234 Pages
English

High-level optimization of performance and power in very deep sub-micron interconnects [Elektronische Ressource] / von Tudor A. Murgan

-

Gain access to the library to view online
Learn more

Description

High-LevelOptimizationofPerformanceandPowerinVeryDeepSub-MicronInterconnectsVom Fachbereich 18Elektrotechnik und Informationstechnik¨der Technischen Universitat Darmstadtzur Erlangung der Wu¨ rde einesDoktor–Ingenieurs (Dr.-Ing.)genehmigte DissertationvonDipl.-Ing.TudorA.Murgangeboren am 7. Februar 1978in Bukarest, Ruman¨ ienReferent: Prof. Dr. Dr. h. c. mult. Manfred GlesnerTechnische Universita¨t DarmstadtKorreferent: Prof. Dr. Mircea R. StanUniversity of Virginia, CharlottesvilleTag der Einreichung: 18.07.2006Tag der mu¨ ndlichen Pru¨ fung: 06.10.2006D17Darmstadt¨ er DissertationenFamiliei mele. . .dedics¸i ma˘ dedic.“Arbor invers am ra˘mas, rupt din sfera˘cu sfera aceasta aidoma, geama˘na˘. . .S¸i totul ˆımi pare s¸tiut, dar nimicadin ce s¸tiu cu ce este nu se aseama˘na˘. . . ”Nichita Sta˘nescu“Nihil tam bene dictum quod non possit depravari.”AbælardusPrefaceThis dissertation is the consequence of the work as a teaching and research assistant atthe Institute of Microelectronic Systems of the Darmstadt University of Technology. Iwould like to sincerely thank my Doktorvater, Prof. Manfred Glesner, not only for his kindadvice and guidance that made this thesis possible, but also for giving me the opportunityto be involved in several teaching activities and consistent research projects funded byimportant companies and scientific foundations.I also express my gratitude towards Prof.

Subjects

Informations

Published by
Published 01 January 2006
Reads 65
Language English
Document size 2 MB

High-LevelOptimizationof
PerformanceandPowerin
VeryDeepSub-MicronInterconnects
Vom Fachbereich 18
Elektrotechnik und Informationstechnik
¨der Technischen Universitat Darmstadt
zur Erlangung der Wu¨ rde eines
Doktor–Ingenieurs (Dr.-Ing.)
genehmigte Dissertation
von
Dipl.-Ing.
TudorA.Murgan
geboren am 7. Februar 1978
in Bukarest, Ruman¨ ien
Referent: Prof. Dr. Dr. h. c. mult. Manfred Glesner
Technische Universita¨t Darmstadt
Korreferent: Prof. Dr. Mircea R. Stan
University of Virginia, Charlottesville
Tag der Einreichung: 18.07.2006
Tag der mu¨ ndlichen Pru¨ fung: 06.10.2006
D17
Darmstadt¨ er DissertationenFamiliei mele. . .
dedic
s¸i ma˘ dedic.
“Arbor invers am ra˘mas, rupt din sfera˘
cu sfera aceasta aidoma, geama˘na˘. . .
S¸i totul ˆımi pare s¸tiut, dar nimica
din ce s¸tiu cu ce este nu se aseama˘na˘. . . ”
Nichita Sta˘nescu“Nihil tam bene dictum quod non possit depravari.”
Abælardus
Preface
This dissertation is the consequence of the work as a teaching and research assistant at
the Institute of Microelectronic Systems of the Darmstadt University of Technology. I
would like to sincerely thank my Doktorvater, Prof. Manfred Glesner, not only for his kind
advice and guidance that made this thesis possible, but also for giving me the opportunity
to be involved in several teaching activities and consistent research projects funded by
important companies and scientific foundations.
I also express my gratitude towards Prof. Mircea Stan from the University of Virginia
in Charlottesville, who kindly accepted to act as a reviewer for this thesis. His comments
and observations have been very valuable for improving the quality of the work, and as
a result of our discussions, new ideas emerged.
Furthermore, I am indebted to Prof. Jur¨ gen Adamy, Prof. Alex Gershman, and Prof.
¨Jurgen Stenzel as members of the examination committee. I would especially like to thank
Prof. Gershman for the fruitful discussions around the topic of the dissertation.
This work could not have been accomplished smiling without a good atmosphere at
the working place. For the provided pleasant environment, I would like to express spe-
cial thanks to all colleagues at the institute, with whom I had the pleasure of carrying out
interesting research projects, producing dozens of papers and reports, sharing various
teaching activities, and solving numerous stringent administrative issues. The friendly
help and support of Petru Bacinschi, Oana Cobianu, Andre Guntoro, Heiko Hinkelmann,
Hans-Peter Keil, Massoud Momeni, and Hao Wang permitted me – especially during the
last months – to concentrate on writing the final manuscript and on preparing the exam.
Moreover, I will never forget the late-night spicy discussions, spanning from scientific and
projects-related to political and historical issues, with the colleagues of the same “genera-
tion” (or not - sic!), with whom I spent at least four years at the institute: Leandro Soares
Indrusiak, Abdulfattah Obeid, Sujan Pandey, and Oliver Soffke. In this context, I would
especially like to mention the collaboration with Mihail Petrov, with whom I shared for
many years office, projects, reports, presentations, meetings, deadlines, worries, and wa-
tering of plants.
I am also greatly indebted to the older, former, and external colleagues who shared,
on various occasions, their experience regarding a multitude of issues like writing pa-
pers and project proposals, finalizing project reports, and understanding of the given
facts (f.i. earth rotation): Jur¨ gen Deicke, Alberto Garc´ıa Ortiz, Thomas Hollstein, Lukusa
iiiiv PREFACE
Kabulepa, Ralf Ludewig, Octavian Mitrea, Juan Jesus´ Ocampo Hidalgo, Thilo Pionteck,
Matthias Rychetsky, Clemens Schlachta, Burkart Voß, Heiko Zimmer, Peter Zipf, Klaus
Koch, Ma´rcio Kreutz, Jose´ Palma, and Gilles Sassatelli.
Further, I would like to thank the former students I have been able to work with
throughout these years, especially Ismail Deflaoui, Mateusz Majer, Oliver Mitea, and
Alexander Werth. Thanks also to the competent and good-humored assistance of the sys-
tem administrator, Andreas Schmidt, as well as to the friendly support of our secretaries,
Silvia Hermann and Irmgard Wackermann.
In particular, I would like to acknowledge the exceptional support received from my
colleagues Alberto Garc´ıa Ortiz, Ralf Ludewig, Massoud Momeni, and Petru Bacinschi.
Alberto and Ralf have been not only my first research mates, but also patient “initiators”
and “real-time online debuggers”. The comments and constructive criticism of Alberto,
Massoud, Petru, and Ralf provided a solid foundation for finding the “red line”, carrying
out the research, and clarifying several key issues. My good old friend Nicolae Statu
gave very generously of his time to read the manuscript and provide essential feedback
regarding the language used.
During the last five years, I had the chance to visit many conferences, institutions, and
universities around the world and thus, to meet many researchers and design engineers.
Those more constant or punctual interactions contributed significantly to finalizing and
improving the results of this work. In this context, I would like to mention the discus-
sions with Prof. Enrico Macii, Prof. Mircea Stan, Dr. Vladimir Zolotov, Prof. Luis Miguel
Silveira, Prof. Radu Ma˘rculescu, Prof. Ricardo Reis, to name only a few.
My stay in Darmstadt would have not been possible without the professional and
friendly support received from several people from the “Politehnica” during and after my
studies in Bucharest. Therefore, I would like to gratefully acknowledge the continuous
encouragement of Prof. Anca Manuela Manolescu, Prof. Anton Manolescu, Prof. Felicia
Ionescu, and Prof. Radu Dogaru.
Old and new friends allowed us a seamless adaptation to the living in a new city
and a new country. For the wonderful time spent together, many thanks to Giuliana and
Leandro, Alina and Radu, Blanca and Alberto, Monica and Nicolae, Oana and Tavi, Mis¸u,
Ralf, Jayjay, Claudia and Burkart, Andrea and Matthias, and all those fantastic friends
which we have been meeting only roughly twice a year when traveling to Bucharest.
Last but not least, I wish to express my heartfelt gratitude to my entire family for all
their efforts, for the received education and opportunities, for their unconditional contin-
uous support, encouragement, trust, and friendship. The deepest thoughts to my inspir-
ing wife, Ilina.
Darmstadt, November 2006Abstract
Interconnect analysis and optimization at high levels of abstraction is extremely attractive
since it offers a much larger room for improvement than optimization at lower levels. The
objective of this thesis is to optimize performance and power consumption in intercon-
nect structures at high levels of abstraction. For this purpose, efficient high-level models
for delay and power consumption in very deep sub-micron interconnects are developed
and employed for constructing and evaluating different low power and throughput im-
proving signal encoding schemes. Moreover, in order to achieve an even higher efficiency,
coding is combined with lower level techniques like spacing, shielding, and buffer plan-
ning.
In order to construct and evaluate encoding schemes at high levels of abstraction, two
conceptually different issues must be solved. On the one hand, bit-level characteristics of
the data transmitted over the interconnect structures need to be extracted during system-
level design and architecture specification. On the other hand, important interconnect-
related very deep sub-micron effects have to be incorporated into high-level models as
well. Delay models able to predict the line delay for each set of input patterns (and not
only for the worst case) are required in order to develop and evaluate coding schemes
tailored for performance improvement. An essential contribution of this work is the de-
velopment of a pattern-dependent delay model. The essence of the so-called extended
linear delay (ELD) model is to incorporate the effects of all possible input patterns in
buses exhibiting not only inter-wire capacitance, but also inductive effects which are in
general more difficult to predict and more daunting because of their long-range nature.
Further, the described power macromodel shows that in order to decrease dynamic power
consumption at high levels of abstraction, one has to reduce not only the self transition
activity but also the so-called coupling transition activity responsible for charging and
discharging the inter-wire capacitances in a bus.
The abovementioned models are employed in order to construct and evaluate sev-
eral low-power and throughput improving codes. Based on the observation that the bit-
level transition activity in typical DSP applications can be accurately described by two
breakpoints, several simple yet very efficient hybrid codes are constructed. Those codes
combine non-redundant and redundant schemes in such a way that the total self and
coupling transition activity are significantly decreased. Moreover, maximum achievable
limits are derived, which show the effectiveness of the developed codes. Further, several
vvi ABSTRACT
low-complex codes are proposed that improve bus performance by avoiding a certain set
of input patterns. In this context, fundamental limits and bounds are derived for state
and transition coding, respectively. Coding is also compared and combined with low-
level interconnect optimization techniques like spacing and shielding. The problem of si-
multaneously addressing coding-based power reduction and performance improvement
is introduced and analyzed.
Finally, an interconnect-centric design flow is presented that integrates signal encod-
ing for power and performance optimization. Signal encoding schemes can be constructed
at high levels of abstraction while analyzing the data that is transmitted through the in-
terconnect system. After interconnect planning and synthesis, when exact information
regarding layout and routing optimization is available, codes can be refined based on
the specific wire topology. Moreover, in order to prove the large optimization opportu-
nities available at high levels, a simultaneous buffer insertion and placement algorithm
is developed. In this context, coding for throughput is appended to the developed algo-
rithm, and it is shown that performance and/or power consumption can be thus further
improved.Kurzfassung
Die Analyse und Optimierung von Verbindungsstrukturen in integrierten Schaltun-
gen auf hohen Abstraktionsebenen ist a¨ußerst attraktiv, da diese im Vergleich zu
niedrigeren Ebenen deutlich mehr Verbesserungsmo¨glichkeiten anbieten. Ziel dieser
Dissertation ist es Leistungsverbrauch und Performanz in Verbindungsstrukturen auf
hohen Abstraktionsniveaus zu optimieren. Fur¨ diesen Zweck werden effiziente Mo-
delle fur¨ Signalverzo¨gerung und Leistungsverbrauch in sub-100 nm (very deep sub-micron)
Verbindungsstrukturen erstellt und anschließend auf die Entwicklung und Bewertung
verschiedenartiger Kodierungsmethoden angewendet, die den Leistungsverbrauch ver-
¨ringern und den Datendurchsatz verbessern. Um eine noch großere Effizienz zu erzie-
len, werden Kodierungsschemen mit Methoden wie Abstandvergro¨ßerung, Abschir-
mung und Einfug¨ en von Leitungstreibern kombiniert, die ub¨ licherweise auf den unteren
Abstraktionsebenen angewandt werden.
¨ ¨Um wirksame Kodierungsmethoden entwickeln und analysieren zu konnen, mussen
zwei grundlegend verschiedene Probleme gelo¨st werden. Auf der einen Seite ist fur¨
die Entwicklung der Kodierungsmethoden notwendig, auf der System- und Architek-
turebene bedeutende Eigenschaften der zu sendenden Daten zu extrahieren. Auf der
anderen Seite ist fur¨ deren korrekte Bewertung erforderlich, wesentliche technologiebe-
dingte Effekte in Makromodellen auf ho¨heren Abstraktionsebenen einzubinden. Die
Evaluierung von Kodierungstechniken, die den Durchsatz erho¨hen, kann nur dann erfol-
gen, wenn die verwendeten Verzo¨gerungsmodelle die von allen mo¨glichen Eingangstran-
sitionen erzeugten Verzo¨gerungen vorhersagen ko¨nnen und nicht nur die ungun¨ stigsten
(worst case) Fa¨lle betrachten. In diesem Zusammenhang wird in dieser Arbeit ein tran-
sitionsabha¨ngiges Verzo¨gerungsmodell entwickelt, das sowohl kurzreichende kapazi-
tive Kopplungen als auch weitreichende und somit unub¨ ersichtlichere induktive Effekte
beuc¨ ksichtigt. Des Weiteren wird auch ein Makromodell fur¨ den Leistungsverbrauch
beschrieben. Dieses Makromodell zeigt im Wesentlichen, dass sich die Optimierung des
dynamischen Leistungsverbrauchs auf hohen Abstraktionsebenen auf die Verringerung
sowohl der Eigenschaltaktivita¨t als auch der sogenannten Koppelschaltaktivita¨t re-
duziert, die fur¨ das Umladen der Koppelkapazita¨ten verantwortlich ist.
Im Laufe der Arbeit werden die oben genannten Modelle fur¨ die Entwicklung
und Bewertung von verschiedenartigen und optimierten Kodierungsmethoden ver-
wendet. Basierend auf der Beobachtung, dass die Schaltaktivita¨t in typischen Sig-
viiviii KURZFASSUNG
nalverarbeitungsarchitekturen mittels zwei sogenannter Grenzpunkte modelliert wer-
den kann, werden verschiedene hybride Kodierungstechniken entwickelt, die nichtre-
dundante und redundante Methoden kombinieren, sodass die Eigen- und Koppel-
schaltaktivita¨t stark reduziert werden. Ferner werden theoretische Schranken fur¨ die
Reduzierung der Schaltaktivita¨t abgeleitet, um die Effektivita¨t der vorgeschlagenen
Kodierungstechniken nachzuweisen. Außerdem werden mehrere durchsatzerho¨hende
Kodierungsmethoden entwickelt, in denen eine bestimmte Menge von Eingangstransitio-
nen ungul¨ tig gemacht wird. In diesem Zusammenhang werden sowohl grundsa¨tzliche
Schranken fur¨ Zustands- und Transitions-Kodierung berechnet als auch Vergleiche
mit Verzo¨gerungsoptimierungsmethoden wie Abstandvergro¨ßerung und Abschirmung
durchgefuh¨ rt, die auf niedrigeren Abstraktionsebenen angewandt werden. Es wird
gezeigt, dass durch das Zusammenlegen von Kodierung und solcher Methoden eine
verbesserte Effizienz erreicht werden kann.
Schließlich wird eine Entwurfsmethodik fur¨ integrierte Schaltungen und Systeme
beschrieben, in deren Mittelpunkt die Optimierung von Verbindungsstrukturen steht.
Kodierungsschemen ko¨nnen im Wesentlichen wa¨hrend der ersten Entwurfsphasen ent-
wickelt und analysiert werden, da die bedeutendsten Eigenschaften der gesendeten
Daten zur gleichen Zeit extrahiert werden ko¨nnen. Nach der Planung und Synthese der
Verbindungsstrukturen und der dazugeho¨rigen Kodierungen sind exakte Details zu der
endgul¨ tigen Geometrie der Verbindungsstrukturen bekannt. Folglich ko¨nnen Kodierun-
gen weiter in einer leitungsspezifischen Weise verfeinert werden. Darub¨ er hinaus wird
ein Algorithmus entwickelt, der die Plazierung und das Einfug¨ en von Leitungstreibern
gleichzeitig durchfuh¨ rt, um somit die beachtlichen Optimierungsmo¨glichkeiten her-
vorzuheben, die auf hohen Abstraktionsebenen vorhanden sind. Die Erweiterung
des entwickelten Algorithmus mit Kodierungsmethoden erlaubt eine Verbesserung des
Durchsatzes und/oder des Leistungsverbrauchs.