0% found this document useful (0 votes)
227 views56 pages

04 Spectrum 2024

The document discusses the pressing need for leaner and greener software to reduce the environmental impact of the software industry, which is responsible for a significant portion of global greenhouse gas emissions. It highlights two articles that advocate for efficient coding practices and the importance of legislation to promote sustainable software development. Additionally, it covers advancements in ultrasound technology through MEMS, showcasing its transformative potential in medical applications.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
227 views56 pages

04 Spectrum 2024

The document discusses the pressing need for leaner and greener software to reduce the environmental impact of the software industry, which is responsible for a significant portion of global greenhouse gas emissions. It highlights two articles that advocate for efficient coding practices and the importance of legislation to promote sustainable software development. Additionally, it covers advancements in ultrasound technology through MEMS, showcasing its transformative potential in medical applications.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Go Ahead, Hack This Ultrasound Gets Small Birth of a Silicon Valley FOR THE

Handheld Ham Radio A tale of MEMS and Unicorn A successful TECHNOLOGY


INSIDER
It’s a feature, not a bug miniaturization startup’s early struggles
P. 16 P. 38 P. 44 APRIL 2024

Software Today’s software is


Goes Lean bloated and inefficient.
There’s a better way
& Green
Higher Fidelities
Faster Workflows
Reach new levels of performance on
your quantum processor with the new
SHF+ line of products.

∏Higher SNR, lower phase noise,


and fast muting functionality
∏Ideal for measurements on even the
most sensitive and long-lived of qubits
∏Scalable software with extensive
example libraries
∏Tested at top labs worldwide

Zurich
Instruments
Unleash your
quantum lab
VOLUME 61 / ISSUE 4 APRIL 2024

44
The Messy Reality
Behind a Silicon
Valley Unicorn
Read the true story of one startup’s push
for the big time. By Benjamin Shestakofsky

Bloat Is Software’s
Biggest Vulnerability
22 Inkjets Are for More
Than Just Printing
32 NEWS
Open-Source Hardware
6

Perplexity’s AI Search
A programmer makes The office tech is great for genomics
Infrared Around Corners
a renewed plea for lean and additive manufacturing.
software. By Bert Hubert By Phillip W. Barth & Leslie A. Field HANDS ON 16
The UV-K5 transceiver is cheap
We Need to How Ultrasound 38 and very, very hackable.
Decarbonize 26 Became Ultra Small 5 QUESTIONS 19
Software MEMS technology might Seth Fraden explains
The green software movement turn ultrasound into the new Brandeis University’s
is tackling the environmental stethoscope. new engineering program.
impact of code. By F. Levent Degertekin
CAREERS 20
By Rina Diane Caballar Zach Rattner’s AI tool
makes moving day easier.
EDITOR’S NOTE 2
ON THE COVER: With software, leaner and PAST FORWARD 52
Illustration by Elias Stein greener is better. No, You Can’t Quantify a Soul

Photo by The Voorhes APRIL 2024 SPECTRUM.IEEE.ORG 1


EDITOR’S NOTE BY HARRY GOLDSTEIN

Software Sucks,
but It Doesn’t
Have To
How to make leaner, greener software

Y
ou can’t see, hear, taste, feel, or smell it, “There’s Hubert, a software developer himself, walks the
but software is everywhere around us. It lean walk: His 3-megabyte image-sharing program
an already
underpins modern civilization even while Trifecta does the same job as other programs that
consuming more energy, wealth, and time
existing use hundreds of megabytes of code.
than it needs to and burping out a significant amount large Lean software should, in theory, be green soft-
of carbon dioxide into the atmosphere. The soft- segment ware: It should run so efficiently that it reduces the
ware industry and the code it ships need to be much of the amount of energy used in data centers and transmis-
more efficient in order to minimize the emissions software- sion networks. Overall, the IT and communications
attributable to programs running in data centers development sectors are estimated to account for 2 to 4 percent
and over transmission networks. Two approaches ecosystem of global greenhouse gas emissions and, according
to software development featured in this issue can that cares to one 2018 study, could by 2040 reach 14 percent.
help us get there. And that study came out prior to the explosion in AI
In his article “Why Bloat Is Still Software’s Biggest
about this applications, whose insatiable hunger for computing
Vulnerability” [p. 22], Bert Hubert pays homage to
space—they resources and the power required to feed the algo-
the famed computer scientist and inventor of Pascal, just haven’t rithms exacerbates an already complicated problem.
Niklaus Wirth, whose influential essay “A Plea for known what Thankfully, several groups are working on solu-
Lean Software” appeared in IEEE Computer in 1995. to do.” tions, including the Green Web Foundation. The
Wirth’s essay built on a methodology first conceived —ASIM HUSSAIN, GREEN GWF was spun up almost 20 years ago to figure out
WEB FOUNDATION
by IEEE Spectrum contributing editor Robert N. how the Internet is powered and now has a goal of
­Charette, who in the early 1990s adapted the Toyota a fossil-free Internet by 2030.
Production System for software development. There are three main ways to achieve that objec-
Hubert points out that bloated code offers giant tive, according to the foundation’s chair and exec-
attack surfaces for bad actors. Malicious hacks and utive director Asim Hussain: Use less energy, use
ransomware attacks, not to mention run-of-the-mill fewer physical resources, and use energy more pru-
software failures, are like the weather now: partly dently—by, for instance, having your apps do more
cloudy with a 50 percent chance of your app crash- when there’s power from wind and solar available
ing or your personal information being circulated and less when there’s not.
on the Dark Web. Back in the day, limited compute “There’s an already existing large segment of the
resources forced programmers to write lean code. software-development ecosystem that cares about
Now, with much more robust resources at hand, this space—they just haven’t known what to do,”
coders are writing millions of lines of code for rela- Hussain told Spectrum contributing editor Rina
PORTRAIT BY SERGIO ALBIAC

tively simple apps that call on hundreds of libraries Diane Caballar. They do now, thanks to Caballar’s
of, as Hubert says, “unknown provenance.” extensive reporting and the handy how-to guide she
Among other things, he argues for legislation includes in “We Need to Decarbonize Software”
along the lines of what the European Union is trying [p. 26]. Programmers have the tools to make soft-
to enforce: “NIS2 for important services; the Cyber ware leaner and greener. Now it’s up to them—and
Resilience Act for almost all commercial software as we’ve seen in the EU, their legislators—to make
and electronic devices; and a revamped Product sustainable and secure code their top priority. Soft-
Liability Directive that also extends to software.” ware doesn’t have to suck.

2 SPECTRUM.IEEE.ORG APRIL 2024 Illustration by Daniel Zender


Power Innovation
in Electrical Design
with COMSOL Multiphysics®
Electrification success calls for smart design innovation and fast-paced
product development. To achieve this, industry leaders are turning to
multiphysics simulation to accurately test, optimize, and predict the
performance of high-voltage equipment and power systems.
» comsol.com/feature/electrical-innovation
CONTRIBUTORS

 PHILLIP W. BARTH
Barth is an alumnus of HP Labs and
Agilent Labs and is now at SmallTech
EDITOR IN CHIEF Harry Goldstein, [email protected] IEEE BOARD OF DIRECTORS
Consulting, which advises on mini-, PRESIDENT & CEO Thomas M. Coughlin, [email protected]
EXECUTIVE EDITOR Jean Kumagai, [email protected]
micro-, and nanotechnology. His +1 732 562 3928 Fax: +1 732 981 9515
MANAGING EDITOR Elizabeth A. Bretz, [email protected]
coauthor, Leslie A. Field, founded and PRESIDENT-ELECT Kathleen A. Kramer
CREATIVE DIRECTOR TREASURER Gerardo Barbosa
leads SmallTech and led work at HP to SECRETARY Forrest D. Wright
Mark Montgomery, [email protected]
improve inkjet refill for large-format PAST PRESIDENT Saifur Rahman
DIRECTOR OF DIGITAL INNOVATION
printers. In their article [p. 32], Barth Erico Guizzo, [email protected] VICE PRESIDENTS
and Field survey how inkjets have EDITORIAL DIRECTOR, CONTENT DEVELOPMENT
Rabab Kreidieh Ward, Educational Activities; Deepak Mathur,
Member & Geographic Activities; Sergio Benedetto,
progressed from printing images to Glenn Zorpette, [email protected]
Publication Services & Products; Manfred J. Schindler,
producing objects and devices. SENIOR EDITORS Technical Activities; James E. Matthews, President, Standards
Evan Ackerman (Digital), [email protected] Association; Keith A. Moore, President, IEEE-USA
Stephen Cass (Special Projects), [email protected]
Samuel K. Moore, [email protected] DIVISION DIRECTORS
 RINA DIANE CABALLAR Tekla S. Perry, [email protected] Yong Lian (I); Kevin L. Peterson (II); Stefano Bregni (III); Alistair
Eliza Strickland, [email protected] P. Duffy (IV); Christina M. Schober (V); Kamal Al-Haddad (VI);
Caballar, a software engineer turned Christopher E. Root (VII); Leila De Floriani (VIII); Aylin Yener (IX);
ART & PRODUCTION
journalist and an IEEE Spectrum Stephanie M. White (X)
DEPUTY ART DIRECTOR Brandon Palacio, [email protected]
contributor, discovered that, ironically, PHOTOGRAPHY DIRECTOR Randi Klett, [email protected] REGION DIRECTORS
Bala S. Prasanna (1); Andrew D. Lowery (2); Eric Grigorian (3);
the U.N.’s climate change conference ONLINE ART DIRECTOR Erik Vrielink, [email protected]
PRINT PRODUCTION SPECIALIST Vickie A. Ozburn (4); Anthony M. Francis (5); Kathy Herring Hayashi
website consumes an unnecessary Sylvana Meneses, [email protected] (6); Thamir F. Murad (7); Vincenzo Piuri (8); Jenifer P. Castillo
amount of energy. On page 26, she ex- MULTIMEDIA PRODUCTION SPECIALIST ­Rodriguez (9); ChunChe Fung (10)
plores how the way we write software Michael Spector, [email protected] IEEE STAFF
is having underappreciated impacts NEWS MANAGER Margo Anderson, [email protected] EXECUTIVE DIRECTOR & COO Sophia A. Muirhead
+1 732 562 5400, [email protected]
on carbon emissions, and explains ASSOCIATE EDITORS
CHIEF OF STAFF TO THE EXECUTIVE DIRECTOR Kelly Lorne
Dina Genkina, [email protected]
how the green software movement is +1 732 562 6011, [email protected]
Willie D. Jones (Digital), [email protected]
GENERAL COUNSEL & CHIEF COMPLIANCE OFFICER
meeting the challenge. Michael Koziol, [email protected]
Anta Cissé-Green +1 212 705 8927, [email protected]
Emily Waltz, [email protected]
INTERIM MANAGING DIRECTOR,
SENIOR COPY EDITOR Joseph N. Levine, [email protected] TECHNICAL ACTIVITIES Kenneth Gilbert
 F. LEVENT DEGERTEKIN COPY EDITOR Michele Kogon, [email protected] +1 732 562 3856, [email protected]
EDITORIAL RESEARCHER Alan Gardner, [email protected] CHIEF MARKETING OFFICER Karen L. Hawkins
Degertekin is the George W. Woodruff DIRECTOR, NEW PRODUCT AND AUDIENCE DEVELOPMENT +1 732 562 3964, [email protected]
Chair in Mechanical Systems at Tim Warder, [email protected] PUBLICATIONS Steven Heffner
+1 212 705 8958, [email protected]
Georgia Tech’s school of mechanical AUDIENCE DEVELOPMENT MANAGER CORPORATE ACTIVITIES Donna Hourican
engineering. On page 38 he chronicles Laura Bridgeman, [email protected] +1 732 562 6330, [email protected]
how microelectromechanical system CONTRIBUTING EDITORS Rina Diane Caballar, Robert N. Charette, MANAGING DIRECTOR, CONFERENCES,
Charles Q. Choi, Tom Clynes, Peter Fairley, Edd Gent, W. Wayt Gibbs, EVENTS AND EXPERIENCES Marie Hunter,
(MEMS) technology has transformed + 1 732 465 5889, [email protected]
Mark Harris, Lucas Laursen, Allison Marsh, Julianne Pepitone, ­
ultrasound machines and opened up Matthew S. Smith, Lawrence Ulrich MEMBER & GEOGRAPHIC ACTIVITIES Cecelia Jankowski
new applications in medicine. “I mostly THE INSTITUTE
+1 732 562 5504, [email protected]
EDUCATIONAL ACTIVITIES Jamie Moesch
work on catheter-based devices, and EDITOR IN CHIEF Kathy Pretz, [email protected]
+1 732 562 5514, [email protected]
ASSOCIATE EDITOR Joanna Goodrich, [email protected]
for those, the road was blocked for like STANDARDS ACTIVITIES Alpesh Shah
SUBSCRIPTION SERVICES
30 years. MEMS enabled us to push IEEE Members [email protected]
+1 732 465 6467, [email protected]
CHIEF FINANCIAL OFFICER Thomas R. Siegert
through that barrier,” he says. Non-members +1 866 363 2304 +1 732 562 6843, [email protected]
DIRECTOR, PERIODICALS PRODUCTION SERVICES Peter Tuohy CHIEF INFORMATION DIGITAL OFFICER Jeff Strohstein
ADVERTISING PRODUCTION MANAGER +1 732 562 5332, [email protected]
 BERT HUBERT Felicia Spagnoli, [email protected] CHIEF HUMAN RESOURCES OFFICER Cheri N. Collins Wideman
+1 732 562 5335, [email protected]
ADVERTISING PRODUCTION +1 732 562 6334
In this issue, Hubert—software MANAGING DIRECTOR, IEEE-USA Russell T. Harrison
EDITORIAL ADVISORY BOARD, IEEE SPECTRUM +1 202 530 8326, [email protected]
engineer, founder of two companies, Harry Goldstein, Chair; Ella M. Atkins, Yedukondala N. Dwith Chenna,
technical advisor to the Dutch govern- Sangyeun Cho, Hugh Durrant-Whyte, Matthew Eisler, Shahin F ­ arshchi, IEEE PUBLICATION SERVICES & PRODUCTS BOARD
Sergio Benedetto, Chair; Derek Abbott, Nazanin Bassiri-Gharb,
ment—laments the current state of Alissa Fitzgerald, Benjamin Gross, Carlos Gutierrez, Lawrence O. Hall,
Stefano Galli, Maria Sabrina Greco, Lawrence O. Hall, James
Daniel Hissel, Jason K. Hui, Benjamin Kroposki, ­Michel M. Maharbiz,
software [p. 22], whose size, he argues, Somdeb Majumdar, Ramune Nagisetty, Paul Nielsen, Sofia Olhede, Irvine, Fabrice Labeau, Yong “Peter” Lian, Aleksandar Mastilovic,
has ballooned beyond all reason. This Amit K. Singh, Christoph Stiller, Mini S. ­Thomas, Wen Tong, Haifeng Paolo Montuschi, Annette Reilly, Anna Scaglione, Gianluca Setti,
Wang, Boon-Lock Yeo Gaurav Sharma, Sarah Spurgeon, Steve Yurkovich, Weihua Zhuang
software bloat is not merely inelegant,
EDITORIAL ADVISORY BOARD, THE INSTITUTE IEEE OPERATIONS CENTER
Hubert says, but also has dire conse- 445 Hoes Lane, Box 1331, Piscataway, NJ 08854-1331 U.S.A.
Kathy Pretz, Chair; Qusi Alqarqaz, Stamatis Dragoumanos,
quences for software security, as it is Madeleine Glick, Harry Goldstein, Francesca Iacopi, Meenakshi Jindal, Tel: +1 732 981 0060 Fax: +1 732 981 1721
virtually impossible to ensure there are Lawrence O. Hall, Shashi Raj Pandey, John Purvis, Chenyang Xu
IEEE SPECTRUM (ISSN 0018-9235) is published monthly by
no bugs in a vast code base. MANAGING DIRECTOR, PUBLICATIONS Steven Heffner
The Institute of Electrical and Electronics Engineers, Inc. All rights
DIRECTOR, BUSINESS DEVELOPMENT, reserved. © 2024 by The Institute of Electrical and Electronics
MEDIA & ADVERTISING Mark David, [email protected] Engineers, Inc., 3 Park Avenue, New York, NY 10016-5997, U.S.A. Volume

 BENJAMIN SHESTAKOFSKY EDITORIAL CORRESPONDENCE No. 61, Issue No. 4. The editorial content of IEEE Spectrum magazine
does not represent official positions of the IEEE or its organizational
IEEE Spectrum, 3 Park Ave., 17th Floor,
Before becoming an assistant profes- New York, NY 10016-5997 TEL: +1 212 419 7555 units. Canadian Post International Publications Mail (Canadian
BUREAU Palo Alto, Calif.; Tekla S. Perry +1 650 752 6661 Distribution) Sales Agreement No. 40013087. Return undeliverable
sor of sociology at the University of Canadian addresses to: Circulation Department, IEEE Spectrum, Box
ADVERTISING INQUIRIES Naylor Association Solutions,
Pennsylvania, Shestakofsky worked Erik Albin +1 352 333 3371, [email protected]
1051, Fort Erie, ON L2A 6C7. Cable address: ITRIPLEE. Fax: +1 212 419
7570. INTERNET: [email protected]. ANNUAL SUBSCRIPTIONS:
for 19 months at a successful startup. REPRINT SALES +1 212 221 9595, ext. 319 IEEE Members: $21.40 included in dues. Libraries/institutions: $399.
His article [p. 44] is excerpted from the REPRINT PERMISSION / LIBRARIES Articles may be photocopied POSTMASTER: Please send address changes to IEEE Spectrum,
book he wrote about his experiences. for private use of patrons. A per-copy fee must be paid to the Copyright c/o Coding Department, IEEE Service Center, 445 Hoes Lane, Box 1331,
Clearance Center, 29 Congress St., Salem, MA 01970. For other copying Piscataway, NJ 08855. Periodicals postage paid at New York, NY,
His colleagues reacted with disbelief and additional mailing offices. Canadian GST #125634188. Printed at
or republication, contact Managing Editor, IEEE Spectrum.
when he returned to his Ph.D. program: 3401 N. Heartland Dr, Liberty, MO 64068, U.S.A. Printed at Spenta
COPYRIGHTS AND TRADEMARKS Multimedia Pvt. Ltd., Plot 15, 16 & 21/1, Village Chikhloli, Morivali, MIDC,
“Who jumps off a rocket ship that’s IEEE Spectrum is a registered trademark owned by The Institute of Ambernath (West), Dist. Thane. IEEE Spectrum circulation is audited
blasting into space?” said one. Electrical and Electronics Engineers Inc. Responsibility for the sub- by BPA Worldwide. IEEE Spectrum is a member of the Association
stance of articles rests upon the authors, not IEEE, its organizational of Business Information & Media Companies, the Association of
units, or its members. Articles do not represent official positions of Magazine Media, and Association Media & Publishing. IEEE prohibits
IEEE. Readers may post comments online; comments may be excerpt- discrimination, harassment, and bullying. For more information, visit
ed for publication. IEEE reserves the right to reject any advertising. https://www.ieee.org/about/corporate/governance/p9-26.html.

4 SPECTRUM.IEEE.ORG APRIL 2024


BE AN INNOVATOR
CHANGE THE WAY WE SEE THE WORLD

NRO - DII RESEARCH & DEVELOPMENT FUNDING PROGRAM


The National Reconnaissance Office (NRO) Director’s DII awards up to $500,000.
Innovation Initiative (DII) funds cutting-edge scientific Visit the website for additional details.
research in a high-risk, high-payoff environment
to discover innovative concepts and creative ideas www.nro.gov/DII | [email protected]
that transform overhead intelligence capabilities
and systems for future national security intelligence
ACCEPTING PROPOSALS
needs. The program seeks the brightest minds and
breakthrough technologies from industry, academia, MARCH 11 – MAY 3
national laboratories, and U.S. government agencies.
THE LATEST DEVELOPMENTS IN TECHNOLOGY, ENGINEERING, AND SCIENCE 

The Power
of Openness
The history of
RISC-V’s rise

Rocket: 64-bit
SEMICONDUCTORS RISC-V microprocessor

Open-Source 2010 2011 2012 2013 2014 2015


Hardware
Secures a Win
Sprung from RISC- RISC-V
instruction set

V’s roots, the first


architecture
(ISA) project
begins
open security chip
hits the market BOOM: Out-
of-order
RISC-V core

BY DINA GENKINA

T
he first commercial silicon By the end of 2023, engineering at ETH Zurich and lead of
chip that includes open-source, more than 10 billion PULP project, the academic open-source
chips that contained
built-in hardware security was RISC-V cores had project from which OpenTitan’s own
announced in February by the shipped to equipment RISC-V core is derived.
makers and end users
OpenTitan coalition. around the world. The
Using a RISC-V-based processor
This milestone represents another RISC-V open-hardware core, the chip, called Earl Grey, includes
step in the growth of the open-hardware movement began in a lab a number of built-in hardware security
at the University of
movement. Open hardware has been California, Berkeley,
and cryptography modules, all working
gaining in popularity and adoption since in 2010. Its evolution together in a self-contained micropro-
the release of the now broadly deployed shows the appeal cessor. The project began back in 2019
of using an agreed-
RISC-V open-source instruction set upon set of hardware
with a coalition that has grown to 10
architecture in 2010. specifications without companies. It was started by Google and
As a prescription for how a computer the restrictions of shepherded by the nonprofit lowRISC
licensing fees. Here
can operate efficiently at the most basic are some noteworthy in Cambridge, England. Modeled after
level, RISC-V provides a compelling highlights on the open-source software projects, Open-
starting point for a hardware parallel to RISC-V timeline. Titan has been developed by contribu-
the open-source software movement. CLOCKWISE FROM TOP LEFT: tors from around the world, both official
UNIVERSITY OF CALIFORNIA,
But OpenTitan goes beyond it. Open­ BERKELEY; WESTERN DIGITAL; affiliates with the project and indepen-
MIPS; OPENTITAN; ESPERANTO
Titan open-sources key aspects of the TECHNOLOGIES; ETH ZURICH/ dent coders. OpenTitan’s new silicon
silicon design itself. Although other UNIVERSITY OF BOLOGNA;
UNIVERSITY OF CALIFORNIA,
represents the culmination of five years
open-source silicon chips have been BERKELEY of work.
developed, this is the first one to include “This chip is very, very exciting,” says
the design-­verification stage and to pro- OpenTitan founder and former director,
duce a fully functional commercial chip, Dominic Rizzo, who is now founder and
the coalition says. CEO of coalition partner zeroRISC. “But
“OpenTitan is a big step forward in there’s a much bigger thing here, which
giving credibility to the open-source is the development of this whole new type
hardware movement, that it can really of methodology. Instead of a traditional…
deliver industrial-strength projects,” command-and-control-style structure,
says Luca Benini, professor of electrical this is distributed.”

6 SPECTRUM.IEEE.ORG APRIL 2024


APRIL 2024

SweRV:
Open-source eVocore: RISC-V chip for
core with heterogeneous compute
commercial Earl Grey: Open-source hardware
design security via OpenTitan
verification

2016 2017 2018 2019 2020 2021 2022 2023 2024

Manticore:
RISC-V
chiplet Esperanto:
architecture High-
for efficient performance
floating-point RISC-V
computing AI processor

The methodology developed is called Rooting for Hardware Root of Trust is that bugs are more likely to get noticed
Silicon Commons, a framework that pro- OpenTitan uses a hardware security pro- and the bug fixes can be verified. “The
vides rules for documentation, pre- tocol known as a root of trust (RoT). The openness is a good thing,” says Subrah-
defined interfaces, quality standards, and idea is to provide an on-chip source of manyan. “Because for example, let’s say
governance structure on making deci- cryptographic keys that is inaccessible a proprietary implementation has some
sions as a collective. remotely. Because it’s otherwise inacces- problem. I won’t necessarily know, right?
These protocols can help, says the sible, the system can trust that it hasn’t I’m at [the designer’s] mercy as to
CEO of lowRISC, Gavin Ferris, because been tampered with, providing a basis to whether they’re going to tell me or not.”
open-source hardware design faces chal- build security on. “Root of trust means Open-sourcing hardware security
lenges that open-source software that at the end of the day, there is some- also has certain drawbacks. “Quality
doesn’t: Among them are greater costs, thing that we both believe in,” and a open-source [intellectual property]
a smaller professional community, and trusted secure connection can be estab- organically becomes part of the global
an inability to supply bug fixes as patches lished, explains Ravi Subrahmanyan, supply chain,” says Todd Austin, profes-
after a product is released. senior director of integrated circuit sor of electrical engineering and com-
Another key to the success of the design at Analog Devices, who was not puter science at the University of
project, Ferris says, was picking a prob- involved in the effort. Michigan. “Thus, when a vulnerability is
lem that all the partners would have an Conventional, proprietary chips can (inevitably) found in OpenTitan, every
incentive to continue solving, in over also include open-source RoT technol- system that uses that root of trust will be
what became five years of development. ogy—to provide an extra layer of trust, vulnerable to attacks.” But, Austin
Hardware security was the right fit for proponents argue. Because anyone can believes the advantages—having more
the job because of its commercial inspect and probe the design, the theory eyes on the security protocol and allow-
importance as well as its particular fit ing a diverse pool of researchers and
to the open-source model. There’s a companies to work together—outweigh
notion in cryptography known as the downsides, especially in the long run.
­Kerckhoffs’s principle, which states that This kind of on-chip security is espe-
the only thing that should actually be cially relevant in devices forming the
secret in a cryptosystem is the secret key Internet of Things (IoT), which suffer
itself. Open-sourcing the entire protocol from unaddressed security challenges.
allows the cryptosystem to conform to ZeroRISC and its partners opened up
this rule. sales to IoT markets via an early-access

APRIL 2024 SPECTRUM.IEEE.ORG 7


NEWS

program, and they anticipate broad


adoption in that sphere. Open source “Once the methodology has been
proven, others will pick it up,” Rizzo

Open Source Marches On


“just takes over says. “If you look at what’s happened
with open-source software, first, people
Rizzo and Ferris believe their chip is a because it has thought it was kind of an edge pursuit,
template for open-source hardware devel-
opment that other collaborations will certain valuable and then it ended up running almost
every mobile phone. It just takes over
replicate. On top of providing transparent
security, open-sourcing saves companies
properties.... because it has certain valuable proper-
ties. And so I think we’re seeing the
money by allowing them to reuse hard- I think we’re beginning of this now with silicon.”
ware components rather than having to
independently develop proprietary ver- seeing the Indeed, multiple open-hardware
projects are gaining ground within aca-
sions of the same thing. It also opens the
door for many more partners to partici-
beginning of this demia and the broader industry. Those
include Benini’s PULP project, which
pate in the effort, including academic now with silicon.” designs RISC-V based ultralow power
institutions such as OpenTitan coalition chips; frameworks and design reposito-
partner ETH Zurich. Thanks to academic —DOMINIC RIZZO, ZERORISC ries like the University of California,
involvement, OpenTitan was able to incor- Berkeley’s Chipyard and the Linux foun-
porate cryptography protocols that are dation’s CHIPS alliance; Shenzhen,
safe against future quantum computers. ­China-based RISC-V research facility

C
hatGPT’s release on 30 November 2022 was
met with much fanfare and plenty of pushback.
It quickly became clear people wanted to ask AI
the same questions they asked Google—and
ChatGPT often wasn’t capable of an answer.
The problems were numerous. ChatGPT’s replies were
out of date, didn’t cite sources, and frequently hallucinated
new and inaccurate details. Emily Bender, director of the
University of Washington’s Computational Linguistics
Laboratory, was quoted at the time as saying that AI search
was “The Star Trek fantasy, where you have this all-­
knowing computer that you can ask questions.”
Founded in August of 2022, Perplexity the startup
stumbled into—and then raced toward—building an
AI-powered search engine that’s updated daily and
responds to queries by citing multiple sources. It now has
over 10 million monthly users and recently received an
investment from Jeff Bezos.
“I think Google is one of the most complicated systems
humanity has ever built. In terms of complexity, it’s prob-
ably even beyond flying to the moon,” says Perplexity.ai
ARTIFICIAL INTELLIGENCE cofounder and chief technology officer Denis Yarats.
Perplexity initially hoped to build an AI-powered text-

Perplexity.ai Shakes to-SQL tool, Yarats says, to let developers query and code
for SQL (structured query language) in natural language.
But something different started brewing in the company’s
Up Search AI startup Slack channels—a chatbot that combined search with
OpenAI’s large language models (LLMs).
challenges Google and Then, in late November of 2022, ChatGPT went public
and became the fastest-growing consumer application in

Bing—and ChatGPT, too history, hitting 100 million users within two months.
People were asking ChatGPT all sorts of questions, many
of which it couldn’t answer. But Yarats says Perplexity’s
Slack bot could.
ISTOCK

“Literally in two days, we created a simple website and


BY MATTHEW S. SMITH hooked it up to our Slack bot’s back-end infrastructure,

8 SPECTRUM.IEEE.ORG APRIL 2024


RIOS Lab; and the OpenHW Group non- The OpenTitan coa­li­
profit where members collaborate on tion says its first
commercial-grade,
open-source hardware and related soft- open-source hardware-
ware and IP. security chip is ready.
Benini’s vision is to open-source the
entirety of the chip-development pro-
cess, from the high-level design, to the
electronic design automation (EDA)
tools, to the foundry’s process design kit
(PDK) and related data. Currently, EDA
tools and PDKs are closed source.
Open-sourcing them would open up the
entire chip-development life cycle. “The
next generation of open-source hard-
ware will be really, from software to
atoms, open source. This is the vision
that we have looking forward. And it’s a
couple of steps forward with respect to
what OpenTitan demonstrated, but I
think it’s achievable.”

and just released it as a fun demo,” says Yarats. “Honestly, every hour,” says Yarats. (Contrast this with the search-­
it didn’t work super well. But given how many people liked engine data itself, which is more frequently updated. Yarats
it, we realized there [was] something there.” says RAG allows Perplexity access to new data without
How did Perplexity, a company founded by four people retraining.)
(now having grown to roughly 40), cut through problems Meanwhile, crawling the Web at Google’s scale also
in less than two years that seemingly made AI terrible for isn’t practical; Perplexity lacks the tech giant’s resources
search? and infrastructure. To manage the load, Perplexity splits
Retrieval-augmented generation, or RAG, is one pillar results into “domains,” which are updated with more or
of the company’s efforts. Invented by researchers at Meta, less urgency. News sites are updated more than once every
the University of London, and New York University, RAG hour. Sites that are unlikely to change quickly, on the other
teams generative AI with a “retriever” coupled to a vector hand, are updated once every few days.
database. The pair together can incorporate new informa- But two decades of failed Google competitors have
tion into the model’s answers without retraining. proven that just being good enough isn’t good enough.
“I do agree RAG [is useful for search],” says Bob van That’s where AI offers a shortcut.
Luijt, cofounder and CEO of AI infrastructure company LLMs are excellent at parsing text to find relevant
Weaviate. “What [RAG] did was allow normal develop- information—indeed, finding patterns is what they do
ers, not just people working at Google, to just build these best. That allows an LLM to produce convincing text in
kinds of AI native applications without too much response to a prompt, but it can also be used to efficiently
hassle.” He points out that the resources for implement- parse and then present information an LLM examines.
ing RAG are available for free on HuggingFace, an AI You can try this yourself by uploading a PDF to ChatGPT,
developer resource. Google Gemini, or Claude.ai. The LLM can ingest the
That’s led to widespread adoption. Weaviate uses RAG documents within seconds, then answer questions about
to help its clients ground the knowledge of AI agents on the document.
proprietary data, van Luijt says. Moreover, Nvidia uses Perplexity essentially does the same for the Web and,
RAG to reduce errors in ChipNeMo, an AI model built to in so doing, it fundamentally alters how search works. It
aid chip designers. And as IEEE Spectrum has previously doesn’t attempt to rank Web pages to place the best page
reported, an LLM called Latimer uses it to combat racial at the top of a list of queries, Yarats says, but instead ana-
bias and amplify minority voices. So, Yarats says, ­Perplexity lyzes the information available from an index of well-
has turned RAG toward search. ranked pages to find what’s most relevant and generate
But for RAG to be of any use at all, a model must have an answer. That, Yarats says, is the big secret.
something to retrieve, and here Perplexity.ai adopts more “You can think of it like the LLM does the final ranking
traditional search techniques. The company uses a Web task,” says Yarats. “[LLMs] don’t care about an [SEO]
crawler of its own design, known as PerplexityBot, to index score. They just care about semantics and information.
OPENTITAN

the Internet. It’s more unbiased, because it’s based on the actual infor-
“When trying to excel in up-to-date information, like mation gain rather than the signals Google engineers
news…we won’t be able to retrain a model every day, or optimize for whatever reasons.”

APRIL 2024 SPECTRUM.IEEE.ORG 9


NEWS

TELECOMMUNICATIONS

Wi-Fi HaLow
Extends Signals
to 3 Kilometers
But fierce
competition
assures a rough
road ahead
BY EDD GENT

M
ost people have probably experienced Startup Morse quency band rather than the 2.4-gigahertz band.
the frustration of weak Wi-Fi signals. Micro has used Lower-frequency signals are able to propagate far-
the HaLow
Even getting a network to cover every standard to ther and are better at penetrating solid objects, both
corner of a fairly modest house can be beam megabits of which enable HaLow to reach longer ranges.
a challenge. That’s not a problem for Wi-Fi technol- of data across The protocol also allows channel bandwidths as
3 kilometers
ogy developed by the Australian startup Morse using Wi-Fi. narrow as 1 MHz, compared with the 20-MHz chan-
Micro, which just demonstrated a Wi-Fi signal with nels that are standard in conventional Wi-Fi. Guda
a 3-kilometer range. And while the range and long says this makes it possible to have many more ded-
battery lifetimes remain impressive, the technolo- icated channels that don’t interfere with one another,
gy’s boosters are now vying for wireless and Internet which is useful for IoT applications in which several
of Things (IoT) developers to adopt it rather than a devices are connected to the same network.
number of other competing low-power and long- The trade-off is lower throughput, says Guda.
range standards. However, Morse Micro says its technology can still
Morse Micro has developed a system-on-chip support data-intensive applications like video. The
(SoC) design that uses a wireless protocol called technology also operates in a lower power range than
Wi-Fi HaLow, based on the IEEE 802.11ah standard. conventional Wi-Fi, as it has built-in features that
The protocol significantly boosts range by using allow devices to remain dormant for long periods and
lower-frequency radio signals that propagate farther wake only when needed for transmitting data. So
than conventional Wi-Fi frequencies. It is also low HaLow devices, Guda says, can run longer on a single
power, and is geared toward providing connectivity battery charge. He says Morse Micro is working with
for Internet of Things (IoT) applications. partners building battery-powered HaLow IoT
To demonstrate the technology’s potential, devices that will need to operate for years.
Morse Micro recently conducted a seaside test in But IoT connectivity is a crowded space, says
San Francisco’s Ocean Beach neighborhood. The Ermanno Pietrosemoli, a scientific consultant at the
company showed that two tablets connected over a International Centre for Theoretical Physics, in
HaLow network could communicate at distances of ­Trieste, Italy, who specializes in wireless networks.
up to 3 km while maintaining speeds of around “[HaLow] is a viable competitor, but there are many
1 megabit per second—enough to support a slightly players in this field,” he says.
grainy video call. The technology has a considerable range advan-
“It is pretty unprecedented range,” says Prakash tage over IoT-specific alternatives like ZigBee or
Guda, vice president of marketing and product man- Bluetooth. And while the LoRa standard is capable
agement at Morse Micro. “And it’s not just the ability of distances of tens of kilometers, its throughput is
to send pings but actual megabits of data.” too low to enable video or even real-time voice, says
MORSE MICRO

The HaLow protocol works in much the same Pietrosemoli. HaLow’s high data rates could be
way as conventional Wi-Fi, says Guda, apart from useful for more data-heavy applications, like wire-
the fact that it operates in the 900-megahertz fre- lessly connecting lots of security cameras, he adds.

10 SPECTRUM.IEEE.ORG APRIL 2024


JOURNAL WATCH

Lights Out, Lessons Learned

Vehicles drive
down Bourbon Street
in New Orleans on
29 August 2021,
during a citywide
power outage caused
by Hurricane Ida.

NEW RESEARCH PUBLISHED in December indicates


that power inverters may not be as grid destabilizing as
is sometimes alleged. These infrastructure components
are used to convert solar and wind-energy systems’ DC
output to the grid’s alternating current. Most blackouts,
the authors found, are caused by extreme weather—
Despite the HaLow standard being approved and not by renewables and their inverters.
published as of 2017, the technology has very low The researchers, from the Karlsruhe Institute of
adoption, says Pietrosemoli, which might give devel- Technology (KIT) in Karlsruhe, Germany, also examined
opers pause due to concerns around whether the whether inverter-based resources might counteract
technology will continue to be supported. In con- blackout mechanisms such as cascading failures that
trast, competing technologies like SigFox, which have plagued grids for decades. The group published their
operates on principles similar to those of HaLow, findings in the IEEE Open Journal of Power Electronics.
and Narrowband IoT (NB-IoT)—a low-power cellu- “In the scientific literature, you often read that
lar technology—are already widely deployed. renewables will cause some sort of problems, and so
Bill Ray, a VP analyst at Gartner, is even more we need to work on their integration. I wanted to know
circumspect. While Morse Micro has built some whether it was real,” says Jan Wachter, a doctoral student
impressive technology, he says HaLow has failed to and research associate at KIT’s Institute for Automation
gain any significant traction in the industry. The fact and Applied Informatics.
that Qualcomm, which was an early HaLow booster, The primary difference between the power plants and
has lost interest is indicative, he adds. electric loads of the past century and those interfaced via
“We are not optimistic about the standard, even inverters is how intimately they couple to the grid. Most
if it can be stretched to 3 km,” says Ray. “Bluetooth rotating generators and machines are physically coupled
offers greater interoperability, while LoRa offers to the grid’s electromagnetic oscillations. As a result, they
greater range, putting HaLow in a middle position react quickly to stabilize the grid during disturbances.
with limited application.” That physical response gives the grid a certain inertia,
Guda puts a different spin on it. By combining buying time for generators to adjust their output and
high data rates and long ranges, he says, the tech- thus stabilize the system, Wachter says. Inverter-based
nology can serve the whole gamut of IoT applica- resources, in contrast, must react to faults through a process
tions. “It’s the best of both worlds,” he says. of signal detection and control—a slower process that relies
on the timeliness and quality of observational data feeding
LUKE SHARRETT/BLOOMBERG/GETTY IMAGES

He acknowledges that the lack of adoption does


raise concerns about interoperability and over­ into control systems and the wisdom of their programming.
reliance on a single vendor, but he says things are Wachter says the complexity of the grid and its possible
changing. California-based Newracom now also fault mechanisms will grow as the number and diversity
produces a HaLow SoC, and in 2021 the Wi-Fi Alli- of grid-connected devices mushroom. But that, he says,
ance launched a certification program for the stan- provides new opportunities to maintain reliability into
dard that Guda says will improve interoperability. the future.
“If you’re trying to make some equipment based “If we’re rebuilding our grids, it provides an opportunity
on a new technology, you want to make sure that to make things better. We’re spending a lot of money. We
there’s interoperability and there’s more than one should take the chance to do this,” says Wachter.
supplier,” he says. “That’s the critical mass that we —Peter Fairley
have now.”

APRIL 2024 SPECTRUM.IEEE.ORG 11


NEWS

COMPUTING

­ iobium titanium nitride wire was cooled


n
Infrared Tech Sees Around to about 2 kelvins (about -271 °C), rendering
the wire superconductive. A single photon

Corners These eye-safe could disrupt this fragile state, generating


electrical pulses that enabled the efficient

lasers could make self-driving detection of individual photons.


The scientists contorted the nanowire

cars safer
in each device into a fractal pattern that
took on similar shapes at various magnifi-
cations. This let the sensor detect photons
of a wide range of polarizations, boosting
BY CHARLES Q. CHOI
its efficiency.
The new detector was as much as three
times as efficient as other single-photon

J
ust because an object is around a corner doesn’t mean it has to be detectors at sensing near- and midinfra-
hidden. Non-line-of-sight imaging can peek around corners and spot red light. This let the researchers perform
those objects, but it has been limited so far to a narrow band of fre- non-line-of-sight imaging, achieving a
quencies. Now, a new sensor can help extend this technique from work- spatial resolution of between 1.3 and
ing with visible light to infrared. This advance could help make autonomous 1.6 centimeters.
vehicles safer, among other potential applications. In addition to an algorithm that recon-
Non-line-of-sight imaging relies on reconstructing images from the faint structed non-line-of-sight images based on
signals of light beams that have reflected off surfaces. The ability to see around multiple scattered light rays, the scientists
corners may prove useful for machine vision—for instance, helping autono- developed a new algorithm that helped
mous vehicles foresee hidden dangers, says Xiaolong Hu, the senior author of remove noise from their data. When each
the study and a professor at Tianjin University, in China. It may also improve pixel during the scanning process was given
endoscopes that help doctors peer inside the body. 5 milliseconds to collect photons, the new
The light that non-line-of-sight imaging depends on is not very bright. denoising algorithm reduced the root mean
Until now, detectors suited to the non-line-of-sight imaging task operated square error—a measure of its deviation
only in visible or near-infrared wavelengths. These portions of the spectrum, from a perfect image—of reconstructed
while somewhat useful for the above applications, nevertheless also face the images by a factor of one-eighth.
challenges, Hu says, of interference with daylight as well as laser wavelengths The researchers now plan to arrange
that could pose risks to the eyes. multiple sensors into larger arrays to
So now Hu and his colleagues have for the first time performed non-line- boost efficiency, reduce scanning time, and
of-sight imaging using 1,560- and 1,997-nanometer wavelengths—putting it extend the distance over which imaging
in what’s called short-wavelength infrared, the next wavelength subdivision can take place, Hu says. They would also
beyond near infrared. “This extension in spectrum paves the way for more like to test their device in daylight condi-
practical applications,” Hu says. tions, he adds.
In the new study, the researchers experimented with superconducting The scientists detailed their findings in
nanowire single-photon detectors. In each device, a 40-nanometer-wide November in the journal Optics Express.

12 SPECTRUM.IEEE.ORG APRIL 2024 Illustration by Greg Mably


ENERGY

Fly a Kite for Portable Power


Wind power, for all its promise, is hardly portable. Yet, the Netherlands-based
company Kitepower aims to convert breezes and gales to electrons using
technology that occupies only a shipping-container-size footprint. The catch
is that its “turbine” is actually a kite—although the 60-square-meter area
perhaps argues for a heftier word. “Winched windfoil” may be closer to the mark.
KITEPOWER

Kitepower says it aims to ship its first 40-kilowatt Hawk systems [whose kite is
pictured] this year.

APRIL 2024 SPECTRUM.IEEE.ORG 13


THE BIG PICTURE

Mechanical
Megatalent
By Willie D. Jones

In late 2023, research-


ers at Georgia Tech
in Atlanta unveiled a
robotic performance
sculpture they named
Medusai. It is modeled
after Medusa, the
mythological Gorgon who
turned anyone ​
who
looked at her to stone.
The bot has snakelike
arms that allow it to
simultaneously play
percussion and strings
and interactively
“dance” with observ-
ers.“The eerie and
uncanny notion of
AI-driven robotic
snakes following and
threatening humans
is balanced by the
sculpture’s potential
to inspire humans to
push the boundaries of
their creativity and
expression,” explains
Gil Weinberg, the
professor at Georgia
Tech’s Center for Music
Technology who heads
the team that
created Medusai.
PHOTOGRAPH BY
GIOCONDA BARRAL-SECCHI

14 SPECTRUM.IEEE.ORG APRIL 2024


APRIL 2024 SPECTRUM.IEEE.ORG 15
TECH TO TINKER WITH

You can add


new features or
even upload your
own boot screen
to the UV-K5
radio with just
a few clicks.

The Most Hackable Handheld Ham


Radio Yet The UV-K5 can be modded
at the click of a mouse
BY STEPHEN CASS

16 SPECTRUM.IEEE.ORG APRIL 2024 Illustrations by James Provost


APRIL 2024

A
ll right, confession time. I don’t Quansheng probably thought of its
use my handheld ham radio for design purely in terms of fixing software
much more than eavesdrop- bugs or adjusting for regulatory
ping on the subway dispatcher changes—it offers a free install tool for
when my train rumbles to a mysterious uploading official firmware releases to
halt in a dark tunnel. But even I couldn’t the radio. But the prospect of an
help but hear the buzz surrounding a new ­updatable radio dangled an irresistible
handheld, Quansheng’s UV-K5. temptation for folks to start reverse engi-
It caught my attention in part because neering the firmware and hardware so
for over a decade, Baofeng has been the they could try writing their own code.
name in Chinese handhelds. In 2012 Modifications to date have generally
Baofeng made waves with its UV-5R taken the form of patches to the official
radio, upending the sleepy hand- firmware, rather than wholesale rewrites.
held-transceiver market. Prior to the 5R, With the official firmware taking up most
the price tag of the cheapest VHF/UHF of the radio’s 64 kilobytes of flash
handheld was a little north of US $100. memory, such mods have to fit into less
The 5R sold for a quarter to a third of that. than 3 KB. And the CPU is not brimming
Hams groused about the 5R’s so-so tech- with compute power—it’s a 48-mega-
nical performance—and then bought a hertz, 32-bit ARM-based processor with
couple anyway, so they’d always have a 8 KB of RAM. Nonetheless, I found the
radio in their car or workplace. results impressive.
Now it’s Quansheng that’s making a For example, one mod installs a fairly
splash. The UV-K5, released last year, sophisticated graphical spectrum ana-
might be the most hackable handheld lyzer: You can adjust the bandwidth, set
ever, with a small army of dedicated Updatable a threshold for tuning into detected
hams adding a raft of software-based
improvements and new features. I had to firmware dangled peaks automatically, and specify fre-
quencies to ignore, among other things.
have one, and $30 later, I did.
Like Baofeng’s 5R, Quansheng’s K5
an irresistible Another mod allows you to exchange text
messages between K5s. Other mods
as a radio transceiver is fine. (I’m using temptation improve the K5’s ability to receive AM
K5 here to refer to both the original K5
and the new K5(8) model.) The key tech- for folks to signals, meaning you can, say, listen in
on aviation bands more clearly. And
nical distinction between the 5R and K5
is a seemingly minor design choice. With
start reverse there are plenty of fun little mods that do
things like change up the system fonts or
Baofeng’s 5R, the firmware resides in engineering… replace the start-up message with a line-
read-only memory. But Quansheng art image of your choice.
stores the K5’s firmware in flash memory Installing many of these mods is ridic-
and made it possible to rewrite that ulously easy. Normally at this point in a
memory with the same USB program- Hands On article that involves hacking
ming cable used to assign frequencies to some consumer electronics, things get
preset channels. pretty heroic as I futz with the hardware
This feature has opened the door for or unravel a software-installation enigma.
improvements to the K5 that are well But not this time.
beyond what Quansheng offers out of the A modder known as whosmatt pro-
box. Hopefully, this design will inspire vides a Web-based patcher/flasher for
other radio makers to offer more support the K5 that lets you pick a selection of
for modders, in turn bringing more inno- mods from a menu. It then combines
vation to the VHF and UHF radio bands. them with the official firmware to create

APRIL 2024 SPECTRUM.IEEE.ORG 17


HANDS ON

Reception amplifiers and filters


68–108 MHz
FM receiver Antenna
Transmission amplifiers and filters

Audio BK4819 DP32060 EEPROM


amplifier RF IC CPU

This block diagram of the


128x64 UV-K5 is based on the work
Keypad of Phil McAllen. Hams have
LCD reverse engineered many
details of the radio’s
hardware and software.
Speaker Microphone

a custom image for uploading (as long Quansheng firmware blocks transmitting
as you don’t exceed the total amount of on the aviation band, to prevent illegal and
memory). hazardous interference. But this block can
In fact, if you’re using Chrome, Edge, be removed by a patch (although to be a
or Opera, you don’t even need to use significant threat, you’d likely need an
Quansheng’s installer to upload the amplifier to boost the K5’s 5-watt signal).
custom firmware: You can update the However, hams have always had the
radio’s flash memory directly from the ability to behave badly, with or without
browser via the built-in Web Serial API firmware blocks. Such blocks are conve-
and the USB programming cable. (The nient for guarding against accidental
instructions say this will work only on abuse, but the truth is that unless prob-
Linux and Windows, but I was able to do lematic signals are persistent enough to
it using a Mac as well.) Web Serial could allow a transmitter’s location to be tri-
do with some improved error handling, angulated, amateur radio must continue
though. The first USB programming cable to rely on an honor system, whether that
I used was a bit flaky, but where Quan- means not jamming a neighbor’s TV or
sheng’s installer would halt and flag a transmitting on forbidden frequencies.
communications error with a failed Many of the most exciting uses of
upload, Web Serial would silently crash ham radio today involve digital process-
and take the whole Windows operating ing, and that processing is normally done
system with it. using a computer connected to a trans-
There are even more K5 mods avail- ceiver. With embedded controllers
able than are in whosmatt’s online becoming ever more powerful, the K5
patcher. If you want to play with those or modding scene points toward a future
start writing your own mods, Python- where more processing happens ­in-radio
based toolchains exist to assist you. and where you can add new functions
Of course, allowing unfettered mod- the way apps are added to smartphones.
ding of the K5’s transceiver does raise the Here’s hoping manufacturers embrace
possibility of abuse. For example, the that future! 

18 SPECTRUM.IEEE.ORG APRIL 2024


Q&A BY MICHAEL KOZIOL

quality. Beyond that, ABET has very well-thought-


out criteria for what defines excellence and leaves
each individual program the freedom to define the
learning objectives, the tools, the assessment, and
how to continuously improve. It’s a set of very-well-
founded principles that we would support, even in
the absence of this certification.

What is the first course you’re offering?


Fraden: We’re doing an introduction to design. It’s a

5 Questions
course in which the students develop prosthetics for
both animal and human use. It’s open to all students
at Brandeis, but it’s still quite substantive: They’re
working in Python, they’re working with CAD pro-
grams, and they’re working on substantive projects

for Seth Fraden


using open-source designs. The idea is to get stu-
dents excited about engineering, but also to have
them learn the fundamentals of ideation—going
from planning to design to fabrication, and then this
How to boot up a new engineering program will help them decide whether or not engineering is
the major for them.

How do you see liberal arts such as history and

S
tarting a new engineering program at ethics being part of engineering?
a university is no simple task. But that’s Fraden: Many of our students want to intervene
just what Brandeis University in Waltham, in the world and transform it into a better place. If
Mass., is doing. By 2026, the university you solely focus on the production of the technol-
will offer an undergraduate engineering degree— ogy, you’re incapable of achieving that objective.
but without creating an engineering department. You need to know the impact of that technology on
Instead, Brandeis aims to lean on its strong liberal society. How is this thing going to be produced? Who
arts tradition, in hope of offering something dif- says what labor is going to go into manufacturing?
ferent from the more than 3,500 other engineering What’s its life cycle? How’s it going to be disposed
programs in the United States accredited by the of? You need to have a full-throttled liberal arts edu-
Accreditation Board for Engineering and Technol- cation to understand the environmental, ecological,
ogy (ABET). IEEE Spectrum spoke with Seth Fraden, economical, and historical consequences of your
one of the new program’s interim cochairs, about intervention as a technologist.
getting a new engineering program up and running.
How will you develop an engineering culture?
What prompted offering an engineering degree? Fraden: We’re not going to have a department. It will
Seth Fraden: We saw that we had 90 percent of all be the only R1 [top-tier research institution] engi-
the elements that are necessary for a vibrant engi- neering major without a department. We see that
neering program—the basic sciences, math, phys- as a strength, not a weakness. We’re going to embed
ics, computer science, life science, all put in a social new engineering faculty throughout all our sciences,
context through the liberal arts. We see our new in order to have a positive influence on the depart-
program as a way of bridging science and society ments and to promote technology development.
through technology, and it seems like a natural fit for Seth Fraden is a That said, we want there to be a strong engi-
us without having to build everything from scratch. professor of physics neering culture, and we want the students to have
at Brandeis a distinctive engineering identity, something that
Brandeis’s engineering degree will be accredited University. He is a scientist like myself—though I am enthusiastic
serving as one
by ABET. Why is that important? about engineering—doesn’t have in my bones. In
of the two interim
Fraden: Being the new kids on the block in engi- cochairs for the order to do that, our instructors will each come from
neering, it’s natural to want to reassure the com- university’s new an engineering background, and will work together
munity at large that we’re committed to outstanding engineering degree. to build a culture of engineering.

Photo-illustration by Stuart Bradford APRIL 2024 SPECTRUM.IEEE.ORG 19


SHARING THE EXPERIENCES OF WORKING ENGINEERS

Careers:
Zach Rattner
His startup’s AI tool makes moving day easier

E
ngineers are used to being experts in their
field, but when Zach Rattner cofounded
his artificial-intelligence startup, Yembo,
he quickly realized he needed to get com-
fortable with being out of his depth. He found the
transition from employee to business owner to be
a steep learning curve. Taking on a host of unfamil-
iar responsibilities like finance and sales required a
significant shift in mind-set.
Rattner cofounded Yembo in 2016 to develop being able to perform three to five times as many
an AI-based tool for moving companies that cre- “You just surveys per day with the same headcount,” he says.
ates an inventory of objects in a home by analyzing need to get “If you compare us to an in-home visit, the savings
video taken with a smartphone. Today, the startup are even more since Yembo doesn’t have drive time.”
comfortable
employs 70 people worldwide and operates in 36
countries, and Rattner says he’s excited to get out of
being In 2016, he quit his job to become a consultant and
bed every morning because he’s building a product horrible at work on his startup idea in his spare time. A few
that simply wouldn’t exist otherwise. some things.” months later, he decided the idea had potential, and
“I’m making a dent in the universe,” he says. he convinced a former Qualcomm colleague, Sid-
“We are bringing about change. We are going into dharth Mohan, to join him in cofounding Yembo.
an industry and improving it.” Rattner admits the responsibilities that come
with starting a new business took some getting used
Rattner has his wife to thank for his startup idea. to. In the early days, you’re not only building the
From 2011 to 2015, she worked for a moving com- technology, he says, you also have to get involved in

JUAN CASIANO/LUMETRY MEDIA; ILLUSTRATION: RITA SUS/EMIT LIGHT


pany, and she sometimes told him about the chal- marketing, finance, sales, and a host of other areas
lenges facing the industry. A major headache for these you have little experience in. (For more on the chal-
companies, he says, is the time-consuming task of lenges of startup life, see “The Messy Reality Behind
taking a manual inventory of everything to be moved. a Silicon Valley Unicorn,” page 44.)
At the time, he was a software engineer in Qual- “If you try to become that rigorous expert at
comm’s internal incubator in San Diego, where everything, it can be crippling, because you don’t have
employees’ innovative ideas are turned into new enough time in the day,” Rattner says. “You just need
products. In that role, he got a lot of hands-on expe- to get comfortable being horrible at some things.”
rience with AI and computer vision, and he realized As the company has grown, Rattner has become
that object-detection algorithms could be used to less hands-on, but he still gets involved in all aspects
automatically catalog items in a house. of the business and is prepared to tackle the most
Rattner reports that his clients are able to com- challenging problems on any front.
plete three times as many daily inspections as they In 2020, the company branched out, developing
would using traditional methods. Also, his custom- a tool for property insurers by adapting the original
ers have increased their chances of getting jobs by AI algorithms to provide the information needed
27 percent because they’re able to get quotes out for an accurate insurance quote. Along with cata-
faster than the competition, often in the same day. loging the contents of a home, this version of the
“Comparing Yembo’s survey to a virtual option AI tool extracts information about the house itself,
like Zoom or FaceTime, our clients have reported including a high-fidelity 3D model that can be used

20 SPECTRUM.IEEE.ORG APRIL 2024


BY EDD GENT

Eugxe H. SpIta
dolorru ptatus
erias ipsa dolut
explam volo beris
restionestis utae
eaturiam, non cus some experience at a major tech company before
magnam expernatia
denimuscilis
striking out on his own.
mos ex ea In 2010, the summer before his senior year, he
nimporeperum interned at Qualcomm. As 4G technology was just
enition secatem rolling out, the company was growing rapidly and
elit audit offered Rattner a full-time job. He joined in 2011
after earning his bachelor’s degree in computer
engineering.
Rattner started out at Qualcomm as a modem
software engineer, working on technology that
measured cellphone signal strength and searched
for the best cell connections. He took algorithms
designed by others and used his coding skills to
squeeze them onto the meager hardware available
on cellphones of the era.
Rattner says the scale of Qualcomm’s operations
forced him to develop a rigorous approach to engi-
neering quality. “If you ship code on something that
has a billion installs a year and there’s a bug, it will
be found,” he says.
Eventually, he decided there was more to life
to take measurements virtually. The software can than signal bars and began looking for new career
also be used to assess damage when a homeowner opportunities. That’s when he discovered Qual-
makes a claim. comm’s internal incubator. After having one of his
“It feels like it’s a brand-new startup again,” ideas accepted and following the project through to
Rattner says. completion, Rattner accepted a job to help to manage
the program. “I got as close as I could to running a
From a young age, Rattner had an entrepreneur- startup inside a big company,” he says.
ial streak. As a 7-year-old, he created a website to
display his stamp collection. By his teens, he was Rattner wrote a book about his journey as a startup
freelancing as a Web developer. founder called Grow Up Fast, which he self-published
“I had this strange moment where I had to con- last year. In it, he offers a few tips for those looking
fess to my parents that I had a side job online,” he to follow in his footsteps.
Employer:
says. “I told them I had a couple of hundred dollars Yembo
He suggests developing concrete skills and
I needed to deposit into their bank account. They obtaining experience before trying to make it on your
weren’t annoyed; they were impressed.” Occupation: own. One way to do this is to get a job at a big tech
When he entered Virginia Tech in 2007 to study Chief technology company, he says, since they tend to have a wealth of
officer and
computer engineering, he discovered his roommate experienced employees you can learn from.
cofounder
had also been doing freelance Web development. It’s crucial to lean on others, he says. Joining
Together they came up with an idea for a tool that Education: startup communities can be a good way to meet
would allow people to build websites without writ- Bachelor’s degree people in a similar situation whom you can turn to
ing code. in computer for advice when you hit roadblocks. And the best way
engineering,
They were accepted into a startup incubator to Virginia Tech to master the parts of the job that don’t come natu-
further develop their idea. But acceptance came with rally to you is to seek out those who excel at them.
an offer of only US $15,000 for funding and the stipu- “There’s a lot you can learn from just observing,
lation that they had to drop out of college. As he was studying, and asking questions of others,” he says.
writing the startup’s business plan, Rattner realized Most important, Rattner advises, is to simply
that his idea wasn’t financially sustainable long term learn by doing.
and turned the offer down. “You can’t think of running a business as if you’re
“That is where I learned there’s more to running at school, where you study, practice, and eventually
a startup than just the technology,” he says. get good at it, because you’re going to be thrown
This experience reinforced his conviction that into situations that are completely unforeseen,” he
betting everything on one great business idea wasn’t says. “It’s about being willing to put yourself out
a smart move. He decided to finish school and get there and take that first step.”

APRIL 2024 SPECTRUM.IEEE.ORG 21


IDEAS FOR CODE: LEAN SOFTWARE

Why Bloat
Is Still
Software’s
Biggest
Vulnerability
APRIL 2024 SPECTRUM.IEEE.ORG 23

A 2024
plea for
lean
software

By Bert Hubert
Illustration by
Daniel Zender
S
IDEAS FOR CODE: LEAN SOFTWARE

The terrible state of software security


Without going all “Old man (48) yells at cloud,” let
me restate some obvious things. The state of soft-
ware security is dire. If we only look at the past year,
if you ran industry-standard software like Ivanti,
MOVEit, Outlook, Confluence, Barracuda Email
Security Gateway, and ­NetScaler Gateway, chances
are you got hacked. Even companies with near-in-
finite resources (like Apple and Google) made trivial
“worst practice” security mistakes that put their
customers in danger. Yet we continue to rely on all
these products.
Software is now (rightfully) considered so dan-
gerous that we tell everyone not to run it them-
S OME YEARS AGO I DID A TALK at a local selves. Instead, you are supposed to leave that to
university on cybersecurity, titled “Cyber and Infor- an “X as a service” provider, or perhaps just to “the
mation Security: Have We All Gone Mad?” It is still cloud.” Compare this to a hypothetical situation
worth reading today since we have gone quite mad where cars are so likely to catch fire that the advice
collectively. is not to drive a car yourself, but to leave that to
The way we build and ship software these days professionals who are always accompanied by pro-
is mostly ridiculous, leading to apps using millions fessional firefighters.
of lines of code to open a garage door, and other The assumption is then that the cloud is some-
simple programs importing 1,600 external code how able to make insecure software trustworthy.
libraries—dependencies—of unknown provenance. Yet in the past year, we’ve learned that Microsoft’s
Software security is dire, which is a function both email platform was thoroughly hacked, including
of the quality of the code and the sheer amount of classified government email. (Twice!) There are also
it. Many programmers know the current situation well-founded worries about the security of the
is untenable. Many programmers (and their man- Azure cloud. Meanwhile, industry darling Okta,
agement) sadly haven’t ever experienced anything which provides cloud-based software that enables
else. And those programmers who do know better user log-in to various applications, got comprehen-
rarely get the time to do a better job. sively owned. This was their second breach within
Let me briefly go over the terrible state of soft- two years. Also, there was a suspicious spate of Okta
ware security, and then spend some time on why it users subsequently getting hacked.
is so bad. I then talk about an actual useful piece of Clearly, we need better software.
software I wrote as a proof of concept that one can
still make minimal and simple yet modern software. Why software security is so bad
Finally, I mention some regulatory and legislative The situation today is clearly working well for com-
things going on that we should use to make software mercial operators. Making more secure software
quality a priority again. takes time and is a lot of work, and the current secu-
I hope that this article provides some mental and rity incidents don’t appear to be impacting the
moral support for suffering programmers and tech- bottom line or stock prices. You can speed up time
nologists who want to improve things. It is not just to market by cutting corners. So from an economic
you; we are not merely suffering from nostalgia. standpoint, what we see is entirely predictable. Leg-
Software really is very weird today. islation could be very important in changing this
equation, and I’ll come back to that later.

200
The security of software depends on two fac-
tors—the density of security issues in the source
code and the sheer amount of code accessible by
This article is dedicated to the memory of Niklaus Wirth, hackers. As the U.S. defense community loved to
a computing pioneer who passed away 1 January 2024.
KILOBYTES point out in the 1980s, quantity has a quality all of
In 1995 he wrote an influential article called “A Plea for Size of Nicholas its own. The reverse applies to software—the more
Lean Software,” published in Computer, the magazine for Wirth’s Oberon you have of it, the more risks you run.
members of the IEEE Computer Society, which I read operating system As a case in point, Apple iPhone users got repeat-
early in my career as an entrepreneur and software in 1995. Today, edly hacked over many years because of the huge
many projects
developer. In what follows, I try to make the same case attack surface exposed by iMessage. It is possible
have more than
nearly 30 years later, updated for today’s computing 200 KB for their to send an unsolicited iMessage to an Apple user.
horrors. A version of this article was originally published configuration The phone will then immediately process that mes-
on my personal blog, Berthub.eu. files alone. sage so it can preview it. The problem is that Apple

24 SPECTRUM.IEEE.ORG APRIL 2024


in its wisdom decided that such unsolicited mes- Perhaps more. The app, meanwhile, likely pulls in
sages needed to support a vast array of image for- hundreds or thousands of helper packages. Many
mats, accidentally including PDFs with weird packages used will also, by default, snitch on your
embedded, compressed fonts using an ancient users to advertisers and other data brokers. Depen-
format that effectively included a programming lan- dencies pull in further dependencies, and exactly
guage. So someone could send an unsolicited mes- what gets included in the build can change on a daily
sage to your iPhone that could probe for weaknesses basis, and no one really knows.
in the rest of the phone. If this app controls anything in your house, it will
In this way, attackers were able to benefit from also connect to a software stack over at Amazon,
security bugs in the phone’s millions of lines of code. probably also powered by Node.js, also pulling in
You don’t need a high bug density to find an exploit- many dependencies.
able hole in millions of lines of code. Apple could But wait, there’s more. We used to ship software
have prevented this situation by restricting previews as the output of a compiler, or perhaps as a bunch of
to a far smaller range of image formats, or even a files to be interpreted. Such software then had to be
single “known good” image format. installed and configured to work right. Getting your
Apple is (by far) not the worst offender in this code packaged to ship like this is a lot of work. But it
field. But it is a widely respected and well-resourced was good work since it forced people to think about
company that usually thinks through what they do. what was in their “package.” This software package
And even they got it wrong by needlessly shipping would then integrate with an operating system and
and exposing too much code. It is not just with local services, based on the configuration.
you; we are Since the software ran on a different computer
Could we not write better code? than the one it was developed on, people really had
There are those who think the biggest problem is
not merely to know what they shipped and think it through. And
the quality of the code, expressed in terms of the
suffering sometimes it didn’t work, leading to the joke where
density of bugs in it. There are many interesting from a developer tells the operations people, “Well, it works
things happening on this front, like the use of nostalgia. on my system,” and the retort “Then back up your
­memory-safe languages like Rust. Other languages Software email, we’re taking your laptop into production!”
are also upping their security game. Fuzzers—test really is very This used to be a joke, but these days we often
tools that automatically modify inputs to computer weird today. ship software as containers, shipping not only the
programs to find weaknesses and bugs—are also software itself but also including operating-system
getting ever more advanced. files to make sure the software runs in a well-known
But many security problems are in the logic environment. This frequently entails effectively
underlying the code. For example, the Barracuda shipping a complete computer disk image. This
email exploit originated in a third-party library that again vastly expands the amount of code being
would actually execute code in Excel spreadsheets deployed. Note that you can do good things with
when they were scanned for viruses. Wiping out all containers like Docker (see below), but there are a
the bugs in your code won’t save you from the deci- lot of images over 350 MB on the Docker Hub.
sion to implement a feature to automatically execute Add it all up and we are likely looking at over 50
code embedded in documents. million active lines of code to open a garage door,
Another problem is that we often don’t know running several operating-system images on mul-
what code we are actually shipping. Software has tiple servers. At least if it’s using Electron and
gotten huge. In 1995 Niklaus Wirth lamented that Docker, which is common industry practice.
software had grown to megabytes in size. In his arti- Now, even if all the included dependencies are
cle “A Plea for Lean Software,” he went on to golden, are we sure that their security updates are
describe his Oberon operating system, which was making it to your garage-door opener app? I
only 200 kilobytes, including an editor and a com- wonder how many Electron apps are still shipping
piler. There are now projects that have more than with the image processing bug that had Google

50
200 KB for their configuration files alone. and Apple scramble to put out updates last year.
A typical app today is built on Electron, a frame- We don’t even know.
work that incorporates both Chromium (“Chrome”) But even worse, it is a known fact that all these
and Node.js, which provides access to tens of thou- MILLION dependencies are not golden. The Node.js ecosystem
sands of software packages for ­JavaScript. Chro- has a comical history of package repositories being
mium is 41 million lines of code by itself, but it in Approximate taken over, hijacked, or resurrected under the same
turn relies on dependencies. Meanwhile, even a number of lines name by someone else, someone with nefarious plans
of code in any
rather trivial Node.js app pulls in 4 million lines of for your security. PyPI (a Python counterpart of
app written
dependencies. Somehow. with the popular Node.js) has suffered from similar problems.
All told, just using Electron entails at least 50 Electron Dependencies always need scrutiny, but no one
million lines of code if you include dependencies. framework. can reasonably be expected to CONTINUED ON P. 50

APRIL 2024 SPECTRUM.IEEE.ORG 25


IDEAS FOR CODE: LEAN SOFTWARE

We Need to
Decarbonize
Software
By Rina Diane Caballar • Illustrations by Elias Stein

The way we
write software
has unappreciated
environmental
impacts

S
S OF T WA R E M AY B E E AT I NG T H E WOR L D, but it is also heating it. • In
December 2023, representatives from nearly 200 countries gathered in Dubai for COP28,
the U.N.’s climate-change conference, to discuss the urgent need to lower emissions.
Meanwhile, COP28’s website produced 3.69 grams of carbon dioxide (CO2) per page
load, according to the website sustainability scoring tool Ecograder. That appears to be
a tiny amount, but if the site gets 10,000 views each month for a year, its emissions would
be a little over that of a one-way flight from San Francisco to Toronto.
This was not inevitable. Based on Ecograder’s analysis, unused code, improperly sized
images, and third-party scripts, among other things, affect the COP28 website’s emissions.
These all factor into the energy used for data transfer, loading, and processing, consuming
a lot of power on users’ devices. Fixing and optimizing these things could chop a whopping
93 percent from the website’s per-page-load emissions, Ecograder notes.
While software on its own doesn’t release any emissions, it runs on hardware in data
centers and steers data through transmission networks, which account for about 1 percent
of energy-related greenhouse gas emissions each. The information and communications
technology sector as a whole is responsible for an estimated 2 to 4 percent of global green-
house gas emissions. By 2040, that number could reach 14 percent—almost as much
carbon as that emitted by air, land, and sea transport combined.
Within the sphere of software, artificial intelligence has its own sustainability
issues. AI company Hugging Face estimated the carbon footprint of its BLOOM large
language model across its entire life cycle, from equipment manufacturing to deploy-

26 SPECTRUM.IEEE.ORG APRIL 2024


public class BasicClientTest {
public static void main(String [] args)
{
ThriftClient client = null;
long ns = -1;
try {
client = ThriftClient.create("lo-
calhost", 38080);
ns = client.namespace_open("test");
// HQL examples
show(client.hql_query(ns, "show
tables").toString());
IDEAS FOR CODE: LEAN SOFTWARE

ment. The company found that BLOOM’s final organizations, including tech giants Google, Intel,
training emitted 50 tonnes of CO2—equivalent to and Microsoft. But the sector will have to embrace
about a dozen flights from New York City to Sydney. these practices even more broadly if they are to pre-
Green software engineering is an emerging dis- vent worsening emissions from developing and
cipline consisting of best practices to build applica- using software.
tions that reduce carbon emissions. The green
software movement is fast gaining momentum. What Is Green Software Engineering?
Companies like Salesforce have launched their own The path to green software began more than 10
software sustainability initiatives, while the Green years ago. The Sustainable Web Design Community
Software Foundation now comprises 64 member Group of the World Wide Web Consortium (W3C)

Tips for Greener AI


1. Train on less data.
Cutting back on
training data may
5. Tolerate some drift.
Once deployed, an AI
model can experience model
drastically reduce energy drift—a decrease in accuracy
costs while keeping the due to the difference between
accuracy of the final model training data and the live data
almost unchanged. used in production. It might
Machine-learning
model be tempting to retrain the
model each day to avoid this
drift, but try a less energy-
and resource-intensive
approach: setting a metric
2. Choose your model
with energy use in
mind. Some models are more
that determines when a model
is accurate enough and
retraining it only when it’s
energy intensive than others,
below that level.
not necessarily resulting in
greater accuracy. Tools like
CodeCarbon and ML CO2
Impact can help estimate
the energy usage and
carbon footprint of training
AI models. 6. Choose your hardware
wisely. Choosing
powerful multicore
processors that enable
parallel computing or
hardware accelerators like

3. Simplify model GPUs or tensor-processing


architecture. This can units optimized for AI
be done by removing workloads will go a long way
parameters, pruning the toward energy efficiency.
number of layers, and picking
the best hyper­parameters—
external model controls—via
a tuning strategy that uses
the least number of iterations.

Learning system
task 1

4. Employ transfer learning. Transfer learning reuses


pretrained models to save on compute time and power.
The algorithms for detecting SMS spam messages, for
instance, can be used for email spam messages as well.
AI engineers can also apply existing best practices, such as
early stopping, to end training once a model’s performance
on a validation dataset starts to degrade. Learning system
task 2

28 SPECTRUM.IEEE.ORG APRIL 2024


was established in 2013, while the Green Web Yet other developers may be motivated by the cli-
Foundation began in 2006 as a way to understand mate crisis itself, wanting to play their part in fostering
the kinds of energy that power the Internet. Now, a habitable planet for the coming generations. And
the Green Web Foundation is working toward the software engineers have tremendous influence on the
ambitious goal of a fossil-free Internet by 2030. actual purpose and emissions of what they build.
“There’s an already existing large segment of the “It’s not just lines of code. Those lines have an
software-development ecosystem that cares about impact on human beings,” says June Sallou, a post-
this space—they just haven’t known what to do,” doctoral researcher specializing in sustainable-soft-
says Asim Hussain, chairperson and executive direc- ware engineering at the Delft University of Technology,
tor of the Green Software Foundation and former in the Netherlands. Because of AI’s societal impact in
director of green software and eco-
systems at Intel.
What to do, according to Hussain,
falls under three main pillars: energy
efficiency, or using less energy; hard- Tips for Greener 5. Select sustainable
Web hosting or

Websites
ware efficiency, or using fewer phys- cloud-computing provid-
ical resources; and carbon-aware ers. The Green Web
Foundation has a tool to

and Apps
computing, or using energy more
intelligently. Carbon-aware comput- check if your website runs
on green energy, as well as
ing, Hussain adds, is about doing a directory of hosting
more with your applications during providers powered by
the periods when the electricity renewable energy.
comes from clean or low-carbon
sources—such as when wind and 1. Optimize video
and image sizes.
Use low-resolution
solar power are available—and doing images and modern
less when it doesn’t. image formats, and
load animations only
The Case for when a user scrolls
Sustainable Software them into view.
So why should programmers care
about making their software sus-
tainable? For one, green software is
efficient software, allowing coders
2. Minimize
­JavaScript.
JavaScript runs on
to cultivate faster, higher-quality every device that loads
the website, often
systems, says Kaspar Kinsiveer, a
inefficiently.
team lead and sustainable-software
strategist at the software-develop-
ment firm Helmes. 3. Make it modular.
Large programs
can be broken down
These efficient systems could
into smaller
also mean lower costs for compa-
components, called
nies. “One of the main misconcep- microservices. These
tions about green software is that microservices can be
you have to do something extra, and called only when
it will cost extra,” Kinsiveer says. “It they’re needed.
Similarly, employ a
doesn’t cost extra—you just have to
function-as-a-service
do things right.” (FaaS) model to divide
Other motivating factors, espe- actions into functions
cially on the business side of soft- that are executed only
ware, are the upcoming legislation when required.
and regulations related to sustain-
ability. In the European Union, for
instance, the Corporate Sustainabil-
ity Reporting Directive requires
4. Develop lean
code. Refactor
old or legacy code that
companies to report more on their might slow down a
environmental footprint, energy program’s runtime,
and delete unused
usage, and emissions, including the code that could lead to
emissions related to the use of their wasted processing
products. time and power.

APRIL 2024 SPECTRUM.IEEE.ORG 29


IDEAS FOR CODE: LEAN SOFTWARE

particular, she adds, developers have a responsibility Creating sustainable websites like Tijgerbrood’s
to ensure that what they’re creating isn’t damaging is a team effort that involves different roles—from
the environment. business analysts who define software require-
ments to designers, architects, and those in charge
Building Greener Websites and Apps of operations—and includes green practices that
The makers of COP28’s website could have taken a can be applied at each stage of the software-devel-
page from directories like Lowwwcarbon, which opment process.
highlights examples of existing low-carbon websites. First, analysts will have to consider if the feature,
The company website of the Netherlands-based Web app, or software they’re designing should even be
design and branding firm Tijgerbrood, for instance, developed in the first place. Tech is often about cre-
emits less than 0.1 grams of carbon per page view. ating the next new thing, but making software sus-
tainable also entails decisions on what not to build,
and that may require a shift in mind-set.
The design stage is all about choosing efficient
algorithms and architectures. “Think about sustain-
ability before going into the solution—and not after,”
Green Software Measurement Tools says Chiara Lanza, a researcher at the Sustainable
Developers can use these tools to measure the impact AI unit of the Centre Tecnològic de Telecomunica-
cions de Catalunya, in Barcelona.
of green software engineering practices
During the development stage, programmers
need to focus on optimizing code. “We need the
AI: Estimate the energy usage and carbon footprint overall amount of energy we’re using to run software
of training AI models with these tools.
to go down. Some of that will come from writing
carbontracker experiment-impact-tracker [code] efficiently,” says Hannah Smith, a sustainable
digital tech consultant and director of operations at
ML CO2 Impact the Green Web Foundation.
Tijgerbrood’s website optimized the company’s
code by using low-resolution images and modern
Cloud: These dashboards give an overview of the estimated
carbon emissions associated with cloud workloads. image formats, loading animations only when a
user scrolls them into view, and removing unnec-
AWS Customer Carbon Google Cloud Carbon Footprint essary code. These techniques help speed up data
Footprint Tool transfer, loading, and processing on a user’s device.
The website also uses minimal JavaScript. “When
Microsoft Azure Emissions Cloud Carbon Footprint (free,
Impact Dashboard open source, provider agnostic) a user loads a website [with] a lot of JavaScript, it
causes them to use a lot more energy on their own
device because their device is having to do all the
Code: Integrate emissions estimation at the code level using work of reading the JavaScript and running [it],”
these tools. explains Smith.
Carbon-Aware SDK CodeCarbon When it comes to operations, one of the most
impactful actions you can take is to select a sustain-
Impact Framework able Web hosting or cloud-computing provider. The
Green Web Foundation has a tool to check if your
Middleware: These energy profilers or power monitors website runs on green energy, as well as a directory
provide APIs (application programming interfaces) to of hosting providers powered by renewable energy.
measure power consumption of apps or track energy You can also ask your hosting provider if you can
metrics of processors. scale how your software runs in the cloud so that
peak usage is powered by green energy or pause or
Intel Performance ­Counter PowerAPI switch off certain services during nonpeak hours.
Monitor
Web: These tools help calculate the carbon footprint AI the Green Way
of websites. Programmers can apply green software strategies
when developing AI as well. Trimming training data
Are my third parties green? CO2.js is one of the major ways to make AI systems greener.
Starting with data collection and preprocessing, it’s
Ecograder Firefox Profiler worth thinking about how much data is really
needed to do the job. It may pay to clean the dataset
Website Carbon Calculator to remove unnecessary data, or select only a subset
of the dataset for training.

30 SPECTRUM.IEEE.ORG APRIL 2024


“The larger your dataset is, the more time and Green estimated carbon emissions associated with cloud
computation it will take for the algorithm to go software workloads, such as the AWS Customer Carbon
through all the data,” hence using up more energy, Footprint Tool and Microsoft’s Azure Emissions
says Sallou.
is efficient Impact Dashboard; energy profilers or power mon-
For instance, in a study of six different AI algo- software, itors like Intel’s Performance Counter Monitor; and
rithms that detect SMS spam messages, Sallou and allowing tools that help calculate the carbon footprint of
her colleagues found that the random forest algo- coders to websites, such as Ecograder, Firefox Profiler, and
rithm, which combines the output of a collection of cultivate Website Carbon Calculator.
decision trees to make a prediction, was the most faster, The vision, Hussain says, is “for every software
energy-greedy algorithm. But reducing the size of higher- team to be able to measure their emissions, and that
the training dataset to 20 percent—only 1,000 data quality every software product in the world is reporting its
points out of 5,000—dropped the energy consump- emissions.” Smith echoes this sentiment, hoping
systems.
tion of training by nearly 75 percent, with only a 0.06 that data passing through the Internet would con-
percent loss in accuracy. tain not only information about what is being sent
Choosing a greener algorithm could also save but also details about the associated carbon
carbon. Measurement tools like CodeCarbon and ­emissions. “That’s the big dream,” she says. “If we
ML CO2 Impact can help make the choice by esti- have that, we’d be able to figure out what’s going on,
mating the energy usage and carbon footprint of make things more optimized, and get to a fossil-free
training different AI models. ­Internet a lot quicker.”

Tools for Measuring Software’s The Future Is Green


Carbon Footprint Green software engineering is growing and evolving,
To write green code, developers need a way of mea- but we need more awareness to help the discipline
suring the actual carbon emissions across a system’s to become more widespread. This is why, in addition
entire life cycle. It’s a complex feat, given the myriad to its Green Software for Practitioners course, the
processes involved. If we take AI as an example, its Green Software Foundation aims to create more
life cycle encompasses raw material extraction, GREEN training courses, some of which may even lead to
materials manufacturing, hardware manufacturing, SOFTWARE certifications. Likewise, Sallou coteaches a graduate
model training, model deployment, and disposal— RESOURCES course in sustainable software engineering, whose
and not all of these stages have available data. The Green Software
syllabus is open and can be used as a foundation for
“We don’t understand huge parts of the ecosys- Foundation offers a anyone looking to build a similar course. Providing
tem at the moment, and access to reliable data is catalog of green this knowledge to students early on, she says, could
tough,” Smith says. The biggest need, she adds, is software patterns ensure they bring it to their workplaces as future
“open data that we can rely on and trust” from big for AI, the cloud, and software engineers.
the Web.
tech data-center operators and cloud providers like In the realm of artificial intelligence, Navveen
Amazon, Google, and Microsoft. The W3C’s Balani, an AI expert and Google Cloud Certified
Until that data surfaces, a more practical Sustainable Web Fellow who also serves on the Green Software
approach would be to measure how much power Design Community Foundation’s steering committee, notes that AI
software consumes. “Just knowing the energy con- Group released a could inherently include green AI principles in the
draft of its Web
sumption of running a piece of software can impact coming years, much like how security consider-
sustainability
how software engineers can improve the code,” guidelines, with both ations are now an integral part of software devel-
Sallou says. tactical and opment. “This shift will align AI innovation with
Developers themselves are heeding the call for technical environmental sustainability, making green AI not
more measurement, and they’re building tools to recommendations just a specialty but an implied standard in the field,”
for business and
meet this demand. The W3C’s Sustainable Web he says.
product strategy,
Design Community Group, for instance, plans to user-experience As for the Web, Smith hopes the Green Web Foun-
provide a test suite to measure the impacts of design, Web dation will cease to exist by 2030. “Our dream as an
implementing its Web sustainability guidelines. development, and organization is that we’re not needed, we meet our
Similarly, the Green Software Foundation wrote a hosting and goal, and the Internet is green by default,” she says.
infrastructure. The
specification to calculate the carbon intensity of Kinsiveer has observed that in the past, software
draft guidelines also
software systems. For accurate measurements, include impact and had to be optimized and built well because hardware
Lanza suggests isolating the hardware in which a effort ratings to give then was lacking. As hardware performance and
system runs from any other operations and to avoid software engineers innovation leveled up, “the quality of programming
running any other programs that could influence an idea of the level of itself went down,” he says. But now, the industry is
difficulty in terms of
measurements. coming full circle, going back to its efficiency roots
implementation
Other tools developers can use to measure the and the level of and adding sustainability to the mix.
impact of green software engineering practices impact in terms of “The future is green software,” Kinsiveer says.
include dashboards that give an overview of the sustainability. “I cannot imagine it any other way.” 

APRIL 2024 SPECTRUM.IEEE.ORG 31


THE AGE OF MEMS  INKJETS

Inkjets Are for More Than Just Printing

THEY CAN BUILD


DNA ARRAYS,
3D STRUCTURES,
AND MUCH MORE

BY PHILLIP W. BARTH & LESLIE A. FIELD


ILLUSTRATION BY EDDIE GUY

32 SPECTRUM.IEEE.ORG APRIL 2024


IN THE EARLY 1980s, offices were noisy places, filled with the INKJET TECHNOLOGY DATES BACK to 1948, when Swedish
INKJETS

sound of metal striking inked ribbons to mark characters on inventor Rune Elmqvist patented a chart recorder wherein a
paper. IBM Selectric typewriters clacked, daisy wheel printers very thin glass tube emitting a continuous jet of ink was steered
clattered, and dot-matrix printers made loud ripping sounds. to make a trace on a moving strip of paper. A couple of years
Today, those noises are gone. And though we do spend more later, he demonstrated his invention in the form of a device for

time reading on screens, we haven’t stopped printing on paper. recording electrocardiograms.


THE AGE OF MEMS

The main reason for the quiet? The inkjet printer. While In 1965, Richard G. Sweet of Stanford University developed
laser printers do the big printing jobs in commercial settings, a chart recorder in which the jet of ink was broken into a uni-
the inkjet has become the printer most of us use at home and form stream of electrically charged droplets. Diverter electrodes
at the office. on either side of the stream could permit the drops to proceed
The printhead of an inkjet printer performs a remarkable straight to the paper, or else deflect them onto an absorbent
task. Even at the coarse resolution of 96 dots per inch (dpi), as pad or into a gutter to be collected and reused.
was typical for the first models in the 1980s, the distance from This technology is called continuous inkjet printing, and by
dot center to dot center is a mere 260 micrometers. To fill a 1976 IBM had incorporated it in a commercial printer, the IBM
standard letter page that has 2.5-centimeter margins would 6640. But continuous inkjets lose ink to evaporation even when
require more than half a million individual ink droplets. Deliv- recycling is used, limiting their appeal.
ery of those tiny droplets involves moving them with very pre- To get around the wastefulness of continuous inkjets, others
cise control, repeated a vast number of times as rapidly as worked on developing drop-on-demand inkjet printers, where
possible. This process is ideally suited for microelectromechan- each orifice on the printhead emits one drop of ink at a time,
ical systems (MEMS), which are electronic devices with micro- avoiding the waste of a continual flow of drops. Surface tension
scopic components that employ movement. holds the ink in place in a tiny open nozzle until a mechanism
As with all microtechnology, the specs of inkjet systems pushes the ink to eject a drop. Each drop hitting the paper
have evolved considerably over time. A typical inkjet printhead creates a dot, and moving the printhead back and forth builds
in the mid-1980s had 12 nozzles working in parallel, each one up an image. A printhead with multiple orifices can emit many
emitting up to 1,350 droplets per second, to print 150 alpha­ drops of ink simultaneously, so each pass of the printhead
numeric characters per second. Today, a high-end inkjet print­ across the page adds a strip of the image, not just a single drop-
head used in a commercial printing press may contain 21,000 thin line.
nozzles, each nozzle printing 20,000 to 150,000 dots per In the late 1970s, Siemens was the first to sell a drop-on-
second. Each drop of ink may be just 1.5 picoliters—a picoliter demand inkjet printer. It came not as a stand-alone device
is one-trillionth of a liter—and measure roughly 14 micrometers like a modern desktop printer, but as an integral part of a
in diameter. computer terminal, the Siemens PT80i (Printer Terminal 80
Surpassing the visions of its creators, the inkjet technology Inkjet). The printer used piezoelectric actuators surrounding
used in these printers has found a host 12 ink tubes, which fed 12 nozzles to
of applications beyond putting dots on shoot ink droplets, printing 270 charac-
paper. These include making DNA ters per second.
microarrays for genomics, creating elec- Piezoelectric devices rely on how
trical traces for printed circuit boards, some materials, such as ceramic lead
and building 3D-printed structures. ­zirconate titanate (PZT), change shape
Future uses could include personalized when subjected to a voltage. This effect
medicine and development of advanced has proved extremely useful in MEMS in
batteries. general, for generating precise forces and
Indeed, a search for patents contain- motion on command. If a layer of PZT is
ing the word “inkjet” today returns more bonded to a nonpiezoelectric material,
than 92,000 results. If there is a way to forming what’s called a bimorph, it will
package something in microscopic drop- bend when exposed to a voltage. In the
lets with the appropriate fluid properties, piezoelectric inkjet nozzle, the bending
chances are someone is looking to adapt of the bimorph pushes ink out of the ori-
inkjet technology to work with it. fice. [For another application of piezo-
electric MEMS technology, see “How
Ultrasound Became Ultra Small,” p. 38.]
IN APRIL 1984, the HP ThinkJet [top] This novel printing technology, how-
ushered in the era of desktop inkjet printing.
ever, was not yet as dependable as proven
The ThinkJet’s ink cartridge [middle]
delivered microscopic droplets on demand impact printers in the 1970s, and the
thousands of times a second from 12 whole Siemens terminal became unusable
nozzles. The MEMS technology to perform if the printer failed, so it didn’t catch on.
that feat was entirely within the thin strip Meanwhile, researchers at both
that formed the printhead. Today, the HP Jet Hewlett-­Packard and Canon noticed that
Fusion 5200 3D printer [bottom] uses an
inkjet process to build parts, up to 38 by ink would boil and splatter when exposed
28 by 38 centi­meters, out of nylon, to a hot element like a soldering iron, and
polypropylene, or polyurethane. they decided to turn that splattering into
HP

34 SPECTRUM.IEEE.ORG APRIL 2024


THERMAL PIEZOELECTRIC

a useful inkjet printing mechanism. They


knew that a resistor could be used as a Bimorph
heating element and could be miniatur- Resistor
ized with the same technology as that
used for integrated circuits. In the print-
ers they built, each ink nozzle contains a Ink
resistor instead of a piezoelectric actua-
tor. An electrical pulse heats the resistor,
which flash-boils a thin layer of the ink,
forming a rapidly expanding vapor
bubble that pushes a droplet of ink out
through the orifice.
This work led to two competing ver-
sions of thermal inkjet technology
coming to market at nearly the same time
40 years ago. (The same year, 1984,
Epson introduced a stand-alone piezo-
electric inkjet printer.)
Hewlett-Packard’s HP ThinkJet was
its first desktop inkjet printer based on
its thermal technology, and it was
designed to connect to a personal com-
puter for everyday printing. It had an
immediate advantage over the recently
developed laser printers: It was much
cheaper. A desktop laser printer from HP
Vapor
cost US $3,500 (about $10,500 today);
HP’s 2225A ThinkJet cost only $495
($1,500 today). Inkjet printers also used
far less power than laser printers did and
were quieter. Admittedly, inkjets didn’t
have great resolution—96 dpi compared
with 300 for laser printers in those early
days—and they were slow.
But the advantages outweighed the
IN THE RESTING STATE [top] of an inkjet nozzle, the ink is held in place
disadvantages (more so as the technol- by surface tension. In a thermal inkjet nozzle [left], a voltage pulse to the
ogy improved), and inkjet printers came heating resistor flash-vaporizes a thin layer of ink, producing an expanding
to dominate the desktop and home vapor bubble [left, middle] that pushes a droplet of ink out through the
printer markets. Today, more than 20 orifice [left, bottom]. In less than a millisecond, the vapor recondenses
and the chamber cools and refills with ink, returning the nozzle to the
companies make inkjet printers, gener-
resting state. A piezoelectric inkjet nozzle [right] is driven by a piezoelectric
ating a market of more than $100 billion bimorph, which bends when a voltage is applied [right, middle] to push out
annually and continuing to grow at more an ink droplet [right, bottom].
than 8 percent per year.

WHILE THE BUSINESS of making inkjet


printers matured and grew, some compa-
nies began exploring what other kinds of
“ink” might be delivered with an inkjet. One of these was Agilent A DNA microarray consists of a substrate, usually glass,
Technologies, a spin-off of ­Hewlett-Packard with a focus on with an array of small regions called spots where DNA strands
life-science and c­ hemical-analysis technologies. Agilent devel- are attached. Agilent produces arrays with as many as a million
oped a way to print strands of DNA from the four nucleic acid spots on a single 2.5-by-7.6-cm slide. An open-source system
bases—cytosine (C), guanine (G), a ­ denine (A), and thymine (T). puts up to 10,000 in a somewhat smaller area. Each DNA
Specifically, the company adapted existing DNA chemistries plus strand is made of sequences of the bases C, G, A, and T. In
inkjet printing techniques to build microarrays of DNA on glass ­double-stranded DNA, the strands have complementary
slides for genomics work, such as measuring which genes are sequences, which join up like rungs of a ladder, C joining with G,
being expressed in an organism under various conditions. Aca- and A with T.
demic researchers have shared open-source methods for con- A DNA microarray uses single-stranded DNA, and each spot
verting existing inkjet printers to build their own microarrays, has millions of strands with a common sequence. When a
albeit with specs that are much more modest than those of the sample with copies of the complementary strand washes over
commercial systems. the spot, those strands bind together with the strands anchored

Illustration by Chris Philpot APRIL 2024 SPECTRUM.IEEE.ORG 35


in the spot. The sample strands are tagged with fluorescent binding agent to each layer of powder in the regions that will
INKJETS

molecules, and the user learns which DNA sequences were form the finished 3D items.
present in the sample by examining which spots light up. The HP Multi Jet Fusion (MJF) line of 3D printers extends
In Agilent’s method for fabricating a microarray, the printer this approach by depositing two types of ink: One is a binding
makes multiple passes over the substrate, each pass adding promoter and the other a detailing agent, which is applied at

one base to each strand in the spots, with intermediate steps the edges of the pattern to prevent the promoter from bleeding
THE AGE OF MEMS

to prepare for the next pass. into the surrounding powder. A printhead carrying a wide array
Adding a base is actually a three-step process. Each of the of inkjet nozzles dispenses these inks, and the array is quickly
growing strands in the microarray spots has a molecular “cap” followed by a lighting bar to heat the powder, fusing it in the
at the end that prevents the indiscriminate addition of more regions where the binding promoter is present. A fresh layer of
bases. So the first step is to remove or deactivate those caps powder is then spread over the entire printing area in readiness
by washing a solution over the nascent microarray. The second for the next cycle of the process. At the end, compressed air
step is analogous to printing a page: At each spot on the micro­ and a vacuum hose remove the unfused powder to reveal the
array, the inkjet adds a dot of liquid containing the next mono- completed 3D objects. The HP MJF printers perform this in a
mer molecule (modified versions of C, G, A, or T) to be added volume of up to 38 by 28 by 38 cm.
to the end of the strand. These monomers each include a new A quite different approach has been taken by Mimaki Engi-
cap so that only one molecule gets added to each strand. neering Co. of Japan, which has introduced 3D printers with
Although the newly added monomers are now attached to the piezoelectric inkjet heads that dispense droplets of resin. The
strands, the connection is not fully stable, and so the third resins are photopolymers that are cured by ultraviolet
step applies an oxidizer solution that modifies the bonds, fully light-emitting diodes after each layer is printed. Instead of using
integrating the new monomers into the DNA structure. Rinse a powder bed that fills the entire build area, the printer deposits
and repeat. the resins on top of the growing structure. To deal with steep
The versatility of the open-source inkjet construction allows overhangs—such as an outstretched arm of a figurine—one of
researchers to rapidly build prototype arrays with whatever the resins produces a water-soluble material, which is used to
sequences they want to try out. A new array can be designed, build supports where needed. After the build is finished, these
synthesized, and used to analyze DNA in a single day. One group supports can be dissolved away.
reported a cycle time of 10 to 20 minutes to attach each base Seven other resins provide colors that can include CMYK—
with their system, or about 13 hours to produce a batch of the familiar cyan, magenta, yellow, and black inks of consumer
arrays, each with about 10,000 spots containing 40-base inkjet printers—as well as white and clear, for a total of 10 mil-
strands. For comparison’s sake, Agilent’s commercial micro­ lion color combinations—comparable to the color depth that
arrays typically have strands up to 60 bases long. the human eye can discriminate. The resulting parts can com-
Agilent also uses its inkjet system to synthesize another bine solid color, colored transparency and translucency, and
genomic workhorse known as an oligonucleotide library. The colorless transparency.
process is the same as for making a microarray, but at the end The printer provides a volume for building that measures
all the strands are cleaved from the substrate, dried, and pack- 51 by 51 by 30 cm. Unlike with a powder-bed machine, small
aged together in a single tube for the customer. Agilent’s inkjet-­ test parts can be made without filling the entire volume. In
printed libraries have strands up to about 230 bases long. general, however, the Mimaki approach is slower than that of
the HP MJF because it uses smaller printheads instead of a
IN ADDITION TO printing two-dimensional pages and building wide one that can cross the entire area in one sweep.
one-dimensional molecular strands, inkjet technology has for
many years been used to produce three-dimensional objects. INKJET PRINTING’S STRENGTH is the ability to pattern
One approach is a variant of powder-bed 3D printing, in which various inks over large areas in short, rapid production runs at
objects are built up by fusing or binding layers of powder in the a reasonable cost. It cannot generally compete with standard
desired pattern. The inkjet printhead applies droplets of a liquid high-volume production approaches, because those will usually
be cheaper. Thus, a car enthusiast, for instance, may embrace
3D inkjet printing to make bespoke parts for repairs or other
tinkering, but a high-volume car-parts manufacturer is not
going to introduce such printers to its factory lines. Similarly,
a company may build individual figurines from a customer’s
design, printed by 3D inkjet, but the same technique won’t be
economical for mass-producing models of the latest superhero.
With many potential applications, it isn’t clear if there is a niche
where the inkjet approach will win.
MIMAKI ENGINEERING CO.

An example is the use of 3D inkjet printing for personalized


medicine. The idea is to produce tablets of a medication cus-
tomized for a specific patient. Such personalized pills can
include simple fine-tuning of the dose for an individual, as well
THIS MODEL OF a handheld vacuum cleaner with colored, translucent,
as adjustments to the drug’s release rate—from very rapid to
and transparent parts was printed in one piece on a Mimaki inkjet 3D
printer. The clear material was polished after printing. Manufacturers slow and sustained—through modifications to the binding
can cut weeks off their design process by using such models. agents and structure of the tablet. Rather than juggling multiple

36 SPECTRUM.IEEE.ORG APRIL 2024


Printhead Protection

G C A T Nucleotide

Glass slide

Spot

Nucleotide
chain

medications on a complicated schedule


each day, a patient could take a single
daily polypill—a 3D-printed tablet con- A DNA MICROARRAY can be fabricated using an inkjet system to build
taining multiple medications, each with custom-designed strands of DNA at each spot of the array. The printhead
a different rate of release. delivers a droplet of “ink” [left] containing modified monomers of one
Researchers are exploring how to adapt nucleotide [G, C, A, or T] to each spot. On the first print cycle, these
monomers attach to the chemically treated glass surface. On subsequent
existing 3D printing techniques, including
print runs [right], a single monomer joins on the end of each growing DNA
inkjet, to make these personalized medi- strand. Each monomer includes a protective cap to prevent other monomers
cations. Inkjet systems are particularly from joining. Next, additional processes [not depicted here] wash away the
suited for printing drugs in the form of thin nucleotide ink, apply a catalyst to complete the monomer bonding, and strip
films, such as transdermal patches to be away the protective caps in preparation for the next printing step.
applied to the skin and buccal films to be
held in the cheek, where drugs can pass
directly to the bloodstream without first
going through the digestive system.
These printed personalized medicines,
however, would be expensive compared to fixed doses rolling structures with good electrochemical properties?
off standard high-volume production lines. Thus the technique We think it is unlikely, however, that inkjet printing could
is likely to be reserved for relatively rare conditions. compete with high-volume manufacturing on price. Inkjet
Another potential application of 3D inkjet printing is in the printing of prototypes, on the other hand, may uncover an opti-
fabrication of advanced lithium-ion batteries. The charging mal battery design that can then be adapted for production by
and discharging of these batteries relies on lithium ions moving conventional techniques.
from the battery’s electrodes to its electrolyte and back again, Inkjet systems have been demonstrated for a wide variety
in the process releasing or absorbing electrons that produce of applications beyond what we have discussed above: Living
the current flow. The energy-storage density of the standard cells can be printed, for instance, to form tissue structures for
electrode design can be increased by using thicker electrodes, in vitro experiments. MEMS such as microscopic motors have
but this compromises the power density—the rate of energy been printed using inks containing nanoparticles of gold and
release—because a smaller proportion of the electrode is in silver as conductors and resin-based inks to act as insulators.
close contact with the electrolyte. Flexible sensors for health care monitoring have been printed
A 3D inkjet could build electrodes with a detailed micro- using an electrically conducting polymer that responds to tem-
structure that allows the electrolyte to penetrate throughout perature differentials. And then there are all the ways inkjets
the electrode volume. This could boost the ability of the are used to create images on media other than office printouts,
active lithium ions and electrons to reach the entire electrode such as printing of textiles, inkjet robots to apply custom auto-
efficiently even when the electrode is larger, thereby increas- motive paint jobs, and the “Giclée” printing of fine art using
ing the energy storage and power density in tandem. For this archival-quality inks and substrates.
vision to become a reality, however, researchers will need Each of these applications is like a colored dot on the vast
to learn more about how to formulate the “inks” for printing canvas of human technology and activity. And while the dots
these electrodes: What are the best particle sizes and sol- from inkjets, powered by MEMS, may be only a single color
vents to make an ink with fluid properties suitable for use among many others on that metaphorical page, the picture
in an inkjet system and that will produce stable printed would be very different without them. 

Illustration by Chris Philpot APRIL 2024 SPECTRUM.IEEE.ORG 37


THE AGE OF MEMS  U LT R A S O U N D

With MEMS technology, all you need is one probe and a smartphone

By F. LEVENT DEGERTEKIN  Photo-illustration by Edmon de Haro


HOW
ULTRASOUND
BECAME
ULTRA
SMALL

APRIL 2024 SPECTRUM.IEEE.ORG 39


A startling change in medical ultra- To be useful for imaging, the ultra- to 2-MHz probe, which can provide 2- to
U LT R A S O U N D

sound is working its way through hos- sound waves need to travel out of the 3-millimeter resolution and can reach up
pitals and physicians’ offices. The slabs and into the soft tissue and fluid to 30 cm into the body. To image blood
long-standing, state-of-the-art ultra- of the patient’s body. This is not a trivial flow in arteries in the neck, physicians
sound machine that’s pushed around on task. Capturing the echo of those waves typically use an 8- to 10-MHz probe.
a cart, with cables and multiple probes is like standing next to a swimming
dangling, is being wheeled aside perma- pool and trying to hear someone speak- The need for multiple

nently in favor of handheld probes that ing under the water. The transducer probes along with the lack of
T
THE AGE OF MEMS

send images to a phone. arrays are thus built from layers of miniaturization meant that
These devices are small enough to fit material that smoothly transition in conventional medical ultra-
in a lab coat pocket and flexible enough stiffness from the hard piezoelectric sound systems resided in a heavy, boxy
to image any part of the body, from deep crystal at the center of the probe to the machine that had to be lugged around
organs to shallow veins, with sweeping soft tissue of the body. on a cart. The introduction of MEMS
3D views, all with a single probe. And the The frequency of energy transferred technology changed that.
AI that accompanies them may soon into the body is determined mainly by Over the last three decades MEMS
make these devices operable by untrained the thickness of the piezoelectric layer. has allowed manufacturers in an array
professionals in any setting—not just A thinner layer transfers higher fre- of industries to create precise, extremely
trained sonographers in clinics. quencies, which allow smaller, sensitive components at a microscopic
The first such miniaturized, hand- ­higher-resolution features to be seen in scale. (See, for example, “Inkjets Are for
held ultrasound probe arrived on the an ultrasound image, but only at shallow More Than Just Printing,” p. 32.) This
market in 2018, from Butterfly Network depths. The lower frequencies of thicker advance has enabled the fabrication of
in Burlington, Mass. Last September, piezoelectric material travel further into high-density transducer arrays that can
Exo Imaging in Santa Clara, Calif., the body but deliver lower resolutions. produce frequencies in the full 1- to
launched a competing version. As a result, several types of ultra- 10-MHz range, allowing imaging of a
Making this possible is silicon ultra- sound probes are needed to image var- wide range of depths in the body, all with
sound technology, built using a type of ious parts of the body, with frequencies one probe. MEMS technology also
microelectromechanical system that range from 1 to 10 megahertz. To helped miniaturize additional compo-
(MEMS) that crams 4,000 to 9,000 image large organs deep in the body or nents so that everything fits in the hand-
transducers—the devices that convert a baby in the womb, physicians use a 1- held probe. When coupled with the
electrical signals into sound waves and
back again—onto a 2-by-3-centimeter
silicon chip. By integrating MEMS
transducer technology with sophisti-
cated electronics on a single chip, these
scanners not only replicate the quality
of traditional imaging and 3D measure-
ments but also open up new applica-
tions that were impossible before.

To understand how
researchers achieved this
T feat, it’s helpful to know the
basics of ultrasound technol-
ogy. Ultrasound probes use transducers
to convert electrical energy to sound
waves that penetrate the body. The Butterfly Network
developed a handheld
sound waves bounce off the body’s soft
ultrasound probe with
tissue and echo back to the probe. The capacitive
transducer then converts the echoed micromachined
sound waves to electrical signals, and a ultrasonic transducer
computer translates the data into an (CMUT) technology.
image that can be viewed on a screen.
Conventional ultrasound probes
contain transducer arrays made from
slabs of piezoelectric crystals or ceram-
BUTTERFLY NETWORK

ics such as lead zirconium titanate


(PZT). When hit with pulses of electric-
ity, these slabs expand and contract and
generate high-frequency ultrasound
waves that bounce around within them.

40 SPECTRUM.IEEE.ORG APRIL 2024


computing power of a smartphone, this
eliminated the need for a bulky cart.
The first MEMS-based silicon ultra-
sound prototypes emerged in the mid-
1990s when the excitement of MEMS as
a new technology was peaking. The key
element of these early transducers was Exo Imaging developed
the vibrating micromachined mem- a handheld ultrasound
probe with
brane, which allowed the devices to piezoelectric
generate vibrations in much the same micromachined
way that banging on a drum creates ultrasonic transducer
sound waves in the air. (PMUT) technology.
Two architectures emerged. One of
them, called the capacitive micro­
machined ultrasonic transducer, or
CMUT, is named for its simple
­capacitor-like structures. Stanford Uni-
versity electrical engineer Pierre
­Khuri-Yakub and colleagues demon-
strated the first versions.
The CMUT is based on electrostatic
forces in a capacitor formed by two con-
ductive plates separated by a small gap.
One plate—the micromachined mem-
brane mentioned before—is made of
silicon or silicon nitride with a metal
electrode. The other—typically a micro-
machined silicon wafer substrate—is
thicker and more rigid. When a voltage
is applied, placing opposite charges on
the membrane and substrate, attractive
forces pull and flex the membrane In smoke alarms these structures are erated. Combining arrays of CMUT
toward the substrate. When an oscillat- typically 4 cm in diameter, and they’re devices with different dimensions into
ing voltage is added, that changes the what generates the shrieking sound of a single probe to increase the frequency
force, causing the membrane to vibrate, the alarm, at around 3 kilohertz. When range also compromised the pressure
like a struck drumhead. the membrane is scaled down to 100 output because it reduced the probe
When the membrane is in contact micrometers in diameter and 5 to 10 μm area available for each frequency.
with the human body, the vibrations send in thickness, the vibration moves up The solution to these problems came
ultrasound frequency waves into the into megahertz frequencies, making it from Khuri-Yakub’s lab at Stanford. In
tissue. How much ultrasound is g­ enerated useful for medical ultrasound. experiments in the early 2000s, the
or detected depends on the gap between Honeywell in the early 1980s devel- researchers found that increasing the
the membrane and the substrate, which oped the first micromachined sensors voltage on CMUT-like structures caused
needs to be about one micrometer or less. using piezoelectric thin films built on the electrostatic forces to overcome the
Micromachining techniques made that silicon diaphragms. The first PMUTs restoring forces of the membrane. As a
kind of precision possible. operating at ultrasound frequencies result, the center of the membrane col-
The other MEMS-based architecture didn’t emerge until 1996, from the work lapses onto the substrate.
is called the piezoelectric micro­ of materials scientist Paul Muralt at the A collapsed membrane seemed disas-
machined ultrasonic transducer, or Swiss Federal Institute of Technology trous at first, but it turned out to be a way
PMUT, and it works like a miniaturized Lausanne (EPFL). of making CMUTs both more efficient and
version of a smoke-alarm buzzer. These more tunable to different frequencies. The
buzzers consist of two layers: a thin A big challenge with efficiency increased because the gap
metal disk fixed around its periphery and CMUTs was getting them to around the contact region was very small,
a thin, smaller piezoelectric disk bonded
A generate enough pressure to increasing the electric field there. And the
on top of the metal disk. When voltages send sound waves deep into pressure increased because the large
are applied to the piezoelectric material, the body and receive the echoes coming doughnut-shaped region around the edge
it expands and contracts in thickness back. The membrane’s motion was lim- still had a good range of motion. What’s
and from side to side. Because the lateral ited by the exceedingly small gap more, the frequency of the device could
EXO IMAGING

dimension is much larger, the piezo disk between the membrane and the sub- be adjusted simply by changing the volt-
diameter changes more significantly and strate. This constrained the amplitude age. This, in turn, allowed a single CMUT
in the process bends the whole structure. of the sound waves that could be gen- ultrasound probe to produce the entire

APRIL 2024 SPECTRUM.IEEE.ORG 41


CMUT PMUT
U LT R A S O U N D

Electrode
Piezoelectric thin film V
V Electrostatic gap

Cavity

Substrate

THE AGE OF MEMS

V
V>Vc

On the manufacturing side, the University of California, Davis, con-

ANATOMY challenges involved finding the right


materials and developing the pro-
nected around 2,500 PMUT elements to
CMOS electronics and placed them
OF MEMS cesses needed to produce smooth sur- under a silicone rubberlike layer. When

ULTRASOUND faces and a consistent gap thickness.


For example, the thin dielectric layer
a fingertip was pressed to the surface,
the prototype measured the amplitudes
PROBES that separates the conductive
­membrane and the substrate must
of the reflected signals at 20 MHz to
distinguish the ridges in the fingertip
withstand about 100 volts at a thick- from the air pockets between them.
Two MEMS ultrasound
architectures have emerged. ness of 1 μm. If the layer has defects, This was an impressive demonstra-
In the capacitive micromachined charges can be injected into it, and the tion of integrating PMUTs and electron-
ultrasonic transducer (CMUT) device can short at the edges or when ics on a silicon chip, and it showed that
design, attractive forces pull and the membrane touches the substrate, large 2D PMUT arrays could produce a
flex the membrane toward the
killing the device or at least degrading high enough frequency to be useful for
substrate. When an oscillating
voltage is added, the membrane its performance. imaging of shallow features. But to make
vibrates like a struck drumhead. Eventually, though, MEMS found- the jump to medical ultrasound, PMUT
Increasing the voltage causes the ries such as Philips Engineering Solu- technology needed more bandwidth,
electrostatic forces to overcome tions in Eindhoven, Netherlands, and more output pressure, and piezoelectric
the restoring forces of the Taiwan Semiconductor Manufacturing thin films with better efficiency.
membrane, causing the membrane
to collapse onto the substrate. In Co. (TSMC), in Hsinchu, developed Help came from semiconductor
the piezoelectric micromachined solutions to these problems. Around companies such as ST Microelectron-
ultrasonic transducer (PMUT) 2010, these companies began produc- ics, based in Geneva, which figured out
architecture, voltages applied to ing reliable, high-performance CMUTs. how to integrate PZT thin films on sili-
the piezoelectric material cause it con membranes. These films require
to expand and contract in thickness
and from side to side. Because the
Early PMUT designs also extra processing steps to maintain their
had trouble generating properties. But the improvement in per-
lateral dimension is much larger,
the piezo disk diameter changes
E enough pressure to work for formance made the cost of the extra
ILLUSTRATION SOURCE: ALESSANDRO STUART SAVOIA

significantly, bending the whole medical ultrasound. But steps worthwhile.


structure. they could bang out enough to be useful To achieve a larger pressure output,
in some consumer applications, such the piezoelectric layer needed to be
as gesture detection and proximity sen- thick enough to allow the film to sustain
sors. In such “in-air ultrasound” uses, the high voltages required for good
ultrasound frequency range needed for bandwidth isn’t critical, and frequen- ultrasound images. But increased thick-
medical diagnostics with high efficiency. cies can be below 1 MHz. ness leads to a more rigid membrane,
From there, it took more than a decade In 2015, PMUTs for medical appli- which reduces the bandwidth.
to understand and model the complicated cations got an unexpected boost with One solution was to use an oval-
electromechanical behavior of CMUT the introduction of large 2D matrix shaped PMUT membrane that effec-
arrays and iron out the manufacturing. arrays for fingerprint sensing in mobile tively combined several membranes of
Modeling these devices was tricky phones. In the first demonstration of different sizes into one. This is similar
because thousands of individual mem- this approach, researchers at the Uni- to changing the length of guitar strings
branes interacted in each CMUT array. versity of California, Berkeley, and the to generate different tones. The oval

42 SPECTRUM.IEEE.ORG APRIL 2024 Illustration by Chris Philpot


membrane provides strings of multiple price is comparable to Butterfly’s latest
lengths on the same structure with its model, the IQ+, priced at $2,999.
narrow and wide sections. To efficiently The ultrasound MEMS chips in
vibrate wider and narrower parts of the these probes, at 2 by 3 cm, are some of
membrane at different frequencies, elec- the largest silicon chips with com-
trical signals are applied to multiple bined electromechanical and elec-
electrodes placed on corresponding tronic functionality. The size and
regions of the membrane. This approach complexity impose production chal-
allowed PMUTs to be efficient over a lenges in terms of the uniformity of the
wider frequency range. devices and the yield.
These handheld devices operate at
In the early 2000s, low power, so the probe’s battery is
researchers began to push lightweight, lasts for several hours of
I CMUT technology for med- continuous use while the device is con-
ical ultrasound out of the lab nected to a cellphone or tablet, and has
and into commercial development. a short charging time. To make the
Stanford spun out several startups output data compatible with cellphones
aimed at this market. And leading med- and tablets, the probe’s main chip per-
ical ultrasound imaging companies such forms digitization and some signal pro-
Ultrasound technology has
as GE, Philips, Samsung, and Hitachi cessing and encoding. historically required bulky
began developing CMUT technology To provide 3D information, these machinery with multiple probes.
and testing CMUT-based probes. handheld probes take multiple 2D slices
But it wasn’t until 2011 that CMUT of the anatomy and then use machine
commercialization really began to make learning and AI to construct the neces-
progress. That year, a team with semi- sary 3D data. Built-in AI-based algo-
conductor electronics experience rithms can also help doctors and nurses means startups and academic groups
founded Butterfly Network. The 2018 precisely place needles in desired loca- now have access to the base technolo-
introduction of the IQ Probe was a tions, such as in challenging vasculature gies that will make a new level of inno-
transformative event. It was the first or in other tissue for biopsies. vation possible at a much lower cost
handheld ultrasound probe that could The AI developed for these probes is than it was 10 years ago.
image the whole body with a 2D imaging so good that it may be possible for pro- With all this activity, industry ana-
array and generate 3D image data. About fessionals untrained in ultrasound, such lysts expect ultrasound MEMS chips to
the size of a TV remote and only slightly as nurse midwives, to use the portable be integrated into many different
heavier, the probe was initially priced at probes to determine the gestational age ­medical devices for imaging and sensing.
US $1,999—one-twentieth the cost of a of a fetus, with accuracy similar to that For instance, Butterfly Network, in
full-size, cart-carried machine. of a trained sonographer, according to ­collaboration with Forest Neurotech, is
Around the same time, Hitachi in a 2022 study in NEJM Evidence. The developing MEMS ultrasound for
­
Tokyo and Kolo Medical in Suzhou, AI-based features could also make the brain-computer interfacing and
China (formerly in San Jose, Calif.), handheld probes useful in emergency ­neuromodulation. Other applications
commercialized CMUT-based probes medicine, in low-income settings, and include long-term, low-power wearable
for use with conventional ultrasound for training medical students. devices, such as heart, lung, and brain
systems. But neither has the same monitors, and muscle-activity monitors
­capabilities as Butterfly’s. For example, This is only the beginning used in rehabilitation.
the CMUT and electronics aren’t for miniaturized ultra- In the next five years, expect to see
­integrated on the same silicon chip,
T sound. Several of the world’s miniature passive medical implants with
JULIAN KEVIN ZAKARAS/FAIRFAX MEDIA/GETTY IMAGES

which means the probes have 1D arrays largest semiconductor ultrasound MEMS chips, in which power
rather than 2D. That limits the system’s foundries, including TSMC and ST and data are remotely transferred using
ability to generate images in 3D, which Microelectronics, now do MEMS ultra- ultrasound waves. Eventually, these
is necessary in advanced diagnostics, sound chip production on 300- and handheld ultrasound probes or wearable
such as determining bladder volume or 200-mm wafers, respectively. arrays could be used not only to image
looking at simultaneous orthogonal In fact, ST Microelectronics recently the anatomy but also to read out vital
views of the heart. formed a dedicated “Lab-in-Fab” in signs like internal pressure changes due
Exo Imaging’s September 2023 Singapore for thin-film piezoelectric to tumor growth or deep-tissue oxygen-
launch of its handheld probe, the Exo MEMS, to accelerate the transition ation after surgery. And ultrasound fin-
Iris, marked the commercial debut of from proofs of concept to volume pro- gerprint-like sensors could be used to
PMUTs for medical ultrasound. Devel- duction. Philips Engineering Solutions measure blood flow and heart rate.
oped by a team with experience in semi- offers CMUT fabrication for CMUT- One day, wearable or implantable
conductor electronics and integration, on-CMOS integration, and Vermon in versions may enable the generation of
the Exo Iris is about the same size and Tours, France, offers commercial passive ultrasound images while we
weight as Butterfly’s IQ Probe. Its $3,500 CMUT design and fabrication. That sleep, eat, and go about our lives. 

APRIL 2024 SPECTRUM.IEEE.ORG 43


44 SPECTRUM.IEEE.ORG APRIL 2024

A VC-backed startup’s push for growth


left little time for actual engineering

THE
MESSY
REALITY
BEHIND
A SILICON
VALLEY
UNICORN
By Benjamin Shestakofsky Photography by The Voorhes
W
cial contributor to the company’s rapid growth. It
was, in the words of two executives, “the magic
behind AllDone.”

IN THE PERIOD IMMEDIATELY following the first


round of funding, AllDone’s founders prioritized two
hen I began my research, kinds of expansion: growing the user base and hiring
AllDone had just secured its first round of venture more staff for the San Francisco team. First, to have
capital funding to fuel its quest to build an “Amazon any hope of success, AllDone would have to bring a
for local services.” The company had built a digital critical mass of users on board. While the company
platform connecting buyers and sellers of local ser- had enrolled 250,000 “sellers” of services, “buyers”
vices—housecleaners, plumbers, math tutors, and were submitting only about 7,000 requests for services
everything in between—across the United States. per month. The team aimed to boost buyer requests
Although the influx of US $4.5 million was cause for by nearly 50 percent over the next quarter, demon-
celebration, it also incited a sense of urgency among strating the kind of explosive growth that would make
employees in the San Francisco office. As Carter, AllDone an attractive target for future VC funding
AllDone’s president, intoned in an all-staff email, rounds. AllDone’s software developers would thus be
mobilized to overhaul the platform and make users’
We know what the future of local services is. experiences more intuitive and engaging.
But we’re not the only people that know this Executives planned to use most of the new
is the future. And, more importantly, there’s money to hire more engineers and designers.
lots of people—smart, scrappy, and well- Recruiting them soon became an all-consuming task
funded people—building our vision. Someone that engaged AllDoners both inside and outside of
is going to do it. And it looks like it’s going to the office, leaving little time for the staff to run the
happen soon. We just have to finish building business. The recruitment effort was led by Peter,
faster than anyone else and we will win. AllDone’s CEO. First, an external headhunter
reviewed résumé submissions and scheduled intro-
Demonstrating AllDone’s potential for explosive ductory phone calls between promising applicants
growth was the founders’ highest priority—and that and Peter. Next came a coding challenge devised by
priority shaped the company’s strategy and struc- the company’s four software engineers, followed by
ture. AllDone faced extraordinary pressure from a phone interview with one of the engineers to fur-
venture capital investors to grow as quickly as pos- ther evaluate each applicant’s technical prowess.
sible, which required finding new ways to attract Those who passed that test moved on to a daylong
users and increase their activity on the platform. At interview in the office, which consisted of 90-minute
the same time, AllDone’s leaders knew the firm one-on-one sessions with each of the four current
would be worthless if it couldn’t keep its product engineers. Candidates would also spend an hour with
functioning properly and provide services to its Josh, the product manager, and finally another hour
ever-expanding user base. with Peter before being sent off in the evening with
So the engineers in San Francisco set out to meet a beer stein emblazoned with the AllDone logo. Each
investors’ expectations by finding new ways to grow member of the hiring committee would write an eval-
the company. Meanwhile, AllDone’s managers hired uation that everyone involved would read before
contractors in the Philippines to perform routine conferring in person to discuss the candidate’s fate.
information-processing tasks. Some of the contrac- For weeks at a time, the hiring team interviewed one
tor work involved operations that software alone or two candidates per day.
was unable to accomplish. But engineers also off- The engineers’ heavy involvement in the labori-
loaded processes that software was technically ous and time-consuming hiring process reduced
capable of handling so employees in San Francisco their productivity, which threatened to slow the
could remain focused on their strategic goals. Man- company’s progress at a time when investors
agers viewed AllDone’s Filipino workforce as a cru- expected precipitous growth. Although I had come

For 19 months, the sociologist Benjamin Shestakofsky embedded himself in an early-stage tech startup to study
its organization and culture. The company went on to become one of Silicon Valley’s “unicorns,” valued at over
US $1 billion. This article is adapted from an excerpt of the author’s new book, Behind the Startup: How Venture
Capital Shapes Work, Innovation, and Inequality (University of California Press, 2024). The names of staff members
and the company have been changed to preserve privacy.

46 SPECTRUM.IEEE.ORG APRIL 2024


to AllDone because of my interest in studying work ALLDONE’S MANAGERS increasingly turned to
and life inside a startup, my field notes reflected my the company’s digital assembly line in the Philippines,
surprise: “Since I began at AllDone, there doesn’t where contractors performed computational work
appear to be much work going on at all, at least as that stood in for or supported software algorithms.
far as software production is concerned.” My obser- AllDone had hired its first work-from-home
vations were later confirmed by Josh, AllDone’s ­Filipino contractor a few months after the compa-
product manager, when he reported that during the ny’s launch. Within a year, the team had grown to
first quarter of the year, AllDone’s four software 125, and during my research it expanded to 200.
engineers had “accomplished very little” in terms Most contractors were college educated and
of their production goals because they had been between the ages of 20 and 40; about 70 percent
“very, very focused on recruiting,” which he said had were women. Executives often called these workers
consumed at least half of their work hours. AllDone’s “human machine.”
How, then, did AllDone run and even grow its Contractors logged in to AllDone’s adminis-
platform when its software developers were fre- trative portals to complete various sets of tasks.
quently too busy with recruiting to do their jobs? Most notably, a division that eventually numbered

APRIL 2024 SPECTRUM.IEEE.ORG 47


Filipino contractors’ wages and work hours were
determined by their jobs: On average, contractors
earned about $2.00 per hour and worked about 30
hours per week. While AllDone paid its Filipino
workers only a tiny fraction of what San Francisco–
based employees earned, their compensation sub-
stantially exceeded the Philippines’ legal minimum
wage. As independent contractors, these workers
didn’t receive paid vacation, sick leave, health insur-
ance, or retirement benefits, nor did they enjoy the
perks (like free food) available to workers in the San
Francisco office. Contractors were also responsible
for providing their own computer equipment and
Internet connections.
Companies seeking workers to do routine infor-
mation processing often post tasks to on-demand
“crowdwork” platforms like Amazon Mechanical
Turk. In AllDone’s case, the importance of its con-
tractors’ tasks to the company’s success meant that
an open call fulfilled by anonymous workers simply
wouldn’t do. AllDone’s staff in San Francisco con-
sidered AllDone Philippines an integral part of the
organization and built enduring relationships with
contractors, who typically performed the same
assigned task for a period of months or even years.
Newly hired contractors watched training videos
to learn how to use AllDone’s proprietary admin-
istrative software. Managers of the Filipino divi-
sions distributed weekly quizzes and offered
coaching to ensure that workers understood All-
Done’s rules and procedures.
Yet at times, even high-ranking managers in the
Philippines were excluded from important decisions
that would affect their teams. In one meeting I had
with Carter, AllDone’s president, he explained that
AllDone’s engineers had recently made a change that
suddenly increased some contractors’ workload by
60 percent. “We should have told them ahead of
nearly 100 people handled the company’s primary time so they would know it’s coming,” Carter said,
function of manually matching buyer requests with wincing a little and shrugging sheepishly, “but it just
sellers from AllDone’s database of service provid- didn’t occur to us.” For most staffers at AllDone San
ers—a process that users likely assumed was auto- Francisco, their Filipino colleagues were effectively
mated. Another division onboarded new sellers by invisible human infrastructure that they could take
classifying the services they provided, running an for granted.
array of checks to verify their trustworthiness, The efforts of AllDone’s Filipino workforce had
and proofreading their profiles. A third division the desired effect. During the first quarter of the year,
was responsible for generating brief descriptions AllDone met its user-growth goal, receiving almost
of AllDone sellers; these blurbs were then com- 50 percent more buyer requests than in the prior
piled on Web pages designed to boost AllDone’s three-month period. During the second quarter, that
position in search-engine rankings. In total, Fili- metric would increase again by 75 percent.
pino contractors executed over 10,000 routine AllDone’s Filipino contractors made these sub-
tasks per day. stantial gains possible by laboring alongside com-

Contractors effectively functioned as artificial


artificial intelligence, simulating the output of software
algorithms that had yet to be completed.

48 SPECTRUM.IEEE.ORG APRIL 2024


puter code. In some instances, their efforts bases, running voluntary criminal-background
complemented software systems because the work- checks on sellers, and sending customized emails
ers’ skills allowed them to perform tasks that algo- apologizing to buyers whose requests received
rithms couldn’t yet reliably manage, like writing zero quotes from sellers.
original blurbs about specific sellers. In other cases,
AllDone relied on workers to imitate software algo- THE SAN FRANCISCO TEAM further reduced
rithms, taking on functions that computers were the engineering burden that came with developing
technically capable of performing but that develop- new product features by having contractors sup-
ers in San Francisco believed would have been too port what AllDone’s software engineers called
costly or time-consuming to code themselves. “quick and dirty” tests. That is, Filipino workers
would manually execute algorithmic tasks that
BECAUSE ALLDONE’S search-engine optimiza- were under consideration for automation, provid-
tion strategy was yielding an ever-increasing volume ing a rough approximation of a project’s potential
of buyer requests, the company had to connect far before developers invested time and resources in
more buyers with sellers than ever before. Indeed, coding the software.
this matching process was AllDone’s core function. In one such case, the product team wanted to
But instead of expending scarce engineering determine whether they should add information
resources on matching buyers with sellers, AllDone from sellers’ profiles on the consumer-review
relied on staff in the Philippines to manually con- website Yelp to their AllDone profile pages. They
struct every introduction. This arrangement allowed theorized that this additional information would
software engineers to devote their energies to exper- enhance the perceived trustworthiness of AllDone
imenting with new projects that could “move the sellers and increase buyer requests. Yelp offers free
needle,” or significantly increase key metrics (such tools that allow software developers to embed Yelp
as the number of buyer requests) that VC investors users’ business information directly into their own
watched to assess the startup’s success. websites. However, Bill, the AllDone engineer in
Members of the Filipino matching team used a charge of the project, preferred not to spend his
Web portal that displayed the details of each new time learning how to use Yelp’s tools without first
buyer request. They began their work by vetting knowing whether the new feature was likely to
requests and deleting those that appeared to be succeed. So he devised a test whereby contractors
fraudulent (for example, a request placed by “Mickey in the Philippines manually searched for 9,000
Mouse”). The portal then provided team members AllDone sellers on Yelp and gathered information
with a rough, algorithmically generated list of local from their Yelp user profiles. Bill then put this
AllDone sellers who might be eligible to fulfill the information on relevant AllDone pages. Upon find-
request because they worked in relevant service ing that it did not have a statistically significant
categories. Workers would select all the sellers effect on buyer behavior, he abandoned the test.
whom they judged to be appropriate matches, and Throughout my research, AllDone had between
the sellers would then be automatically notified so four and eight software engineers on staff. Without
they could provide quotes for the service. The Fili- the Filipino team, the startup would have been
pino contractors effectively functioned as artificial forced to abandon some functions of its website
artificial intelligence, simulating the output of soft- and to reallocate some of its engineering resources
ware algorithms that had yet to be completed. toward building software infrastructure. The
AllDone’s users never knew that human workers, ­Filipinos’ reliable performance of important tasks
rather than a computer algorithm, had handcrafted helped the company achieve the precipitous
each introduction. To keep up with the rapid rise in growth demanded by venture capital investors to
request volume, the matching team more than dou- rapidly increase the company’s valuation. While
bled in size during the first phase of my research, the team in San Francisco threw parties for new
increasing from 30 to 68 people. Additionally, local recruits, enjoyed catered meals, and created the
managers cross-trained members of another divi- impression of technological wizardry, Filipino
sion on the matching function so that when user contractors were toiling behind the scenes.
activity peaked, more workers could be immediately AllDone’s story highlights the unseen but ongo-
mobilized to assist. ing role of human workers on the frontiers of auto-
There were many other processes that AllDone’s mation, and it demonstrates why it’s too soon to
engineers agreed could have been automated yet forecast a future of full automation or a world
were instead handled by contractors. These included without work. The interdependence between gen-
screening out sellers whose names appeared on the erously compensated software engineers in San
U.S. Department of Justice’s national sex-offender Francisco and low-cost contractors in the Philip-
registry, adding badges to seller profiles that passed pines suggests that advances in software automa-
a series of verifications, checking sellers’ profes- tion still rely not only on human labor, but also on
sional license numbers against relevant state data- global inequalities. 

APRIL 2024 SPECTRUM.IEEE.ORG 49


IDEAS FOR CODE: LEAN SOFTWARE

Why Bloat Is Still The response to Trifecta has been rather inter-
esting. The most common response to Trifecta so

Software’s Biggest far has been that I should use a whole bag of
Amazon Web Services to deploy it. This is an

Vulnerability exceedingly odd response to a project with the


clearly stated goal of providing stand-alone soft-
CONTINUED FROM P. 25 ware that does not rely on external services. I’m
not sure what is going on here.
Another reaction has been that I treat Docker
check thousands of them frequently. But we prefer software containers unfairly, and that you could
not to think about this. (Note that you should also definitely use containers for good. And I agree
not overshoot and needlessly re-implement every- wholeheartedly. But I also look at what people are
thing yourself to prevent dependencies. There are actually doing (also with other forms of containers
very good modules that likely are more secure than or virtual machines), and it’s not so great.
what you could type in on your own.)
The world is shipping far too much code, where A legislative push
we don’t even know what we ship, and we aren’t Recently, the European Union has launched three
looking hard enough (or at all) at what we do know pieces of legislation that should have software com-
we ship. panies on high alert: NIS2 for important services;
the Cyber Resilience Act for almost all commercial
You can write lean code today software and electronic devices; and a revamped
Writing has been called the process by which you The world Product Liability Directive that also extends to soft-
find out you don’t know what you are talking about. ships too ware. Legislation is always hard, and it remains to
Actually doing stuff, meanwhile, is the process by much code, be seen if they got it right. Incidentally, the E.U.’s
which you find out you also did not know what you Cyber Resilience Act explicitly tells vendors to min-
were writing about.
most of it by imize the attack surface.
In a small reenactment of Wirth’s Oberon Proj-
third parties, I want to end with an observation from Niklaus
ect, I too wrote some code to prove a point, and to sometimes Wirth’s 1995 paper:
reassure myself I still know what I am talking and unintended, “Time pressure is probably the foremost reason
writing about. Can you still make useful and modern most of it behind the emergence of bulky software. The time
software the old way? I decided to try to create a uninspected. pressure that designers endure discourages careful
minimalistic but full-featured image-sharing solu- planning. It also discourages improving acceptable
tion that I could trust. solutions; instead, it encourages quickly conceived
Trifecta is the result. It is actual stand-alone soft- software additions and corrections. Time pressure
ware that lets you use a browser to drag and drop gradually corrupts an engineer’s standard of quality
images for easy sharing. It has pained me for years and perfection. It has a detrimental effect on people
that I had to use Imgur for this purpose. Not only as well as products.”
does Imgur install lots of cookies and trackers in my If it is on the shoulders of software engineers to
browser, I also force these trackers onto the people make software lean, we should perhaps demand
who view the images that I share. If you want to more time for it.
self-host a Web service like this, you also don’t want The world ships too much code, most of it by
to get hacked. Most image-sharing solutions I found third parties, sometimes unintended, most of it
that you could run yourself are based on huge frame- uninspected. Because of this, there is a huge attack
works that I don’t trust too much for the reasons surface full of mediocre code. Efforts are ongoing
outlined above. to improve the quality of code itself, but many
I ended up with a grand total of 3 megabytes of exploits are due to logic fails, and less progress has
software. been made scanning for those. Meanwhile, great
To contrast, one other image-sharing solution strides could be made by paring down just how
ships as a 288-MB Docker image, although admit- much code we expose to the world. This will increase
tedly it looks better and has some more features. But time to market for products, but legislation is around
not 285 MB worth of them. Another comparison is the corner that should force vendors to take security
a Node.js-based picture-sharing solution, which more seriously.
clocks in at 1,600 dependencies, apparently totaling Trifecta is meant as a proof that you can deliver
over 4 million lines of JavaScript. a lot of functionality even with a limited amount
Note that Trifecta is not intended as a public site of code and dependencies. With effort and legisla-
where random people can share images, as that does tion, maybe the future could again bring sub-50-
not tend to end well. It is however very suitable for million-line garage-door openers. Let’s try to make
company or personal use. it happen. 

50 SPECTRUM.IEEE.ORG APRIL 2024


TAP.
CONNECT.
NETWORK.
SHARE.
Connect to IEEE–no matter where you are–with the IEEE App.
Stay up-to-date Schedule, manage, or Get geo and interest-based
with the latest news join meetups virtually recommendations

Read and download Create a personalized Locate IEEE members by location,


your IEEE magazines experience interests, and affiliations

Maintenance Engineer
in Port Angeles, WA
Require BS in Mechatronics.,
and training in
Harness the paper manufacturing
publishing Send Resume To:
power of McKinley Paper Company
One of the IEEE Access.
®

7850 Jefferson St NE
most influential Albuquerque, NM 87109
reference resources Ad paid by an Equal Opportunity Employer
for engineers • Receive high-quality,
around the world. rigorous peer review in only
For over 100 years, Proceedings of the IEEE 4 to 6 weeks
has been the leading journal for engineers
looking for in-depth tutorial, survey, and
review coverage of the technical
• Reach millions of global users
developments that shape our world.
Offering practical, fully referenced articles,
through the IEEE Xplore®
Proceedings of the IEEE serves as a bridge digital library by publishing Connect with thousands
to help readers understand important
technologies in the areas of electrical open access of your colleagues in the
engineering and computer science.
IEEE AuthorLab today!
Learn more at
To learn more and start ieeeaccess.ieee.org
your subscription today, visit
ieee.org/proceedings-subscribe

collabratec.ieee.org

APRIL 2024 SPECTRUM.IEEE.ORG 51


HISTORY IN AN OBJECT BY ALLISON MARSH

The Scientology Technically, the e-meter is a modified


ohm meter that detects changes in
But his applications never mentioned
the e-meter’s actual use as a religious
Machine the electrical resistance of the user’s
skin while the user holds two metal
screening device that supposedly
surfaced the user’s past-life traumas—
cylinders and a small electrical current even if those events were trillions of
Starting in the 1950s, L. Ron Hubbard, flows through them. Although he didn’t years ago—and exploited those read-
the founder of Scientology, promoted invent the e-meter, Hubbard inspired its ings to persuade potential converts.
the electropsychometer as a means creation, came up with a transistorized,
of measuring the souls, spirits, and battery-powered unit, and received FOR MORE ON THE E-METER, see
minds of potential church members. several patents for later versions. spectrum.ieee.org/pastforward-apr2024

WHIPPLE MUSEUM OF THE HISTORY OF SCIENCE/


UNIVERSITY OF CAMBRIDGE

52 SPECTRUM.IEEE.ORG APRIL 2024


YOU’VE
INSURED
YOUR
HOUSE
AND CAR.
NOW
WHAT
ABOUT
YOUR LIFE?
1-800-493-IEEE (4333)
To learn more*, visit IEEEinsurance.com/TL LIFE INSURANCE
*For information on features, costs, eligibility, renewability, limitations and exclusions.
Group Term Life Insurance is available only for residents of the U.S. (except territories), Puerto Rico and Canada (except Quebec). This is
underwritten by New York Life Insurance Company, 51 Madison Ave., New York, NY 10010 on Policy Form GMR. AMBA does not as a broker
with respect to Canadian residents and acts solely as an Administrator on behalf of New York Life.
Association Member Benefits Advisors, LLC.
In CA d/b/a Association Member Benefits & Insurance Agency
CA License #0I96562 • AR License #100114462
Program Administered by AMBA Administrators, Inc.
103137 (4/24) Copyright 2024 AMBA. All rights reserved
MATLAB
FOR AI
Boost system design and simulation with explainable and
scalable AI. With MATLAB and Simulink, you can easily train
and deploy AI models.

mathworks.com/ai

© The MathWorks, Inc.

You might also like