Professional Documents
Culture Documents
Internet Autonomous Vehicles Hardware: Quantum Computing
Internet Autonomous Vehicles Hardware: Quantum Computing
Computing
Internet
Autonomous
Vehicles
Hardware
Evolving Career
Opportunities
Need Your Skills
Explore new options—upload your resume today
STAFF
Editor Publications Portfolio Managers
Cathy Martin Carrie Clark, Kimberly Sperka
Circulation: ComputingEdge (ISSN 2469-7087) is published monthly by the IEEE Computer Society. IEEE Headquarters, Three Park Avenue, 17th
Floor, New York, NY 10016-5997; IEEE Computer Society Publications Office, 10662 Los Vaqueros Circle, Los Alamitos, CA 90720; voice +1 714 821 8380;
fax +1 714 821 4010; IEEE Computer Society Headquarters, 2001 L Street NW, Suite 700, Washington, DC 20036.
Postmaster: Send address changes to ComputingEdge-IEEE Membership Processing Dept., 445 Hoes Lane, Piscataway, NJ 08855. Periodicals Postage
Paid at New York, New York, and at additional mailing offices. Printed in USA.
Editorial: Unless otherwise stated, bylined articles, as well as product and service descriptions, reflect the author’s or firm’s opinion. Inclusion in
ComputingEdge does not necessarily constitute endorsement by the IEEE or the Computer Society. All submissions are subject to editing for style,
clarity, and space.
Reuse Rights and Reprint Permissions: Educational or personal use of this material is permitted without fee, provided such use: 1) is not made for
profit; 2) includes this notice and a full citation to the original work on the first page of the copy; and 3) does not imply IEEE endorsement of any third-
party products or services. Authors and their companies are permitted to post the accepted version of IEEE-copyrighted material on their own Web
servers without permission, provided that the IEEE copyright notice and a full citation to the original work appear on the first screen of the posted copy.
An accepted manuscript is a version which has been revised by the author to incorporate review suggestions, but not the published version with copy-
editing, proofreading, and formatting added by IEEE. For more information, please go to: http://www.ieee.org/publications_standards/publications
/rights/paperversionpolicy.html. Permission to reprint/republish this material for commercial, advertising, or promotional purposes or for creating new
collective works for resale or redistribution must be obtained from IEEE by writing to the IEEE Intellectual Property Rights Office, 445 Hoes Lane,
Piscataway, NJ 08854-4141 or pubs-permissions@ieee.org. Copyright © 2021 IEEE. All rights reserved.
Abstracting and Library Use: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy for private use of patrons,
provided the per-copy fee indicated in the code at the bottom of the first page is paid through the Copyright Clearance Center, 222 Rosewood Drive,
Danvers, MA 01923.
Unsubscribe: If you no longer wish to receive this ComputingEdge mailing, please email IEEE Computer Society Customer Service at help@
computer.org and type “unsubscribe ComputingEdge” in your subject line.
IEEE prohibits discrimination, harassment, and bullying. For more information, visit www.ieee.org/web/aboutus/whatis/policies/p9-26.html.
2469-7087/21 © 2021 IEEE Published by the IEEE Computer Society September 2021 1
SEPTEMBER 2021 � VOLUME 7 � NUMBER 9
Long-distance
The distances between connectivity and the
quantum computers diversity of qubits
within a quantum farm are expected.
are relatively short.
The distances Qubits are most likely
Distance and Heterogeneity
8 14 34
mps Computer
Coplanar Waveguide
ier Transition to MCM
Time
Internet
20 Principles and Elements of Governance of Digital
Public Services
VIRGILIO ALMEIDA, FERNANDO FILGUEIRAS, AND FRANCISCO GAETANI
Autonomous Vehicles
31 Right Code
GERARD J. HOLZMANN
Hardware
42 High-Level Synthesis-Based Approach for Accelerating
Scientific Codes on FPGAs
RAMSHANKAR VENKATAKRISHNAN, ASHISH MISRA, AND
VOLODYMYR KINDRATENKO
Departments
4 Magazine Roundup
7 Editor’s Note: Moving Quantum Computing Forward
60 Conference Calendar
T he IEEE Computer Society’s lineup of 12 peer-reviewed technical magazines covers cutting-edge topics
ranging from software design and computer graphics to Internet computing and security, from scien-
tific applications and machine intelligence to visualization and microchip design. Here are highlights from
recent issues.
4 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
AI, signal processing, and crypto-
graphic applications. Interleaved-
Optimal Finite-Horizon AI-Driven Provisioning in the multithreading (IMT) processor
Perturbation Policy for 5G Core cores are interesting to pursue
Inference of Gene energy efficiency and low hard-
Regulatory Networks Network slicing enables com- ware cost for edge computing,
munication service providers to yet they need hardware acceler-
A major goal of systems biology is partition physical infrastructure ation schemes to run heavy com-
to model accurately the complex into logically independent net- putational workloads. Following
dynamical behavior of gene regu- works. Network slices must be a vector approach to accelerate
latory networks (GRNs). Despite provisioned to meet the service- computations, this article from
several advancements that have level objectives (SLOs) of dispa- the March/April 2021 issue of IEEE
been made in inference of GRNs, rate offerings. Network orches- Micro explores possible alterna-
two main issues continue to make trators must customize service tives to implement vector copro-
the problem challenging: (1) non- placement and scaling to achieve cessing units in RISC-V cores,
identifiability of parameters and the SLO of each network slice. In showing the synergy between IMT
(2) limited amounts of data. Thus, it this article from the March/April and data-level parallelism in the
becomes necessary to experimen- 2021 issue of IEEE Internet Com- target workloads.
tally perturb or excite the system puting, the authors discuss the
into different states. This perturba- challenges encountered by net-
tion process disrupts the expres- work orchestrators in allocating
sion of genes from active to inac- resources to disparate 5G net- Feature-Guided Spatial
tive, or vice versa, at each time work slices. They propose the use Attention Upsampling
point. Another issue is the partial of artificial intelligence to make for Real-Time Stereo
observability of the gene states, core placement and scaling deci- Matching Network
which must be inferred indirectly sions that meet the requirements
from noisy gene expression mea- of network slices deployed on In this article from the January–
surements. In this article from the shared infrastructure. March 2021 issue of IEEE MultiMe-
January/February 2021 issue of IEEE dia, the authors propose an end-
Intelligent Systems, the latter issue to-end real-time stereo matching
is accounted for by employing the network (RTSMNet). RTSMNet
partially observed Boolean dynami- Klessydra-T: Designing consists of three modules. The
cal system signal model for the data Vector Coprocessors global and local feature extrac-
and applying optimal state estima- for Multithreaded Edge- tion (GLFE) module captures the
tion. Then, the optimal finite-hori- Computing Cores hierarchical context information
zon perturbation policy is derived and generates the coarse cost
to achieve the highest possible Computation-intensive kernels, volume. The initial disparity esti-
expected performance for the such as convolutions, matrix multi- mation module is a compact 3D
maximum a posteriori estimator plication, and Fourier transform, are convolution architecture aim-
under a small perturbation cost. fundamental to edge-computing ing to produce the low-resolution
www.computer.org/computingedge 5
MAGAZINE ROUNDUP
Moving Quantum
Computing Forward
2469-7087/21 © 2021 IEEE Published by the IEEE Computer Society September 2021 7
EDITOR: Erik P. DeBenedictis, Zettaflops, LLC, erikdebenedictis@gmail.com
Quantum Access
Erik DeBenedictis, Zettaflops, LLC
Quantum computers are available via the Internet for students and small-scale
research. What if similar access could be extended to quantum hardware?
I
took a class on integrated circuit design when in an IEEE Quantum Initiative working group, and I’m
I was a student in 1981. The class project was for writing about it here as a noncommercial concept
each student to design and test a CMOS circuit. that IEEE Members could champion in the public
I designed a CMOS arbiter, a circuit used to reliably interest.
detect which of two input signals arrives first. The chip
was fabricated as part of a multiproject wafer during A BRIEF HISTORY OF
the first run of ARPA’s (now DARPA’s) MOS Imple- INTEGRATED CIRCUITS
mentation Service (MOSIS) and delivered to me as a Integrated circuits were originally designed by cut-
chip bonded into a package—like a simplified version ting plastic such as Rubylith and handcrafting what
of the one shown in Figure 1—and I tested it with an were essentially photographic negatives. The meth-
oscilloscope. od’s scale-up limitations are obvious. At some point,
However, the version shown in Figure 1 is a quan- companies started writing their own computer-aided
tum module built by a professional research team1 design (CAD) software, which codified design rules
out of components created by governmental and such as the minimum dimensions of wires or tran-
commercial fabricators (fabs) that could become sistors and the various spacings between them. The
the quantum equivalent of my 1981 CMOS chip. The design processes were proprietary because they
central gray box in Figure 1 contains qubits bonded gave the electronics manufacturer that sold the chips
on top of a somewhat larger classical electronics a competitive advantage in speed, density, and reli-
chip. Unlike my 1981 chip, some of the leads in Figure ability. Proprietary design processes preserved this
1 carry microwave signals, and the module operates competitive advantage but meant that the employ-
at millikelvins. Although the module in Figure 1 is not ees had to be trained on the job.
a complete quantum computer, its structure and the To make the semiconductor industry scalable,
process of creating it would give students hands-on universities championed relaxed “least common
experience or support experimental research proj- denominator” design rules that were independent of
ects within the bounds of the module’s external any specific process.2 Process-independent design
interfaces. rules would work on proprietary fab lines, thus fore-
Quantum technology is roughly as mature as shadowing the emergence of a separate CAD industry,
CMOS was in 1981, making it feasible to offer students semiconductor foundries, and a systematic way to
and small research groups access to quantum hard- train the workforce that has engineered all of the chips
ware prototyping. This concept has been discussed in use today.
ARPA’s MOSIS made chip design widely accessible.
The main idea was that students from many universi-
Digital Object Identifier 10.1109/MC.2020.3011079 ties would create chip designs using generic design
Date of current version: 5 October 2020 rules as class projects or for theses. Student designs
8 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
Silicon Qubit Printed Circuit
MCM Chip Board
dc Wiring
Harness RF Wiring Harness
Wire Bonds
Ribbon Bonds
Microbumps
Coplanar Waveguide
Metal Carrier Transition to MCM
FIGURE 1. A hybrid classical-quantum module.1 The small square indicated contains two qubits, and the larger square below it
contains single-flux quantum logic. The test fixture carries both dc and microwave signals. RF: radio frequency; MCM: mul-
tichip module.
are typically so small that the overhead of commercial ›› fabricates quantum components like those
contracting is burdensome; therefore, many student shown in Figure 1, using unique quantum fea-
designs could be combined into a single fab run. The tures such as cryogenic packaging, microwave
MOSIS operational model is one in which the layout signals, qubits, and control electronics, extend-
from multiple student projects would be combined by ing the suite of available features over time as
the MOSIS operator into a single set of masks and then discussed later in this article
manufactured by any foundry that could support the ›› at the operational level, the operator would
generic design rules. contract with fabs, design tool suppliers, and
MOSIS is in operation today (https://mosis.com) companies that can perform certain assembly
and the main chip fab option available to students. But activities, essentially aggregating funds from
it is also a practical option for start-ups and, in fact, small users to create a fab run that is compatible
any researcher who wants to create a handful of small with standard industrial processes
chips to test a hardware idea. ›› organizes, but does not create, educational
materials, process design kits (PDKs), and
QUANTUM ACCESS generic intellectual property (IP) that enables
Quantum computer systems are now accessible users to get up to speed quickly.
via the Internet for quantum software training and
small-scale research. Extending access to quantum The quantum access user would be a student,
hardware could follow and then extend the semicon- researcher, or start-up company that wants to try out
ductor MOSIS paradigm. The IEEE Quantum Initiative a hardware idea in quantum information technology.
is using the name quantum access for a potential non- Student interest would start by taking classes, perhaps
profit service that those specifically developed for quantum access. All
www.computer.org/computingedge 9
REBOOTING COMPUTING
Students Hands-on experience to University uses educational Educational modules where indicated; a
augment classwork materials that highlight the use of standard PDK and generic IP from others;
quantum access; submits simple little interest in commercializing the IP
components from the class project students create
for fab
University Foundry for trying out Fab using standard processes with May modify the PDK or create a new
research novel ideas for a thesis new designs; researcher may go on one; interest in IP for publication or
or for faculty research site commercialization through university
projects licensing
Commercial users Prototyping hardware Fab using standard processes with May modify the PDK; the design may be
concepts, R&D new designs; employees may go on commercially sensitive and must be kept
missions, and derivative site from other user; new IP will be owned by
technology the user
Government Exploring high-risk or Fab request directly from a Standard or government-supplied PDKs;
research labs niche technology that government entity government may create a fab; government
may not be destined for IP may be sensitive but may not be
commercialization intended for commercialization
users would then require access to a capital-intensive cryogenic). This would enable the user (or the user’s
fab and help construct a “test harness,” that is, a university or employer) to place an order for commer-
simple system, to test an idea. Table 1 summarizes the cially available equipment and have some assurance
different user classes. that the purchased parts would work together and
To try out the idea quickly and at low cost, quan- also work with the custom parts created by quantum
tum access would also curate information about access’s foundries. Unlike the situation I experienced
the design of standard components, called IP, which in 1981, it may be possible for a cryogenic test sys-
tem to be constructed on a centralized location and
accessed via the Internet.
UNLIKE THE SITUATION I Quantum access should anticipate that a few
EXPERIENCED IN 1981, IT MAY BE projects will become unusually successful and provide
POSSIBLE FOR A CRYOGENIC TEST a path to volume manufacturing. Thus, start-ups, the
SYSTEM TO BE CONSTRUCTED ON government, and government labs would have a role
A CENTRALIZED LOCATION AND that eventually outgrows the quantum access model;
ACCESSED VIA THE INTERNET. these are listed near the bottom of Table 1.
QUANTUM ACCESS
users could include with no modification and hence MATURATION PROCESS
with less effort or risk of error. For example, users with It didn’t occur to me in 1981 to ask which semiconduc-
ideas for a new microwave amplifier can prototype tor process would be used for my CMOS MOSIS chip.
their innovative amplifier circuit but use open source, I was simply glad there was one option instead of no
tested designs for filters, drivers, and so on. options. Quantum access could start out similarly by
Quantum access would also be a source of reliable supporting a single quantum-system type (for exam-
information about physical test apparatus (including ple, a qubit type) and expand the number of types over
time. As an example of where this might go, the semi- are cryo CMOS and superconducting electronics
conductor MOSIS service currently supports 20 semi- based on classical JJs.
conductor processes from two foundries. For example, Intel and Microsoft fund quantum
control electronics using existing CMOS processes
QUBITS of a modest linewidth (for example, 28 nm) and that
Superconducting qubits (transmons) seem to have are in commercial use for the Internet of Things
reached the maturity level necessary for quantum and other timely business opportunities. 3,4 Many
access to be viable. Transmons are physically large and room-temperature semiconductor processes work
do not require many mask layers, making current designs at cryogenic temperatures, although some work bet-
mature enough for general production. Although ter than others. It is expected that, over time, special
research on superconducting qubits continues, the cryo CMOS processes will be developed for quantum
quantum industry is growing and expected to special- information, further differentiating quantum access
ize as it matures, so some students and new corporate from MOSIS. Lincoln Labs, SeeQC, Skywater, and
or university entrants may want to avoid well-trodden others use JJ-based control circuitry,1,5 which has
areas, such as making their own qubits. Instead, they different properties in terms of density, heat dissipa-
might use generic qubits for research projects in other tion, and so forth.
areas under the assumption that their results would Unlike the simple fab processes used by transmon
be combined with the latest qubits prior to production.
Even though specific plans are yet to be deter-
mined, ion traps may be the second qubit type EXPERIMENTAL ION TRAPS USED
offered. Experimental ion traps used for quantum FOR QUANTUM COMPUTER-TYPE
computer-type applications are manufactured by gov- APPLICATIONS ARE MANUFACTURED
ernment labs. Some of these labs fabricate for other BY GOVERNMENT LABS.
parties, although foundry access will need to mature
before ion traps will be viable in quantum access.
qubits, classical control systems will have roughly the
CLASSICAL CONTROL SYSTEMS same process complexity (for example, the number of
Transmon systems are almost universally controlled layers) as that of today’s CMOS, implying an expensive
by room-temperature lab equipment interfacing via fab. JJ processes are destined to be equally complex
electronics at various temperature stages. The elec- but are still under development, including through
tronics may be passive, transistor, or Josephson junc- government R&D funding. When it is time to support
tion (JJ) based. ion traps, quantum access would need to embrace a
Just as I tested my CMOS chip in 1981 using an different fab process, electronics with different oper-
oscilloscope, the idea is that quantum access would ating voltages, optical (laser) signaling, and optical
engage with test equipment and cryogenic equip- components.
ment manufacturers from the start. Although quan-
tum technology is not ready for formal standards, DESIGN TOOLS AND IP
researchers need information on signal flow through Integrated circuits have not been designed directly by
cryogenic stages and what to expect from test equip- humans for decades, but rather by electronic design
ment. Test equipment and cryogenics will inevitably automation (EDA) software that synthesizes circuits,
evolve; therefore, these manufacturers are expected translates them into physical layouts, and then simu-
to be interested in feedback from users. lates their performance. Thus, new quantum-relevant
As quantum technologies mature, there is the design principles will need to be embedded into design
belief that room-temperature lab equipment will be tools. It will be essential that industry fosters the devel-
supplemented by electronics close to the qubits for opment of quantum-specific design tools and dissem-
reasons of signal latency, bandwidth, and thermal inates to the quantum workforce the knowledge for
backflow. Broadly speaking, the two leading options how to use them.
www.computer.org/computingedge 11
REBOOTING COMPUTING
The Internet just turned 50: five decades that shaped the world we live in. But what
comes next, the so-called Quantum Internet, will be even more revolutionary, likely in
ways we can’t imagine yet.
O
n 29 October 1969, the first successful mes- to participate in the so-called quantum race. Several
sage was exchanged over the Arpanet, the start-up companies also have been founded to join
predecessor to what we now know as the in this monumental endeavor. A very significant mile-
Internet. In the five decades since, the Internet has stone was achieved at the end of 2019 by a group of
revolutionized communications to the extent that its researchers at Google, which announced quantum
impact on our lives is not only technological but rather supremacy by solving a classically intractable prob-
has affected almost every facet of business and life- lem with its quantum processor7,8 (see “The Quantum
style, throughout the structure of society. Supremacy”).
The Internet itself evolved amazingly during these Immense interest in the future of quantum tech-
decades, from a network comprising a few static nodes nologies is not only displayed by industry but also
in the early days to a leviathan interconnecting half of by governments around the world. To mention some
the world’s population through billions of devices. Yet initiatives, in April 2017, the European Commission
the fundamental underlying assumption—the Inter- launched a 10-year, €1 billion flagship project to accel-
net’s primary purpose of transmitting messages that erate European quantum technologies research.9
can be successfully encoded in a sequence of classical Meanwhile, across the Atlantic, in September 2018, the
bits—has been unchanged since the beginning. U.S. House of Representatives unanimously approved
The advent of the engineering phase of quantum the establishment of a National Quantum Initiative
technologies is challenging the Internet’s fundamen- funded with US$1.25 billion over 10 years.10
tal assumption because quantum devices require—as Within this context of a real quantum revolution,
communication primitives—the ability to transmit the ultimate vision is to build a quantum network infra-
quantum information. Hence, research groups structure—also known as the Quantum Internet—to
throughout the world, and ours as well, are investing interconnect remote quantum devices so that quan-
their efforts to design and engineer the Quantum tum communications among them are enabled.2,3 The
Internet.1–6 But there’s still a long way to go and no reason behind this vision is that the Quantum Internet
guarantee of getting there very soon. is capable of supporting functionalities with no direct
counterpart in the classical Internet—ranging from
THE QUANTUM REVOLUTION secure communication5 to blind computing 11 through
Quantum technology advances have successfully distributed quantum computing 1,2—as recently over-
enticed tech giants, such as IBM, Google, and Intel, viewed by the Internet Engineering Task Force.12
Although it is too early to tell when and how this
quantum network will be deployed, our goal here is
Digital Object Identifier 10.1109/MC.2020.2984871 to describe how the Quantum Internet differs from
Date of current version: 4 June 2020 the current Internet. For this, we introduce the very
14 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
basic idea of the Quantum Internet and its underlying
foundation, and we highlight the necessary steps as
well as the novel challenges we will face on our journey THE QUANTUM
toward the Quantum Internet design and deployment. SUPREMACY
www.computer.org/computingedge 15
REBOOTING COMPUTING
Sender
Original
Quantum State Entanglement”), in 1993 Bennett
et al.13 showed that it is possible to
Classical instantaneously transfer the quan-
Communication tum state encoded in a qubit at a
certain sender to a qubit stored at
a certain receiver without, surpris-
EPR Pair ingly, the physical transfer of the
qubit at the sender.3 This quantum
Receiver communication protocol, already
experimentally verified, is known
as quantum teleportation.
In a nutshell, the teleportation
process, portrayed in Figure 1 for a
Two-Qubit Operations One-Qubit Operations Measurement single qubit, requires 1) the genera-
tion and distribution of a maximally
FIGURE 1. A general schematic of quantum teleportation protocol, where the entangled pair of qubits (referred
standard bra-ket notation |· is adopted for describing quantum states. Notice in the to as an EPR pair) between the
figure that after quantum teleportation, the original qubit and the entanglement are source and destination, and 2)
destroyed. As weird as it seems, quantum teleportation fully obeys the fundamen- a classical transmission to send
tal principles of quantum mechanics. Therefore, the cost of transmitting quantum two classical bits. Consequently, a
information can be exchanged with entanglement and classical communications. classical link for sending classical
Because the entanglement is always destroyed after every single teleportation, information and a quantum link for
it constitutes the primary consumable resource in the Quantum Internet, which entanglement generation and dis-
means it needs to be generated continuously. tribution need to be established in
advance.
Moreover, each teleportation
process destroys the entanglement-pair member at
the source. A successive teleporting requires the
INTRODUCING generation and distribution of a new entangled pair
ENTANGLEMENT between source and destination. This, in turn, implies
E
radically new challenges from a network design per-
ntanglement is one of the most distinguishing spective, completely changing the classical concepts
quantum phenomena with no counterpart in
of network connectivity and throughput. Indeed, the
the classical world, in which the quantum states of
connectivity between two quantum nodes is strictly
two or more particles become inextricably linked
even if they are separated by a great distance. The
determined by the availability of a shared entangled
entanglement of quantum particles demonstrates pair, and it inherently varies in time as a consequence
a relationship between their fundamental proper- of the depletion of the entanglement-pair member at
ties that cannot happen arbitrarily. When a mea- the source.
surement is performed on one of the particles, the The challenges are not limited to the above-
other particle will be instantly influenced. mentioned ones. In fact, long-distance entanglement
distribution still constitutes a key issue due to the
decay of the entanglement distribution rate as a func-
tion of the distance.1,3 And because qubits cannot be
BEYOND DIRECT QUBIT copied due to the no-cloning theorem, classical signal
TRANSMISSION amplification techniques cannot be employed. In this
By using a unique feature of quantum mechan- context, quantum teleportation relies on intermediate
ics, known as entanglement (see “Introducing nodes, known as quantum repeaters, that are capable
FIGURE 2. The entanglement swapping portrait. (a) Each quantum device shares an EPR pair with an intermediate node, the
quantum repeater. The repeater performs Bell-state measurement on the two qubits in its possession, which results in the col-
lapse of their quantum states into classical bits. The repeater sends the classical bits obtained from the measurement opera-
tion to the quantum devices. Finally, based on the received bits, the quantum devices perform local operations to complete the
swapping process. (b) The result is that the entanglement between the quantum devices is created over a longer distance.
www.computer.org/computingedge 17
REBOOTING COMPUTING
www.computer.org/computingedge 19
EDITOR: Virgilio Almeida, virgilio@dcc.ufmg.br
This article originally
appeared in
DEPARTMENT: INTERNET GOVERNANCE
vol. 23, no. 6, 2019
Digital technologies can be valuable tools to create better public services that benefit society
as a whole. Governance is the way to align digital public services with the interests of society.
In this article, we discuss principles and elements of governance of digital public services.
D
igital technologies, especially algorithms, and operated in a decentralized way and distributed
Internet, artificial intelligence, voice inter- over different parts of the public administration. The
face technologies and IoT, are transforming governance ecosystem of digital public services has
the world, modifying how we communicate, live, and many players with very different legal statuses, which
work. In many countries, better and more efficient depend on the formal organization of government
public services are being demanded by society, due to as well as on the informal power structure, based on
economic pressures, ageing populations, new forms of political interests. Common principles could be the
social services, and digital transformation initiatives.1 element that will put together different interests in
Digital technologies can be valuable tools to create government to create digital public services that are
better public services that benefit society as a whole inclusive, effective, and legitimate.
and in particular the most vulnerable groups. Mecha- A set of high-level principles could be useful to
nisms and procedures are needed to assure that digiti- shape the governance structure of digital public
zation of public services meets society's needs and is services. Principles should be formulated by a multi-
not adopted because of the technological fads pushed stakeholder group, involving different sectors of gov-
by the technology industry. ernment, civil society, and businesses. For example, an
Thus, a natural question that arises is: How can we overarching principle to guide the digitization of public
use digital technologies to improve public services? services could be one that emphasizes the inclusive
Governance is the answer. Governance is the way to nature of public services in a simple statement, such
align digital public services with the interests of soci- as: Assure that digital public services leave no one
ety. A governance framework for digital public services behind.10 Good governance must also adhere to gen-
should be able to oversight the process of procure- eral principles such as transparency, accountability,
ment, development, delivery, and evaluation of digital openness, fairness, and responsiveness, to ensure
public services. In this article, we discuss principles that digital public services meet the legitimate expec-
and elements of governance of digital public services. tations and needs of citizens.
Digital public services are usually developed There are elements that we consider important
for the governance of digital public services. They can
be applied to the governance framework, depending
Digital Object Identifier 10.1109/MIC.2019.2936928 on the stage of digitization of the public services. The
Date of current version 17 January 2020. elements that follow provide practical support for the
20 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
INTERNET GOVERNANCE
www.computer.org/computingedge 21
INTERNET GOVERNANCE
political, or social—or where systems are subject to the subject have emphasized the importance of
cybersecurity issues. handling strategically the processes through which
Protecting citizens' privacy is key to building trust governments are going through. Public servants
in the institutional functioning of digital public ser- should develop technology skills and develop more
vices.8 To this end, regulatory authorities need to be proactive and analytical skills that can support
created, or strengthened if they exist, to constitute a a profile of greater service experimentation and
set of standards that are enough to delimit the appli- innovation.14
cability of protection, the scope of regulation and the Digital public services impact governments and
penalties, which should apply to different companies. their management tools. In this case, the process
Personal data processing must be regulated from end of digital transformation of public services depends
to end in order to avoid problems of misuse and inse- on building skills that enable the transformation and
curity. At the same time, regulation of data protection proper use of digital tools for governments. These skills
must be dynamic in order to meet the challenges of are related to dynamic learning ability in organizations
organizations going through digital transformation. so that they can learn from mistakes and successful
Compliance with data protection regulation should experiences.
result in governments’ provision of a set of standards
for citizens’ action, as well as channels for them to ELEMENT #5: HAVE
communicate with the regulatory authority for pri- PROCUREMENT GUIDELINES
vacy protection.9 Initiatives such as the General Data The quality of public expenditure is directly related to
Protection Regulation (GDPR) of the European Union procurement processes—a key public management
are recommended in order to exercise extra-territorial policy. Procurement governance has been deeply
regulation, and to define authority and principles impacted by the acceleration of the path of techno-
within the scope of personal data protection. logical transformations. The current context has no
precedent in terms of speed, volume, and value. Pro-
ELEMENT #4: DEVELOP DIGITAL curement policy is central to providing infrastructure
AND DATA-RELATED SKILLS for government digital technologies.
Lacking digital expertise in its workforce can lead to On one side of the table—the government—there
serious consequences for governments’ provision are political appointees, high-level officials, middle
of digital public services. Human resources’ recruit- managers, public lawyers, and contractors. On the
ment and retention are on-going challenges for gov- other side of the table, you find marketing people from
ernments, which are in short supply of qualified per- the big tech companies, such as Amazon, Microsoft,
sonnel, due to uncompetitive rates of pay and unclear Google, Facebook, and Apple—all globally connected
career progression rules. and on the frontier of digital innovation. This formal
The use of digital technologies in governments procurement policy relationship creates an informa-
transforms the government workforce. Algorithms tion asymmetry that impacts the technology infra-
gradually replace street-level bureaucrats with structure available to governments.
street-level algorithms.12 Algorithms gradually replace The digital capacity deficit in the public sector is
people and make discretionary decisions about rights not usually trivial nor easily solvable. Third parties can
and benefits. The result of the use of digital tech- play an important role in reducing informational asym-
nologies is the reduction of the size of the workforce, metries. Universities, think-tanks, and consultants
suppression of clerk activities, adoption of big data should be called upon to support the decision-making
practices, redefinition of the profile of civil servants, process despite the unfairness of the default nego-
introduction of artificial intelligence and others. Public tiation. It is not only a matter of resources—human,
service work shifts to the back office, requiring new financial, technological—but of governance arrange-
skills from public servants.13 ments. Together they can establish guidelines to
Human resources’ policies in the public sector guide government procurement of infrastructure,
have faced radical changes. OECD's reports 9 on making the most of available technology.
ELEMENT #6: DEVELOP area have already proposed and developed governance
ALGORITHMIC DECISION- frameworks, policies, and principles for guiding the
MAKING POLICIES use of algorithmic decision-making.17–19 One example
A working definition of decision-making algorithms proposed by Copeland15 is: ”Every algorithm used by a
is given by European Commission20: ”A software sys- public sector organization should be accompanied with
tem—including its testing, training, and input data a description of its function, objectives, and intended
as well as associated governance processes—that, impact, made available to those who use it.” Before
autonomously or with human involvement, takes deci- decision-making algorithms and artificial intelligence
sions or applies measures relating to social or physical technologies are deployed in large scale in government,
systems on the basis of personal or nonpersonal data, a key question should be answered by public administra-
with impacts either at the individual or collective level.“ tors17: Who should be responsible for governing, audit-
Decision-making algorithms in public services have ing, and assuring algorithmic decision-making systems?
been used in many areas such as education, healthcare,
transportation, security, and judicial system. Examples CONCLUSION
of government applications include risk-assessment The rapid evolution of digital technologies requires
algorithms that estimate the likelihood of recidivism for flexible and adaptive governance framework for dig-
criminals and customer service chatbots used by local ital public services. Flexible governance approaches
governments to answer questions from their constitu- should support multistakeholder dialogues and part-
ents. The use of algorithms has the potential to improve nerships with other government organizations, indus-
the quality of decision-making.17,19 They have possibil- try, and civil society in order to create sustainable,
ity to increase speed, reduce cost, improve accuracy, inclusive, and principle-based solutions for digital pub-
support new types of applications, and extend the lic services.11 The process of digital transformation of
reach of public services, depending on the availability public services is underway in several countries. Gov-
of communication infrastructure. Although algorith- ernance frameworks, with their principles and ele-
mic decision-making can improve efficiency of public ments, is key in the digital transformation of public
services, it also faces some challenges, such as errors, services and should be created. In this article, we pre-
biases, discrimination, and other undesirable behavior sented a number of principles and elements that point
patterns. Its impact needs to be monitored. As algorith- in the direction of adaptive governance models.
mic decision-making is used in an ever-growing number
of public service applications, it is necessary to protect REFERENCES
citizens against negative side effects of algorithmic 1. P. Dunleavy, H. Margetts, S. Bastow, and J. Tinkler.
decisions. As pointed out by Brauneis and Goodman,16 Digital Era Governance: IT Corporations, the State, and
“In the public sector, the opacity of algorithmic deci- E-Government. Oxford, U.K.: Oxford Univ. Press, 2007.
sion making is particularly problematic, both because 2. D. Linders, “From e-government to we-government:
governmental decisions may be especially weighty Defining a typology for citizen coproduction in the age
and because democratically elected governments of social media,” Government Inf. Quart., vol. 23, no. 6,
have special duties of accountability.” It is clear that pp. 446–454, 2012.
governance bodies should provide government-wide 3. M. Janssen and G. Kuk, “The challenges and limits
oversight of services that rely on algorithmic deci- of big data algorithms in technocratic governance,”
sions. Governance mechanisms and measures need Government Inf. Quart., vol. 33, no. 3, pp. 371–377, 2016.
to be specified to achieve transparency accountabil- 4. T. Bovaird, “Beyond engagement and participation-user
ity and fairness in digital public services. and community co-production of public services,”
In order to minimize risks and negative conse- Public Admin. Rev., vol. 67, pp. 846–860, 2007.
quences of automated decision-making in public 5. A. Meijer, “Co-production in an information age: Individual
services, governments have to create governance and community engagement supported by new media,”
frameworks to deal with the development and use of VOLUNTAS, vol. 23, no. 4, pp. 1156–1172, 2012.
this kind of technology. Frontrunner countries in this 6. H. Margetts and P. Dunleavy, “The second wave of
www.computer.org/computingedge 23
INTERNET GOVERNANCE
digital-era governance: A quasi-paradigm for govern- 18. C. Castelluccia and D. Le Métayer, “Understanding
ment on the Web,” Philos. Trans. Roy. Soc. A, Math., algorithmic decision-making: Opportunities and
Phys. Eng. Sci., vol. 371, no. 1987, pp. 1–17, 2013. challenges,” Study, Panel Future Sci. Technol., Eur.
7. International Telecommunication Union (ITU), “Uni- Parliamentary Res. Service, Mar. 2019.
versal service fund and digital inclusion for all,” 2013. 19. S. Chignard and S. Penicaud, “With great power comes
[Online]. Available: https://www.itu.int/en/ITU-D great responsibility: Keeping public sector algorithms
/Digital-Inclusion/Documents/USF_final-en.pdf accountable,” RightsCon, 2019, Tunis, Tunisia. [Online].
8. The Royal Society and the British Academy, “Data Available: rightscon2019.sched.com/event/Pvg5/with
management and use: Governance in the 21st Century,” -great-power-comes-great-responsibility-keeping
The Royal Society and the British Academy, London, U.K., -public-sector-algorithms-accountable
2017. [Online]. Available: https://royalsociety.org/~/media 20. European Commission (procured by) State-of-the-Art
/policy/projects/ data-governance/data-management Report | Algorithmic decision-making, delivered by
-governance.pdf Optimity Advisors, Dec. 2018. [Online]. Available:
9. OECD, “Strengthening digital government,” OECD https://www.algoaware.eu/state-of-the-art-report/
Going Digital Policy Note, OECD, Paris, Italy. [Online].
Available: www.oecd.org/going-digital/strengthening
-digital-government.pdf VIRGILIO A. F. ALMEIDA is a Faculty Associate at the Berk-
10. A. Young-Powell, “The challenge to ensure digital pub- man Klein Center for Internet and Society, Harvard Univer-
lic services leave no one behind,” The Guardian, Mar. 5, sity and an Emeritus Professor in the Computer Science
2019. [Online]. Available: https://www.theguardian Department, Federal University of Minas Gerais, Brazil. His
.com/society/2019/mar/05/challenge-ensure-digital research interests include cyber policies and governance,
-public-services-leave-no-one-behind large-scale distributed systems, the Internet, and social
11. I. Linkov, B. Trump, K. Poinsatte-Jones, and M. Florin, computing. He received the Ph.D. degree in computer sci-
“Governance strategies for a sustainable digital world,” ence from Vanderbilt University. Contact him at virgilio
J. Sustain., vol. 10, no. 2, p. 440, 2018. @dcc.ufmg.br or valmeida@cyber.harvard.edu.
12. A. Alkhatib and M. Bernstein, “Street-level algorithms:
A theory at the gaps between policy and decisions,” FERNANDO FILGUEIRAS is the Director of Research and
in Proc. CHI Conf. Human Factors Comput. Syst. Proc., Graduate Studies, National School of Public Administration,
2019, Paper no. 530. Brazil. He is an Associate Professor at the Department of
13. OECD, Building on Basics. Paris, Italy: OECD Publish- Political Science, Federal University of Minas Gerais. He
ing, 2015. is a researcher of the Digital Democracy Institute and a
14. OECD, “HR and Leadership strategies for building innova- researcher at the Institute of Advanced Studies, University of
tive public sector organisations,” Paris, Italy. [Online]. São Paulo. His research interests include digital governance,
Available: https://oecd-opsi.org/wp-content digital public policy and accountability. He received the
/uploads/2019/03/HR-and-Leadership-Catalyst-for Ph.D. degree in political science from Rio de Janeiro Univer-
-Innovation-Capabilities.pdf sity Research Institute. Contact him at fernando.filgueiras
15. E. Copeland, “10 principles for public sector use of @enap.gov.br or fernandofilgueiras@hotmail.com.
algorithmic decision making,” Nesta Blog, Feb. 20, 2018.
[Online]. Available: www.nesta.org.uk/blog/10-principles FRANCISCO GAETANI is Professor of Public Management of
-for-public-sector-use-of-algorithmic-decision-making/ the Getulio Vargas Foundation. He was the Deputy Minister
16. R. Brauneis and E. Goodman, “Algorithmic Transpar- for Planning, Budget and Management, Brazilian Federal
ency for the Smart City” (August 2, 2017). 20 Yale J. of Government. He was the Deputy Minister for the Environ-
Law & Tech. 103 (2018); GWU Law School Public Law ment, Brazilian Federal Government. His research interests
Research Paper; GWU Legal Studies Research Paper. include governance, public management. He received the
17. The Centre for Data Ethics and Innovation (CDEI), Ph.D. degree in political science from London School of
“Review into bias in algorithmic decision-making,” Jul. Economics and Political Science. Contact him at francisco
2019, U.K. [Online]. Available: www.gov.uk/cdei .gaetani@fgv.br or fgaetani@gmail.com.
DEPARTMENT:
vol. 17, no. 5, 2019
SOCIOTECHNICAL SECURITY AND PRIVACY
H
uman failure is the weakest link in many, In one common version of the launchpad attack,
if not most, security systems. As a result, the criminal compromises an email account of a per-
criminals are increasingly relying on social son whose job involves financial transactions—for
engineering as a way to circumvent security controls. example, a person who invoices clients of a contract-
To improve their yield, criminals constantly experiment ing company. Based on the emails this launchpad user
with methods aimed at making their attacks harder to has sent and received, the criminal determines where
detect—both to security systems and to the end users invoices are sent and sends additional invoices—or
behind them. Naturally, an attack that successfully just a request for a change of bank accounts—to
evades detection by both human and machine has the these unfortunate targets. These emails are typically
potential to make criminals very wealthy. Therefore, sent from the compromised account of the launchpad
once discovered and successfully tested, such attacks user, making them instantly trustworthy to the target
exhibit dramatic growth and are commonly copied and due to the already established trust relationship.
tweaked by other criminals spotting an opportunity However, the criminal knows that a slightly suspicious
when they see it. What we term the launchpad attack target user may respond to the email from the launch-
is the newest example of such an attack. This attack pad user to ask for a confirmation before updating the
is also commonly referred to as a business email com- bank account information to be used to pay invoices.
promise (BEC) that uses account takeover. Although These are messages that the criminal wants to receive
it was virtually unheard of just a few years ago, 44% but that he or she does not want the launchpad user to
of organizations have now experienced this type of get to see. Therefore, criminals often set up selective
attack according to a recent industry report.1 forwarding and deletion rules (e.g., based on an email
In a launchpad attack, a first user—the launchpad thread or a subject line).
user—is compromised by the criminal. This typically
means that the criminal gains access to this user’s WHY ARE LAUNCHPAD ATTACKS
email account, enabling the criminal to review all SO SUCCESSFUL?
emails sent and received and identify valuable con- There are two principal reasons why launchpad
tacts based on their affiliations and the conversations attacks are as successful as they are. First of all, the
between the launchpad user and these contacts. deceptive emails sent to the target users are rarely
Based on this, the criminal selects one or more target blocked by automated email filters. For one thing,
victims—the valuable contacts—and sends them these emails are not sent from users that have been
messages that, based on the scanned emails, will most reported as being abusive, as a spammer or a phisher
likely be both credible (to the victim) and profitable (to might have been. See “A Brief History of Online Social
the criminal). Engineering: The Early Years” for a short history of
Internet attacks. The emails are also not sent from
unknown users with display names that are decep-
Digital Object Identifier 10.1109/MSEC.2019.2922865 tively similar to users the target user has a trust rela-
Date of publication: 3 September 2019 tionship with (an otherwise common deception
2469-7087/21 © 2021 IEEE Published by the IEEE Computer Society September 2021 25
SOCIOTECHNICAL SECURITY AND PRIVACY
strategy). Therefore, common solutions to detect tra- highly similar to legitimate messages, and they do not
ditional BEC attacks also do not apply. Moreover, while contain keywords indicative of spam (such as Viagra),
the emails are deceptive, they are not spoofed. There- nor do they contain high-volume URLs associated with
fore, Domain-Based Message Authentication, Report- malicious behavior (such as a phishing URL). In other
ing, and Conformance (DMARC)2 does not detect words, today’s email filters simply do not block these
them. For more about DMARC, see “A Brief History of deceptive messages. Otherwise, based on their cur-
Online Social Engineering: The Recent Years.” rent detection strategies, they would also have blocked
Launchpad attacks are also so successful because countless benevolent messages. Finally, it is known
the deceptive emails are sent either from users with that some security technologies adapt the rules based
whom the target users have a trust relationship on the actions of the recipient, thereby becoming less
(namely, the launchpad user) or from strangers with- likely to block emails of the types a recipient responds
out an abusive history known to the system. to.3 This unfortunately weakens the protections of the
Furthermore, considering the content of the decep- most vulnerable users. The problem is not that tradi-
tive emails, this also does not cause the messages tional security controls are flawed. Rather, they simply
to be blocked. Typically, the deceptive messages are do not address launchpad attacks.
Turning now to the human recipients of the decep- has inherited vast fortunes). Instead, the email mes-
tive emails, we note that the contents are not unex- sages are mostly business as usual and sometimes
pected (e.g., nobody is claiming that an unknown rela- even expected. This logical fit is made possible by
tive of the target user has died and that the target user the criminal’s use of detailed contextual information
www.computer.org/computingedge 27
SOCIOTECHNICAL SECURITY AND PRIVACY
for the targeting of the intended victims, in terms of not have artifacts. The system can to a very large extent
crafting the deceptive messages and what accounts maintain the visual appearance of the modified emails
these are sent from. Indeed, it has been shown4 that (e.g., by replacing an attached document with a hyper-
the success rate of deceptive emails can be increased linked image that appears like the thumbprint of the
from single-digit percentages to more than 70% by attached document). When a user interacts with a ref-
using contextual information for targeting. erence to a removed artifact (e.g., by double-clicking
on the thumbprint representing the artifact), a request
THE BUILDING BLOCKS TO DETECT is made for the cloud-hosted artifact. However, before
LAUNCHPAD ATTACKS this is served, the system characterizes the requester.
To understand the attack and how to counter it, it is
not sufficient to understand how existing countermea- Step 2: Characterize Requesters
sures are circumvented, whether these are computa- Every time a user clicks on an artifact reference to
tional or psychological. It is critical to also understand load the corresponding artifact, the system charac-
the behavior of a successful attacker. Launchpad terizes the requester in terms of three dimensions:
attacks always start with information collection. At the device, environment, and automation. The device
identifier corresponds to a stored state (e.g., using
HTML cookies or flash cookies) and user agent infor-
EVERY TIME A USER CLICKS ON AN mation (e.g., information relating to the operating
ARTIFACT REFERENCE TO LOAD system, screen size, and application name and ver-
THE CORRESPONDING ARTIFACT, sion). The stored state of a given device may change
THE SYSTEM CHARACTERIZES THE over time, but it typically does not undergo dramatic
REQUESTER IN TERMS OF THREE changes. The environmental identifier corresponds to
DIMENSIONS: DEVICE, ENVIRONMENT, information about the requester’s geographical and
AND AUTOMATION. service context, such as his or her geolocation, server
name, and carrier or Internet provider. Like the device
identifier, the environmental identifier may change—
heart of the problem is that traditional security controls but typically not in a dramatic manner and rarely at
do not identify from where (i.e., what locations or what the same time as the device identifier undergoes dra-
computers) actions are initiated and, therefore, do not matic changes. A third identifier indicates the extent
detect when an attacker rummages through the mail- to which automation was used for accessing an arti-
box of a launchpad user to identify suitable target vic- fact or sending an email. This can be determined from
tims or when the attacker remotely sends emails from script and application programming interface indi-
the launchpad user’s account. We describe an approach cations in the headers or from the timing of multiple
that addresses this problem, based on tracking the requests. Most email users never use automation.
access to artifacts. By artifacts, we mean attachments Some (such as e-commerce email servers) always use
(such as invoices and purchase orders) and URLs. For it. Very few email users switch back and forth.
concreteness, we focus on attachments here. A given user is associated with a profile, which cor-
responds to one or more sets of identifiers described
Step 1: Replace Artifacts With Links previously. As a new user is observed by the system, it
The security system scans incoming and outgo- has no profile, but as the user request artifacts, the sys-
ing emails of protected users, detects artifacts in tem gradually builds a profile. An attacker is detected
these emails, and replaces them with references to and classified based on the types of mismatch that
cloud-hosted copies of the artifacts. In this way, the result from his or her artifact access requests.
messages in the mailbox of a protected user will not
have artifacts but instead will have links to cloud-hosted PUTTING EVERYTHING TOGETHER
copies of these. For the same reason, an email from a The functionality of the system is perhaps best
protected user in the mailbox of its recipient will also described using an example. The transmission of an
Cloud Storage
email from sender S to recipient R
?
is described in Figure 1. At step 1, R =
a filter replaces artifact A with ref-
X = A
erence X and reports to the cloud
storage that A and X are associ-
(Step 2) A
ated with R. When a user asso- X
ciated with recipient R initiates A
a request for the artifact corre-
X
sponding to X, the cloud storage Sender S Recipient R
A X
characterizes the requestor (step R
2). If the characterization iden- Email Modified
(Step 1) Email
tifies a likely attack, the system
takes a security action. Otherwise,
it transmits artifact A to R. FIGURE 1. The transmission of an email with artifact A.
We will now consider an
attacker that corrupts a user to
use his or her account as a launch- Device ID Environmental ID Automation ID
pad in the attack against another Legit User Good Match Good Match Good Match
user. Let us start by assuming
Legit User, New Device Poor Match Good Match Good Match
that the launchpad victim is a
protected user. To collect intel- Phisher Poor Match Poor Match
ligence from the launchpad user, Remote Access Trojan Poor Match
the attacker accesses one or more Automated Forwarding Poor Match Poor Match Poor Match
attachments. Depending on the
manner in which the attacker has FIGURE 2. The extent to which the three types of identifiers—device, environmen-
gained access to the launchpad tal, and automation—would typically be matched in a variety of scenarios.
user’s account, the profile match-
ing generates different results, as
shown in Figure 2. Green corresponds to a likely good the attacker’s computer. Therefore, the device match
match, while red corresponds to a likely poor match. will be poor. Moreover, the attacker is also likely to
(For the white fields, the match depends on the level of be associated with a different environment, making
sophistication of the attacker.) that match poor as well. If the attacker uses a script
Consider the various cases shown in Figure 2. The to request and render attachments, this will show
legitimate user is expected to have a very high degree up as an anomaly related to automation. The same
of match. A legitimate user on a new device would kinds of mismatches will also be observed—without
naturally have new device identifiers and would trig- any interaction between the attacker and the cloud
ger a verification. A phisher gaining access to a user’s server—when the attacker uses the launchpad user’s
account after having stolen the user’s password would email account to send email to intended targets. In
typically be very easily distinguished from a legitimate other words, indicators similar to those that can be
user. A remote access Trojan could be more difficult observed when a user makes artifact requests will also
to detect, as it involves a compromised device of the be automatically encoded in the headers of the emails
legitimate user. An automated forwarding rule, on the that this user sends.
other hand, is very easy to detect.
If we assume that the attacker has stolen the Step 3: Reacting to an Attack
launchpad user’s password (e.g., the attacker is a For each artifact request, the system computes a risk
phisher), then the attacker will not access the email score that depends on the three types of identifiers
account from the launchpad user’s computer but from and on the profile of the legitimate user. The score also
www.computer.org/computingedge 29
SOCIOTECHNICAL SECURITY AND PRIVACY
depends on the number of artifacts requested from attacks and help consumers and enterprises protect
this party, the pattern of requests (such as the inter- themselves.
arrival times of the requests and whether they appear
to correspond to a particular search term), and the his- REFERENCES
torical request patterns associated with the profile. 1. Osterman Research, “Best practices for protecting
If the risk score is low, access is permitted, and the against phishing, ransomware and email fraud,” Black
requested artifact is transmitted to the requester. If Diamond, WA, White paper, 2018.
the risk score is intermediate, the system may request 2. M. Moorehead, “How to explain DMARC in plain English,”
additional authentication, such as two-factor authen- Return Path, July 20, 2015. [Online]. Available: https://blog
tication, before transmitting the requested artifact to .returnpath.com/how-to-explain-dmarc-in-plain-english/
the requester. Finally, if the risk score is high, the sys- 3. M. Jakobsson and T.-F. Yen, “How vulnerable are we to
tem may respond with a synthetic artifact, that is, a scams?” Black Hat, 2015. [Online]. Available: https:
modification of the original artifact or an automati- //www.blackhat.com/docs/us-15/materials/us-15
cally generated replacement. Moreover, the system -Jakobsson-How-Vulnerable-Are-We-To-Scams-wp.pdf
may notify affected users. Any email sent by an identi- 4. T. N. Jagatic, N. A. Johnson, M. Jakobsson, and F.
fied attacker from the launchpad account to contacts Menczer, “Social phishing,” Commun. ACM, vol. 50, no.
of the launchpad user may be blocked or augmented 10, pp. 94–100, 2007. doi: 10.1145/1290958.1290968.
with a warning.
GET PUBLISHED
www.computer.org/cfp
T
hese days, when you see the headline “How Safe In a review like this, it is always tempting to try to
Are Airplanes?” the text more likely refers to the come up with a single remedy that should fix all prob-
spread of disease than the likelihood of a crash. lems. The NASA study team that is investigating the
With steady safety improvements over at least the last issues may be leaning in that direction as well, with
five decades, airplane crashes have become exceed- a recommendation that all software from NASA con-
ingly uncommon, and they are rarely caused by flaws tractors would need to comply with a new restriction
in the design of the airplane itself. The most common on the cyclomatic complexity of functions.4 The cyclo-
remaining causes are the usual suspects of human matic complexity metric, which was popular for a while
error, extreme weather events, and mechanical failure. in the 1980s, measures the number of independent
That perception of safety for air travel may have paths through the control flow graph of a function.
changed with the recent crashes of two new Boeing Would it not be great if one could indeed demonstrate
airplanes. The 737 MAX was introduced as a relatively the existence of a correlation between cyclomatic
small upgrade in a long line of successful airplane complexity and the postrelease fault density of func-
designs that, in this case, was primarily meant to tions? Quite a few researchers have tried to look for
improve fuel efficiency.1 The cause—or, rather, the such an effect, but, alas, most have found that there is
causes—of the two crashes has been sufficiently no such correlation. So, is there no hope?
fleshed out by now, for instance, in the detailed con-
gressional report that was released in September BUILDING RELIABLE SYSTEMS
2020, “The Design, Development, and Certification In most disciplines, system reliability is achieved
of the Boeing 737 MAX.”2 As the title of this report through the judicious use of redundancy. Virtually all
indicates, the failures cannot be attributed to a single spacecraft, for instance, have multiple ways of com-
mistake, but were found to have been caused by a municating with Earth, using independent transmit-
sequence of shortcuts and errors combined with a ters and receivers. Some spacecraft, such as the Cas-
lack of meaningful oversight by the U.S. Federal Avia- sini spacecraft that orbited Saturn, are designed with
tion Administration. two separate main engines and two sets of thrusters
to avoid single points of failure that could end a mis-
SOFTWARE PROCESS sion prematurely.
The lack of rigor in the vetting of new or revised designs To make data transmissions over noisy transmis-
has led to a series of problems, including in the develop- sion channels more reliable, we can add redundant
ment of the CST-100 Starliner system, Boeing’s design information in the form of error-correcting codes so
for a reusable crewed spacecraft. A serious test failure that small bursts of errors are recoverable. The big
of the Starliner craft in December 20193 was found in a question is, of course, how we could do something
review by a NASA panel to have been caused by multi- similar in software design. The answer is not to
ple software bugs. The panel concluded, among many merely duplicate a code base and then run it on two
other findings, that NASA should “go beyond merely computers. This may protect against hardware faults,
correcting the cause of the anomalies” and “scrutinize but it cannot do the same for the software. If there is
Boeing’s entire software testing processes.” a bug in the code, the same bug will obviously affect
both executions, which then provides no protection
Digital Object Identifier 10.1109/MS.2020.3028759 at all. This is also the scenario that led to the failure
Date of current version: 22 December 2020 of the maiden flight of the Ariane 5 launch system in
2469-7087/21 © 2021 IEEE Published by the IEEE Computer Society September 2021 31
RELIABLE CODE
June 1996. 5 The same software was running on both provides a form of software redundancy that can
the main CPU and backup CPU, which meant that indeed make a system more reliable.
a single floating-point exception error could crash
both, causing the costly loss of the vehicle. WHAT ABOUT TESTING?
Another method for using redundancy in the design I have not mentioned testing in the discussion of reli-
of software systems, called N-version programming, ability so far. Some say that, since the cyclomatic com-
was popular for a while, but it also failed to deliver the plexity metric measures the number of paths in the
promised increase in reliability. The thought was to control-flow graph of a function, this gives the num-
have multiple independent teams write software in ber of tests that are required to test all of those paths.
parallel and compare the results of the executions Fewer required tests should, then, make it possible to
of all systems to detect inconsistencies caused by test code more rigorously. This ignores, though, the
bugs. As Knight and Leveson 6 pointed out in the fact that most of the complexity in a program execution
late 1980s, all teams working on the design still does not originate in the structure of the control-flow
begin from a common set of requirements and are graph but in the data that are processed. A single path
likely to make similar types of mistakes. Splitting a in the control-flow graph can be executed in an astro-
team into N subteams can also lead to an increase nomical number of different ways, which means that a
in cost and loss of productivity as well as create single test of each path in the control-flow graph has
tension among them. That is, how does one avoid little chance of revealing all of the bugs that may hide
cross contamination of key findings during develop- on that path.
ment, for example, if one team discovers significant A function with a lower cyclomatic complex-
omissions in the design requirements all teams ity does not necessarily require fewer tests to vet
work from? thoroughly. All code is written to satisfy some set of
requirements and provide a desired functionality. We
SELF-CHECKING CODE can spread that functionality across multiple modules
Although all of this sounds discouraging, there is, in and functions to create a large assembly of intercon-
fact, an effective technique for using redundancy to nected and mutually dependent small functions, but
improve system reliability, although we often do not that will rarely be the right way to structure the code,
think of it in quite that way. The method is to increase and neither will it be able to simplify testing.
the use of assertions throughout a code base. As many
software developers will tell you, these assertions are REQUIREMENTS TESTING
redundant and can safely be removed after testing. Do To put it most succinctly: tests should not be derived
they help to improve reliability? The sobering news, at from the structure of a control-flow graph but from
least for those developers, is that it can be shown that the requirements that prompted the creation of the
the number of assertions that are retained in the code code. To check if a requirement is met, the tester or
after testing correlates strongly with postrelease fault developer should define a suite of tests that can eval-
density. More assertions means fewer faults. It is as uate the adequacy of the code to handle a range of
simple as that. This was first shown in a study done by cases, including expected cases, boundary cases, and
Microsoft researchers, who studied postrelease faults even cases that outright violate the design assump-
in the Office suite of tools and compared it with the tions. If, after running that suite of tests for all require-
assertion density of the failing code.7 ments, it turns out that parts of the code base were
For the mission code that is developed at NASA/ not reached, this indicates a flaw in either the require-
Jet Propulsion Laboratory, we require that the aver- ments or in the code. The most common situation is
age assertion density for each module is 2% or more. that the requirements are found to be incomplete
This means that 2% of the code performs self-checks and fail to cover things that the code must be able to
at key steps in the computations performed to make handle. It can, however, also be that the code itself is
sure that integrity is maintained, even in anomalous redundant and that parts of it serve no legitimate pur-
execution scenarios. The use of assertions thus pose to satisfy the requirements.
Code coverage metrics serve to check the ade- 4. T. J. McCabe, “A complexity measure,” IEEE Trans.
quacy of a test suite for a specific set of requirements. Softw. Eng., vol. 2, no. 4, pp. 308–320, Dec. 1976. doi:
Adding test cases to an existing test suite that serve 10.1109/TSE.1976.233837.
only to improve the coverage metric, without actually 5. J. L. Lions, “Ariane 5: Flight 501 failure,” 1996. [Online].
testing anything, is common practice, but it cannot Available: http://sunnyday.mit.edu/nasa-class/Ariane5
improve a system’s reliability. -report.html
6. J. C. Knight and N. G. Leveson, “An experimental evalua-
REFERENCES tion of the assumption of independence in multi-version
1. M. R. Jacobson. “What’s wrong with Boeing?” Defense programming,” IEEE Trans. Softw. Eng., vol. SE-12, no. 1,
One. https://www.defenseone.com/ideas/2019/11 pp. 96–109, 1986. doi: 10.1109/TSE.1986.6312924.
/whats-wrong-boeing/161245/ (accessed Oct. 14, 2020). 7. G. Kudrjavets, N. Nagappan, and T. Ball, “Assessing the
2. “Boeing 737 MAX investigation,” U.S House of Repre- relationship between software assertions and faults:
sentatives, Washington, D.C. Accessed: Oct. 14, 2020. an empirical investigation,” in Proc. 17th Int. Symp.
[Online]. Available: https://transportation.house.gov Softw. Rel. Eng. (ISSRE), Nov. 2006, pp. 204–212. doi:
/committee-activity/boeing-737-max-investigation 10.1109/ISSRE.2006.14.
3. J. Roulette. “Boeing’s botched Starliner test flirted with
‘catastrophic’ failure: NASA panel.” Reuters. https: GERARD J. HOLZMANN is a consultant and
//www.reuters.com/article/us-space-exploration researcher at Nimble Research, Monrovia,
-boeing/boeings-botched-starliner-test-flirted-with California, USA. Further information about
-catastrophic-failure-nasa-panel-idUSKBN20106A him can be found at http://spinroot.com
(accessed Oct. 14, 2020). /gerard. Contact him at gholzmann@acm.org.
ADVERTISER INFORMATION
Midwest US:
Advertising Sales Contacts Dave Jones
Email: djones@computer.org
Mid-Atlantic US: Phone: +1 708-442-5633 Fax: +1 888-886-8599
Dawn Scoda Cell: +1 708-624-9901
Email: dscoda@computer.org
Phone: +1 732-772-0160
Cell: +1 732-685-6068 | Fax: +1 732-772-0164 Jobs Board (West Coast and Asia), Classified Line Ads
Northeast, Europe, the Middle East and Africa: Jobs Board (East Coast and Europe), SE Radio Podcast
David Schissler
Email: d.schissler@computer.org Marie Thompson
Phone: +1 508-394-4026 Email: marie.thompson@computer.org
Phone: +1 714-813-5094
www.computer.org/computingedge 33
DEPARTMENT: IT TRENDS This article originally
appeared in
In response to claims that 5G will “change the world,” 5G mobile communications systems
have drawn considerable attention, from consumers and practitioners, to policymakers and
academics. Amid mass-marketing campaigns and public posturing by early 5G pioneers
in the telecom industry, cutting through the hype to understand the actual implications
of this new technology is difficult. This article examines the differences between 5G
and earlier generations of wireless technology while debunking some of the myths and
misconceptions surrounding this new technology. It further discusses 5G’s potential
business applications, as well as the opportunities and challenges IT professionals and
general managers may face in incorporating 5G within their business operations.
A
fter years of hype surrounding the A key feature of 5G is that it utilizes millimeter-wave
lightning-fast speeds promised by this new technology to tap into the very high end of the wire-
generation of mobile network technology, 5G less spectrum, where large amounts of bandwidth
was finally introduced to the market in 2019. Yet, like remain unused. While millimeter-wave signals enable
many new technologies, such as artificial intelligence faster network speeds, they are less reliable at long
(AI) and blockchain, 5G has been labeled both as hype distances and are more susceptible to blockage.6
and hope. On the one hand, the arrival of 5G has been Whereas a 4G tower can offer a range as far as 30
compared to past paradigm shifting technologies, miles, a 5G tower can, on average, only offer a range of
such as electricity and automobiles. Such bold claims less than half a mile.7 As a result, 5G networks require
have bolstered expectations that 5G will usher in a a greater number of access points than 4G networks,
new era of similarly innovative applications, such as meaning that substantial investments are required
fully immersive entertainment, autonomous vehicles, in entirely new infrastructures and technologies to
and the Massive Internet of Things (MIoT`).1–3 Adding truly harness the revolutionary use cases, such as
to the hype, 5G is predicted to contribute an astound- self-driving cars and robotic telesurgery, envisioned
ing $13.2 trillion to the global economy by 2035.2,4 by 5G proponents.3,8 Beyond these costs, first- and
On the other hand, various industry experts and early-movers must navigate a plethora of technical,
academics have argued that the current hype sur- organizational, regulatory, and social challenges in
rounding 5G is premature, particularly given the costs deploying 5G technology. 2,9,10
associated with its implementation and the limited In light of the differing viewpoints concerning 5G
range of viable use cases available in the short-term. 5 technology, this article aims to provide managers
with a clear understanding of 5G, its benefits and
potential applications, while also debunking some
Digital Object Identifier 10.1109/MITP.2020.2972139 of the myths and misconceptions surrounding this
Date of current version 6 November 2020. emergent technology. To better contextualize the 5G
34 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
IT TRENDS
FIGURE 1. Evolution of wireless communication systems from first (1G) to fifth (5G) generation.
revolution, we begin with a brief timeline of the previ- messages.11 As pioneering technologies, both 1G and
ous generations of mobile technology leading up to 2G networks improved the accessibility of voice com-
the development of 5G. Next, we compare 5G with its munications across much larger population bases.
predecessors across its generation-defining features: Moreover, by enabling messaging services and offer-
speed, bandwidth, and latency. We then review the use ing incrementally faster speeds, 2G prompted the
cases and business opportunities that are enabled by mass-adoption of mobile phones on a scale never
5G's enhanced capabilities, as well as the challenges before seen. From 2G to 4G, each successive mobile
IT professionals and general managers are likely to network generation has focused on increasing speed
face in deciding whether and which 5G technologies and efficiency to improve mobile experiences for
to incorporate within their business operations and end-users.
revenue models. Drawing on digital broadband technologies, 3G and
4G mobile networks offered advanced mobile data ser-
FIVE GENERATIONS OF vices and introduced the mobile internet respectively.
WIRELESS TECHNOLOGY While 3G gave rise to new services such as mobile TV,
Since the arrival of the first-generation network in the video conferencing, and voice over IP (VOIP) services
early 1980s, four subsequent generations of wireless like Skype, 4G enabled improved user experiences to
communication systems have been introduced. With these services through its increased speeds, ability to
a new mobile generation emerging every decade, each handle far greater data loads, and reductions in net-
new generation has brought improved performance work latency. Consequently, 4G allowed for near real
and a wider range of capabilities compared to its pre- time, intensive data exchange between end-users and
decessor. The evolution from 1G to today's nascent 5G ushered in an “anytime, anywhere, in anyway era” for
technology is illustrated in Figure 1. mobile network use.11,12
While the first mobile network generation (1G) was Over the last decade, faster and more responsive
strictly analog and only allowed voice calls, 2G was a data transfer capabilities have led innovators to
digital network and the first to provide data services envision a range of new and revolutionary applica-
for mobile devices, including SMS and multimedia tions, such as the tactile internet, fully autonomous
www.computer.org/computingedge 35
IT TRENDS
vehicle fleets, and next-generation industrialization will open the door to a wide range of new industrial
including configurable factories and mobile robots. applications and opportunities to improve efficien-
Yet, despite massive performance improvements over cies within existing applications. For instance, experts
previous generations, today's 4G networks still lack in the manufacturing industry envision putting video
adequate throughput capacity, as well as sufficiently cameras and sensors throughout factories to allow
low network latency to power these ambitious future for the real-time monitoring of product quality. Simi-
applications.13 Fortunately, a new generation of net- larly, if we imagine a world where wireless speeds can
work technology has arrived. exceed wired speeds, industries such as energy and
mining can “cut the cord” and move to more flexible
5G TECHNOLOGY: WHAT IS IT AND and autonomous operations in order to increase pro-
HOW IS IT DIFFERENT? ductivity and reduce costs.14
Whereas the evolutions from 2G and 3G to 4G focused
on increasing data throughput and data transfer Ultra-Low Latency
speeds, as well as expanding mobile service areas While telecommunications carriers and marketers
to support connectivity between users, today's have focused their 5G messaging on high data speeds,
fifth-generation wireless technology offers more than most technical experts argue that the true benefits of
an incremental performance improvement over its 5G derive from its reduced latency.10,13 As compared
predecessor.5 Rather, 5G is an expansion of mobile with its predecessors, 5G has significantly lower net-
networking into new regulatory and technological work latency, which allows it to deliver more instanta-
paradigms where 5G technologies are not only being neous, real-time data. Where high latency connections
designed to enable interactions between users, but common today can cause web pages to load slowly,
also to support advanced user-to-machine interactions and can affect user experiences in applications that
and, most importantly, machine-to-machine interac- require real-time data exchange, the reduced latency
tions.3,6 To support such ubiquitous connectivity, 5G offered by 5G is expected to enable improved multi-
brings three new features to the table: unprecedented person video conferencing and gaming, as well as AR
levels of speed; significant reductions in latency; and and VR. Moreover, 5G's very low latency will be partic-
greatly expanded system capacity, allowing for the ularly important for a range of cutting-edge technol-
simultaneous connections of many more devices. ogies and applications that require high responsive-
ness, such as fully autonomous vehicles and robotic
Very High Data-transfer Speeds surgery. It is in these applications where the benefits
The multigigabit speeds of 5G are due to more effi- of 5G's reduced latency become apparent, by allow-
cient encoding and the use of high frequency ing real-time remote control at scale and across dis-
millimeter-waves not utilized by 4G networks.7 With tances, as well as across devices and services. With an
peak data rates expected to reach 20 Gbps, 5G tech- actual latency of around 50 ms, 4G networks are inca-
nology promises to bring speeds 600 times faster than pable of providing autonomous vehicles with reflexes
the 4G LTE (i.e., Long Term Evolution, a network stan- comparable to the 2 ms decision-making speeds of
dard that offers slightly faster connection than a basic humans.15 In this context, the ultra-low latency of a 5G
4G connection) speeds seen in today's mobile phones.2 network would allow for mobile technology to be used
To put things into perspective, 5G users will be able more safely in autonomous vehicles.
to download an entire movie of 2–3 gigabytes (GB) of
data in less time than it would have taken a 4G user Massive System Capacity
to download a 30-megabyte (MB) application. While 4G allowed for the introduction of more high-
the current speeds attained by 4G are often sufficient bandwidth applications through improvements in
for the typical mobile user,5 5G's higher speeds bring data transmissions speeds and network latency com-
the potential for improved augmented reality (AR) and pared to 3G. As such, 4G device penetration rates
virtual reality (VR) experiences when combined with increased between 2006 and 2016, as evidenced by
5G's very low latency. Moreover, 5G's gigabit speeds a 4000-fold increase in global data traffic.6 With 4G
close to reaching the technical limits of how much service classes. In doing so, we highlight the key 5G
data it can transfer across blocks of the wireless capabilities supporting these service classes.
spectrum, 3 a new wireless technology is needed if
we are to continue our upward trajectory of improve- Enhanced Mobile Broadband
ment of current mobile applications and the intro- As the first commercial application of 5G, eMBB is
duction of new ones. 5G continues this momentum centered on driving efficiencies at the core network
through another one of its features: an expansion in level to pave the way for greater data consumption
bandwidth which promises to significantly reduce through a lower cost-per-bit for data and enhanced
network congestion. Specifically, by relying on dif- user experiences.3 The increased use of broadband
ferent radio frequencies from the ones used today by applications on mobile networks will be driven by 5G's
4G, 5G is set to expand into a much broader range of improved speeds, latency, and bandwidth, as well as
the wireless spectrum, allowing for greater network the promise of ubiquitous coverage and energy effi-
capacity as well as increased flexibility and scalabil- ciency to address the high-power consumption rates
ity of new applications. Most notably, 5G is expected that have hindered 4G-driven technologies.2,3 Envi-
to provide a 100-fold increase in traffic capacity as sioned eMBB applications include: improved outdoor
compared to 4G, while supporting up to 1 million wireless broadband; ultrahigh definition or UHD (4K
connected devices per 0.38 square miles, as com- and 8K) broadcast and mobile streaming; 3D multi-
pared to around 2000 connected devices with 4G. 2,6 player gaming; and virtual meetings with 360° video
Though enhanced bandwidth and connection density and even real-time language translation. 5G is also
will certainly benefit consumer end-users, the true expected to enable a range of new VR/AR applica-
benefit of 5G's massive capacity and connectivity tions, including: virtual dressing rooms; augmented
improvements lie in its ability to create a truly inter- reality for physical therapy; and fully immersive enter-
connected world. tainment experiences in 100 000 seat venues using
5G-backed hotspots that are characterized by very
HOW WILL 5G BE USED? high user-density and traffic capacity but that have
A common misconception, particularly among the low mobility requirements. eMBB use cases are those
general public, is that 5G will only really impact the most likely to have near-term economic and social
telecommunications industry. As we have begun to impacts as eMBB applications are mostly likely to be
demonstrate, this is only part of the story. Rather, 5G incremental improvements of existing 4G use cases.4
is a General Purpose Technology (GPT) that, through
continuous improvements over time, will come to be MASSIVE INTERNET OF THINGS
widely adopted and offer a vast range of uses and Although the introduction of 4G made IoT possi-
technological complementarities. Through their pro- ble, device-to-device applications have generally
liferation across multiple industries, GPTs are innova- remained limited due to the inability of these networks
tion catalysts that lead to transformational, and often to reliably handle simultaneous connections. With its
disruptive, changes that reach beyond industries, to high connection density, 5G technology is set to spur
impact entire economies and societies.4 the evolution from today's Machine Type Commu-
In building on earlier mobile technologies, 5G is nication (MTC) paradigm towards Massive Machine
expected to bring about new classes of advanced Type Communication (mMTC), a paradigm more com-
applications, promote business innovation, and drive monly known as the Massive Internet of Things (MIot).
economic growth.4 Though a universal set of speci- Enabled by inexpensive, long-life modules with con-
fications defining 5G is still under development, 5G stant connectivity, MIoT applications are being devel-
technology is being purpose-built to support three oped across a wide array of industries with use cases
classes of services: enhanced mobile broadband including: in-store beacon technology to support
(eMBB); the Massive Internet of Things (MIoT); and proximity marketing and connected shopping expe-
mission-critical services (MCS).2 We now briefly dis- riences (e.g., iBeacon); precision agriculture; smart
cuss select potential use cases within each of these metering in the utilities sector; and remote industrial
www.computer.org/computingedge 37
IT TRENDS
monitoring. On a broader scale, mMTC has the poten- form of a 5G offering, network coverage is still heav-
tial to dramatically impact our daily lives via the devel- ily limited.5 Moreover, all 5G devices to be launched
opment of “smart cities.” Building on earlier IoT appli- over the next year are set to rely on existing 4G infra-
cations, MIoT use cases are likely to experience wider structures, a situation known as a “non-standalone”
consumer uptake in the near to mid-term.3,4 Given its network (NSA), until telecoms implement a stand-
low power requirements, uniform coverage, capacity, alone network.7 Given that 5G phones are expected
and robustness to connect up to 100 times the num- to cost substantially more than current 4G phones
ber of devices per unit area as compared to 4G, 5G is and are likely to become outdated quickly as stan-
expected to bring substantial economies of scale to dards coalesce and the technology is improved, 5G
MIoT applications that will drive adoption across all smartphones (and other devices) are not expected to
sectors. Through their widescale proliferation, experts become mainstream for at least a few years.
predict that MIoT use cases will showcase the truly Second, because many MIoT and MCS use cases
transformative impact of 5G.4 will require a wireless spectrum range far greater than
those offered by today's 4G networks, a 5G standalone
Mission Critical Services (SA) spectrum range will likely need to be available
Mission-critical services refer to applications that before many of the new capabilities of 5G will be
require high reliability and ultra-low latency in data brought to bear.5 Relatedly, given that requirements
transfer. Otherwise known as ultra-reliable and for 5G use cases will differ both within and across
low-latency communications, examples of possi- service classes, there may be complications in allocat-
ble 5G mission-critical services include: fully autono- ing the appropriate network resources for each use
mous vehicles; smart grids and the remote control of case, which today is achieved through the creation of
critical infrastructures; many drone applications; and private networks or segmenting portions of the wire-
medical applications, such as remote patient moni- less network for groups of users.3 Although network
toring and telemedicine. Contrary to MIoT use cases segmenting, or slicing, is already a common practice
that are characterized by a very large number of con- in 4G, 5G slicing is expected to be significantly more
nected devices typically transmitting a relatively low complex and challenging due to its unknown behavior
volume of nondelay-sensitive data, MCS use cases and higher data throughput. Thus, similar to the pat-
involve the transmission large volumes of data where a terns seen historically in the adoption of previous
high-latency connection and/or a loss in connectivity mobile generations, we do not expect 5G to become
could have serious, potentially life-threatening, con- ubiquitous for at least several years.
sequences. With near-zero latencies and network reli-
ability of 99.999% uptime as key minimum technical Managerial Challenges
requirements,3,16 5G has the potential to enable a wide And Considerations
range of MCS applications that can transform indus- Due to the lack of real-world 5G applications currently,
tries and entire economies. That being said, MCS have Gartner has recently placed 5G services at the “Peak of
strong security requirements, and most applications Inflated Expectations” in its 2019 Hype Cycle for Enter-
are still considered nascent. Experts predict that while prise Networking.5 Where Gartner considers 5G to be
the potential impact from 5G-enabled MCS use cases five to ten years away from mainstream adoption,5 a
stands to be profound, MCS growth will take longer to key challenge facing business and IT managers will be
gain traction than eMBB and MIoT applications.4 in identifying which 5G use cases present feasible and
valuable business opportunities in the near-term while
5G REALITY CHECK simultaneously preparing for profitable long-term use
Like many technologists, we are hopeful about 5G's cases. Moreover, organizations must also be aware of
transformative potential. Nevertheless, we also the risks and challenges associated with the adoption
believe that much of the early hype surrounding 5G is of 5G technologies and innovations.
premature. First, it bears noting that although the four To help organizations meet this challenge, we
major U.S. carriers have all recently announced some suggest that managers make use of Iansiti and
www.computer.org/computingedge 39
IT TRENDS
applications involve broad, and often public, use cases uncertainty while orchestrating a range of ecosystem
that require a high level of coordination among ecosys- players with potentially competing objectives and a
tem participants. An example of a substitution case is lack of understanding of the new technology.
in-vehicle infotainment (IVI). Following advances in
computing power and display technologies in the early CONCLUSION
2000s, localized use cases of in-vehicle infotainment As we enter into the next generation of mobile net-
became possible; early IVIs included touchscreen works, 5G represents an important “inflection point”
displays to play music and to use GPS navigation from consumer to industry applications,3 where
assistance. Today, 5G offers an opportunity to extend advanced user-to-machine and machine-to-machine
mature IVIs to include applications such as in-car mar- interactions are set to revolutionize the way we live
keting, in-car 5G streaming, and AR/VR-augmented and work. In the years it will take for a widescale
navigation systems.19 While these applications are rel- adoption, we will likely see more providers praise 5G
atively low in novelty, they require coordination across as a “silver bullet.” We urge IT professionals and gen-
car manufacturers, as well as a range of service provid- eral managers to be critical of the hyped expecta-
ers and vendors. Moreover, IVI systems must comply tions and focus instead on the reality of how very high
with laws and safety regulations that can differ across data-transfer speeds, ultra-low latency, and massive
markets. Given their scope and complexity, business system capacity can actually offer opportunities for
managers should approach substitution cases with single use cases, localization, substitution, or trans-
caution. Careful planning and strong change manage- formation of their businesses.
ment strategies are necessary when developing and
implementing substitute applications as “existing REFERENCES
solutions may be difficult to dislodge.”17 1. D. Rosenberg, “How 5G will change the world,” World
Transformation use cases involve applications that Economic Forum, 2018. Accessed: Jan. 15, 2020.
are both high in novelty and complexity. Where suc- [Online]. Available: https://www.weforum.org/agenda
cessful, transformation use cases have the potential /2018/01/the-world-is-about-to-become-even-more
to bring about fundamental economic, political, and -interconnected-here-s-how/
social change. Smart highways are a notable example 2. Qualcomm Technologies, Inc., “Everything you need to
of a transformative use case. By allowing for the inte- know about 5G,” 2019. Accessed: Jan. 15, 2020. [Online].
gration of self-driving cars and sensors, along with Available: https://www.qualcomm.com/invention/5g
technologies such as solar panels to generate power, /what-is-5g
dynamic digital signage, and structural maintenance 3. S. Fulton III, “What is 5G? The business guide to
monitoring systems, smart roads can dramatically next-generation wireless technology,” ZD Net, 2019.
transform the transportation sector. Despite their Accessed: Jan. 19, 2020. [Online]. Available: https:
potential to deliver tremendous value, transformation //www.zdnet.com/article/what-is-5g-the-business
use cases, such as smart highways, require extensive -guide-to-next-generation-wireless-technology/
coordination between a vast array of ecosystem play- 4. K. Campbell, D. J. B. Flanagan, B. O’Neil, and F. Sideco,
ers, such as regulators and infrastructure developers. “The 5G economy: How 5G technology will contribute
As such, transformation use cases will take substan- to the global economy,” IHS Econ. IHS Technol., IHS
tially longer to realize and gain market traction than Markit, London, U.K., 2017.
single use and localization use cases. Thus, similar to 5. A. Lerner and D. Young, “Hype cycle for enterprise
substitution applications, we recommend that IT pro- networking,” Gartner Inc., Stamford, CT, USA (ID
fessionals and general managers consider transforma- G00377401), 2019.
tion use cases very carefully. Certainly, early prepara- 6. Making 5G NR a Reality. San Diego, CA, USA: Qual-
tion can enable forward-thinking organizations to comm Technol. Inc., 2016.
leverage transformation use cases to create new, and 7. S. Segan, “What is 5G?” PC Mag, 2020. Accessed: Jan.
potentially disruptive, business models. Neverthe- 15, 2020. [Online]. Available: https://www.pcmag.com
less, first-movers must be prepared to weather high /news/what-is-5g
8. K. Finley, “ The WIRED guide to 5G ,” Wired, 2019. NURA JABAGI is currently a Ph.D. candidate in business
Accessed: Jan. 15, 2020. [Online]. Available: https: technology management with the John Molson School
//www.wired.com/story/wired-guide-5g/ of Business, Concordia University, Canada. Her research
9. S. Li, L. Da Xu, and S. Zhao, “ 5G internet of things: A interests include the digital economy and understanding
survey,” J. Ind. Inf. Integr., vol. 10, pp. 1 – 9, 2018 . how technological advances coupled with macroeconomic
10. D. O’Halloran, “ What you need to know about 5G ,” trends are disrupting industries, businesses, and jobs. She
World Economic Forum, 2019. Accessed: Jan. 17, 2020. is the corresponding author of this article. Contact her at
[Online]. Available: https://www.weforum.org/agenda n_ jaba@jmsb.concordia.ca .
/2019/12/what-you-need-to-know-about-5g/
11. Techworld, “A timeline of 5G development: From 1979 ANDREW PARK is currently a Ph.D. candidate with the Beedie
to Now,” Int. Data Group (IDG), 2019. Accessed: Jan. 17, School of Business, Simon Fraser University, Vancouver, BC,
2020. [Online]. Available: https://www.techworld.com Canada. His research interests include innovation and tech-
/picture-gallery/tech-innovation/timeline-of-5g nology management. Contact him at andrew_park@sfu.ca .
-development-3654794/
12. E. Ezhilarasan and M. Dinakaran, “A review on mobile JAN KIETZMANN is an associate professor in Management
technologies: 3G, 4G, and 5G ,” in Proc. IEEE Conf. Information Systems, University of Victoria, BC, Canada.
Recent Trends Challenges Comput. Models, 2017, pp. His research interests include organizational and social
369 –373. perspectives related to emerging technologies. He received
13. J. Edwards, “ 5G versus 4G: How speed, latency and the Ph.D. degree in innovation and information systems from
application support differ,” CIO, 2019. Accessed: Jan. London School of Economics and is an associate editor for
17, 2020. [Online]. Available: https://www.cio.com Business Horizons. Contact him at jkietzma@uvic.ca .
/article/3491393/5g-versus-4g-how-speed-latency
-and-application-support-differ.html
14. Ericsson, “A case study on automation in mining | 5G
value,” 2018 . [Online]. Available: https://www.ericsson
.com/en/reports-and-papers/consumerlab/reports
Cutting Edge
/a-case-study-on-automation-in-mining
15. M.-A. Russon, “ Will 5G be necessary for self-driving
stay
cars? ” BBC News, 2018 . Accessed: Jan. 27, 2020.
[Online]. Available: https://www.bbc.com/news on the
of Artificial Intelligence
/business-45048264
16. K. Mendis and F. Li, “Achieving ultra reliable commu-
nication in 5G networks: A dependability perspective
availability analysis in the space domain,” IEEE Com-
J a n ua ry/ f E b r ua ry 2 016
mun. Lett., vol. 21, no. 9, pp. 2057–2060, Sep. 2017. Also in this issue:
aI’s 10 to Watch
real-Time Taxi Dispatching
56
68
IEEE
January/FEBruary 2016
2017.
that perceive, reason, learn, and
18. C. Shea , S. Jacobs, D. Esserman, K. Bruce, and B.
Weiner, “Organizational readiness for implement- act intelligently.
VOLuME 31
nuMBEr 1
www.computer.org/intelligent
https://iot-analytics.com/the-leading-5g-iot-use
-cases-2019/
www.computer.org/computingedge 41
EDITORS: Volodymyr Kindratenko, kindr@ncsa.uiuc.edu
Anne Elster, anne.elster@gmail.com
This article originally
appeared in
High-Level Synthesis-Based
Approach for Accelerating Scientific
Codes on FPGAs
Ramshankar Venkatakrishnan, Ashish Misra, and Volodymyr Kindratenko,
National Center for Supercomputing Applications (NCSA), University of Illinois
T
raditionally, hardware description Languages capable each tool is in terms of optimization, without
(HDLs), such as Verilog or VHDL, have been using tool-specific attributes or pragmas, as well as
used for programming Field-Programmable how well it utilizes the available hardware. We find
Gate Arrays (FPGAs). However, this approach requires that HLS tools are easy to learn and the time to design
an advanced knowledge of digital design and computer is much shorter compared to the HDL approach.
architecture. Recently emerged high-level design However, a good knowledge of digital design and the
tools make it easier for the programmers to code underlying FPGA architecture is still needed to deliver
complex designs in C/C++. High-level synthesis (HLS)1 a high-performance implementation.
and OpenCL2 are the two leading high-level design
platforms that are becoming widely used for program- OVERVIEW OF THE TOOLS
ming FPGAs. Their proponents claim that these tools The OpenCL kernel was implemented using the Xilinx
require little to no knowledge of the hardware design SDAccel environment and the C-based HLS kernel was
principles and can significantly improve developer's implemented in Xilinx Vivado HLS, and then brought
productivity. In this article, we explore these two into the SDAccel environment to integrate with the
high-level design approaches from the point of view of host side of the application.
a software developer. We use Xilinx Vivado HLS C/C++
ver. 2019.13 and Xilinx SDAccel OpenCL ver. 2019.14 to Xilinx Vivado HLS
implement a cross-correlation operation from scratch The Vivado HLS compiler is used to convert code
and synthesize it for a Xilinx u250 Alveo FPGA board.5 written in C/C++ or SystemC into a register transfer
The selected operation is at the core of convolutional level (RTL) representation, which can then be synthe-
neural networks and is generally nontrivial to imple- sized to run on an FPGA. From the point of view of the
ment using a traditional HDL methodology, but is end-user, the HLS compiler is similar to other language
rather simple to implement using a programming lan- compilers that are available for application develop-
guage, such as C. We opted not to focus on the design ment. This tool is simple to learn and use, as long as the
optimization, but rather getting a design that works designer has a good knowledge of C or C++. The overall
on the FPGA with the minimal time spent on its imple- design approach consists of the following steps.
mentation. We use the Xilinx SDAccel platform that
provides support for implementing both OpenCL- and 1. Compile, execute, and debug the algorithm
HLS-based kernels to run on an FPGA using OpenCL written in C/C++.
drivers on the host platform. We also took note of how 2. Synthesize the C/C++ algorithm into an RTL
implementation, optionally using user-guided
optimization directives.
Digital Object Identifier 10.1109/MCSE.2020.2996072 3. Generate comprehensive reports and analyze
Date of current version 19 June 2020. the design.
42 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
NOVEL ARCHITECTURES
4. Optimize the design to meet application the Xilinx xocc compiler. The overall design approach
requirements. consists of the following steps.
5. Verify the Register Transfer Language (RTL)
implementation using a pushbutton flow. 1. Code the desired kernel in OpenCL.
6. Package the RTL implementation into a selec- 2. Run software emulation to ensure functional
tion of Intellectual Property (IP) formats. correctness.
3. Run hardware emulation to generate host and
The last step is to package the RTL output files as kernel profiling data.
an IP core. There are three formats they can be packed 4. Optimize kernel for performance using direc-
into depending upon the way in which the IP is going to tives and code restructuring techniques.
be used further in the design process. In our case, we 5. Run hardware synthesis to generate the FPGA
packed IP as a Vivado IP catalog format and exported kernel bit file.
the generated .xo file into SDAccel. This .xo file, which 6. Write host code using the OpenCL API to inter-
contains the actual kernel, will be integrated with the face with the kernel executed on the FPGA.
OpenCL FPGA wrapper by the SDAccel tools and can
be loaded onto the FPGA from the code executed on Note that nearly the same host code can be used
the host. to interface with the FPGA kernel regardless of the way
the kernel was produced, as long as it was imported
Xilinx SDAccel into SDAccel. Therefore, we use the same host code to
The OpenCL-based design flow utilizes SDAccel, an test both our designs.
Eclipse-based integrated development environment.
We use C/C++ with OpenCL API calls for the host soft- OpenCL IMPLEMENTATION
ware implementation, and OpenCL for hardware ker- The source code of the cross-correlation kernel
nel development. The host application is built through implemented in OpenCL is shown in Figure 1. The
using the standard gcc host compiler, and the FPGA host code writes the data required by the kernel
binary is built through a separate process that uses into a memory bank on the FPGA board. The kernel
www.computer.org/computingedge 43
NOVEL ARCHITECTURES
code then performs the computation by accessing as Advanced eXtensible Interface (AXI) memory
the data from that memory. Once the kernel execu- interfaces and scalar parameters called by the value
tion completes and the data are written back to the are mapped to an AXI4-Lite interface. When creat-
FPGA-attached memory, the host code reads the ing interfaces, it is important to specify the depth of
data back from the global memory and continues pro- the AXI ports––the wrong depth size will result in a
cessing as needed. C/RTL simulation mismatch.
The input image, filter mask, and output image are Once the kernel design is complete, the C simula-
declared as global variables in our kernel. Along with tion tool can be used to verify the design. A simple
these variables, the image and mask dimensions are testbench can be written to run the C simulation.
also declared as function arguments. The values for Once the functionality of the design is verified, the C
the function arguments are written into the global Synthesis tool is used to synthesize the design to an
memory by the host code and are read in during kernel RTL implementation. A top-level function is required
execution. The design consists of four nested loops in the tool. The C Synthesis tools have several options
that are responsible for moving the mask across the to make this simpler, such as echoing the progress of
image horizontally and vertically. The design uses a the synthesis project to console and a GUI interface
stride and padding equal to 1. that provides enhanced information hyperlinks in the
Compared to a straight OpenCL code, our kernel output messages, which provide more information on
uses one additional line of code, __attribute__((reqd_ the source of design issues and how to resolve them.
work_group_size(1, 1, 1))), which specifies working size When synthesis is complete, a report for the top-level
of the kernel to be just one copy. function is generated. The report provides details
Once the host code and kernel code are created, on both the performance and the area of the RTL
the next step is to build the application. The build design and can be used to guide further performance
target can be chosen by the programmer. The SDAc- optimizations.
cel provides three such targets: software emulation, After successfully synthesizing the design, the
hardware emulation, and system. Our design was com- next step is to verify if the RTL is correct. This is done
piled and built using the system option targeting the using the C/RTL co-simulation tool. When the verifica-
Xilinx u250 Alveo FPGA board. The SDAccel generates tion completes, the last step in the Vivado HLS design
reports that can be used to analyze the timing, latency, flow is to package the RTL output as an IP. This is done
and area information of a design. using the Export RTL tool that packages the RTL as a
From the timing information we can, for example, Xilinx object (.xo) file that can then be included in the
verify that the estimated frequency is same as the SDAccel project. Thus, when creating a new project in
target frequency. The area information can be used to the Vivado HLS, it is important to select the SDAccel
further guide the optimization of the design. Bottom Up Flow option, which allows one to run HLS
kernels in SDAccel. A host code similar to the one used
VIVADO HLS in the OpenCL implementation can be used to run the
C/C++ IMPLEMENTATION HLS kernel on the FPGA.
In the HLS implementation, the hardware kernel
is created in C/C++ using the Vivado HLS tool. The DISCUSSION
optimization and performance validation are done We have shown how to implement the cross-correlation
in HLS. The major difference between OpenCL and operation with the two programming frameworks. The
Vivado HLS C/C++ kernel implementation is the biggest challenge was to learn how to use the tools and
use of HLS pragmas instead of OpenCL attributes. how to implement a host code using API calls for cre-
These pragmas are used to declare appropriate ating platforms, attributes, and contexts for executing
interfaces for scalar and vector variables (see Fig- a kernel on the FPGA hardware. The design tools are
ure 2). The kernel acts as an accelerator in SDAc- rather complex and the code implementation method-
cel and is required to be modeled using the guide- ology is tailored more toward hardware designers that
lines provided by SDAccel. Interfaces are modeled software programmers.
The kernel code very closely resembles the origi- significantly longer compilation time than software
nal C code, with some additional directives for guid- designers are used to in more common environments.
ing the compiler to properly implement the intended Even though we have not applied any manual
design. However, the kernels we have implemented optimizations, the compilers attempt to flatten, unroll,
were not optimized, and only basic attributes or prag- and pipeline loops on their own, albeit with limited suc-
mas were applied to enable proper functionality. Our cess. For example, as reported by HLS synthesis tools,
OpenCL implementation consists of 29 lines of kernel loops 3 and 4 are flattened and pipelined with a depth
code and 131 lines of host code, whereas the HLS of 21 cycles and a loop initiation interval of 2. The limit-
kernel consists of 40 lines of kernel code and 154 lines ing factor for the loop initiation interval is the fact that
of host code. (Our host code also included a CPU ver- the data is coming from a single memory port that can
sion of the kernel for verification.) Compared with a only provide one input value on each clock cycle. Both
traditional CPU-only implementation, this is a similar designs are reported to work at 300 Mhz.
code size. The kernel could be further optimized using attri-
A typical compilation time to generate the com- butes or pragmas and code restructuring. The OpenCL
plete FPGA design even for such a small kernel is well approach is somewhat easier for software program-
over three hours on a multicore system where multiple mers who are not familiar with hardware design, but
cores are used by the Xilinx tools. This is, of course, a even this approach is not entirely friendly for those
www.computer.org/computingedge 45
NOVEL ARCHITECTURES
not familiar with the hardware design terminology. UG902 (v2019.1) July 12, 2019. [Online]. Available:
The application developer needs to be able to read https://www.xilinx.com/support/documentation
hardware synthesis reports and correlate the informa- /sw_manuals/xilinx2019_1/ug902-vivado-high-level
tion from these reports with their design. The tools -synthesis.pdf
produce a substantial amount of reports in each phase 4. SDAccel development environment: Release notes,
of the code compilation process, and each of these installation, and licensing guide, UG1238 (v2019.1) July
reports contains information instrumental in guiding 26, 2019. [Online]. Available: https://www.xilinx.com
the kernel optimization. These reports are oriented /support/documentation/sw_manuals/xilinx2019_1
toward hardware designers. /ug1238-sdx-rnil.pdf
Can software developers without knowledge of 5. Getting started with alveo data center accelerator
hardware design principles use these platforms to cards, UG1301 (v1.4) December 18, 2019. [Online]. Avail-
develop codes for FPGAs? Yes, but they will have to able: https://www.xilinx.com/support/documentation
learn a lot in the process. The Vivado HLS approach is /boards_and_kits/accelerator-cards/1_4/ug1301
quite involved with all the verification steps, IP genera- -getting-started-guide-alveo-accelerator-cards.pdf
tion, integration with SDAccel project, etc., whereas
OpenCL is easier for software developers. However,
even a naïve implementation, which is what we have RAMSHANKAR VENKATAKRISHNAN is a Research Pro-
done here, is relatively involved and time-consuming grammer at the National Center for Supercomputing Appli-
regardless of the framework used. And of course, any cations. He received the master’s degree from University of
performance optimizations will require an understand- Illinois at Chicago in 2015. Contact him at rvenka21@illinois
ing of the specific FPGA board's architecture. For .edu.
example, efficient utilization of the available memory
bandwidth on the u250 Alveo FPGA board requires ASHISH MISRA is a Postdoctoral Research Associate at
distributing data across four DDR memory banks as the National Center for Supercomputing Applications. He
well as defining global pointers as 512-bit data types received the master’s and Ph.D. degrees from Birla Institute
to sustain the maximum data bandwidth between the of Technology and Science, India, in 2008 and 2017, respec-
kernel and DDR memory. This requires changes to the tively. Contact him at ashishm@illinois.edu.
host and kernel code to specify which bank is to be
used for which data buffer as well as splitting data into VOLODYMYR KINDRATENKO is a Senior Research Scientist
several buffers. Software developers are almost never at the National Center for Supercomputing Applications. He
concerned with such issues. received the Specialist degree from the Vynnychenko State
Pedagogical University, Kirovograd, Ukraine, in 1993 and the
ACKNOWLEDGMENTS D.Sc. degree from the University of Antwerp, Belgium, in
This work was supported by the National Science 1997. Contact him at kindrtnk@illinois.edu.
Foundation’s Major Research Instrumentation pro-
gram under Grant 1725729, as well as the University of
Illinois at Urbana-Champaign.
REFERENCES
1. G. Martin and G. Smith, “ High-level synthesis: Past,
present, and future,” IEEE Des. Test Comput., vol. 26, no.
4, pp. 18 –25, Jul./Aug. 2009, doi: 10.1109/MDT.2009.83.
2. J. Stone, D. Gohara , and G. Shi, “OpenCL: A parallel
programming standard for heterogeneous computing
systems,” Comput. Sci. Eng., vol. 12, no. 3, pp. 66 –72,
WWW.COMPUTER.ORG/COMPUTINGEDGE
May/Jun. 2010, doi:10.1109/MCSE.2010.69.
3. Vivado design suite user guide: High-level synthesis,
A
clear differentiation between logic, memory, Many SRAM-CiMs in 2-D configurations suffer
and process information in conventional von from a large area overhead due to the need for addi-
Neumann computing schemes necessitates tional transistors.3–6 This has encouraged research-
the frequent movement of data between the memory ers to develop CiM in 3-D technologies for suppressing
and processor. Thus, much of the execution time and the overhead in density or performance of the memory
energy consumption is spent in the movement of data, system when used for CiM.
a barrier referred to as the “von Neumann bottleneck” Multilayer 3-D IC technologies can be divided into
or “memory wall.” 1,2 two categories: through silicon vias (TSV)-based 3-D
Computing-in-memory (CiM) is a promising (TSV-3D) and monolithic 3-D (M3D). The large dimen-
approach to deal with the memory wall problem by sions of TSVs7 and difficulties in establishing vertical
executing computation tasks wherever the data interconnects between layers greatly limits the den-
resides in order to minimize data access operations. sity of TSV-3D devices [see Figure 1(a)]. In contrast,
Among various on-chip memories, SRAM is the best sequential manufacturing can be used to fabricate
working memory due to its high speed and high endur- M3D devices with ultrahigh density vertical cross-layer
ance. A number of silicon-verified SRAM-based CiM connections,8,9 as shown in Figure 1(b).
(SRAM-CiM) devices have been reported.3–6 However, M3D integration has been a technology enabler
most previous implementations have been shown to in the development of CiM-capable SRAM devices,
affect the fundamental performance of the constitu- thanks to affinity for fine-grained vertical intercon-
ent memory devices (i.e., density, speed, and robust- nects with small parasitic load.8–14 CiM designs that
ness). This has prompted research into process tech- involve M3D integration cannot be implemented in
nology aimed at advancing the integration of SRAM TSV-based 3-D devices, due to high pitch and large
within CiM devices. parasitic load for vertical interconnect.
This article also reviews recent advances and
trends in the implementation of CiM using M3D IC
Digital Object Identifier 10.1109/MM.2019.2946489 technology as well as the challenges researchers face
Date of current version 8 November 2019. as they progress in furthering this technology.
2469-7087/21 © 2021 IEEE Published by the IEEE Computer Society September 2021 47
EXPERT OPINION
Multi-channel
Device FDSOI Epi-likeSi FDSOI Multi-finEpi- Junction-
Epi-like Si NC-FinFETs
architecture MOSFET UTB-MOSFETs MOSFET likeSiFinFETs lessdevices
UTB-MOSFETs
Challenges in Developing 2-D CiM Researchers have the sensitivity of SRAM cells in CiM to read
developed a number of 2-D CiM structures, includ- disturbance.
ing an algorithm-dependent SRAM-CiM for binary 5. Long routing: Intense transistor integration
fully connected neural networks,3 a 10T SRAM for in SRAM bitcells requires that the computing
multiple-bit input and binary weight CNN,4 a 6T-SRAM results from CiM operations undergo long
CiM for multiple weight-value multiplication and aver- routing to peripheral read-out circuits when
aging with on-chip training,5 and a twin-8T SRAM-CiM dealing with large-capacity CiM macros.
for multibit input and multibit weight CNN.6 Nonethe- Extended routing tends to increase parasitic
less, all of the above 2-D CiM solutions face a number capacitance, degrade operational speeds, and
of challenges as follows. boost power consumption.
www.computer.org/computingedge 49
EXPERT OPINION
www.computer.org/computingedge 51
EXPERT OPINION
Structure
Bit cell
6T+3T SRAM Standard 6T SRAM 6T+2T SRAM
structure
VDDmin 0.4 V 0.25 V 0.2 V
Boolean: NAND/AND, Boolean: NAND/AND, OR/NOR, XOR/XNOR
Function Arithmetic: Full Adder
OR/NOR, XOR/XNOR Multi Row Multiply and Accumulate (MAC)
CONCLUSION
CiM based on M3D technology is an attractive solu-
tion for data-intensive and AI applications. In this arti- ACKNOWLEDGMENT
cle, we explored the benefits of CiM based on recent The authors gratefully acknowledge support from TSRI,
M3D technologies. We also examined the trends in MOST-Taiwan, and the SRC JUMP Center for Research
M3D-based SRAM-CiM. With improvements in VDD- on Intelligent Storage and Processing-in-memory
min, on–off current ratio, and yield, it is expected that (CRISP). This work was supported in part by the
M3D-based SRAM-CiM will be able to handle func- National Science Foundation Expeditions in Comput-
tions of high complexity with high energy efficiency. ing Program: Visual Cortex on Silicon CCF 1317560.
www.computer.org/computingedge 53
EXPERT OPINION
Proc. IEEE Int. Electron Devices Meeting, San Francisco, Pennsylvania State University, University Park, PA, USA.
CA, USA, 2017, pp. 20.3.1–20.3.4. He received the BTech degree in electrical and electronics
23. S. Srinivasa, et al., “Compact 3-D-SRAM memory with engineering from SASTRA University, Thanjavur, India. He is
concur-rent row and column data access capability a student member of the IEEE.
using sequential monolithic 3-D integration,” IEEE
Trans. Very Large Scale Integr. (VLSI) Syst., vol. 26, no. 4, MENG-FAN CHANG is currently a Full Professor with
pp. 671–683, Apr. 2018. National Tsing Hua University, Hsinchu, Taiwan. Before 2006,
24. H. H. Nho, et al., “A high-speed, low-power 3D-SRAM he has worked in industry for more than ten years. From 1996
architecture,” in Proc. IEEE Custom Integr. Circuits to 1997, he designed memory compilers in Mentor Graphics,
Conf., San Jose, CA, USA, 2008, pp. 201–204. NJ, USA. From 1997 to 2001, he designed embedded SRAMs
25. P. Bhattacharyya, B. Kundu, S. Ghosh, V. Kumar, and and Flash in Design Service Division (DSD) at TSMC, Hsinchu,
A. Dandapat, “Performance analysis of a low-power Taiwan. During 2001–2006, he was a co-founder and a Direc-
high-speed hybrid 1-bit full adder circuit,” IEEE Trans. tor in IPLib Company, Taiwan, where he developed embedded
Very Large Scale Integr. (VLSI) Syst., vol. 23, no. 10, pp. SRAM and ROM compilers, flash macros, and flat-cell ROM
2001–2008, Oct. 2015. products. His research interests include circuit designs
26. W.-H. Chen, et al., “A dual-split-controlled 4P2N 6T for volatile and nonvolatile memory, ultralow-voltage sys-
SRAM in monolithic 3D-ICs with enhanced read speed tems, 3-D memory, circuit–device interactions, spintronics
and cell stability for IoT applications,” IEEE Electron circuits, memristor logics for neuromorphic computing,
Device Lett., vol. 39. no. 8, pp. 1167–1170, Aug. 2018. and Computing-in-Memory for artificial intelligence. He
received the MS degree from The Pennsylvania State Uni-
versity, University Park, PA, USA, and the PhD degree from
ZHIXIAO ZHANG is currently with Fuzhou University, the National Chiao Tung University, Hisnchu, Taiwan. He
Fuzhou, China. His research interests include SRAM- and is an Associate Editor for the IEEE Transactions on Very
NVM-based Computing-in-Memory design. He received Large Scale Integration (VLSI) Systems and the IEEE Trans-
the BS degree in electrical engineering in 2006 and the MS actions on Computer-Aided Design of Integrated Circuits
degree in microelectronics and solid-state electronics in and Systems. He is also a Guest Editor of the IEEE Journal
2009 from Fuzhou University. He is currently pursuing the of Solid-State Circuits, the IEEE Transactions on Circuits
PhD degree in electrical engineering with the National Tsing and Systems—II: Express Briefs, and the IEEE Journal on
Hua University, Hsinchu, Taiwan. Emerging and Selected Topics in Circuits and Systems. He
is also on technical program committees for ISSCC, IEDM
XIN SI is currently with National Tsing Hua University, Hsin- (Ex-Com and MT chair), DAC (Sub-Com Chair), ISCAS (Track
chu, Taiwan. His research interests include analog circuit Co-Chair), A-SSCC, and numerous international confer-
design, memory, and Computing-in-Memory circuit designs. ences. He has been a Distinguished Lecture speaker for the
He received the BS degree in integrated circuit design and IEEE Solid-State Circuits Society (SSCS) and the Circuits and
integration system in 2016 from the University of Electronic Systems Society (CASS), a technical committee member of
Science and Technology of China (UESTC), Chengdu, China, CASS, and the administrative committee (AdCom) member
where he is currently working toward the PhD degree with of the IEEE Nanotechnology Council. He has also been the
UESTC. He is a student member of the IEEE. Program Director of Micro-Electronics Program of Ministry
of Since and Technology (MOST) in Taiwan during 2018–2020,
SRIVATSA SRINIVASA has been with Pennsylvania State Associate Executive Director for Taiwan’s National Program
University, State College, PA, USA, since spring 2015 and is cur- of Intelligent Electronics (NPIE) and NPIE bridge program
rently working toward the PhD degree with the Electrical Engi- during 2011–2018. He is the recipient of several prestigious
neering Department. He received the BE degree in electronics national-level awards in Taiwan, including the Outstanding
and communication engineering from VTU, Belgaum, India. Research Award of MOST-Taiwan, the Outstanding Electrical
Engineering Professor Award, the Academia Sinica Junior
AKSHAY KRISHNA RAMANATHAN is currently working Research Investigators Award, and the Ta-You Wu Memorial
toward the PhD degree in electrical engineering at The Award. He is a Fellow of the IEEE.
www.computer.org/computingedge 55
CAREER OPPORTUNITIES
ONLINE
BUSINESS INFORMATION SYSTEMS
CYBERSECURITY
CYBERSECURITY MANAGEMENT
INFORMATION
HEALTH INFORMATION TECHNOLOGY
INNOVATION IN AI/ML
NETWORKING
VTMIT.VT.EDU
Quantum Advantage. It’s the advantage that comes with harnessing the
quantum universe, and applying that knowledge to computing, sensing,
communications and more. These discoveries promise to change the world
for the better.
QUANTUM
ORCHESTRATION
PLATFORM
The new paradigm for
quantum control
systems, across all
qubit platforms
QUANTUM-MACHINES.CO
www.computer.org/computingedge 57
Quantum Algorithm Design
Our software automatically synthesizes
optimized quantum circuits from high-
level functional models.
www.classiq.io
IEEE SOFTWARE
January/February 2015
March/April 2015
INTERNETWARE AND BEYOND
RELEASE ENGINEERING
Volume 32 Number 1
Volume 32 Number 2
www.computer.org/computingedge 59
Conference Calendar
I EEE Computer Society conferences are valuable forums for learning on broad and dynamically shifting top-
ics from within the computing profession. With over 200 conferences featuring leading experts and thought
leaders, we have an event that is right for you. Questions? Contact conferences@computer.org.
60 September 2021 Published by the IEEE Computer Society 2469-7087/21 © 2021 IEEE
4 November I n f o r m a t i o n Te c h n o l o g y 9 December
• SLIP (ACM/IEEE Int’l Workshop and Eng. Conf.), Windhoek, • BIBM (IEEE Int’l Conf. on Bio-
on System-Level Interconnect Namibia informatics and Biomedicine),
Prediction), virtual 24 November virtual
6 November • SNPD (IEEE/ACIS Int’l Conf. on 12 December
• SmartCloud (IEEE Int’l Conf. on Soft ware Eng., Artificial Intel- • HOST (IEEE Int’l Symposium
Smart Cloud), Newark, USA ligence, Networking, and Par- on Hardware-Oriented Secu-
7 November allel/Distributed Computing), rity and Trust), Washington,
• IISWC (IEEE Int’l Symposium Taichung, Taiwan DC, USA
on Workload Characteriza- 29 November 13 December
tion), virtual • ISM (IEEE Int’l Symposium on • PST (Int’l Conf. on Privacy,
12 November Multimedia), virtual Security, and Trust), virtual
• ICEBE (IEEE Int’ l Conf. on 30 November 20 December
e-Business Eng.), Guangzhou, • ICRC (IEEE Int ’ l Conf. on • MCSoC (IEEE Int’l Sympo -
China Rebooting Computing), virtual sium on Embedded Multicore/
14 November Many-Core Systems-on-Chip),
• SC (Int’l Conf. for High-Perfor- DECEMBER Singapore
mance Computing, Network- 1 December
ing, Storage, and Analysis), St. • PRDC (IEEE Pacific Rim Int’l
Louis, USA Symposium on Dependable
15 November Computing), virtual
• ASE (IEEE/ACM Int’l Conf. on 6 December
Automated Sof tware Eng.), • BDCAT (IEEE Int’l Conf. on
Melbourne, Australia Big Data Computing, Appli-
• BigMM (IEEE Int’l Conf. on Mul- cations, and Technologies),
timedia Big Data), Taichung, Leicester, UK
Taiwan • Blockchain (IEEE Int’l Conf. on
16 November Blockchain), virtual
• AVS S (IEEE Int’ l Conf. on • UCC (IEEE Int’l Conf. on Utility
Advanced Video and Signal- and Cloud Computing), Leices-
Based Surveillance), virtual ter, UK Learn more
17 November 7 December
• ICMU (Int’l Conf. on Mobile • CSASE (Int’l Conf. on Com-
about IEEE
Computing and Ubiquitous puter Science and Software Computer
Networking), Tokyo, Japan Eng.), Duhok, Iraq Society
22 November • ICDM (IEEE Int’l Conf. on Data
• ATS (IEEE Asian Test Sympo- Mining), virtual
conferences
sium), virtual • RTSS (IEEE Real-Time Sys-
computer.org/conferences
23 November tems Symposium), Dortmund,
• IMITEC (Int’l Multidisciplinary Germany
IEEE COMPUTER SOCIETY ELECTION
Volunteer Leadership
Is Vital
Vote by Monday, 20 September at 12PM EDT
www.computer.org/election2021