peak substrate
current during stress.3 The transistor lifetime under
stress conditions that generate inte rface states is
given by
D
Figure 4
Medium Gate Voltages
1
-
7
Nss
m
= A · /8
'
Vol. 4 No. 2
Degradation as a Function of the
Gate Voltage
10
Cii
�
0
8
�
6
()
LiJ
LiJ
::<
�
4
C)
2
u.
::::;
g
-5
-4
-3
LOG Is (AMPERES)
(2)
-
2
KEY:
where TNss is the transistor lifetime for stress i n
region I I o f Figure 3 , and !8 i s the substrate current.
The constants A and m are established by fitting
the experimental data; the value of m is usually
about 2.9. Figure 5 shows the dependence of TN."
Digital Technical journal
Is
THRESHOLD DEG RADATION AFTER STRESS
THRESHOLD DEG RADATION AFTER ELECTRON
INJECTION
Spring 1992
-·
6-·
--
DATA
FIT TO EQUATION (2)
Figure 5
Transistor Lifetime as a Function
of the Substrate Current
1 03
Semiconductor Technologies
Low Gate Voltages
The second type of damage
( 5)
resu lts from the injection of avalanche-generated
hot holes into the oxide. These i njected holes can
create neutral electron traps and hole traps. s-7. 15
The hole traps do not significantly affect transistor
l ifetime estimates and wil l not be discussed exten­
sively in this paper. It was not evident in previous
work that electron traps were being created. Since
the traps were neutral, their effects on the current
versus voltage characteristics were not observed.
However, if the stressed transistors of Figure 4 are
injected with electrons (brief stress with high l�;·s
where
T'""·'"
is the transistor l ifetime under stress in
region I I of Figure 2. I0 and Ic. are the drain and gate
currents during stress. The constant s c and I arc
determined by fits to the experimental data; the
value of I is usually about 2.9. Figure 7 ind icates
that the data for these stresses is linear, where �,1
is equ a l to l"0 1 , al lowing for the extrapolation and
prediction of transistor l ifetimes at different drain
voltages for this type of stress damage.
equal to VDS ) the neutral electron traps become
,
charged and contribute to the degradation, as
shown i n curve (c). The l ifetime clue to electron
6
traps created at low gate voltage obeys
(3)
where
T.v, ,_,
is the transistor l ifetime for stress in
region I of Figure 3 7 18 and 10 are the substrate and
drain currents during stress. The constants B and n
are determined by fits to the experimental data; the
value of n is usually in the range of 7 to 12. Figure 6
shows the transistor l ifetime for oxide trap damage
created a t low gate voltages as a fu nction of the
ratio of the substrate current to the drain current.
Using equat ion (3), the transistor lifetime for elec­
tron trap damage at low gate voltage can be fou nd
at any given drain voltage.
(j)
a
5
u
4
w
�
::9
•
x
0
z
2
.c:
.!:.
(.9
g
0
-2 �----�----_L_
-3
KEY:
--
High Gate Voltages Tlw third type of stress dam­
age in :'\MOS transistors occurs at high gate volt­
ages, under conditions that inject electrons into
-2
fl - -
LOG (Is f lo)
-1
0
DATA
FIT TO EQUATION (3)
Figure 6
the oxide (region I I I of Figure 3). This damage,
caused by electron trapping, was first ident ified
Transistor Lifetimefor Oxide
Trap Damage Created at
Low Gate Voltage
through the u ifferen t gradient obtained when plot­
Process Variations
ting degradation against time, as shown in Figure 6.
Similar to equation ( J ), equation (4 ) for the time
applied voltages on the transistor, i t is important to
behavior is
In addition to u nderstand ing the effects of different
consider the effect of manufacturing process vari­
(4)
where the constants D and k are determined by
fi tting the data; the value of k is usual ly between
abil i t y on hot carrier rel iability For a given source/
drain design, the major cause of that variability is
variation i n the transistor effective channel length
L ff" Of secondary importance is variation in the
e
dict a transistor l ifetime for this type of damage.
gate oxide thickness tax · From a performance stand­
point, it is desirable to make L �1 and tox as smal l as
That is, apply stress to a n umber of devices at difter­
curren t /DSAP ancl thereby the highest speed. How­
0.15 and 0.35. The same procedure used to obtain
the resu lts presented in Figure 2 was used to pre­
ent drain voltages (with �,_1 equal to
Vas ), and carry
out the extrapo lat ion/interpolation to obtain the
transistor l ifetimes at the operating drain voltage.
The transistor l ifetime can be plotted as
1 04
c
possible in order to get the highest saturated drain
ever, as
Lerr and t0x decrease,
the transistor l ifetime
rapidly decreases.
Because the damage region remains roughly con­
stant in size, as the channel length decreases a
Vol. 4 No. 2
Spring 1992
Digital Tecbllical journal
Transistor Hot Carrier Reliability Assurance in CMOS Technologies
u nder normal circu i t operation. Thus, although
-3
accurate transistor hot carrier lifetime models exist
for static bias conditions, dynamic bias concli­
tions must also be taken into accou n t l;. I6 Initial
0
attempts to predict transistor dynamic stress l ife­
0
times were based on quasi-static sums of transistor
static stress !ifetimes n. I H However, these initial pre­
0
ID
x
0
z
-t-
dictions were much longer than transistor l ifetime
measured u nder dynamic stress, because they only
(9 -6
g
considered the contribution from interface state
-
�
g---------�
8---------�
7----�
6
? L-_
K EY :
0
-
-
-
LOG (IG / Io)
degradation has generated much debate and numer­
ous explanations, the effect can be explained by
DATA
considering a quasi-static sum of the three damage
mechanisms detailed in the previous section. 19 21
FIT TO EQUATION (5)
Figure 7
Transistor Lifetimefor Oxide Trap
Damage Created for Equal Gate
and Drain Voltage Stress
larger fraction of the channel is damaged and the
transistor l ifetime is shorter. The direct effect of L err
on transistor l ifetime is approximately
(6)
where n1 is usua l ly abou t 3. 14 There is also a n
impl ied transistor l ifetime dependence on
L err
through the increase i n I8 that resu lts from a
smal ler L e w Tak ing this into account, the net depen­
dence of transistor l i fetime on effective channel
length is approximately
T cc
generation and omitted the effects of bu l k oxide
trapping. I 9. 2 0 W11 i le the enhanced dynamic stress
1 I L �rr ·
A n Accurate Dynamic Hot Carrier Lifetime Model
Our model, based on transistor static stress I ife­
times, d iffers from previous static-based models in
that i t takes into account all three types of damage.
D uring dynamic stress, with the drain voltage con­
stant at some high voltage, a dynamic gate voltage
subjects the MOSFET to the three types of damage
shown in Figure 8. (The stress waveform shown in
Figure 8 is for il lustrative purposes only and does
not necessarily reflect a n actual circu i t waveform.)
If the instantaneous values of ID ' I8, and Ic are
known for the dynamic stress, we can calcu late
the contributions of the three damage mechanisms
by integrating equations (2), (3), and (5) over the
time period T of the dynamic stress waveform. The
(7)
The effect of variations in t0x is accounted for i n
the changes in 18 , fc , and In that resu lt.
One of the keys to designing the source/drain
process is to optimize the moderately doped drain
(MOD) region for minimum peak substrate current
under operating conditions and thus m aximize
the hot carrier rel iabil ity. However, once the MOD
implant and diffusion process is fixed, the variation
in transistor hot carrier l ifetime for a given L e rr i s
VGs
process, transistor l ifetime variation is less than a
KEY:
relatively small. Typically, we find that for a stable
factor of two from one integrated circu i t manufac­
turing Jot to another.
LOW GATE VOLTAGE REGION
II
MEDIUM GATE VOLTAGE REGION
Ill
HIGH GATE VOLTAGE REGION
Dynamic Hot Carrier Effects
An accurate transistor l ifetime pred iction model
for hot carrier reliabil ity must consider the actual
stress conditions to which MOSFETs are subjected
Digital Technical journal
Vol. 4 No. 2
Spring 1992
Figure 8
Schematic Representation of
Dynamic Stress Waveform and
Damage Regions
1 05
Semiconductor Technologies
resulting expressions for the l ifetimes due to the
three physical mechanisms are given by
1
T\.-'·'
-
AJrI"' · dt
=
T
-
I)
B
(8)
'
(jJ
0
z
0
0
w
�
(9)
1 07
1 Q6
w
�
i=
1 05
a:
0
1(f)
iii
1 04
w
u_
:::;
(10)
Treating 1 I T as a da mage fu nction, the total
dynamic stress damage can be modeled as
I
--=
7Uynamic
1
-+
TN.,\,.
1
1
- + -- ·
7
Nox.IJ
7
NrH.('
z 1 03
�
a:
1-
1 02
(11)
Example of the Dynamic Hot Carrier Lifetime
Model To demonstrate the application of equa­
tion ( 1 1 ), consider the dynamic stress waveform
in Figure 8 applied to n-channel MOSfETs. 1-(,1 was
pu lsed between zero volts and �Js with 2-micro­
6
5
Vos (VOLTS)
7
8
KEY
0
D
QUASI-STATIC INTERPRETATION
OF N55 DAMAGE
ACTUAL MEASUREMENT UNDER STRESS
CONDITIONS FOR A 50 PERCENT DUTY CYCLE
Figure 9
Transistor Dynamic Stress Lifetime
as a Function of Drain Voltage
seconcl rise/fall times, a 20-microsecond period,
and 10 and 50 percent duty cycles. Stresses were
performed for different val ues of Vm , which was
held constant during each stress.
stress waveform chosen allows us to make simplify­
ing approximations in the calcu lations of
TN"'.·"·
Since v05 is constant during the dynamic stress,
Figure 9 shows the dynamic stress results, as a
fu nction of Vns , compared to the transistor dynamic
( 1 2)
stress I i.fetime:s predicted by a quasi-static interpre­
tation of equation (2) and data similar to that shown
in Figure 5. The measured transistor dynamic l ife­
time is noticeably shorter than that predicted by the
quasi-static sum. The transistor dynamic stress l ife­
times can be as mu ch as two orders of magnitude
lower than T\',,· These results demonstrate that con­
sideration of only one damage mechanism, in this
where T,._-,,..,.(s•a•ic) is the transistor static stress l ife­
time when J.(;-5 equals V,JS> and d is the effective duty
cycle in percent. In our example, we assume that
the time during which VGs equals V,J 1 is large com­
pared to the t i me spent during 1-(;s transi tion in the
high Vcs region. Additional ly, lc, drops off quickly
with reduced Vcs · Thus, d may be approximated by
case the interface states created by medi u m gate
voltage stress. is insufficient for predicting transis­
the fraction of the time period that vcs equals V0s
tor dynamic stress lifetimes. The contribu tions of
cycles were 7 and 45 percent for settings of 10 and
N""'· " and N""'· " must also be included, as indicated
by equation ( 11).
50 percent, respectively.
The quasi-static contributions o f t h e three types
of damage, N,., , Nnx f:J ' and N · • are shown in Figures
,., , e
.
10 and 1 1 , with 10 and 50 percent duty cycles,
As measured on an oscil loscope, the effective duty
The dynamic stress model of equation ( 1 1 )
was appl ied w i t h the data from F igures 10 and 1 1
to yield the transistor dynamic stress lifetime
curves in Figures 12 and 13. Figures 12 and 13 also
respectively. The curves shown were calcul ated
include the measured transistor dynamic stress l ife­
from the static stress results of the previous section
time cu rves for both 10 and 50 percent duty cycles
and equations (8)- ( 10) , except for ,v, , .c' The cal­
and reveal an excel lent match between the pre­
culation of the quasi-static contribution of Nux.e
requires knowledge of the instantaneous gate cur­
d icted and measu red transistor I ifetimes. These
rent which, in this case, requires time-consuming
figures show that the inc lusion of N,._, , Nox.e' and
Nox. IJ fu l ly accounts fo r the enhanced dynamic
measurement techniques. Fortunately, the dynamic
stress degradation.
1 06
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Transistor Hot Carrier Reliability Assurance in CMOS Technologies
Cii
0
z
0
u
Cii
0
z
1 07
0
u
w
w
<:!2.
w
:::E
f=
w
u..
::;
a:
0
I(/)
en
<(
a:
z
I-
1 07
<:!2.
w
1 06
:::E
f=
w
u..
::::;
a:
0
I(/)
en
1 05
<(
a:
I-
z
1 04
1 03
6
5
8
7
Vos (VOLTS)
1 06
1 05
1 04
1 03
KEY:
6
0
D
0
u
w
<:!2.
INTERFACE STATES
TNss
0
•
OXIDE-I NJECTED HOLES TNox, h
w
u..
::;
a:
0
I(/)
en
z
<(
a:
I-
8
7
CALCULATED
MEASURED
Figure 12
Calculated Contributions from
the Three Types ofDynamic Stress
Damage (10 Percent Duty Cycle)
1 07
Cii
0
z
1 06
<:!2.
Measured and Calculated
Transistor Dynamic Stress
Lifetimes (10 Percent Duty Cycle)
1 07
0
u 106
w
w
w
:::E
f=
Vos (VOLTS)
KEY:
TNo x.e
OXIDE-INJECTED ELECTRONS
Figure 10
Cii
a
z
6
5
:::E
f=
w
u..
::::;
1 05
1 05
a:
0 1 04
I(/)
en
1 04
z
<(
a: 1 03
I-
1 03
1 02
6
5
Vos (VOLTS)
7
8
1 02
KEY:
6
0
D
5
6
Vos (VOLTS)
7
8
KEY:
OXIDE-INJECTED HOLES TN ox ,h
I NTERFACE STATES
TNss
OXIDE-I NJECTED ELECTRONS
Figure 1 1
0
•
TNox, e
Calculated Contributionsfrom
the Three Types of Dynamic Stress
Damage (50 Percent Duty Cycle)
CALCULATED
MEASURED
Figure 13
Measured and Calculated
Transistor Dynamic Stress
Lifetimes (50 Percent Duty Cycle)
forms. Figures 10 and 1 1 indicate that Nox e has the
,
Fur­
lowest transistor lifetime at high v05. Thus, the tran­
ther examination of the contributions of each of the
sistor dynamic lifetime is sensitive to the duration
three types of damage illustrates their effect o n
of stress in the region where Yes equals v05; the con­
transistor dynamic l ifetimes rel a t ive to stress wave-
tribution of Nox e decreases with the duty cycle.
Discussion of the Dynamic Lifetime Model
Digital Technical journal
Vol. 4 No. 2
Spring 1992
,
1 07
S cm..i conductor Tec hnologies
ln the case of a 10 percent duty cycle, the contri­
bution of N,.u is reduced to the contribution level
of A0_,·.f· · As shown in Figures 12 and 13, the mea­
sured and calculated transistor dynamic stress life­
times an: 3 to l O times greater for the lO percent
duty cycle than for the 50 percent duty cycle.
The results presented in this section are for one
CMOS technology. The relative importance of the
three types of damage can be different in other
CMOS technologies. Thus, different technologies
can have d ifferent dependencies on the gate and
drain voltage waveforms. These d ifferences can be
analyzed through the use of a SPICE postprocessor,
which calcul ates the damage integrals of equations
(8)-(10) from the gate and drain voltage waveforms.
Circuit Considerations
The
transistor lifetime integrals of equations
(8)- ( 10), along with equation ( 1 1 ) for the combined
transistor l ifetime, contain the essen tial transistor
physics. However, a number of circuit-related con­
siderations arc important i n assessing the hot car­
rier reliabi l ity of a process technology. These
include the actual waveforms experienced by t he
transistor, the impl ications of speed binning, and
the amoun t of degradation that will cause a circuit
to fai l . These topics are discussed in this sectio n .
ations, the on-chip postttve and negative power
supplies, V00 and V15 , experience ringing clue to
inductance i n the package. The ringing worsens
with increases in the package inductance and the
rate at whicl1 current is drawn into the chip (dl/dt).
With advances in technology, increases in clock fre­
quency will resul t in more severe ringing u n less
accompanied by a reduction in the package induc­
tance or an increase in the amoun t of on-chip
decoupl i ng capacitance. The combined effects of
power supply ripple and inductive ringing can be
modeled by a sine wave superimposed on the static
supply with the appropriate ampl itude and a fre­
quency several times greater than the clock rate
(depending on the number of clock phases).
Table 1
Contri butions to t h e M aximum
Node Vo ltage
Contribution
I ncrement
(volts}
Subtota l
(volts}
N o m i nal Power Supply
3.30
3.30
5 Percent Power Supply
To lerance
0 . 1 65
3.465
On-ch ip R i ng i ng
0.1 75
3.64
0.66
4.30
20 Percent Capacitive
Coupling
In-circuit Waveforms
The technique described in the previous section for
determining hot carrier transistor l ifetime u nder
dynamic bias conditions can be applied to tran­
sistors in integrated circuits, if the gate and drain
voltage waveforms arc accurately known. The wave­
form shapes depend on the circuit type of which
the device is a part, and a lso on the magnitude of
switching transients and power supply noise,
which can elevate the drain voltag<:: above I'JJo• the
on-chip posit ive power supply voltage. The foJ low­
ing d iscussion of the factors contributing to i nte­
grated circuit waveform amplitude and timing uses
CMOS microprocessor design as an example.
Maximum Node Voltage Switching transients
and power supply ringing on an i ntegrated circuit
can raise internal node voltages wel l above the
nominal positive power supply voltage V00. Table I
categorizes these effects for a microprocessor chip
with a nominal power supply voltage of 3.3 volts,
and shows that voltages as high as 4.3 volts are
expected . In addition to the 5 percent power
supply tolerance based on ripple and drift consider-
1 08
Although power supply tolerance and power
supply ringing cause increases in the on-chip
power supply voltage to which a l l nodes are
exposed, some internal nodes can be booted above
V00 by capacitive coupling of voltage transitions o n
nearby nodes. The magnitude o f the voltage rise
above V00 depends on the ratio of coupl ing capaci­
tance to total node capacitance, and on the mag­
nitude of the nearby transitio n . Two factors limit
the maximum voltage: clamping by p +/n diodes in
p-channel p u l l-up transistors, when the drain volt­
age exceeds V00 by a forward diode drop; and turn­
ing on of p - channel p u l l-up transistors, when the
node voltage exceeds VfJD by a threshold voltage
drop . Simulations of capacitive coupl ing events
i n CMOS microprocessors have shown that n-well
resistance severdy l imits the diode clamping abi l­
ity, and that p -channel conduction is typically
weak. Furthermorc, some circuit types, e.g. , virtual
ground circuits, conta i n nodes without p -channel
pull-up transistors.
The voltage rise above VDD on intern a l nodes is
often l imited primarily by the amoun t of capacitive
Vol. 4 Nn. 2
)jJring J'J')2
D igital Technical journal
Transistor Hot Carrier Reliability Assurance in CMOS Technologies
coupling. Therefore, it is important to know the
max imum possible coupling. Examination of the
circuits in the CMOS microprocessor example
shows that not a l l nodes experience capacitive
coupl ing, and only a small fraction of the nodes
are expected to reach 4.3 volts during the part of
their operating cycle during which hot carrier clam­
age occurs. In some cases, capacitive coupling is
restricted by noise margin considerations. For the
microprocessor example, the ratio of coupling
(a) Circuit Diagram
capacitance to total nodal capacitance is l imited
to a maximum of about 20 p ercent. A noise margin
4.5
limit to coupl ing capacitance provides an upper
4.0
l imit for the node voltage. In general, though, it is
necessary to independently determine the extent
3.5
of coupling for each node in the circ u i t that may be
affected by hot carrier degradation.
Waveform Timing
The relative timing of gate
and drain voltage waveforms and the extent of
capacltlve coupl ing are both circuit dependent.
Therefore, it is necessary to find the worst-case
combination of these factors in order to determine
the circuit hot carrier l ifetime fo r a particu lar chip.
Analysis of fou r circuit types used in CMOS micro­
processors (complementary drivers, pass transistors
connected to storage nodes, precharge drcuits, and
virtual ground circuits) shows that the worst-case
- - - -: -,.
I
I
I
3.0
(J) 2.5
f-
0 2.0
>
1 .0
0.5
0.0
0
the coupling event to produce the maximum drain
voltage. The factors l isted in Table I increase the
n-channel drain voltage during the low and med ium
gate voltage part of the waveform. This is the region
in which hole-generated electron traps N0x h and
interface states Ns,· are created in the gate oxi de .
Figure 15 shows a waveform for a pass transistor
connected to a storage node (capacit ive coupling
and ringing effects not included) that differs from
I
2
4
6
8
10
TIME (NANOSECONDS)
12
14
16
NODE A (TRI -STATED)
NMOS VGs
Vc
V oo
(b) Wavefo1-m
node at voltage Vc that switches from I-Ss to v00 . The
and phasing between the power supply noise and
I
I
I
KEY:
coupling between the output (node A) and a nearby
capacitive coupling for that circuit type (20 percent)
I
I
circuits and in tri-state driver circuits. Figure 14(a)
the circuit simu lator SPICE, assuming worst-case
I
I
I
,
,
,
I
is an example of a precharge circ u i t with capacitive
Figure 14(b). This waveform was constructed using
I
I
'
hot carrier cond i tions typically occur in precharge
worst-case waveform for this circuit is shown in
I
I
'
1 .5
-0.5
I
I
I
-'
Figure 14
Circuit and Waveform ofa Precharge
Circuit with Capacitive Coupling
I t is possible to establish broad circuit categories
that are particularly susceptible to hot carrier dam­
age. However, the complex dependence of wave­
form shape on circ u i t layout and timing requ ires
that many nodes be examined individually to deter­
m ine if they present hot carrier problems. An auto­
mated software tool such as a SPICE postprocessor
can be very useful for this p u rpose, given that al. l
aspects of the waveforms are modeled accurately.
the precharge circuit described in the previous
Lifetime Criteria
paragraph in the timing between the gate and drain
The previous discussion applies to any transistor
signals. The longer duration of simu ltaneously high
lifetime criterion used . To determine if a process
gate and drain voltages in this circ u i t make elec­
tron-generated trap damage Nox, e potential ly worse
technology produces transistors with acceptable
hot carrier degrad ation, it is necessary to establish
than for the precharge circu it.
reasonable criteria from a circuit perspective.
Digital Technical journal
Vol. 4 No. 2
.VJring 1992
1 09
Semiconductor Technologies
4.0
3.5
3 .0
(/)
2.5
1....J
0 2.0
>
1 .5
(a) Circuit Diagram
Figure 15
Circuit and Waveform of a Pass
Gate to a Storage Node Showing
the Timing of the Gate and Drain
Voltage Signals during a Read
Operation
These criteria depend on the circuit design and the
stated performance. For typical microprocessor cir­
cuits, the performance is stated in terms of speed.
Thus for microprocessors, the most important con­
sideration is usually how much transistor degra­
dation will cause the speed of a part to fal l below
the minimum of the speed bin into which it would
initial ly be placed. A secondary consideration is a
degradation of parameters that results in loss of
functionality, independent of speed.
Circuit Speed From the standpoint of hot carrier
degradation, the single most important parameter
affecting circuit speed is transistor saturated drain
current. Therefore, the key hot carrier l imit is that
the degradation in transistor saturated drain cur­
rent not exceed some value, usually stated as a per­
cent of the unstressed saturated drain current.
As shown in this section, the limiting case is the
slowest part in the fastest speed bin. If the fastest
speed bin has a tested speed of 0.5 nanosecond
faster than the stated performance, then the frac­
tional degradation of the slowest transistor falling
into the bin must be small enough that the overal l
circuit speed drops by less then 0.5 nanosecond .
For a typical 12-nanosecond complex instruction
set computer (CISC) m icroprocessor, this would be
a few percent change in saturated drain current.
Hot carrier degradation of transistors results in a
change in reverse saturated drain current that is dif­
ferent from the change in forward saturated drain
current. Reverse saturated drain current is the tran­
sistor saturated drain current measured with the
source and drain interchanged relative to the stress
configuration. Thus, an additional criterion is nec­
essary for reverse saturated drain current degrada-
1 10
1 .0
0.5
0.0
-0.5
0
2
4
6
8
10
12
14
16
TIME (NANOSECONDS)
KEY:
(b) Waveform
tion. The sensitivity to reverse saturated drain cur­
rent degradation is seen only in cases where the
transistor operates in both forward and reverse
directions, e.g., in pass transistors. Although gener­
ally not used in speed-sensitive situations, these
transistors are sized to provide extra margin, if they
are so used. Thus, the allowable percent degrada­
tion in reverse saturated drain current is typically
several times that permitted for forward saturated
drain current.
Circuit Functionality
In addition to parameters
that affect circuit speed, hot carrier stress also
causes degradation i n p arameters that can affect
circuit functional ity, independent of speed. Most
important among these are threshold voltage and
transistor off-state current.
Transistor threshold voltage shifts can alter
inverter noise margins and indirectly affect off-state
current. Transistor threshold voltages are also
in1portant in cases in which matched performance
of a pair of devices is required, e.g., in a memory
sense amplifier. Because the threshold voltage of
n-channel transistors usually increases with time
(perhaps after an initia l decrease of a few m il l i­
volts), off-state current does not increase. For a
3.3 -volt nominal power supply and transistors
with ± 0.5 -volt threshold voltages, a threshold shift
of a few tens of m i l l ivolts is acceptable. (An
increase in threshold voltage also affects transistor
saturated drain current, but this is taken into
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Transistor Hot Carrier Reliabili�y Assurance in CMOS Technologies
account by a separate l imit on transistor saturated
drain current.)
The degradation in p-channel devices is usual ly
not
important
for technologies
with
channel
lengths larger than 0.5 f.Lm. However, very short
more rapid ly, relat ive to the initial 1/JSA T • than a tran­
sistor that began with a lower IDSA r · However, even
after degradation, the cu rves never cross, i . e . , a
transistor that initially had a h igher ID.\A T continues
to have a higher IDsA T · I f these two parts are p u t into
channel length p-channel devices usual ly su ffer a
the same speed bin, the faster part will take longer
decrease in threshold vo ltage, which leads to
to fai l , because it wil l take longer for IDSA T to drop
increased off-state current and reduced noise mar­
below the critical value for that speed bin. Thus,
gins. These effects will be of increasing importance
very fast chips w i l l be rel iable as long as they are not
put into a very fast b i n . The circuit hot carrier l ife­
in sub-0.5-f.Lm technologies.
Table 2 sum marizes the important transistor
time of a particular speed bin is limited by the cir­
degradat ion parameters for circu it l ifetime of a
cuit l ifetime of the most marginal, i .e . , the slowest,
high-performance microprocessor.
chip in that speed bin.
Table 2
Typical Permitted Degradation
i n Device Pa rameters for H igh­
performance M icroprocessor
Circu its
Para m eter
Forward losAr
Reverse fosAT
4.4
4.2
Tra nsistor Lifet i m e Drifting
3 -1 0% Sh ift
'iJ) 4.0
LJ.J
a:
LJ.J
a_
::1:
<(
::::i
-'
1 0 -25% Sh ift
Threshold
1 0 -1 00 mV Sh ift
Off-state Cu rrent
Absolute Limit
Note t h a t the off-state current absolute l i m i t m a y be difficult
to measure, particularly a t operating temperature. This limit
can be i ncorporated into a limit in t h e decrease in magnitude
�
3.8
TYPICAL TRANSISTOR
"'
..:
tj)
..9 3 . 6
of p-channel threshold voltage. Off-current degradation is
usually not i mportant for n-channel devices.
3.4
Speed Binning Implications
It is customary to bin the chips into one or more dis­
crete speed categories or bins. The circu it l ifetime
3.2
SLOW TRANSISTOR
0
of a given speed bin is equal to the time i t would
2
4
6
TIME (YEARS)
10
8
take for a worst-case device in that bin to fail.
Because slower bins have a lower IDSA T > i . e . , a longer
effective channel length, and therefore have rela­
Figure 16
Predicted Saturated Drain Current
as a Function of Time
tively long transistor l ifet imes, the l ifetime issue is
more of a constraint fo r fast bins. For the fas t bins,
One approach to verifying the rei iabi lity of a fast
the choice of where i n the Ins,1 r distribution the hot
speed bin is to compare the d istribution of the IDSA T
concern. This section d iscusses how to determ i ne
relationship between this transistor and the circu it
the next slower bin. This d istribution shou ld be
compared w i th loMttx• which is the highest IDSAT
value that just meets the reliabi l i ty requ irements
hot carrier l ifetime of the bin, and the selection of
for a discrete transistor. Tf chips with an
carrier reliability shou ld be assessed is thus a major
the worst-case transistor for a given speed bin, the
values fo r chips in that bin with that of chips in
ID
T
SA "
ID5
equal
AT
to loMAX norma lly fal l into the fastest speed bin,
then that speed bin will be reliable even if i t a lso
is determined by the fastest chip, which degrades
contains parts with an
more rapid ly, in fact, the slowest chip in a bin will
Although the faster parts degrade more quickly,
Although it might appear that the c ircu it l ifetime
DSt1 T greater than
I
ID,
HA X
'
be the first to fai l. Figure 16 shows the expected
because they have more margin to degrade, they do
as a function of time for several starting
IDSAT
not limit t l1e rel iability of the speed bin. Thus, chip
degrades
and transistor l ifetimes may not be the same; if the
IDSAT
values. A transistor with a higher
Digital Tecbnicaljourttal
Vol. 4 No. 2
I SAT
D
Spring 1992
111
Semiconductor Technologies
chip has extra margin in the speed bin, the chip hot
2. Characterize transistor degradation under the
carrier l ifetime wil l be longer than the transistor
three gate voltage ranges indicated in Figure 3
l ifetime.
to determine the coefficients for the damage
The above argument applies provided there are
integrals, equations (8)- (10), for each of the pre­
no systematic changes in the parasitic capacitances.
viously determined transistor lifetime criteria.
However, such changes, if they occur, do not pre­
The results must then be scaled for transistors
sent a serious problem because the key thicknesses
with IDSA 'f • or L d f and to.,· determined i n step 3.
Al ternatively, if step 3 has a lready been per­
and widths that determine the capacitances are nor­
mally monitored in production. As a result, circuit
formed , use transistors that match either the
rel iability is limited by the slowest device with the
desired IDSA 'f ' or the Ierr and tox ·
lowest losA T in the fastest speed bin. Therefore,
transistors used for hot carrier evaluation should
have an IDSA T corresponding to the slowest chip in
3. Determine either the IosA T > or the Len and tox'
corresponding to the slowest transistor in the
fastest speed bin. This can be done by simu­
the fastest bin, and the l ifetime criteria are deter­
lation, or by processing a characterization lot
mined by that chip.
Conclusions
This paper presents a physical model for hot carrier
with transistor parameters deliberately varied
in order to determine the speed impact.
4. Choose several representative parts of the
chip circu itry to determine worst-case, time­
degradation incorporating three damage modes:
•
Med ium Gate Voltages, Nss
•
Low Gate Voltages, Nox.fl
•
dependent transistor biases. For each part, using
the coefficients for the damage integrals deter­
mined in step 2, calcu late tbe transistor l ife­
times. There will be several, one for each
criterion. The shortest transistor l ifetime is the
High Gate Vol tages, Nox, e
A quasi-static sum of the contributions of each
of these damage mechanisms accurately predicts
the transistor hot carrier l ifetime for any speci­
fied waveform and technology. In addition, models
were developed that account for the effect on tran­
sistor hot carrier lifetimes of the dominant man­
ufacturing variations. These models show that
circuit waveforms, including power supply ringing,
can have a su bstantial effect on circuit hot carrier
l ifetime. The l imiting case from the standpoint of
speed binning is the slowest transistor fal l ing into
the fastest speed bin. The paper gives an example
of circuit l ifetime criteria for a high-performance
microprocessor. From these conclusions, it is possi­
ble to ou tline the fol lowing procedure for transis­
tor hot carrier rel iability assurance for a CMOS
technology used to fabricate a particular chip:
1 . For the chip to be fabricated, determine the tran­
sistor l ifetime criteria. The m icroprocessor is
only one example; other circuits may have quite
difterent lifetime criteria. In practice, it may be
better to specify the transistor degradation prior
to c ircu i t design and incorporate the l ifetime cri­
teria into the design process; i . e . , as a part of the
design process, make sure the design works,
given the specified degradation.
1 12
l imiting case.
The two key elements in this procedure are the
three damage integrals with the combined transis­
tor dynamic stress l ifetime and the close connec­
tion with the circuit design process. In practice,
several iterations of the procedure wil l p robably
take place as the transistor design and production
process are optimized for maximum circuit perfor­
mance and yield, consistent with the necessary reli­
ability goals. For the best resu lts, this needs to be
done concurrently with circuit design to be sure
the appropriate criteria are optimized.
Acknowledgments
The contribut ions of Kaizad Mistry to the u nder­
standing of the physics of degradation under
dynamic stress, and Frank Fox to determining the
circ u i t impl ications of hot carrier degradation
have been indispensable. In addition, the contribu­
tions of Chris Conran, Hamid Partovi, Shu-Fung
Hsia, Lynn Mohan, and Ann Chen are gratefu l ly
acknowledged.
References
1.
P Correll,
R . Troutman, and
T. Ning "Hot­
Electron Emission i n N-Channel IGFET's," IEEE
Vol. 4
No. 2
Spring 1992
Digital Technical journal
Transistor Hot Carrier Reliability Assurance in CMOS Technologies
Transactions on Electron Devices, vol . ED-26
(1979): 520.
2.
T. Ninget et a l . , " 1 J.Lm MOSFET VLSI Tech­
the IEEE International Electron Devices
Meeting (IEDM) (1984): 100- 104.
12.
on Electron
tron Devices, vol . ED-34 (1987): 378-385.
E. Takeda and N. Suzuki, "An Empirical Model
13.
Characterizat ion of Electron and Hole Traps
EDL-4 (1983): 1 1 1 - 1 13.
Created by Hole Injection D u ring Low Gate
Voltage Hot Carrier Stressing in n-MOS Transis­
C. Hu et al., " Hot-Electron-Induced MOSFET
tors"' by B. Doyle et al., IEEE Transactions on
IEEE
Transactions
on
Electron Devices, vol. ED-39 (1992) : 460-464.
Electron
14.
times of n-Channel MOSFETs;' IEEE Electron
Device Letters, vol. EDL-10 (1989): 500-502.
MOS Transistors Under Alternate Stress Condi­
tions," IEEE Electron Device Letters, vol. EDL-9
15.
IEEE Electron Device Letters, vol. 12 (1991 ) :
B. Doyle, M. Bourcerie, ]. Marchetaux, and
492-494.
A . Boudou, " I n terface State Creation and
Charge Trapping in the Medium-to-High Gate
(v,i 12> 'i > v,t)
16.
Du ring Hot Car­
rier Stressing of n-MOS Transistors," IEEE
Transactions on Electron Devices, vol . ED-37
national Electron Devices Meeting (IEDM)
(1990): 744-755.
Technical Digest (1991 ): 727-730.
B. Doyle and K. Mistry, "The Generation and
17.
Degradation-Model, Monitor, and Improve­
Created by Hole Injection During Low Gate
ment,"
Vol tage Hot Carrier Stressing in n-MOS Transis­
Devices, vol. 32 (1985): 375-385.
18.
IEEE
Tra nsactions
on
Electron
E. Takeda and N. Suzuki, "An Empirical Model
for Device Degradation Due to Hot-Carrier
B. Doyle and K. Mistry, "A Lifetime Prediction
Injection," IEEE Electron Device Letters, vol. 4
Method for Oxide Electron Trap Damage
(1983): 1 1 1 - 113.
Created During Hot Electron Stressing of
n-MOS Transistors," IEEE Electron Device
19.
IEEE Electron Device Letters, vol. 5 (1984):
518.
T. Hori and H. Iwasaki, "Improved Hot- Carrier
Immunit y in Submicrometer MOSFET's with
Reoxidized
Nitrided
Oxides
Prepared by
20.
International
E. Nico l l ian and ]. Brews, MOS Physics and
1982): 534- 537.
Electron
Devices
Meeting
(IEDM) Technical Digest (1986): 390.
21.
W
Weber,
"Dynamic
Stress
Experiments
for Understanding Hot Carrier Degradation
T. Poorter and P Zoestbergen, " Hot Carrier
Effects in MOS Transistors," Proceedings of
Digital Technical journal
W Weber, C. We rner, and A. Schwerin, " Life­
times and Substrate Currents in Static and
Dynamic Hot - Carrier Degradation," IEEE
Rapid Thermal Processing," IEEE Electron
Device Letters, vol. EDL-120 (1989): 64-67.
Technology (New York: John Wiley and Sons,
W Weber, C. Werner, and G. Dorda, " Degrada­
tion of N-MOS-Transistors after Pulsed Stress,"
Letters, vol . EDL-12 (1991): 178- 180.
11.
C. Hu et a l . , " Hot -Elect ron-Induced MOSFET
Characterization of E lectron and Hole Traps
tors," IEEE Transactions on Electron Devices,
10.
K. Mistry, B. Doyle, A. Phil ipossian, and
D. Jackson, "AC Hot Carrier Lifetimes in Oxide
and ROXNOX n-Channel MOSF ET's," IEEE Inter­
vol . ED-37 (1990): 1869- 1879.
9.
K. Mistry and B . Doyle, "A Model for AC Hot­
Carrier Degradation in n-Channel MOSFET's,"
(1988): 232-234.
Voltage Range
K. Mistry and B. Doyle, "An Empirical Model
for the Ldf Dependence of Hot Carrier L ife­
P Bel lens, P Heremans, G. Groeseneken, and
H. E. Maes, " Hot Carrier Effects in n-Channel
8.
Reply to " Comments on 'The Generation and
Injection," IEEE Electron Device Letters, vol.
ment,"
7.
"Two­
Linear Region," IEEE Transactions on Elec­
Devices, vol. ED-32 (1985): 375 - 385.
6.
Cristoloveanu,
Short- Channel MOSFET's Operating in the
Transactions
IEEE
Degradation-Model, Monitor, and Improve­
5.
and S.
straints,"
for Device Degradation Due to Hot- Carrier
4.
Haddara
Dimensional Model ing of Loca l ly Damaged
Devices, vol . ED-26 (1979): 346-353.
3.
H.
nology: Part TV-Hot-Electron Design Con­
Vol. 4 No. 2
Spring 1992
Phenomena," IEEE Transactions on Electron
Devices, vol. 35 (1988): 1476.
113
]. Joseph Clement
Eugenia M. Atakov
James R. Lloyd
Electromigration Reliability
of VLSI Interconnect
Increased speed, reduced line widtbs, larger cbip size, and additional levels of inter­
connect are all factors tbat contribute significan tly to tbe improved pe1Jormance
and jitnctionality of VLSI circuits. At tbe same time, these factors place growing
demands on interconnect reliability. Tberefore, careful cbaracterization of the inter­
connect reliability is important in acbieving VLSI performance and reliability goals.
A scaling model was developed and used to examine factors essen tial to assuring
electromigration reliability in Digital's CM05�4 technology and in the Alpba 21064
microprocessor, which uses tbis technology.
Background
these col l isions creates a net flux of metal atoms in
For complex, very l a rge-scale integration
the direction of the electron tlow and thus biases
(VLSI) cir­
c u its, the individual components that make up the
the normal random atomic diffusion. At sites of
circuit must be extremely reliable to assure accept­
atomic flux divergence, where the number of metal
able overa l l rel iability. Since a chip is expected to
atoms coming in does not equal the number going
operate for many years, testing to characterize the
out, either material depletion or material accumu la­
reliability of circuit components such as in ter­
tion occurs. Corresponding voids or hil locks form
connects must be performed under accelerated test
in the metal l ine and cause open or short circuits
conditions. To evaluate circuit reliability by extrap­
and, u ltimately, circu it failure.
olating from data on components tested at acceler­
For typical circuit operating temperatures, the
ated rates, it is essential to have dependable models.
d iffusivity of metal atoms is much higher in the
E lectromigration is one of the primary failure
grain boundaries than through the lattice of the
mechanisms in the polycrystal l ine alu minum-al loy
grain itself. (A grain is an individual crystal in a poly­
thin films that are widely used for interconnects on
crystall ine film.) Hence, the atomic mass transport
Vl.SI circuits. ' Although much has been written on
occurs primarily along grain boundaries. There­
the effect of stress conditions o n in terconnect fail­
fore, microstructural inhomogeneities, such as vari­
ure clue to electromigrat ion, little information is
ations in grain size, can cause a flux divergence and
available on the use of l i fetime measurements on
thus be sites of potential failure.
simple test structures to evaluate the complex com­
bination of interconnects on a VLSI chip.
Enhanced mass transport along grain boundaries
is one reason that electromigration is a more impor­
In this paper, we describe work performed to
tant fa ilure mechanism for thin films than for b u l k
characterize the rel iabi l ity of the interconnect
conductors. In thin films, the film thickness dictates
CMOs-4 technol­
that the grain size be much smal ler than the grain
in chips manufactured in Digital's
ogy. We place particular emphasis on a model fo r
size tor bu l k materia l . Thus, a greater proportion of
scaling test structure data to chip level. Before
the thin-film cross section can be composed of
presenting the model, we discuss the physics of
high-cl iffusivity grain boundaries.
electromigration and electromigration testing, as
background in.tormation for the reader.
Electromigration
A second reason electromigration is a concern
for thin films is related to current density. The thin­
film in terconnect on integrated circuits is in inti­
mate thermal contact with the underlying sil icon
Electromigration i s the mass transport of metal
substrate, which acts as a big heat sink. There­
atoms from col lisions with the current conduction
fore, thin fi lms can withstand higher current den­
electrons. The momentum exchange resulting fro m
sities than b u l k wires, without incurring thermal
1 14
Vol. 4 No. 2
Spring 1992
Digital Tee/mica/ journal
Electromigration Reliability of VLSI !nterconnect
damage. Maximum operating cu rrent densities in
VLSI
circu its are as much as 100 to 1 ,000 times
temperature dependence, a temperature gradient
itself can cause a flux d ivergence and lead to fa ilure.
higher than normal for bu l k wires. With reduced
l ine widths and increased speed of operati o n , the
Stress Acceleration Model
trend
A model proposed by Shatzkes and Lloyd relates the
is toward increased
c u rrent densities
in
interconnects. This combination of high c u rrent
time-to-failure t to the temperature T and the cur­
densities and
rent density j by
boundaries
high-diffusivity paths a long gra in
promotes
electromigration
in
1
poly­
(1)
crystal l in e thin-film interconnects.
The addi tion of s m a l l amounts of a l loying ele­
ments
can
migrat i o n
(Al)
substant i a l ly
performance
improve
of
the
thin-film
electro­
where A is a constant t h a t depends o n mat erial
properties, k is Boltzmann's constant, and H is
alu min u m
the activa tion energy4 This form u l a t ion d iffe rs
conductors. I n particu lar, a l loying w i t h copper
sl ightly from an earlier, largely empirical model
(Cu) has proven to be a pop u l ar, cost- effective way
proposed by Black, which did not include the pre­
to improve lifetimes.
exponential T term.s
Electromigration Testing
a n d aluminum-al loy t h i n fi l m s i s typically i n the
2
The activation energy measured fo r a l u m i n u m
Testing to eva luate electromigration rel ia b i l i t y is
accelerated using stress cond itions for temperature
and current that are higher than the expected oper­
ating condi tions. Thus, an accurate physical model
is essential for extrapolating test resu lts to actual
operating conditions.
Due to random microstructural variations, nomi­
range of 0.5 to 0.8 electron volt (eV) . This is signifi­
cantly less than the activation energy for l a t tice d if­
fu sion measured in single-crystal aluminum, which
is approximately 1 .4 ev. These measurements indi­
cate that the mass transport takes place along grain
boundaries.
n a l ly identical interconnect test structures wil l not
Length/ Width Scaling Model
a l l fail a t the same time. Typically, several identical
To characterize the electromigration rel iabi l i ty of
samples are stressed together, resu lting i n a d istri­
the interconnect on VLSI chips, we developed a scal­
b u t ion of fai l u re times. Knowledge of the correct
ing model that considers l i ne length, l i ne width,
failure d istribut ion is critical for predicting electro­
and tu ngsten-filled vias.
migration re l iability.
Electromigration rel iabil ity testing is usually per­
The Lognormal Failure Distribution
formed by st ressing packaged test lines in an oven
Electromigration failure times are generally repre­
with a constant current, whil e monitoring the
sented as a lognormal d istribu t i o n , whereby a plot
voltage in situ. Depending on the met a l l ization,
of the logarithms of the fai l u re times relative to the
fai l u re may result from open circuits o r an increase
cumulative percent age order on a normal probabil­
in resistance caused by voiding, or from the fo r­
ity scale approximates a straight line, as il lustrated
mation of extrusions that short-circuit to adjacent
in Figure 1. The lognormal distribution is character­
lines. General guideli nes fo r electromigrat i o n test
ized by two parameters: the median time-to-fa i lure,
structure design and lifet ime measurements are
o r logarithmic mean, 8, and the slope, or logarith­
available. 2·.l
mic standard deviation,
Care must be taken i n the selection of the stress
current and temperature, so that the failure mecha­
F(t)
nism that operates u nder stress condi tions is the
<P(x)
same one that works a t expected operating condi­
where
tions. The temperature must be kept low enough to
tribution h.Inction.
remain in the range that gra i n boundary d iffusion
dominates. At very high temperatures, lat tice d iffu­
sion becomes significant.
y.
The lognormal cumu la­
tive density hmction F(t) is
If
p
=
<P (( ln t
-
8) /y)
(2)
is the standard normal c u m u l a t ive d is­
represents the cumu lat ive fraction fa iled ,
a n d tp is the corresponding time, then
ln tp = 8 + y<P- 1 (p)
In addition, the current must be l i mited to pre­
<P - 1 (x)
(3)
vent excessive joule heating. joule heating creates
where
temperature gradients along the metal l i n e . Since
c u m u lative d i stribution hmction. For p approxi­
the d iffusivity of the metal atoms has a n Arrhe nius
m ately equal to 0 . 1 6 o r 0.84,
D igital Technical journal
Vol. 4 No. 2
Spring 1992
is the inverse of the standard normal
<P- 1 (jJ)
equa l s - 1 and
1 15
Semiconductor Technologies
0.99
If F(t) is lognormal, then G(t) cannot be, and vice
0
0.95
0.90
�
::J
-
0.80
ili
<(
0.70
a:
Q_
0.50
CD
0
LlJ
>
i=
:5
:;)
:::;
:;)
u
0.60
-
- - - - - -- - -- - - -- - - - --- -- �
0.20
by
0
- - - - -- -- - -- - --- ----
G(t) = 1 - (1 - cf> ((Lnt- 8)1-y))"'
0
0
0.30
-- - - - - - - - - - - - - - 0
0.10
0
0.01
Figure 2 contains a plot of the the MLN dist ribu­
I
I
I
I
I
I
is a lognormal curve fitted to the t 16 and t84 points
of the MLN curve at N = 100. The logarithmic stan­
dard deviation of this fitted curve,
I
I
I Iso
I
1 116
I
0.99
KEY:
SIMULATED DATA
LOGNORMAL FIT (-y
0.95
=
1)
0.90
�
Figure 1
::J
ili
<(
Lognormal Failure Distribu tion
+ 1 , respectively; for p equal to 0.5,
cp- • (jJ) equals 0.
Therefore,
ln t 16
=
is a fu nction of
1 0.00
FAILURE TIME (ARBITRARY UN ITS)
0
fJ,
y and N.
ls4
1 .00
0.10
(6)
tion for N = 1 and N = 100. Also shown i n this figure
I
I
0.05
For the particular case where F(t) is lognormal,
G(t) has become known as the muft ilognorma,l or
multiple lognormal (MLN) distribution, and is given
0
0
0.40
versa.
CD
0
a:
Q_
LlJ
>
i=
:5
8 - -y
These three points in time are indicated in Figure 1 .
:;)
::2
:;)
u
I
I
I
0.80
0.70
I
I
0.60
I
I
0.50
I
0.40
0.30
0.20
0. 1 0
0.05
Length Scaling Effect
A major conceptual problem with the lognor m a l
1 .00
0.01
1 0.00
FAILURE TIME (ARBITRARY UNITS)
d istribution, however, is t h a t i t does n o t scale w i t h
l i ne length. That i s , if the failure t imes tor a given
KEY:
N= 1
length l ine are lognormally d istribu ted, the fa i l ure
N = 1 00
times of shorter or longer l ines cannot also be log­
LOGNORMAL FIT TO
A N D 184 F O R N = 1 00
normal ly d i!Stribu tecl. This problem has significant
l16
repercussions in extrapolating data on test devices
to predict the rel iability of VLSI circu its, particu­
larly fo r early fai l ures at and below the l percent
Figure 2
level.
Consider an ensemble of nomina l l y identical
l i nes of length I that fail over t i me with a distribu­
tion described by F(t). We could, i n principle, com­
bine these l i nes to for m new l i nes of length L
N1 ,
where N is the number of original units that make
=
up the new l i ne. The failure of each new l ine would
be determined by the earliest fa ilure time of the
component elements, assuming that the unit ele­
ments behave independently. The failure distribu­
tion of the new ensemble is given by
G(t)
1 16
=
1 - (1 - F(t))"'
Multilognormal Failure Distribution
There are
some
is, over the range of cumulative percent failures
between 0.05 and 0.95, the MLN distribu tion is
nearly indistinguishable from the lognormal . H ow­
ever, the d ifference between the two distributions
is more significant for early fa i l u res, at and below
the 1 percent level. These early fa ilures are of
concern for rel i abil ity. A'lso note in Figure 2 that
as
N
increases,
the
decreases, and the
(5 )
i m portant characteristics to
glean from this figure. For s m a l l sample sizes, that
fJ
median
t i me-to-failure
t50
of the lognormal fitted curve
als o decreases. This behavior of the model agrees
lk>f. 4 No. 2
Spri11g 1992
D igital Techt1ical jourrtal
Electrornigration Reliability of VLSI Interconnect
with experimental observations on the effect of
line length on t50 and
a-. 6
as the line width decreases. For l ine widths some­
what greater than the average grain size, the l ife­
M oreover, recent work model ing electromigra­
time gradually increases with line width. There is,
tion failure in fine l ines yields a failure d istribution
therefore, a minimum in lifetime in relation to the
that is wel l approximated by the MLN distribution 7
l i ne width when the l ine width is comparable to the
The agreement of the modeling resu lts with experi­
average grain size.
mental find ings strongly supports the use of the
Grain size in polycrysta l line thin films is i nflu­
MLN distribution for length scal ing.
The scal ing model as described by the MLN distri­
enced by a number of factors including the sub­
strate, the method by which the film is deposited,
bution in equation (6) has three adjustable parame­
the deposition conditions, and the grain growth
ters: e and ')' of the failure distribution for the
due to postdeposition anneal ing. In addition, grain
elemental failure unit, and N, which is the total l ine
growth and grain boundary movement during post­
length L divided by the characteristic elemental
patterning annealing can be even more significant
length effect has been demonstrated, some or all of
in fine I ines. 1 1
In terms o f the scaling model, t h e expected
these parameters must be a funct ion of l ine width,
improvement in lifetime with reduced l ine width is
length l. Since a l ine-width dependence of the
as weJJ.6
a resul t of two effects on the model parameters.
Research suggests that the susceptibi l ity of a
Width Scaling Effect
For l ine widths somewhat greater than the average
grain size, we can expect a continuous "network "
of grain boundaries a long the length of the l ine.
Since the microstructural defects leading to electro­
migration failure are associated with the grain
boundaries, when the line width is much larger than
the grain size, we might expect m u ltiple defects to
grain cluster to failure decreases as the length of
the cluster decreases. 9. 12- 13 If the time-to-failure of
the unit element is related to the length of the non­
bamboo grain clusters, e increases as the l ine width
decreases.
More important ly, as the l i ne width decreases,
bamboo segments constitute a larger percentage
of the l ine. Thus, the number of microstructural
a l ign at a failure site. The l ifetime of the l ine, there­
defects per unit length decreases; that is, the char­
fore, is determined by the least severe defect exist­
acteristic unit length l increases. In our model, for a
ing along the width. Thus, as the width increases,
l ine of a given length L, as l increases, the number of
the probabil ity of al igning with less severe defects
elemental failure units N decreases. Consequently,
rises, and an increase in l ifetime is expected.
However, the expectation that the l ifetime will
t50 and a- increase, provided 'Y does not change
significant ly. The increase in t50 and a- with decreas­
row lines when a " bamboo" microstructure devel­
ing l ine width given by the model is in complete
agreement with the results obtained from electro­
ops, in which a single grain boundary traverses the
migration l ife tests. Because the slope of the failure
width of the l ine. As the l ine width decreases, the
d istribut ion is increasing together with t50 as the
l ine width decreases, it is generally recognized that
improvement in the early fai lure times a t and below
decrease with line width does not hold for very nar­
l i kelihood of having a single grain span the entire
width of the l i ne increases. When the l ine width
is comparable to the average grain size, the l i n e
consists o f bamboo and nonbamboo segments, as
shown in Figure 3. As the l ine width decreases
further, the fraction of the l ine that is bamboo
increases, that is, the length and number of non­
bamboo segments decreases. Since electromigra­
tion proceeds primarily along grain boundaries,
failure correlates with the length and number of
nonbamboo segments. Hence, as the l ine width
decreases below the average grain size, the electro­
migration lifetime improves.
the 1 percent level is much less substantial than the
increase in t50 . 10
The chip interconnect rel iabil ity is determined
by the time-to-first-failure of its component l ines of
various lengths and widths. Therefore, the probabil­
ity of chip failure can be written as
H(t) = l - TI(l -Fn (t)J LA,
n
(7)
where L is the total length of the l ine of a particu­
,.
lar width, and in and Fn(t) are the characteristic unit
length and unit failure distribu t ion for that l ine
The dependence of electromigration l ifetime on
width. If the total interconnect length on the chip is
l ine width has been wel l establ ished 8.9. IO In the
d istributed fa irly evenly between narrow and wide
bamboo region, the l i fetime increases very rapid ly
l i nes, the chip rel iab i l i t y will be dominated not by
D igital Technical journal
Vol. 4 No. 2
Spring 1992
1 17
Semiconductor Technologies
( a)
(b)
Figure 3
Comparison ofBamboo and Nonbamboo Segments:
(a) A Bamboo Microstructure in a 1.25-micron-wide Line and
(b) A Nonbamboo Orain Cluster Network Found in a 4. 25-micron-wide Line
(Note that gb indicates the location of a grain boundary.)
narrow ( bamboo) l i nes, but by the l ifetime of the
for
i n tervals bounded by the choice of co nfidence
size)
l im i t . This statistical uncertainty is unavoidable b u t
eq u iv al e n t current densities. Thus, a qualifi c a­
c a n b e reduced b y using larger sample sizes. Never­
wide l i nes (one to two t i mes the average grain
tio n plan based on t e st i ng minimu m-width l i nes
theless, we can determine the model parameters
exclusively is inadequate for advanced VlSI circ u i t
withi n the constraints of these l imits.
ability estimates.
with the same li ne width but with different lengths,
technologies a n d m ay l e a d to overly optimistic reli­
such that L2 = JOL1 (i.e . , N2 = JON,). For each sample
Length/ Widtb Scaling Model
Parameters
group, a range for
width, knowing t50 and
u
e,
u
and t50 can be extracted from
the resulting fa i l u re-t i me dist ribu tions. In Figure 4,
In principle, it should be possible to uniquely deter­
mine the three model parameters
For example, consider testing two groups of lines
y, l for a given
from lifetime measure­
ments on two different length l i nes. However, in
the calcul ated y fo r the elemental fa i l ure distribu­
tion is shown as a function of N1 fo r the upper and
lower confidenu: l i m its of
u.
The band of parame­
ter values lying between these two cu rves repre­
practice, this is not the case because of the statisti­
sents the al lowable combinat ions of y and N that
cal uncertainty in estim a t i ng t50 and
w i l l fit the esti mates of
of t511 and
1 18
u
u.
Estim ates
extrapolated from test data range over
u.
Simil arly, two cu rves
were calcu lated fo r the upper and lower confi-
Vol. 4 No. 2
5/Jring
1992
D igital Technica l ]our11 al
Electromigration Reliability of VLSI Interconnect
bars on the data represent the 90 percent confi­
dence l imits on the ratio of the t50 values for the
two l ines. The parameter space defined by the inter­
dence l imits. Also shown in Figure 5 is the increase
section of these two bands, as shown by the cross­
in
hatched area in Figure 4, indicates the combination
of model parameters that will fit the test data.
model parameters were determined from this test
data; the ranges of these parameter values are given
0'
observed as the l ine length increases. The
in Table 1 . The solid l i nes in Figure 5 were calcu­
lated using the model.
2.6
(>-
2.4
1 os ,.---,----.�--.--, 4 . 0
2.2
3.6
2.0
3.2
1 .8
2.8
1 .6
2.4
1 .4
2.0
b
1 .2
1 .6
1 03
1 .2
0.8
0.6
0.8
1
10
1 00
1 ,000
0.4
NUMBER OF FAILURE UNITS, N1
1 02 '-----�-----'--'---' 0.0
KEY:
0.1
CONFIDENCE LIMITS FOR a
CONFIDENCE LIMITS ON t50 RATIO
1 .0
1 0.0
1 00.0
LENGTH (MILLIMETERS)
KEY:
•
Figure 4
tso
a
•
Allowed Values ofModel Parameters
MODEL FIT
90 PERCENT CONFIDENCE LIMIT
l
'Y and N Resultingfrom Statistical
Uncertainty in t 50 and cr from
Lifetime Measurements
Figure 5
t 50 and cr as Functions ofLength
for 1.25-micron-wide Lines
To determine appropriate model parameters for
our met a l l ization, we systematically studied the
effect of interconnect length and width on l ife­
Tab l e 1
Model Pa rameters as a Function
of Line Width
time . 14 The test samples were processed through
in the first-level metal were tested . Each level of
Parameter
,---- Line Width
1 . 25JLm
4.25 J.l.m
metal was an aluminum alloy containing 1 weight­
IJLm
21 00-2700
26-1 24
7.92-8.1 8
6.00-7.1 5
0.975-1 .05
0.975-1 .05
two levels of metall ization; however, only l ines
percent copper (Al: 1 %Cu), 7500 angstrom units (A)
thick, capped with 400 A of titanium nitride (TiN) .
The interlevel dielectric was a planarized plasma­
enhanced tetraethylorthosilicate (PE-TEOS) oxide
e In
'Y
(hou rs)
process 7500 A thick, and the wafers were passi­
vated with a layer of u ndoped oxide 7100 A thick.
the increase in l ifetime with decreasing li ne width
The average metal grain size was determined from
expected in the bamboo region. This data shows
transmission electron microscopy (TEM) analysis to
that t50 decreases by roughly a factor of 100 as
the l i ne width decreases from 4.25 to 1 .25 J.l.m for
be approximately 3.5 microns (JLm).
The effect of line length on l ifetime is demon­
The test results shown in Figure 6 demonstrate
1 .0 5 -mm- long l i nes. The increase in
0'
observed as
strated in Figure 5. t10 decreases by roughly a factor
of 20 as the line length increases from 1 .05 to 18.9
the l ine width decreases is also shown in Figure 6.
mill imeters (mm) for 1 .25-JJ.m-wide l ines. The error
lengths at line widths other than 1.25 J.l.m, all t hree
D igital Technical ]emma/
Vol. 4 No. 2
Spring 1992
Since we do not yet have data on different
1 19
Semiconductor Technologies
4.0
1 05
1 04
Cii
a:
::::>
0
�
�+
3.6
vias with vertical sidewalls have been developed in
recent years using tungsten from low-pressure
3.2
chemical vapor deposition (LPCYD). 15
2.8
2.4
2.0 b
1 03
1.6
0
...'!'
1 .2
1 02
0.8
The electromigration lifetime of via chains for
nonfil led vias decreases with reduced hole diame­
ter. 1 6 This is a direct consequence of poor metal
sidewall coverage for smal l-diameter, high-aspect­
ratio vias. 16· 1 7
Although fi l ling the via w i t h LPCVD tungsten
improves sidewa l l coverage, the presence of the
tungsten " plug" between the two aluminum l i nes
0.4
introduces a discontinuity in the electromigration­
1 0 �------�--��-�� 0.0
3
4
1
2
5
6 7 8
induced mass transport of a lu minum that may lead
WIDTH (MICRONS)
KEY:
•
•
terize the electromigration reliabil it y of tungsten­
fi l led vias.
lso
a
MODEL FIT
90 PERCENT CONFIDENCE L IMIT
Figure 6
to fai l u re. Therefore, it is very important to charac­
Testing was performed on metal-1 to metal-2
tungsten-fil led via chains (M 1 and M 2) fabricated
using standard CMOS-4 processing through all three
metal levels. Both levels of metal were 7500 A thick
t ,0 and a as Functions of Width
for 1. 05-millimeter-long Lines
Al:l %Cu with a 400 A TiN cap. The interlevel dielec­
tric was a planarized PE-TEOS oxide 7500 A thick.
The total dielectric t hickness over the M 2 was
model parameters could not be independently
determined at other widths. Since e is expected to
decrease and l to increase as the length and number
of nonbamboo segments decreases with l ine width,
we chose to fix y. Parameter val ue ranges were
extrapolated from the 4.25 -J.Lm-wide l ine data and
are included in Table 1. The l ines in Figure 6 are
calculated from the model using the parameters
extracted from the 1 .25- and 4.25-J.Lm l ine widths
and interpolating for intermediate li ne widths. The
test results for intermediate l ine widths of 2.0 and
2.75 J.Lm are shown in Figure 6 to lie on the calcu­
lated curve, as well .
For li ne widths l arger t h a n 4 . 2 5 J.L m , t h e l ifetime
does not continue to decrease. Lifetime measure­
ments on 5.75-J.Lm-wide l ines are comparable to the
4.25 -J.Lm-wide l ines. The l ifetime gradually increases
as the l ine width increases, for l ine widths that are
greater than the grain size.
Tungsten-filled Vias
approximately 2.5 J.Lm. The via diameter was
0.75 J.Lm, and there were eight vias in each chain.
The widths of the M1 and M 2 interconnections
were identical, nominally 1 .88 J.Lm , as were the
lengths of the metal l i nks, which measured approxi­
mately 170 J.Lm. The l inks were long enough to elim­
inate the chance of lifetime enhancement due to
the effect of back pressure in short lines. 13. ts
E lectromigration l ifetime measurements were
carried out at temperatures of 200 and 220 degrees
Celsius ("C) and currents of 6 and 8 m i l liamperes
(rnA). A 10 percent change in resistance was the fail­
ure criterion. The l ifetime is found to be approxi­
mately proportional to the inverse of the square of
the current density. This dependence is i n agree­
ment with other test results on tungsten-filled vias
and with observations for single-level l ines. 18
The extrapolated activation energy is nearly
1.2 eY. This value is considerably h igher than
that expected for grain boundary transport, and
approaches the value for lattice d iffusion in single­
crystal aluminum (about 1 .4 eV). However, since the
I n CMOS technology development, the lateral
l ine width is much smal ler than the average grain
dimensions of the vias have shrunk more quickly
size for these thin films, the Ml and M 2 inter­
than the vertical th ickness of the interlevel dielec­
connects will have bamboo microstructures. There­
tric. It is difficu lt to reliably sputter-deposit alu­
fore, we would expect very l it t le mass transport via
minum meta l l i zation to cover the steep, vertical
grain boundaries. Instead, lattice diffusion, or possi­
sidewal l s of sma l l - d iameter, high-aspect-ratio vias.
bly diffusion along the aluminum-oxide interface,
As a result, techniques for fill ing high-aspect-ratio
would be the primary transport mechanism.
1 20
Vol. 4 No. 2
Spring 1992
Digital Technical journal
Electromigration Reliability of VLSJ Interconnect
We also considered the possibility of local heat­
l ine width) that carry currents near the electro­
ing at the via because of the higher resistivity of the
migration design rule l i mit. Interconnects with
via-fill material and the reduced cro ss-sectional
currents much less than the design rule l i mit have
area for current flow. Three- dimensional finite ele­
negligible impact on the electromigration reliabil­
ment models of the via structure were used in
ity. With this information, we can then calculate the
simu lations to calculate the current density distri­
appropriate performance requirements for the test
bution and temperature profiles resulting from
structures to meet our chip reliability goal.
tions indicate that in the entire structure, the tem­
cent probability of chip fa ilure in 10 years under
perature increase above ambient is less than 3 oc
worst-case operating conditions. The s tress accel­
to the good heat conductivity of aluminu m , the
extrapoLate from accelerated stress conditions to
temperature grad ients are small, and the effect on
worst-case operating conditions. Scaling of test
Joule heating, as shown in Figure 7. These simula­
under the worst-case current stress of 8 rnA. Due
Our reliability goal is to assure a less than 1 per­
eration model described in equation (1) is used to
data to the chip level is accomplished by use of
electromigration lifetime is negl igible.
Figure 8 shows a micrograph of a scanning elec­
tron microscope (SEM) image of a stressed via that
has been cross -sectioned using a focused ion beam
the m u lti lognormal approximation to the fa ilure
distribution.
The entire circu it can be considered to consist of
(FIB). The transport of alum inum away from the alu­
several component groups,
minum-tungsten interface in the direction of elec­
includes a particular type of interconnect element,
tron current flow is apparent. The FIB/SEM analysis
e.g. , a certain via type or l ine wid th. The lognormal
where
each
group
of stressed via chains shows voiding at both inter­
unit failure distribution for the ith group F1 (t) is
faces, with no indication of preferential fa ilure at
characterized by y1 and
either the top or the bottom in terface.
in a slightly more compact form, the cumu lat ive
Clearly, via failure is caused by void formation
H(t) = 1 - I1S1 (t)
the tungsten plug. In the future, the reliability of
submicron via interconnects may be improved by
Rewriting equation (8)
probabil ity of chip failure
resu lting from the migration of aluminu m from
developing processes to replace the tungsten with
()1 .
I
where
(10)
aluminum to fi l l the via.
Assuming that the failure distribution for a single
via is lognormal, the fai l u re distribution for the via
chain is also MLN. In this case, since there is no
apparent l ife time dependence on the direction of
current flow, i . e . , from M1 to M2 or vice versa, the
value of N is the number of vias in the chain, i . e . ,
N = 8 . For a single via, () a n d y can then b e deter­
mined from via chain test results using the MLN dis­
tribution as given in equation (6).
The chip scal ing model can be easily modified to
include vias as separate failure elements. The result­
ing probability for chip fa ilure is given by
f
H(t) = 1 - I1 (1 - F" (t)] L n 11 ui1
m (1 - Fm (t) m
n
(8)
and Fm (t) is the fa ilure distribution for a single via.
group on the chip. For vias, N1 is the number of
vias; for l ines, N1 is the total length divided by the
characteristic unit length, which depends o n l i ne
width. Since the reliability goal is fo r H(t) to be less
than 0.01 for 10 years under worst-case operating
conditions, then for each group, F;(t) is much less
than 0.01.
As mentioned previous ly,
values for y,
e,
u ncertainty in the
and l stems from the uncertainty
inherent in estimating t50 and
CT
to a given level of
ing to examine the impact of this inherent uncer­
tainty has on the pe rformance requirements fo r the
test structures to meet the chip reliability goal .
Consider the simple case of only one component
group, namely a single l i ne width. The required test
Electromigration Reliability
Qualification
structu re t50.test relative to the time until 1 percent
The Alpha 21064 microprocessor is the largest and
fastest chip built using the CMOS-4 technology.
Careful analysis of this chip determined the number
of vias and the total length of l i nes (as a function of
Vol. 4 No. 2
and N1 is the number of components from the ith
confidence from experimental data. It is i l l u m inat­
where Nm is the number of vias of a particular type,
Digital Tecb11ical ]om-nal
(9)
Spring 1992
cumul ative failure for the chip occurs t0 1.chi is plot­
p
ted in Figure 9. This ratio is graphed as a fu nction
of the ratio of the number of units in the test struc­
ture to that of the chip, Nres /Nchtp ' for a given combi­
nation of Nrest and y. For this i l l u stration, we use
121
Semiconductor Technologies
CU RRENT DENSITY (mA/I-'m2)
[==:J
(;;),,, ,,q
ff$%1
0.235 -
5.244
5.244 - 1 0 . 254
111111
-
1 5.264 - 20.274
20.274 - 25.283
1 0 .254 - 1 5 . 264
(a)
'
·1
-� - ·
..
:::.,...--···"
TEMPERATURE IN CREASE ("C)
111111
P:':\t<::':l
1 .836 - 1 .9 1 5
1 .9 1 5 - 1 .993
1">·'-h>l
[==:J
2.072 - 2. 150
2 1 50 - 2.229
1 . 993 - 2.072
(b)
Figw·e 7
122
Finite Element Simulations of Current Density and Temperature Distributions
in 7itngslenjilled Vias
Vol. 1 Nu. 2
Spring 19')1
Digital Tecbnicaljournal
Electromigration Reliability of VLSJ lnterconnect
1 03
a.
£�
;;
.:::::_ 1 02
"'
�
0
�
10
Nresr/ Nchip
KEY:
--B-8--v-
-B-
Figure 9
Figure 8
=
=
=
=
=
Requirements on t ,0 for the Test
Structure Rela tive to t 0 1 for the
Secondary Electron Micrograph
Chip
of Electromigration Damage
as
a Function of the Ratio
of the Nu mber ofElements in the
at a Tungsten jilted Via
Test Structure to That of the Chip
points from the example shown in Figure
4 for com­
set by the confid ence l i mits on
cr
and
t50 . The most
stringent criterion for meeting the test perfor­
parison.
Points A and c both l ie on the edge of the al lowed
parameter space in Figure
l a rger for
1 .5, N 1 0
1 .5, N 200
-y = 2.09, N
10
-y = 2.09, N
200
-y =
-y
4, but
y and
N are both
C. Figure 9 shows that the requirements
on t50, ,est are greater for C than for A. However,
the relative effect of increasing N o r y separately is
mance requ irements is to use the lower l imit of the
confidence interval for t50 from the test data, which
corresponds to the minimu m e val ue consistent
with the values of y and N.
Electromigration l ifetimes, and thus the scal ing
N alone, that is, comparing
model parameters, are a sensi t ive function of the
point Bl to A , actually results in a decrease in the
m icrostructure of the conductor film . t9 Therefore,
test requirements, as shown i n Figure 9. This effect
si nce the effect of normal microstructural varia­
not clear. Increasing
N1e5/Ncbip ' ratio approaches u n i ty.
tions cannot be u nambiguously determined a pri­
Thus, the test l i ne shou ld be as long as practical to
ori, a number of lots must be tested to assess the
mitigate the impact of uncertainty in /. The increase
effects of lot-to-lot variations.
weakens as the
in
t5o, tes/t01, ,bip' therefore, is a result of increasing y.
The sensitivity to increasing y alone can be seen in
Each
lot
undergoes
a
statistical test of the
hypothesis
Figure 9, by comparing point 82 to A.
H(10 years) :os 0. OJ
To assure that the chip rel iabil i ty goals are met,
(11)
H(IO years) denotes the cumu lative proba­
10 years of continu­
the most conservative combination of parameter
where
estimates shou ld be used . These val u es are the
bil ity of chip failure i n the first
most stringent i n setting and meeting the test per­
ous operation under worst- case conditions. To pass
formance requirements. The most rigorous test per­
the test, the probabi l i ty of accepting the hypothesis
formance requirements are set by using the largest
when
possible value for y. In the example presented i n
Figure 9, this value o f y impl ies the l argest value of
N, i . e . , the smal lest value for l, within the ranges
Digital Technical journal
Vol. 4 No. 2
Spring 1992
H(lO years) is greater than 0.01 must be less
than 0.1 . This criterion gives at least 90 percent
confidence that a test group passing the statistical
test, i . e . , for which the hypothesis is accepted,
123
Semiconductor Technologies
•
Does indeed come from a lot that meets the
Twentieth A nnual International Reliability
Physics Symposium ( 1982): 292- 299.
requirement of not fa iling more than 1 percent of
the time in 10 years
•
3.
Is not a statistical fluke
Devices, vol . ED-34,
673 - 680.
ment this model in the electromigration qualifi­
cation testing are coded in a soft ware tool. The
extracts
the
most
conservative
MLN
model parameters from the fa ilure-time distribu­
4.
tion measured for every type of structure tested for
currently
tical test just described to verify that the rel iabi l i ty
5.
6.
We have presented a scaling model for relating
minum Thin-film
7.
8.
of the fol lowing individuals: Barbara Mi ner, Steven
Bill, and Jamie Rose for TEM work, Aldo Pe l i l lo for
A nnual International Reliability Physics
Symposium (1980): 1 65 - 170.
9.
sor Carl Thompson of the Massachusetts Institute
m i num Thi n-film Stripes," Applied Pbysics
Letters, vol. 3 6, no. 12 Oune 1980): 9 68- 970.
10.
). Cho and C. Thompson, "Grain Size Depen­
dence of Electromigration-induced Fa ilures
T. Kwok and P. Ho, " E iectromigration i n Th in
in Narrow Interconnects," Applied Physics
Letters, vol. 54, no. 25 Oune 1989): 2577-2579.
Film s." Dif
fusion Phenomena in Thin Films
and Micmelectronic J\t[aterials, edited by
1 24
E. Kinsbron, "A Model for the Width Depen­
dence of Electromigration Lifetimes in Alu­
of Technology.
in VlSI Interconnects," Proceedings of the
S. Va idya, D. Fraser, and A . Sinha, " E lectro­
Applications," Proceedings of the Eighteenth
FIB/SEM work, Ahsan Enver fo r finite element simu­
P. Ghate, " E lectromigration- induced Fa ilures
) . Lloyd and J. Kitch i n , "The Electromigration
migration Resistance of Fine- l ine AI for VlSI
We would like to acknowledge the contributions
2.
Conductors," journal of
journal of Applied Physics, vol. 69, no. 4
(February 1991 ): 2117- 2127
the Alpha 2106 4 mi croprocessor.
references therein.
Electromigration-induced
Failure Distribution: The Fine Line Case,"
assurance of the CMOS-4 process technology and
D. Gupta and P Ho (Park Ridge, New jersey:
of
Applied Physics, val . 41, no. 10 (September
1970): 3954- 3960.
on test structures to the overall chip reliability. This
Noyes Publications, 1988): 369-431 , and the
Symposium
Failure Time on Length and Width of Alu­
the resu lts of accelerated electromigration life tests
1.
Physics
B. Argarwala, M. At tardo, and A. Ingraham,
" Dependence
chip reliabi l i ty.
References
]. Black, " Mass Transport of Al u m inum by
national Reliability
(1967): 148- 159.
able models to relate test device data to long-term
lations, john Kitchin, Bill Marti n , Dave Dunn e l l ,
Resulting
of 2," journal of
Momentum Exchange with Conducting Elec­
to rigorously characterize the various components
Dave Foggo, Te rry Spooner, Lesley E l l iott, and Profes­
Electromigration
Exponent
trons," Proceedings ofthe Sixth A nnual Inter­
tion of CMOS technology. Therefore, it is necessary
Acknawledgments
with
3890-3893.
inc reasingly important with each step i n the evolu­
model was used as the basis for formu lating qual ifi­
3 (March 1987):
Applied Physics, vol. 59, no. 11 Oune 198 6) :
Interconnect electromigration rel iabi l i ty becomes
cation requirements for electromigration reliabil ity
no.
M. Shatzkes and ]. Lloyd, "A Model for Con­
i n a Current
goal has been met.
of the circu it metall ization and to develop depend­
Mea­
ductor Failure Considering Diffusion Con­
each lot. This tool can be used to perform the statis­
Summary
of Electromigration
surements," IEEE Transactions on Electron
The statistical analysis procedures used to imple­
software
H. Schafft, T. Staton, ]. Mandel, and J. Shott,
" Reproducibi l it y
11.
D. Walton, H . Frost, and C . Thompson, " Com­
puter Simulation of Grain Growth i n Thin­
film Interconnect Line s," Materials Research
Society
Symposia
Proceedings,
vol .
225
(1991): 219-224.
Vol. 4 No. 2
Sp-rirzg 1992
Digital Techr�ical]ournal
Electromigration Reliability of VIS/ Interconnect
12.
E. Arzt and W Nix, "A Model fo r the Effect
1 6.
Electromigration
Failure
IEEE Transactions on Electron Devices, vol.
of Interconnects
ED- 37, no. 3 (March 1990): 5 62 - 5 67.
with ' Near-bamboo' Grain Structures," jour­
nal of Materials Research, vol. 6, no. 4 (April
1991): 731 -73 6 .
13.
15.
17.
H . Rathore, "Via Resistance as a Technique to
Control the Electromigration of Non-overlap
I. Blech, " E lectrom igration in Thin Aluminum
Via Holes,"
Films o n Titanium Nitride," journal · of
Applied Physics, vol. 47, no. 4 (Ap ril 1976):
A nnual International Reliability Physics
1203- 1208.
14 .
F. Matsuoka et a l . , " E lectromigration Reliabil­
ity for a Tungsten-filled Via Hole Structure,"
of Line Width and Mechanical Strength on
Proceedings of the Twentieth
Symposium (1982): 77- 80.
18.
]. Estabi l, H. Rathore, and F. Dorleans, "The
E. Atakov and ). Clement, " Electromigration
E ffect
Failure Time Distribution: Scaling from a Test
Structure to the VLSI Chip," Proceedings of
migration-induced Shorts in Submicron Tech­
nology,"
the Nin th VLSI Multilevel Interconnection
A n n ual International Reliability Physics
Conference (1992): 366 - 371 .
Symposium (1991): 57- 6 3.
M . Garver et a l . , "CMOS- 4 Back-end Process
Development for a VLSI 0.75 -J.Lm Triple-level
19.
of
Metal
Thickness
on
Electro­
Proceedings of the Twenty-ninth
P. Smith, ). Lloyd , and G. Prokop, "Lot-to- lot
Va riations in Electromigration Performance
Interconnection Technology," Digital Tech­
for
nical journal, vol. 4 , no. 2 (Spring 1992, this
Vacuum Science and Technology, vol. A2,
issue): 51 -72.
no. 2, (April -June 198 4): 220- 223.
Digital Technical journal
Vol. 4 No. 2
Spring 1992
Thin
Film
Microcircuits;' journal
of
125
I
Further Readings
The D igital Technical journal
pu blishes papers that e::tplore
the technologicalfoundations
ofDigital:� majo rproducts. Each
Journal focuses on at least one
Software Productivity Tools
Vol. I, No.
6, February 1988
VAXcluster Systems
Vol. I, No. 5, September I987
product area andpresents a
VAX 8800 Family
compila tion ofpapers written
Vol. 1, No.
by the engineet·s who developed
N etworking Products
the product. The contentfor
the jo urnal is selected by the
Jo urnal Advisory Board.
Vol.
4, February I987
I, No. 3, September I986
MicroVAX II System
Digital engineers who would
Vol. 1, No. 2, March 1986
like to contribute a paper
VAX 8600 Processor
to the Journal should contact
Vol. I, No. I, August I985
the editor at RDVAX::BlA KE.
Topics covered i n previous issues of the
Digital Technicaljoumal are as fo l lows:
Subscriptions to the Digital Technicaljournal are
ava i l able on a yearly, prepaid basis. The subscrip­
tion rate is
$40.00 per year (four issues).
Requests
should be sent to Cathy Phi l l ips, Digital Equipment
PATIIWORKS: PC Integration Software
Vol.
4, 'Vo. I, Winter I992
Image Processing, Video TerminaJs,
and Printer Technologies
Vul.
3, No.
Corporat ion,
MA 01754,
�·ILO I - 3/B68, 146
Main Street, Maynard,
U.S.A. Subscriptions must be paid in
U.S.
dol lars, and checks should be made payable to
Digital Equipment Corporation.
Single copies and past issues of the Digital
4 Fal/ I99I
,
Availab ility in VAXcluster Systems/
Network Performance and Adapters
Technicaljournal can be ordered from D igital
Press at a cost of $ 16.00 per copy.
ltb/. 3, No. 3, Summer I9fJI
Technical Papers by Digital Authors
Fiber Distributed Data Interface
C. Brench, "Opti mizing EM! Shield Design with
Vol.
3, No.
2, Spring 191JI
Transaction Pro cessing, Databases,
and Fault- tolerant Systems
Vol.
3, No. I, Winter I99I
Society (March
1992).
]. Clement and ]. Lloyd , " N u m erical Investigations
of the Electromigration Boun dary Value Problem,"
VAX 9000 Series
jo urnal ofApplied Ph-ysics (February
Vol. 2, No. 4, Fall I990
DECwindows Program
Vol. 2, No.
Numerical Techniques," Conference Proceedings
of the Applied Comp u tational Electromagnetics
3, Summer I990
1992).
S. Das, "The MR Head : An Emerging Trend in Hard
Disk Drive Head Development," Electrochemical
Society In ternational Symposium on Magnetic
VAX 6000 Model 400 System
Ma terials, Processes and Devices (1991 ).
Vol. 2, No. 2, Spring I990
S. Das, " Selective Laser Microetching," Electro­
Compound Document Architecture
Vol. 2, No. I, Win ter I990
1991 ).
C . Derry, "Transition from Qual ity C ircles t o Total
Distributed Systems
Vol. I, No. 9, ju ne I989
Storage Technology
Vol. I, No. 8, February
chemical Society international Symposium on
Electrochemical Microfabrication (October
1989
Involvement in Semiconductor Manufacturing,"
Quality and Reliability Engineering Interna­
7,
tional, vol
no.
6 (November-December 1991 ) .
B. Doyle, "Recovery o f Hot- Carrier Damage in
CVAX-based Systems
Reox idized Nitride Oxide MOSFETs," IEEE Electron
Vol. I, No. 7, A ugust I988
Device Letters Oanuary
1 26
VtJ/. 4 \'u. 2
.
1992).
Spring 1992
Digital Technical journal
I
B. Doyle, B. Fishbein, and K. Mist ry, "NBTI­
Enhanced Hot -Carrier Dam age in p-Channel
D. Uttley and S. Coverdale, " ReTAB for Test and
Failure Analysis," Proceedings of the Seventeenth
MOSFETs," IEEE International Electron Devices
International Symposium for Testing and Failure
Meeting (December 1991 ).
A nalysis (November 1991 )
A. Garde! and P. Deosthali, "Using Simu lation to
Test the Robustness of Various Existing Produc­
Applications and Limitations in an Industrial
N. Veisfeld, " Microbeam Surface Analysis Methods:
tion Control Policies," 1991 Winter Simulation
Environment," Proceedings of the Microbeam
Conference (December 1991 ).
A nalysis Society ofA merica (Fa l l 1991 ).
S. Heng, "Temperature Mapping of Local ized Hot
A. Vitale, "An Algorithm for High Accuracy Name
Spots on Microelectronic Chip Surfaces," ASME
Pronunciation by Parametric Speech Synthesizer,"
Transactions: journal of Electronic Packaging
journal of Computational Linguistics (September
(September 1991 ).
1991 ).
S. Heng and H. Pei, "Air Impingement Cooled
Pin-Fin Heat Sink for Multi-Chip Unit," National
Electronic Packaging and Production Conference
(February 1991 ).
A. Mason, M . Zurko, C. Kahn, P. Karger, and
X. Wang, "Test Generation Based on Dynamic
Search Space Reductions," International Phoenix
Conference on Computers and Communications
(March 1990).
W Zahavi, "Capacity Planning/Needs Analysis:
D. Bonin, "A VMM Security Kernel for the VAX
The Economy Version," Computer Measurement
Architecture," IEEE Symposium on Research
Group Conference (December 1991).
in Secu rity and Privacy (May 1990).
K. Mist ry, B. Doyle, A. Phil ipossian, and D. Jackson,
"AC Hot Carrier Lifetimes i n Oxide and ROXt"\IOX
N-Channel MOSFETs," IEEE International Electron
Devices Meeting (December 1991).
Digital Press
Digital Press is the book publ ishing group of
Digital Equipment Corporation. The Press is an
i n ternational publisher of computer books and
K. Ramakrishnan, "A Model of Naming for Fine
jou rnals on new technologies and products for
Grained Service Specification in Distributed
Systems," Proceedings of the Eleventh Interna­
users, system and network managers, program­
tional Conference on Distributed Computing
Systems (May 1991).
for books in these and related areas are welcomed.
mers, and other professionals. Proposals and ideas
The fol l owing book descriptions represent a
K. Ramakrishnan, "A Naming System for Feature­
sample of recent publ ications available from
based Service Specification in Distributed
Digital Press.
Operating Systems," Proceedings of the 1991
ACM SIGSM!I LL/PC Symposium on Small Systems
(June 1991 ).
J. Rose, B. Minor, and A. Pel i l lo, "TEM : A Review
MOTIF PROGRAMMING:
The Essentials . . . and More
Marshall Brain, 1992, softbound, 632 pages,
Order No. EY-J816E-DP-EEB ($ 29.95).
and a Comparison with H igh Resolution SEM,"
International Symposium for Testing and Failure
A straightforward and easy-to-understand i ntro­
A nalysis (November 1991 ).
duction to Motif application development, this
K. Springer, "A Forward Error Correcting Code for
Gigabit Fiber Optic Links," Proceedings ofSPIE
­
The International Society of Optical Engineering
(September 1991) .
book will ease you into Motif program ming as
smoothly and quickly as possible. I t starts with an
introduction to event-d riven programming and
proceeds to discuss three concepts essential to
Motif programming: resources, cal I backs, and
R. U l ichney, " O n t h e Manipulation o f Power
containers. Advanced topics will expose the
Spectra of Halftone Patterns," IST (SPIE) Seventh
reader to a l l of the Motif widgets, the capabilities
International Congress on Advances in Non­
of the X and Xt layers, the X drawing model, and
Impact Printing Technology (October 1991) .
the process of application design in Mot if.
Digital Tech71ical jour11al
Vol. 4 No. 2
Spring 1992
1 27
Further Readings
VMS FOR ALPHA PlATFORMS
INTERNALS AND DATA STRUCTURES:
work with ALL-IN-I . Based on the author's ten years
of experience in developing ALL-IN-1 subsystems
Preliminary Edition, Volume 1
and in customizing its application to specific
Ruth E. Goldenberg and Saro Saravanan, 1992,
softbound, 41 6 pages, Order No. EY-L466E-P l-EEB
the product documentation to explore the deep
customer sites, the presentation extends beyond
($ 30.00).
and distant corners of the product. The wealth of
This volume is the first publication to explain
examples of actual installation and customization
experiences help communicate how to best use
VMS operating system support for Alpha, Digital
Equipment Corporation's recently announced
6 4-bit RISC architecture. Volume 1 contains twelve
chapters and an appendix. Subsequent volumes
will be published as more chapters become
available.
computers.
RELIABLE COMPUTER SYSTEMS:
Design and Evaluation, Second Edition
Daniel P Siewiorek and Robert S. Swarz, 1992,
THE X WINDOW SYSTEM SERVER:
X Version 1 1 , Release
ALL-IN-I on VAX, DOS PC, and Apple Macintosh
5
hardbound, 908 pages, Order No. EY-H880E-DP-EEB
($ 6 4.95) .
Elias Israel and Erik Fortune, 1992, softbound,
This major revision of the 1982 publ ication,
534 pages, Order No. EY-L518E-DP-EEB ($45.95).
The Theory and Practice of Reliable System
This technical reference covers every aspect of the
sample server developed by the M IT X Consortium.
It provides essential information to knowledgeable
X users who want to learn about the basic inter­
actions between cl ient and server-including
developers who want to port, extend, tune, or
test a server. Examples, guidelines, and tutorials
reinforce material on theory and practice.
SOFTWARE IMPLEMENTATION TECHNIQUES:
VMS, UNIX, OS/2, and MS-DOS
Donald E. Merusi, 1992, softbound, 743 pages,
Order No. EY-J82 2E-DP-EEB ($39.95).
Written for programmers and software developers,
Design, provides an up-to-date and comprehen­
sive guide to the design, evaluation, and use of
reliable computer systems. It includes case studies
of systems from manufacturers such as Tandem,
Stratus, IBM, and Digital as well
as
the special
Gal ileo faul t protection and AT&T tekphone
switching processor systems.
ALPHA ARCHITECTURE REFERENCE MANUAL
Edited by Richard L. Sites, 1992, softbound,
600 pages, Order No. EY-L520E-DP-EEB ($ 34.95)
.
Written by the engineers who developed the
Alpha specifications, this book contains complete
descriptions of the common architecture required
this book eases the transition from one operating
for all implementations and the interfaces required
to support the OSF/ 1 and OpenVMS operating sys­
system to another. Topics help you u nderstand
tems. Of great interest to software developers,
the capabilities needed to design applications that
system managers, system analysts, and system
run on one or more of these systems or to port
existing applications to a new system. This book
will reduce the time you need to gather informa­
tion to peform a specific task, and diminish your
frustration level by showing you where to turn for
programmers, this is the authoritative reference
on Alpha, Digital Equipment Corporation's new
6 4-bit architecture.
additional help.
To receive a copy of the latest catalog or further
information on these or other pub I ications from
ALL-IN-1: A Technical Odyssey
Digital Press
Tony Redmond, 1992, softbound, 550 pages,
Department EER
Order No. EY-H952E-DP ($ 44.95)
Digital Press, please write:
.
This extensive treatment of Digital Equipment
1 Burl i ngton Woods Drive
Burlington, MA 01803-4597
the needs of system managers, application pro­
Or, you can order a Digital Press book by calling
DECdirect at 800-DIGITA.L (800 -344-4R2S) . When
grammers, and technically oriented users who
ordering be sure to refer to Catalog Code EEB.
Corporation's office automation tool addresses
1 2H
Vol. 4 No. 2
Sprin� 19')2
Digita/ 1echnical journal
ISSN 0891-90 1 X
">
/
다운로드
친절한 알림입니다. 여기에서 문서를 볼 수 있습니다. 그러나 가장 중요한 것은 AI가 이미 읽었다는 것입니다. 복잡한 내용을 간단하게 설명하고, 모든 언어로 질문에 답변하고, 가장 길거나 복잡한 문서도 빠르게 탐색할 수 있도록 도와줍니다.
광고