Learning to Code With A Particle Accelerator What to Do With Spent FOR THE

a Copilot AI upends the for Chipmaking Could Solar Panels A startup TECHNOLOGY
INSIDER
computer science classroom it extend Moore’s Law? explores clean recycling
P. 5 P. 28 P. 34 JULY 2024

Going Vertical
How we’ll get to a 1 trillion transistor GPU
 VOLUME 61 / ISSUE 7 JULY 2024

Could a Particle
Accelerator Save
Moore’s Law?
Future fabs will need superbright sources
of extreme ultraviolet light. By John Boyd 28
A Trillion-Transistor
GPU
22 The Psychiatrist in
Our Pockets
42 HANDS ON
Monitor unseen migrations
16

with a mic and AI.

AI won’t get what it needs Apps track mood swings
without new chip technology. by passively collecting
By Mark Liu & H.-S. Philip Wong smartphone data. CAREERS 19
By Gwendolyn Rak Aakhilesh Singhania speeds
Spent but
Not Trashed
34 hybrid race cars to the
finish line.

A startup recovers valuable EDITOR’S NOTE 2

materials from old solar panels. 5 QUESTIONS 21
When chip devices can’t scale
By Emily Waltz & Scott Best preps for
down, the only way to go is up.
Luigi Avantaggiato postquantum cryptography.

NEWS 5
PAST FORWARD 48
Teaching Coding With AI
A Work of Design Fiction
ON THE COVER: Spacecraft Derring-Do
KEK

Illustration by Optics Lab Cheap, Portable MRI

JULY 2024 [Link] 1

 EDITOR’S NOTE BY HARRY GOLDSTEIN

Moore on
Chip Scaling
Scaling compute to satiate AI’s appetite
will take extreme measures

F
ifty years ago, DRAM inventor and IEEE “I believe While the industry’s ability to affordably make
Medal of Honor recipient Robert Dennard Wong and smaller devices has certainly slowed, Moore believes
created what essentially became the semi- that scaling has a few tricks up its sleeve yet. In
Liu want
conductor industry’s path to perpetually addition to brighter light sources like the one KEK
increasing transistor density and chip performance.
young, is working on, future complementary field-effect
That path became known as Dennard scaling, and technically transistors (CFETs) will build two transistors in the
it helped codify Gordon Moore’s postulate about minded space of one.
device dimensions shrinking by half every 18 to 24 people to In the shorter term, Moore says stacking
months. For decades it compelled engineers to push understand chips is the most effective way to keep increasing
the physical limits of semiconductor devices. the the amount of logic and memory you can throw at
But in the 21st century, when Dennard scaling importance a problem.
began running out of juice, chipmakers have turned of keeping “There are always going to be functions in a
to exotic solutions like extreme ultraviolet (EUV) CPU or GPU that don’t scale as well as core pro-
lithography systems to try to keep Moore’s Law
semi- cessor logic. Increasingly, it doesn’t make sense to
on pace. On a visit to GlobalFoundries in Malta,
conductor try to keep building all these parts using the core
N.Y., in 2017 to see the company install its first EUV advances logic’s bleeding-edge chip processes,” Moore says.
system, Senior Editor Samuel K. Moore asked one going and “It makes more sense to build each part with its
expert what the fab would need to achieve even to make best, most economical process, and put them back
smaller device dimensions. “We’d probably have them want together as a stack, or at least in the same package.”
to build a particle accelerator under the parking lot,” to be part of To meet the demands of the booming AI sector,
the man joked. The idea seemed so fantastic that it that effort,” makers of GPUs will need to stack up. When
stuck with Moore. Moore says. former Taiwan Semiconductor Manufacturing
So when Tokyo-based tech journalist John Boyd Co. chairman Mark Liu and TSMC chief scientist
recently pitched a story about an effort to harness Robert Dennard H.-S. Philip Wong wanted to get their message out
a linear accelerator as an EUV light source, Moore passed away about the future of CMOS, they approached Moore.
was excited. Boyd’s visit to the High Energy Accel- this past April. The result is “The Path to a 1-Trillion-Transistor
erator Research Organization, known as KEK, in To read our profile GPU” [p. 22]. In addition to Wong’s corporate
of Dennard upon
Tsukuba, Japan, became the basis for “Is the Future role, he’s also an academic. One of the worries he’s
his receipt of
of Moore’s Law in a Particle Accelerator?” [p. 28]. As the IEEE Medal of repeatedly expressed to Moore is that AI and soft-
he reports, KEK’s system generates light by “boost- Honor in 2009, ware generally are pulling talent away from semi-
PORTRAIT BY SERGIO ALBIAC

ing electrons to relativistic speeds and then deviat- use this QR code. conductor engineering.
ing their motion in a particular way.” “I believe Wong and Liu want young, techni-
So far, KEK researchers have managed to cally minded people to understand the importance
blast a 17-megaelectron-volt electron beam to of keeping semiconductor advances going and to
produce bursts of 20-micrometer infrared light, make them want to be part of that effort,” Moore
a ways away from the needed 13.5 nanometers. says. “They want to show that semiconductor engi-
But the KEK team is optimistic about their neering has a career-long future despite much talk
technology’s prospects. of the death of Moore’s Law.”

2 [Link] JULY 2024 Illustration by Optics Lab

Simulate real-world
designs, devices,
Innovate
and processes faster.
with COMSOL Test more design iterations

Multiphysics® before prototyping.

[Link]/feature/multiphysics-innovation
Innovate
smarter.
Analyze virtual prototypes and
develop a physical prototype
only from the best design.

Innovate with
multiphysics
simulation.
Base your design decisions
on accurate results with
software that lets you study
unlimited multiple physical
effects on one model.
 CONTRIBUTORS

 LUIGI AVANTAGGIATO
Avantaggiato is a photographer
based in Rome. For the photo essay
EDITOR IN CHIEF Harry Goldstein, [Link]@[Link] IEEE BOARD OF DIRECTORS
“Spent but Not Trashed” [p. 34] PRESIDENT & CEO Thomas M. Coughlin, president@[Link]
EXECUTIVE EDITOR Jean Kumagai, [Link]@[Link]
he spent hours inside a shipping +1 732 562 3928 Fax: +1 732 981 9515
MANAGING EDITOR Elizabeth A. Bretz, [Link]@[Link]
container photographing a PRESIDENT-ELECT Kathleen A. Kramer
CREATIVE DIRECTOR TREASURER Gerardo Barbosa
pilot-scale solar-panel-recycling SECRETARY Forrest D. Wright
Mark Montgomery, [Link]@[Link]
system. “It was very hot, and PAST PRESIDENT Saifur Rahman
DIRECTOR OF DIGITAL INNOVATION
there was no space to turn around,” Erico Guizzo, eguizzo@[Link] VICE PRESIDENTS
says Avantaggiato. “So that was EDITORIAL DIRECTOR, CONTENT DEVELOPMENT
Rabab Kreidieh Ward, Educational Activities; Deepak Mathur,
Member & Geographic Activities; Sergio Benedetto,
a limitation but also an opportunity Glenn Zorpette, [Link]@[Link]
Publication Services & Products; Manfred J. Schindler,
to exercise this kind of photography.” SENIOR EDITORS Technical Activities; James E. Matthews, President, Standards
Evan Ackerman (Digital), ackerman.e@[Link] Association; Keith A. Moore, President, IEEE-USA
The constraints inspired him to
Stephen Cass (Special Projects Director), cass.s@[Link]
focus on the minutest details of the Samuel K. Moore, [Link]@[Link]
DIVISION DIRECTORS
Yong Lian (I); Kevin L. Peterson (II); Stefano Bregni (III); Alistair
process and the photovoltaic Tekla S. Perry, [Link]@[Link]
P. Duffy (IV); Christina M. Schober (V); Kamal Al-Haddad (VI);
Eliza Strickland, [Link]@[Link]
materials recovered. Christopher E. Root (VII); Leila De Floriani (VIII); Aylin Yener (IX);
ART & PRODUCTION Stephanie M. White (X)
DEPUTY ART DIRECTOR Brandon Palacio, [Link]@[Link]
REGION DIRECTORS
PHOTOGRAPHY DIRECTOR Randi Klett, [Link]@[Link]
 JOHN BOYD ONLINE ART DIRECTOR Erik Vrielink, [Link]@[Link]
Bala S. Prasanna (1); Andrew D. Lowery (2); Eric Grigorian (3);
Vickie A. Ozburn (4); Anthony M. Francis (5); Kathy Herring Hayashi
PRINT PRODUCTION SPECIALIST
Boyd, based in Kawasaki near Tokyo, Sylvana Meneses, [Link]@[Link]
(6); Thamir F. Murad (7); Vincenzo Piuri (8); Jenifer P. Castillo
is a longtime contributor to IEEE Rodriguez (9); ChunChe Fung (10)
MULTIMEDIA PRODUCTION SPECIALIST
Spectrum. On page 28, he describes Michael Spector, [Link]@[Link] IEEE STAFF
EXECUTIVE DIRECTOR & COO Sophia A. Muirhead
an effort to ensure the future of NEWS MANAGER Michael Koziol, [Link]@[Link]
+1 732 562 5400, [Link]@[Link]
chipmaking that is based at the SENIOR ASSOCIATE EDITOR CHIEF OF STAFF TO THE EXECUTIVE DIRECTOR Kelly Lorne
Margo Anderson, [Link]@[Link] +1 732 562 6011, [Link]@[Link]
Japanese particle accelerator lab,
ASSOCIATE EDITORS GENERAL COUNSEL & CHIEF COMPLIANCE OFFICER
known as KEK. He was impressed by Dina Genkina, [Link]@[Link] Anta Cissé-Green +1 212 705 8927, [Link]-green@[Link]
the scale of the research complex, Willie D. Jones (Digital), [Link]@[Link] INTERIM MANAGING DIRECTOR,
TECHNICAL ACTIVITIES Kenneth Gilbert
which spans 5 square kilometers. Emily Waltz, [Link]@[Link]
+1 732 562 3856, [Link]@[Link]
“If a particle-accelerator-based SENIOR COPY EDITOR Joseph N. Levine, [Link]@[Link]
CHIEF MARKETING OFFICER Karen L. Hawkins
COPY EDITOR Michele Kogon, [Link]@[Link]
system is going to become key to EDITORIAL RESEARCHER Alan Gardner, [Link]@[Link]
+1 732 562 3964, [Link]@[Link]
MANAGING DIRECTOR, PUBLICATIONS Steven Heffner
the chip industry, then KEK is surely ADMINISTRATOR Jose Zambrano, [Link]@[Link] +1 212 705 8958, [Link]@[Link]
the place where it can be made to DIRECTOR, PRODUCT AND AUDIENCE DEVELOPMENT STAFF EXECUTIVE, CORPORATE ACTIVITIES Donna Hourican
happen,” he says. Tim Warder, [Link]@[Link] +1 732 562 6330, [Link]@[Link]
MANAGING DIRECTOR, CONFERENCES,
AUDIENCE DEVELOPMENT MANAGER
EVENTS AND EXPERIENCES Marie Hunter,
Laura Bridgeman, [Link]@[Link]
+ 1 732 465 5889, [Link]@[Link]
 MARK LIU & CONTRIBUTING EDITORS Rina Diane Caballar, Robert N. Charette, MANAGING DIRECTOR, MEMBER & GEOGRAPHIC ACTIVITIES
Charles Q. Choi, Tom Clynes, Peter Fairley, Edd Gent, W. Wayt Gibbs,
H.-S. PHILIP WONG Mark Harris, Lucas Laursen, Allison Marsh, Julianne Pepitone,
Cecelia Jankowski +1 732 562 5504, [Link]@[Link]
MANAGING DIRECTOR, EDUCATIONAL ACTIVITIES
Liu was, until recently, chairman Matthew S. Smith, Lawrence Ulrich Jamie Moesch +1 732 562 5514, [Link]@[Link]
THE INSTITUTE MANAGING DIRECTOR, IEEE STANDARDS Alpesh Shah
of Taiwan Semiconductor +1 732 465 6467, [Link]@[Link]
EDITOR IN CHIEF Kathy Pretz, [Link]@[Link]
Manufacturing Co. From 2013 to ASSOCIATE EDITOR Joanna Goodrich, [Link]@[Link]
CHIEF FINANCIAL OFFICER Thomas R. Siegert
+1 732 562 6843, [Link]@[Link]
2018 he was president and co-CEO SUBSCRIPTION SERVICES CHIEF INFORMATION AND DIGITAL OFFICER Jeff Strohschein
of the foundry giant. H.-S. Philip IEEE Members member-services@[Link] +1 732 562 5332, [Link]@[Link]
Non-members +1 866 363 2304
Wong is a professor in the school of CHIEF HUMAN RESOURCES OFFICER Cheri N. Collins Wideman
DIRECTOR, PERIODICALS PRODUCTION SERVICES Peter Tuohy +1 732 562 5335, [Link]@[Link]
engineering at Stanford University
ADVERTISING PRODUCTION MANAGER MANAGING DIRECTOR, IEEE-USA Russell T. Harrison
and chief scientist at TSMC. On page Felicia Spagnoli, [Link]@[Link] +1 202 530 8326, [Link]@[Link]
22, the pair argue that semiconductor ADVERTISING PRODUCTION +1 732 562 6334 IEEE PUBLICATION SERVICES & PRODUCTS BOARD
technology is the unsung hero of EDITORIAL ADVISORY BOARD, IEEE SPECTRUM Sergio Benedetto, Chair; Derek Abbott, Nazanin Bassiri-Gharb,
Stefano Galli, Maria Sabrina Greco, Lawrence O. Hall, James
today’s AI boom. “Advances in Harry Goldstein, Chair; Ella M. Atkins, Yedukondala N. Dwith Chenna,
Sangyeun Cho, Hugh Durrant-Whyte, Matthew Eisler, Shahin Farshchi, Irvine, Fabrice Labeau, Yong “Peter” Lian, Aleksandar Mastilovic,
artificial intelligence are now and Paolo Montuschi, Annette Reilly, Anna Scaglione, Gianluca Setti,
Alissa Fitzgerald, Benjamin Gross, Carlos Gutierrez, Lawrence O. Hall,
will continue to be gated by progress Daniel Hissel, Jason K. Hui, Benjamin Kroposki, Michel M. Maharbiz, Gaurav Sharma, Sarah Spurgeon, Steve Yurkovich, Weihua Zhuang

of the foundational semiconductor Somdeb Majumdar, Ramune Nagisetty, Paul Nielsen, Sofia Olhede, IEEE OPERATIONS CENTER
Amit K. Singh, Christoph Stiller, Mini S. Thomas, Wen Tong, Haifeng 445 Hoes Lane, Box 1331, Piscataway, NJ 08854-1331 U.S.A.
technologies,” says Wong. Wang, Boon-Lock Yeo Tel: +1 732 981 0060 Fax: +1 732 981 1721
EDITORIAL ADVISORY BOARD, THE INSTITUTE
Kathy Pretz, Chair; Qusi Alqarqaz, Stamatis Dragoumanos, IEEE SPECTRUM (ISSN 0018-9235) is published monthly by
 GWENDOLYN RAK Madeleine Glick, Harry Goldstein, Francesca Iacopi, Meenakshi Jindal, The Institute of Electrical and Electronics Engineers, Inc. All rights
Lawrence O. Hall, Shashi Raj Pandey, John Purvis, Chenyang Xu reserved. © 2024 by The Institute of Electrical and Electronics
Rak is a freelance journalist and Engineers, Inc., 3 Park Avenue, New York, NY 10016-5997, U.S.A. Volume
MANAGING DIRECTOR, PUBLICATIONS Steven Heffner
former intern at Spectrum. In this EDITORIAL CORRESPONDENCE
No. 61, Issue No. 7. The editorial content of IEEE Spectrum magazine
does not represent official positions of the IEEE or its organizational
issue she writes about smartphone IEEE Spectrum, 3 Park Ave., 17th Floor, units. Canadian Post International Publications Mail (Canadian
sensors that passively track a New York, NY 10016-5997 TEL: +1 212 419 7555 Distribution) Sales Agreement No. 40013087. Return undeliverable
BUREAU Palo Alto, Calif.; Tekla S. Perry +1 650 752 6661 Canadian addresses to: Circulation Department, IEEE Spectrum, Box
person’s mood [p 42]. “I’m really
ADVERTISING INQUIRIES Naylor Association Solutions, 1051, Fort Erie, ON L2A 6C7. Cable address: ITRIPLEE. Fax: +1 212 419
interested in sensors for health, and Erik Albin +1 352 333 3371, ealbin@[Link] 7570. INTERNET: spectrum@[Link]. ANNUAL SUBSCRIPTIONS:
I wanted to dive into applications REPRINT SALES +1 212 221 9595, ext. 319 IEEE Members: $21.40 included in dues. Libraries/institutions: $399.
POSTMASTER: Please send address changes to IEEE Spectrum,
in mental health,” she says. “It’s REPRINT PERMISSION / LIBRARIES Articles may be photocopied ℅ Coding Department, IEEE Service Center, 445 Hoes Lane, Box 1331,
interesting how wide the range of for private use of patrons. A per-copy fee must be paid to the Copyright Piscataway, NJ 08855. Periodicals postage paid at New York, NY,
Clearance Center, 29 Congress St., Salem, MA 01970. For other copying
approaches are–voice monitoring and additional mailing offices. Canadian GST #125634188. Printed at
or republication, contact Managing Editor, IEEE Spectrum. 3401 N. Heartland Dr, Liberty, MO 64068, U.S.A. Printed at Spenta
and tracking typing or location are all COPYRIGHTS AND TRADEMARKS Multimedia Pvt. Ltd., Plot 15, 16 & 21/1, Village Chikhloli, Morivali, MIDC,
very different kinds of sensors. IEEE Spectrum is a registered trademark owned by The Institute of Ambernath (West), Dist. Thane. IEEE Spectrum circulation is audited
But they all tackle the same issue.” 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. [Link]

4 [Link] JULY 2024

 THE LATEST DEVELOPMENTS IN TECHNOLOGY, ENGINEERING, AND SCIENCE JULY 2024

I
ARTIFICIAL INTELLIGENCE t’s clear that generative AI is trans-
forming the software-development

Professors Rethink How industry. AI-powered coding tools

are assisting programmers in their
workflows, while the number of jobs in
They Teach Coding Students AI continues to increase. But the shift is
also evident in academia—one of the
embrace AI copilots, teachers major avenues through which the next
generation of software engineers is

shift to problem solving learning how to code.

Computer science students are embrac-
ing the technology, using generative AI to
help them understand complex concepts,
ALAMY

BY RINA DIANE CABALLAR summarize complicated research papers,

JULY 2024 [Link] 5

 NEWS

brainstorm ways to solve a problem, come the School of Computing at the National at Clarkson University in Potsdam, N.Y.
up with new research directions, and, of University of Singapore. “Given that Another vital expertise is problem
course, learn how to code. large language models are evolving rap- decomposition. “This is a skill to know
“Students are early adopters and have idly, we are still learning how to do this.” early on because you need to break a large
been actively testing these tools,” says problem into smaller pieces that an LLM
Johnny Chang, a teaching assistant at The fundamentals and skills themselves can solve,” says Leo Porter, an associate
Stanford University pursuing a master’s are evolving. Most introductory computer teaching professor of computer science
degree in computer science. He founded science courses focus on code syntax and at the University of California, San Diego.
the AI x Education conference in 2023, a getting programs to run, and while know- “It’s hard to find where in the curriculum
virtual gathering of students and educators ing how to read and write code is still that’s taught—maybe in an algorithms
to discuss the impact of AI on education. essential, testing and debugging—which or software engineering class, but those
Educators are also experimenting aren’t commonly part of the syllabus—now are advanced classes. Now, it becomes a
with generative AI. But they’re grappling need to be taught more explicitly. priority in introductory classes.”
with techniques to adopt the technology “We’re seeing a little upping of that As a result, educators are modifying
while still ensuring students learn the skill, where students are getting code their teaching strategies. “I used to have
foundations of computer science. snippets from generative AI that they need this singular focus on students writing
“It’s a difficult balancing act,” says to test for correctness,” says Jeanna Mat- code that they submit, and then I run test
Ooi Wei Tsang, an associate professor in thews, a professor of computer science cases on the code to determine what their

The State of AI in 2024 Institute for Human-Centered Artificial Intelligence (HAI),

covers trends in R&D, technical performance, responsible AI,
the economy, education, policy and governance, diversity, and

E
ach year, the AI Index lands on virtual desks with a public opinion. Here we’ve compiled four charts that speak to
louder virtual thud. This year, its 502 pages are a tes- the state of AI in 2024. —eliza strickland
tament to the fact that we’re in the midst of an AI For our full selection of 15 charts, see [Link]
boom. The report, published in April by the Stanford org/ai-index-2024

PRIVATE INVESTMENT ESTIMATED TRAINING COST

IN GENERATIVE AI, 2019–23 OF SELECT AI MODELS, 2017–23
TRAINING COST, US $, MILLIONS
TOTAL INVESTMENT, US $, BILLIONS

25 2023: GEMINI ULTRA: 191,400,000

2023: GPT-4: 78,352,034

20 2022: PaLM: 12,389,056

2021: MEGATRON-TURING NLG: 6,405,653

15
2020: GPT-3: 4,324,883

10 2023: LLAMA 2: 3,931,897

2019: ROBERTA LARGE: 160,018

5
2018: BERT-LARGE: 3,288

2017: TRANSFORMER: 930

0
2019 2020 2021 2022 2023 0 40 80 120 160 200

SOURCE: QUID, 2023 | CHART: 2024 AI INDEX REPORT SOURCE: EPOCH, 2023 | CHART: 2024 AI INDEX REPORT

GENERATIVE AI INVESTMENT SKYROCKETS: While FOUNDATION MODELS HAVE GOTTEN SUPEREXPENSIVE: Foun-
corporate investment was down overall last year, dation models—big generative AI models that allow for a
investment in generative AI went through the roof. variety of language or art tasks—are predominately coming
Nestor Maslej, editor in chief of this year’s from industry, because training one takes very deep
report, says the boom is indicative of a broader pockets. AI companies rarely reveal the expenses involved
trend as the world grappled with the new capa- in model training, so the AI Index estimated costs by
bilities and risks of generative AI systems like analyzing information gleaned from publications, press
ChatGPT and DALL-E 2. “Whether it’s in policy, releases, and technical reports. Particularly noteworthy:
in public opinion, or in industry,” says Maslej, Google’s 2017 Transformer model, which introduced the
people have responded to generative AI “with a lot architecture that underpins almost all of today’s large
more investment.” language models, was trained for only US $930.

6 [Link] JULY 2024

grade is,” says Daniel Zingaro, an associ- for us to teach higher-level thinking—for a model built off other people’s code and
ate professor of computer science at the example, how to design software, what is we’d recognize the ownership of that,”
University of Toronto Mississauga. “This the right problem to solve, and what are Porter says. “We also have to recognize
is such a narrow view of what it means the solutions. Students can spend more that models are going to represent the
to be a software engineer, and I just felt time on optimization, ethical issues, and bias that’s already in society.”
that with generative AI, I’ve managed to the user-friendliness of a system rather Adapting to the rise of generative AI
overcome that restrictive view.” than focusing on the syntax of the code.” involves students and educators working
Zingaro, who coauthored a book on together and learning from one another.
AI-assisted Python programming with But educators are also cautious given the For her colleagues, Matthews’s advice is
Porter, now has his students work in tendency of LLMs to hallucinate. “We need to “try to foster an environment where
groups and submit a video explain- to be teaching students to be skeptical of you encourage students to tell you when
ing how their code works. Through the results and take ownership of verifying and how they’re using these tools. Ulti-
these walk-throughs, he gets a sense of and validating them,” says Matthews. mately, we are preparing our students
how students use AI to generate code, Matthews adds that generative AI for the real world, and the real world is
what they struggle with, and how they “can short-circuit the learning process of shifting, so sticking with what you’ve
approach design, testing, and teamwork. students relying on it too much.” Chang always done may not be the recipe that
“It’s an opportunity for me to assess agrees that this overreliance can be a pitfall best serves students in this transition.”
their learning process of the whole and advises his fellow students to explore Porter is optimistic that the changes
software development [life cycle]—not possible solutions to problems by them- they’re applying now will serve students
just code,” Zingaro says. “And I feel like selves so they don’t lose out on that criti- well in the future. “There’s this long his-
my courses have opened up more, and cal-thinking or effective learning process. tory of a gap between what we teach in
they’re much broader than they used to “We should be making AI a copilot—not academia and what’s actually needed as
be. I can make students work on larger the autopilot—for learning,” he says. skills when students arrive in the indus-
and more advanced projects.” Other drawbacks include copyright try,” he says. “There’s hope on my part
Ooi echoes that sentiment, noting issues and bias. “I teach my students that we might help close the gap if we
that generative AI tools “will free up time about the ethical constraints—that this is embrace LLMs.”

CO2-EQUIVALENT EMISSIONS, TONNES EMPLOYMENT OF NEW AI PH.D.s IN THE UNITED

BY SELECT MACHINE LEARNING MODELS STATES AND CANADA BY SECTOR, 2010–22
AND REAL-LIFE EXAMPLES, 2020–23 NUMBER OF NEW AI PH.D. GRADUATES

400
 Industry  Government  Academia
GPT-3 (175B)

GOPHER (280B)

LLAMA 2 (70B)

OPT (175B)

CAR, AVERAGE INCLUDING FUEL, 1 LIFETIME

200
BLOOM (176B)

GRANITE (13B)

AMERICAN LIFE, AVERAGE OVER 1 YEAR

STARCODER (15.5B)

LUMINOUS-EXTENDED (30B)

HUMAN LIFE, AVERAGE OVER 1 YEAR

0
0 100 200 300 400 500 2010 2022

SOURCE: AI INDEX, 2024; LUCCIONI ET AL., 2022; SOURCE: CRA TAULBEE SURVEY, 2023 | CHART: 2024 AI INDEX REPORT
STRUBELL ET AL., 2019 | CHART: 2024 AI INDEX REPORT

LARGE LANGUAGE MODELS HAVE A HEFTY CARBON FOOTPRINT: NEW PH.D.S FLOCK TO INDUSTRY: That more grad-
The AI Index team estimated the carbon footprint of uates of AI programs are drawn to industry is
training certain large language models, including such hardly a surprise, given the amount of private
factors as model size (measured in billions of param- investment for generative AI, and the fact that
eters), data-center energy efficiency, and the carbon industry dominates in the production of AI foun-
intensity of energy grids. This chart does not include dation models. In 2022 (the most recent year for
emissions related to models’ “inference”—when they do the which the AI Index has data), 70 percent of new
work they’re trained for. While “per-query” emissions may AI Ph.D.s in North America took jobs in indus-
be low, the report notes, the total impact can surpass try. It’s a continuation of a trend over the last
training emissions if models are queried millions of decade, and it indicates an intensifying brain
times daily. drain from universities.

JULY 2024 [Link] 7

 NEWS

The pair of
spacecraft making
up the Proba-3
mission will test
high-precision
group flying,
as one craft
will create tiny
eclipses for the
other to study
the sun.

AEROSPACE 1868 total solar eclipse in India. Fifty years

later, British astronomers Frank Dyson

Satellites Team Up to Make and Arthur Eddington performed mea-

surements of the apparent shift of stars
during a 1919 solar eclipse as an early test
Mini-Eclipses Proba-3 will of Einstein’s general theory of relativity.
ESA will use Proba-3 to study solar
demonstrate formation flying astrophysics without any interference
from Earth’s atmosphere. The mission’s
coronagraph (aboard the craft of the
same name) will help to discern why the
BY ANDREW JONES solar corona is significantly hotter than
the sun itself, which could improve solar-
weather predictions.

T
he European Space Agency spacecraft will be positioned roughly 144 However, Proba-3’s precision-for-
will launch a mission in the meters apart. The aim, ESA says, is to use mation flying could be the mission’s
coming months to create arti- the Occulter to block out the sun from the most significant contribution toward
ficial solar eclipses—using two Coronagraph’s point of view. unlocking future aerospace break-
spacecraft that will position themselves Achieving this formation will allow throughs. During each orbit, as the two
in a precise formation while in orbit. the Coronagraph to study our star’s craft approach their farthest point from
ESA’s Proba-3 (PRoject for On-Board highly ionized, extremely hot atmo- Earth, they will align to create a roughly
Autonomy) includes a pair of spacecraft: sphere. More broadly, the two craft will 6-hour mini-eclipse from Coronagraph’s
the 300-kilogram Coronagraph space- demonstrate precision flying as a precur- point of view. The two spacecraft will use
craft and the 250-kg Occulter. The pair sor for more ambitious formation-flying radio links to communicate with one
are slated to launch on a Polar Satellite endeavors in the future. another and star trackers to determine
Launch Vehicle rocket developed by Solar eclipses, aside from being their attitudes.
P. CARRIL/ESA

the Indian Space Research Organiza- moments of stark beauty, are fleeting Additional coordination occurs as the
tion. After entering a highly elliptical opportunities to conduct science. Helium spacecraft approach the closest point to
orbit—600 kilometers at its closest point was first detected in the sun’s chromo- Earth in their orbit. Global navigation
and 60,350 km at its farthest—the two sphere by a French astronomer during a satellite system (GNSS) receivers will

8 [Link] JULY 2024

 pick up GPS signals, which, when com- JOURNAL WATCH that they vibrate only in a perpen-
bined with a dedicated relative naviga- dicular manner. This design
tion algorithm, will help determine the Sea Turtles Inspire results in a high sensitivity.
relative position of the two craft. Optical Zang and his colleagues
sensors on the Occulter will detect puls-
Heart-Monitor created a similarly T-shaped
ing LEDs on the Coronagraph to provide Design heart-sound sensor using a
data for finer measurements. tiny microelectromechanical
After all that, one final bit of techno- RESEARCHERS ARE KEEN to systems (MEMS) cantilever
logical wizardry will allow for millime- develop alternative, lower- beam sensor. As sound hits the
ter-scale precision between the two craft. cost solutions to monitor heart sensor, the vibrations cause
The Occulter will ping a laser at a corner health that people can use at deformations in its beam, and
cube retroreflector mounted on the Coro- home, in lieu of doctors using the fluctuations in the voltage
nagraph spacecraft, which will bounce stethoscopes or echocardio- resistance are then translated
the beam back. Using the data gathered, grams. A team in China has now into electrical signals.
Coronagraph can, the ESA says, control created a heart-monitoring The researchers first tested
and maintain millimeter-level accuracy system inspired by the auditory the sensor’s ability to detect
using 10-millinewton-scale cold gas systems of sea turtles. sound in lab tests, and then
thrusters aboard the craft. Junbin Zang, a lecturer at the tested its ability to moni-
“Guidance, navigation, and control has North University of China, and his tor heartbeats in two human
undergone a lot of development, and this is colleagues were intrigued when volunteers in their early 20s.
what we all want to demonstrate,” Damien they learned about the inner The results, described in a
Galano, Proba-3’s project manager, said workings of the sea turtle’s ears, study published 1 April in IEEE
during a press briefing on 3 April. which are able to detect low-fre- Sensors Journal, show that the
“We have an actual application, which quency signals, especially in the sensor can effectively detect
is observation of the corona. So, by 300- to 400-hertz range. the two phases of a heartbeat.
achieving really good observations of the “Heart sounds are also “The sensor exhibits excel-
corona, we would definitely demonstrate low-frequency signals, so the lent vibration characteristics,”
that all this equipment is working and low-frequency characteristics Zang says, noting that it has
that the technology is delivering actual of the sea turtle’s ear have a higher vibration sensitivity
science data,” Galano said, adding that provided us with great inspira- compared with other acceler-
Proba-3 formation-flying control algo- tion,” explains Zang. ometers on the market.
rithms and metrology systems could be At first glance, sea turtles However, the sensor
applied to future missions. don’t seem to have ears—they currently picks up a significant
For example, precisely controlled lack external ear canals. Instead amount of background noise,
Occulter-like spacecraft could be used their auditory system lies under which Zang says his team
with space telescopes to block light from a layer of skin and fat through plans to address in future work.
a distant star in order to directly detect any which the turtle picks up vibra- Ultimately, they are inter-
potential orbiting exoplanets. Constella- tions. As with humans, a small ested in integrating this novel
tions of spacecraft could, through interfer- bone vibrates as sounds hit it bioinspired sensor into devices
ometry, create large-scale observatories by and converts those vibrations to they have previously created—
collectively achieving large apertures and electrical signals that are sent to including portable handheld and
long focal lengths that would be impossi- the brain for processing. wearable versions and a rela-
ble with large solo satellites. But sea turtles also have a tively larger version for use in
Other applications include Earth unique slender, T-shaped conduit hospitals—for the simultaneous
observation, space-based gravitation- that encapsulates their ear bones, detection of electrocardiogram
al-wave detection, and a range of mis- restricting the movement of the and phonocardiogram signals.
sions in which two or more spacecraft similarly T-shaped ear bones so —Michelle Hampson
need to interact, such as rendezvous,
docking, and in-orbit servicing missions. Researchers in China
Before such complex projects can built a T-shaped heart-
sound sensor, inspired
be undertaken, Proba-3 needs to prove by the internal ear
NORTH UNIVERSITY OF CHINA

it can nail the basics. “That’s clearly the bones of sea turtles.
operational challenge,” said Dietmar Pilz,
ESA director of technology, engineering,
and quality. “To see how far we can get
the formation flying, what are the dis-
tances that we can achieve. This needs a
lot of operational expertise and software
from all the partners in this project.”
 NEWS

COMPUTING human brain—a significant improve-

ment over the previous record holder,

Brain-Emulating
Hala Point, which can emulate 1.15 bil-
lion neurons.
SpiNNcloud Systems, a startup based

Computer Reaches in Dresden, Germany, was founded in

2021 as a spin-off of the Dresden Uni-
versity of Technology. Its original chip,

the Market AI the SpiNNaker1, was designed by Steve

Furber, the principal designer of the
Arm microprocessor—the technol-

accelerators meet ogy that now powers most cellphones.

The SpiNNaker1 chip is already in use

neuromorphic models
by 60 research groups in 23 countries,
SpiNNcloud Systems says.
Neuromorphic computers promise
vastly lower energy consumption for
computations and better performance on
BY DINA GENKINA certain tasks, such as training recurrent
neural networks and processing sensor
data in real time. “The human brain is
the most advanced supercomputer in
the universe, and it consumes only 20
watts to achieve things that artificial

I
n May, at the ISC High Performance intelligence systems today only dream
conference in Hamburg, Germany, of,” says Hector Gonzalez, cofounder and
SpiNNcloud Systems announced that co-CEO of SpiNNcloud Systems. “We’re
its neuromorphic (or brain-emulat- basically trying to bridge the gap between
ing) supercomputer, the SpiNNcloud brain inspiration and artificial systems.”
Platform, was available for sale. The Aside from the sheer number of neu-
machine, which combines traditional AI rons that the SpiNNcloud Platform can
accelerators with neuromorphic com- simulate, a distinguishing feature of
puting capabilities, is the first commer- the system is its flexibility. Most neuro-
cially available neuromorphic computer. This board features 48 morphic computers emulate the brain’s
The largest version of the SpiNNcloud SpiNNaker2 chips working spiking nature: Neurons fire off electrical
together to enable both neuro-
Platform can simulate 10 billion neurons, morphic computing and tradi- spikes to communicate with the neurons
about one-tenth the number in the tional AI processing. around them. The actual mechanism of
these spikes in the brain is quite com-
plex, and neuromorphic hardware often
implements a specific simplified model.
However, the SpiNNcloud Platform can
implement a broad range of brain-spik-
ing models, as specific models are not
hardwired into its architecture.
Instead of looking at how each neuron
and synapse operates in the brain and
trying to emulate that from the bottom
up, Gonzalez says, SpiNNcloud’s
approach involves implementing key
performance features of the brain. “It’s
more about taking practical inspiration
from the brain, following particularly fas-
cinating aspects such as how the brain is
SPINNCLOUD SYSTEMS

energy proportional and how it is simply

highly parallel.”
To build hardware that is energy pro-
portional (meaning each portion draws
power only when it’s actively in use)
and highly parallel, the company started

10 [Link] JULY 2024

 with the SpiNNaker2 chips that compose
the computer. Each SpiNNaker2 chip
hosts 152 processing units, with each
processing unit having an Arm-based
microcontroller. Unlike its SpiNNaker1
predecessor, the SpiNNaker2 also comes
equipped with accelerators for use on
neuromorphic models and traditional
neural networks.
The processing units can stay
switched off unless a spike from a con-
nected unit or an external sensor event
triggers them to turn on. This enables
energy-proportional operation because
each unit draws power only as needed.
There is no central clock to coordinate
movements; instead, messages are routed
between units and across chips when
needed without having to wait their turn
in a clock cycle. Multiple messages can be
sent through different parts of the chip
at the same time, enabling parallelism.
Each chip is connected to six other chips,
and the whole system is connected in the
shape of a torus to ensure all connecting
wires are equally short.
The largest commercially offered
system is not only capable of emulating
10 billion neurons but also performing
0.3 billion billion operations per second BIOMEDICAL
of more traditional AI tasks, putting it on

Compact MRI Ditches

the scale of the top 10 largest supercom-
puters today.
Among the first customers of the
SpiNNaker2 system is a team at Sandia
National Laboratories in New Mexico,
Superconducting Magnets
which plans to use it for further research
on how neuromorphic systems can out-
The whole-body scanner can
perform traditional computer archi-
tectures and how they can perform plug into a standard wall outlet
fundamentally different computational
tasks, such as implementing various
models of the human brain. BY CHARLES Q. CHOI
For example, Fred Rothganger, senior
member of technical staff at Sandia, says

M
the flexibility of the SpiNNaker2 sys- agnetic resonance imaging artificial intelligence, the new scanner
tems will allow Sandia researchers to (MRI) has revolutionized requires only a compact 0.05-T magnet
study different types of artificial neural health care by providing and can run off a standard wall power
SCIENCE PHOTO LIBRARY/GETTY IMAGES

networks that may have relevance to radiation-free, noninvasive outlet, requiring only 1,800 watts during
the human brain. “[Older neuromor- 3D medical images. However, most MRI operation. The researchers say their new
phic systems], of course, can run on a scanners are room-size, complex, and machine can produce clear, detailed images
general-purpose computer. But those power hungry—consuming 25 kilowatts on a par with those from high-power MRI
general-purpose computers are not nec- or more to produce magnetic fields up to scanners currently used in clinics.
essarily designed to efficiently handle the 1.5 teslas strong. As a result, scanners are “It’s the beginning of a multidisci-
kind of communication patterns that go typically limited to specialized depart- plinary endeavor to advance an entirely
on inside a spiking neural network. With ments in hospitals. new class of simple, patient-centric, and
[the SpiNNaker2 system] we get the ideal A University of Hong Kong team has computing-powered point-of-care diag-
combination of greater programmability unveiled a low-power, highly simplified, nostic imaging device,” says Ed Wu, a
plus efficient communication.” full-body MRI scanner. With the help of professor and chair of biomedical engi-

JULY 2024 [Link] 11

 NEWS

MRI receive coil

This simplified MRI
scanner uses weak
0.05-tesla magnets.
As a result, it
Sources doesn’t require
of EMI expensive, bulky
shielding for either
its own magnets or
for external sources
of electromagnetic
interference (EMI).
Coils around the
scanner detect EMI,
and deep-learning
220-volt, 20-amp models filter the
power supply EMI interference out
sensing
coils so that the weak
magnets can still
generate sharp
images. The entire
scanner can be
powered by a stan-
dard wall outlet.

neering at the University of Hong Kong. the work, in a review of the new study. those of conventional MRI scanners.
Wu and the rest of the research team pub- The Hong Kong team’s whole-body The researchers used standard off-
lished their work on 10 May in Science. ULF MRI scanner places a patient the-shelf electronics. All in all, they
More than 150 million MRI scans are between two permanent neodymium fer- estimate hardware costs at about US
conducted worldwide annually, according rite boron magnet plates. Although these $22,000. (Entry-level MRI scanners
to the Organization for Economic Cooper- permanent magnets are far weaker than can start at $225,000, and advanced
ation and Development. However, despite superconductive magnets, they are low- machines can cost $500,000 or more.)
five decades of development, clinical MRI cost, readily available, and don’t need to The prototype scanner’s magnet
procedures remain out of reach for more be cooled to frigid temperatures using assembly is heavy, weighing about 1,300
than two-thirds of the world’s population, expensive liquid helium. kilograms. (This is lightweight, how-
especially in low- and middle-income The new machine consists of two ever, compared with a typical clinical
countries. For instance, the United States units, each roughly the size of a hospital MRI scanner, which can weigh up to 17
has about 40 scanners per million people; gurney. One unit houses the MRI device, tonnes.) The scientists note that optimiz-
as of 2021, Tanzania had fewer than one while the other supports the patient’s ing the hardware could reduce the magnet
scanner per 10 million people. body as it slides into the scanner. assembly’s weight to about 600 kg, which
These disparities largely stem from the To account for radio interference from would make the entire scanner mobile.
high costs and specialized settings required both the outside environment and the The researchers say their new device
for standard MRI scanners. They use ULF MRI’s own electronics, the scientists is not meant to replace conventional
powerful superconducting magnets that deployed 10 small sensor coils around the high-magnetic-field MRI. A 2023 study
require a lot of space, power, and special- scanner and inside the electronics cabi- notes that next-generation MRI scanners
ized infrastructure. They also need rooms net to help the machine detect potentially using powerful 7-T magnets could yield a
shielded from radio interference, which disruptive radio signals. The researchers resolution of just 0.35 millimeters. ULF
restricts their mobility and hampers their also employed deep-learning AI methods MRI could instead complement existing
availability in nonshielded environments. to help reconstruct images even in the MRI by going to places that can’t host
Scientists have already been explor- presence of strong noise. They say this standard MRI devices, such as inten-
ing low-cost MRI scanners that operate eliminates the need for shielding against sive-care units and community clinics.
at ultralow-field (ULF) magnetic-field radio waves, making the new device far In an email to IEEE Spectrum, Anazodo
strengths (less than 0.1 T). Current ULF more portable than a conventional MRI. says this new Hong Kong work is just one
MRI scanners often rely on AI to help In tests on 30 healthy volunteers, the of a number of exciting ULF MRI scan-
reconstruct images from what signals device captured detailed images of indi- ners under development. She also points
they gather using relatively weak mag- viduals’ brains, spines, abdomens, hearts, to researchers at the University of Sas-
netic fields. Until now, these devices lungs, and extremities. Scanning each of katchewan, in Canada, who are develop-
were limited to solely imaging the brain, these targets took 8 minutes or less for ing a device that is potentially even lighter,
extremities, or single organs, notes Udunna image resolutions of roughly 2-by-2-by-8 cheaper, and more portable than the Hong
Anazodo, an assistant professor of neurol- cubic millimeters. In Anazodo’s review, Kong machine, which they are researching
ogy and neurosurgery at McGill Univer- she notes that the new machine pro- for use inwhole-body imaging on the Inter-
sity in Montreal who did not take part in duced image qualities comparable to national Space Station.

12 [Link] JULY 2024 Illustration by Chris Philpot

 m b e r
Me
o u n t s
Disc

EXCLUSIVE
OFFERS &
DISCOUNTS.
ANOTHER BIG
PLUS TO YOUR
IEEE MEMBERSHIP.

[Link]/discounts
 THE BIG PICTURE

Robotic Bees’
Swarm
Autonomy
By Willie D. Jones

The latest buzz in

biomimetic robots
is a robotic bee
like this one. The
BionicBee, produced
by the automation-
technology company
Festo, based in
Esslingen am Neckar,
Germany, is an order of
magnitude bigger than
a real bee. For a sense
of a BionicBee’s size,
imagine a bluejay or an
oversized hummingbird,
albeit one whose
wings flap forward and
back (closer to the
movements of a real
bee) rather than up and
down. The company says
its swarm technology
will enable the drone
to operate autonomously
while it coordinates
its flight path
with other “bees,”
using ultrawideband
beacons whose signals
let it keep track
of its position in
space and avoid
midair collisions.
PHOTOGRAPH BY FESTO

14 [Link] JULY 2024

JULY 2024 [Link] 15
 TECH TO TINKER WITH

Inexpensive parts
can be assembled in
a bucket to make an
effective acoustic
bird detector.

Birding at Night Detect

B
irding is booming. You may
realize your local nature spots
are especially busy during sea-
migrating birds with a plastic sonal migrations, when birds
move between their summer and winter
dish and a cheap microphone grounds. Species that you had been
noticing disappear may have been
replaced by ones that hadn’t been there
before. Or you may have seen migrating
BY DAVID SCHNEIDER birds on the wing—say, a flock of geese
flying in their famous V-formation. Even
if you’re not a dedicated birder, you’ve

16 [Link] JULY 2024 Illustrations by James Provost

 JULY 2024

probably made such observations Incoming sounds are amplified using a parabolic dish made from
throughout your life. So it might come as a plastic bird-feeder cover [top]. A microphone attached at the
focal point of the dish is connected to a preamplifier [middle
a surprise to learn that you’ve been miss- left], which in turns feeds an external sound card [middle right],
ing out on most of this action, which which connects to a host computer via USB. A large gel-acid battery
takes place at night. But, as I discovered, [bottom] provides plenty of power for long-term monitoring.
with some simple electronics and the
right software, you can identify noctur-
nal migrators with ease!
Birds migrate at night for a few rea-
sons. One is that it helps them to avoid
predators. Also, it allows them to use the
stars for navigation. A less obvious
reason is that traveling at night helps
these birds avoid heat stress. And the
night air tends to be less turbulent,
making flying easier.
These nighttime flights are largely
invisible. If you’re lucky, you might view
telltale silhouettes by training a tele-
scope on the moon. But during the
Second World War, scientists realized
that they could detect migrating birds
using radar. Since then, ornithologists’
radar studies, particularly those that use
modern weather radar, have proved
immensely successful in showing where
and when birds migrate at night.
Radar echoes cannot, however, iden-
tify species. But there is another tech-
nique that can: recording the calls that
birds make during their nocturnal
travels.
When ornithologist Richard Graber
and electrical engineer William Cochrane
made the first systematic recordings of
nocturnally migrating birds in 1957, they
used a microphone attached to a 2-meter-
wide upward-facing parabolic dish. But
you can get by today with a far more
modest setup.
You could, for example, reproduce the
gear designed by Bill Evans. On his web-
site he sells a microphone and preamp
for this purpose along with guidance on
how to package the equipment so that it
will hold up to the elements. I explored
a different approach, though, one that
seemed easier and cheaper.
Evans’s preamp is designed to be
insensitive to low frequencies, as these

JULY 2024 [Link] 17

 HANDS ON

20
18
16
14
Kilohertz

12
10
8
6
4
2
0
0 5 10 15 20
Seconds from start

This audiogram reveals the presence of bird calls. I uploaded the data to a server maintained by Cornell
University that then uses AI to quickly identify the species.

aren’t of interest when you’re recording for weeks at a time. Following Evans’s out bird chirps. There is no shortage of
bird calls. I figured that this feature advice, I housed everything in a 2-gallon local birds chirping during the day, but
wasn’t that important, so after testing a paint bucket, stretching some plastic after sunset their ornithological cacoph-
few inexpensive options for the micro- wrap over the top to keep rain out. ony abates, returning again some time
phone and preamplifier, I chose one on I placed my bucket o’ electronics on before dawn.
Amazon for just US $9. the roof of my porch, running a USB The interval in between is where I
This circuit uses the venerable NE5532, cable from the sound card, out the side went hunting for the sound of migrating
a low-noise, low-distortion dual op-amp of the bucket, and into my office through birds. And after 10 days or so, I found my
design that’s been used in professional a window. Then I plugged it into a Win- quarry: chirping that started shortly after
recording equipment since 1979. To make dows laptop onto which I had installed midnight, rising in volume for a few min-
it directional, I unsoldered the condenser Raven Lite, acoustic-spectrogram soft- utes before fading away.
microphone from the board, attached a ware made available for free by the Cor- Using Audacity, a free audio editor, I
short length of audio cable to nell Lab of Ornithology. extracted a few seconds of the loudest
it, and mounted it at the focal Using Raven Lite to com- chirping and uploaded the file to Birdnet,
point of an 8-inch-diameter After 10 days pute spectrograms showed where the good folks at the Cornell Lab
parabolic dish—or, well, a or so, I found just how sensitive this of Ornithology provide a tool for identi-
reasonable approximation of my quarry. arrangement is. I could easily fying bird calls. It indicated that the spe-
a parabolic dish, as it’s actu- view, for example, the effect cies I had recorded was the killdeer, a
ally a rain guard for bird feed- of completely inaudible type of bird found throughout the conti-
ers. You could also purchase a sounds created by rubbing my thumb nental United States, some populations
16-inch-diameter one, but the 8-inch and forefinger together a couple of of which are migratory.
dish served me admirably. meters away from the microphone. Additional nights of recording and
I found the focal point of this dish With the gear in place outside, I started scanning spectrograms turned up other
through trial and error and ran the recording at night, beginning in early sounds that appeared to be from other
output of the preamp into an old Creative March, arranging the Raven Lite software kinds of birds on the move, including
Labs Sound Blaster external sound card, to record a series of 1-hour sound files. such migratory species as the dark-eyed
which had been collecting dust on my The great thing about Raven Lite is that junco and Kentucky warbler.
shelf. I suspect that just about any exter- you can review hours of recordings just I’ve never been an accomplished bird
nal sound card would work fine for this by scanning through spectrograms visu- watcher: I’d be hard pressed to distin-
application, as long as it has an input that ally. Checking out a 1-hour-long sound file guish a sparrow from a wren. So it’s
accepts line-level signals (typically takes just a few minutes. rather satisfying to discover that, with
marked “line in”). These files, of course, picked up a lot some simple electronics and the right
To power the preamp, I used a of sounds: rumbling traffic, screeching software, I am able to pick out different
7-ampere-hour, 12-volt gel-cell battery, cats, wailing sirens, and who knows what species of migratory birds flying high
which is overkill. But the big battery else. But once you’ve looked at spectro- overhead through the inky darkness of
would allow me to leave the thing running grams for a while, it becomes easy to pick the night. 

18 [Link] JULY 2024

 SHARING THE EXPERIENCES OF WORKING ENGINEERS BY EDD GENT

Careers
Aakhilesh
Singhania
This Bosch engineer speeds hybrid
race cars to the finish line

W
hen it comes to motorsports, the need
for speed isn’t only on the racetrack.
Engineers who support race teams
also need to work at a breakneck pace
to fix problems, and that’s something Aakhilesh
Singhania relishes.
Singhania is a senior applications engineer at
Bosch Engineering, in Novi, Mich. He develops
and supports electronic control systems for hybrid Aakhilesh Singhania is on the job in a race-car
race cars, which feature combustion engines and pit at the 2024 Acura Grand Prix of Long Beach,
battery-powered electric motors. Calif., which features hybrid-powered race cars.
His vehicles compete in two iconic endurance
races: the Rolex 24 at Daytona in Daytona Beach,
Fla., and the 24 Hours of Le Mans in France. He rience really motivated me to dive further into
splits his time between refining the underlying motorsports.”
technology and providing trackside support on One incident in particular shaped Singhania’s
competition day. Given the relentless pace of the career trajectory. In 2013, he was leading Manipal’s
racing calendar and the intense time pressure when Formula Student team and was one of the drivers
cars are on the track, the job is high octane. But for a competition in Germany. When he tried to
Singhania says he wouldn’t have it any other way. start the vehicle, smoke poured out of the battery,
“I’ve done jobs where the work gets repetitive and the team had to pull out of the race.
and mundane,” he says. “Here, I’m constantly chal- “I asked myself what I could have done differ-
lenged. Every second counts, and you have to be ently,” he says. “It was my lack of knowledge of
very quick at making decisions.” the electrical system of the car that was the prob-
lem.” So, he decided to get more experience and
Growing up in Kolkata, India, Singhania picked up education.
a fascination with automobiles from his father, a
car enthusiast. After graduating in 2014, Singhania began work-
In 2010, when Singhania began his mechanical ing on engine development for Indian car manu-
engineering studies at India’s Manipal Institute of facturer Tata Motors in Pune. In 2016, determined
Technology, he got involved in the Formula Student to fill the gaps in his knowledge about automotive
program, an international engineering competi- electronics, he left India to begin a master’s degree
tion that challenges teams of university students program in automotive engineering at the Univer-
to design, build, and drive a small race car. The cars sity of Michigan in Ann Arbor.
EDWARD SACKLEY

typically weigh less than 250 kilograms and can He took courses in battery management, hybrid
have an engine no larger than 710 cubic centimeters. controls, and control-system theory, parlaying this
“It really hooked me,” he says. “I devoted a lot background into an internship with Bosch in 2017.
of my spare time to the program, and the expe- After graduation in 2018, he joined Bosch full-time

JULY 2024 [Link] 19

 CAREERS

as a calibration engineer, developing technology for “I’m constantly challenged.

hybrid and electric vehicles.
Transitioning into motorsports required perse-
Every second counts, and
verance, Singhania says. He became friendly with you have to be very quick at
the Bosch team that worked on electronics for race making decisions.”
cars. Then in 2020 he got his big break.
That year, the U.S.-based International Motor
Sports Association and the France-based Auto-
mobile Club de l’Ouest created standardized rules the different components interact and how altering
to allow the same hybrid race cars to compete in their configuration would affect performance.
both the Sportscar Championship in North Amer-
ica, host of the famous Daytona race, and the global Technology development is only part of Singha-
World Endurance Championship, host of Le Mans. nia’s job. On race days, he works as a support engi-
The Bosch motorsports team began prepar- neer, helping troubleshoot problems with the hybrid
ing a proposal to provide the standardized hybrid system as they crop up. Singhania and his colleagues
system. Singhania, whose job already included cre- monitor each team’s hardware using computers on
ating simulations of how vehicles could be electri- Bosch’s race-day trailer, and a mobile nerve center
fied, volunteered to help. hardwired to the organizers’ control center.

2
The competition organizers selected Bosch as “We are continuously looking at all the telemetry
lead developer of the hybrid system that would be data coming from the hybrid system and analyzing
provided to all teams. Bosch engineers would also [the system’s] health and performance,” he says.
be required to test the hardware they supplied to If the Bosch engineers spot an issue or a team
each team to ensure none had an advantage. notifies them of a problem, they rush to the pit stall
ICONIC
“The performance of all our parts in all the cars to retrieve a USB stick from the vehicle, which con-
ENDURANCE
has to fall within 1 percent of each other,” Singha- tains detailed data to help them diagnose and fix
EVENTS FOR
nia says. the issue.
HYBRID RACE
After Bosch won the contract, Singhania offi- After the race, the Bosch engineers analyze the
cially became a motorsports calibration engineer, CARS—ROLEX 24 telemetry data to identify ways to boost the standard-
responsible for tweaking the software to fit the AT DAYTONA ized hybrid system’s performance for all the teams. In
idiosyncrasies of each vehicle. IN DAYTONA motorsports, where the difference between winning
In 2022 he stepped up to his current role: BEACH, FLA., and losing can come down to fractions of a second,
developing software for the hybrid control unit AND 24 HOURS that kind of continual improvement is crucial.
(HCU), which is essentially the brains of the vehi- OF LE MANS Customers “put lots of money into this program,
cle. The HCU helps coordinate all of the different IN FRANCE and they are there to win,” Singhania says.
subsystems such as the engine, battery, and elec-
tric motor and is responsible for balancing power Many engineers dream about working in the fast-
requirements among these different components paced and exciting world of motorsports, but it’s
to maximize performance and lifetime. not easy breaking in. The biggest lesson Singhania
Bosch’s engineers also designed software learned is that if you don’t ask, you don’t get invited.
known as an equity model, which runs on the “Keep pursuing them because nobody’s going
HCU. It is based on historical data collected from to come to you with an offer,” he says. “You have
the operation of the hybrid systems’ various com- to keep talking to people and be ready when the
ponents, and controls their performance in real opportunity presents itself.”
time to ensure all the teams’ hardware operates at Demonstrating that you have experience con-
the same level. tributing to challenging projects is a big help. Many
In addition, Singhania creates simulations of the of the engineers Bosch hires have been involved in
race cars, which are used to better understand how Formula Student or similar automotive-engineering
programs, such as the EcoCAR EV Challenge,
says Singhania.
Employer: Education: The job isn’t for everyone, though, he says. It’s
Bosch Engineering Bachelor’s degree in mechanical demanding and requires a lot of travel and working
engineering, Manipal Institute
Occupation: of Technology, India; master’s on weekends during race season. But if you thrive
Senior applications degree in automotive engineering, under pressure and have a knack for problem solv-
engineer University of Michigan, Ann Arbor ing, there are few more exciting careers.

20 [Link] JULY 2024

Q & A BY RAHUL RAO

5 Questions of computer is not going to exist for another eight

years. So, any sort of commercial system that’s
being deployed now can be upgraded in that time
frame. That is only because [smartphones and other
cloud-touching consumer electronics] don’t have a
terrifically long lifetime, and it’s not expected that
any sort of cryptographically relevant quantum
computer is going to be available until 2030 or 2033.

Are there things that need to be switched now?

Best: Defense-facing systems have a 10-year time
frame. Those absolutely need to be upgraded now
because they are going to still be in operation in that
future where quantum computers exist. Automo-
tive is another leading edge, because cars last long
enough that they’re going to cross into that horizon.

How difficult will the transition be?

Best: It is a very heavy lift. There are billions of

Scott Best
devices that connect to the cloud. I think my toaster
has firmware upgrades occasionally. It’s every-
where, and every single one of those devices could
be potentially compromised.
Prepping today’s systems for tomorrow’s [The NSA] wants critical networking infrastruc-
ture to be fully on the way to transitioning by next
post-quantum cryptography year. The NSA is pressing [all the major cloud ven-
dors] to get things done right now. This is critically
important to domestic Internet infrastructure,
municipal fire and safety, and security.

T
oday’s cryptographic protocols rely on “If there is
mathematical techniques like finding the a shortcut, Do you think the transition will happen in time?
prime factors of very large numbers. But Best: What you’re racing against is physics. You’re
large enough quantum computers would
it’s a math trying to solve the problem of the decoherence of
have a powerful tool called Shor’s algorithm, which we haven’t logical qubits—[today] they’ve got 100 or 120 or 140
can quickly factor colossal integers. invented yet.” functional, logical qubits that could be used. Once
Quantum computers can’t wield that tool yet. that number scales up into the 400s, that’s when a
The U.S. National Institute of Standards and Tech- state-funded actor could absolutely just start mali-
nology is developing a quantum-computer-proof ciously breaking digital signature and secure socket
standard based on lattice cryptography, which finds technology. We’re sort of in a race: on one side, just
the shortest repeating vector in a many-dimensional the grindstone work of upgrading the entire world’s
lattice. And the U.S. National Security Agency (NSA) cloud-based infrastructure; on the other hand, it’s a
has laid out a plan to switch the country’s network race against physics.
infrastructure to lattice-based encryption.
IEEE Spectrum spoke to Scott Best, a senior engi- Could the protocols be cracked before the
neer at chip-design company Rambus, on what transition is complete?
needs to happen now to prepare for future quantum Best: Nobody looks at lattice [cryptography] and
computers. says, “you know, there’s a gap in the math there.” If
there is a shortcut, it’s a math we haven’t invented
When does the switch to quantum-resistant yet. And that gives a lot of people confidence that
Scott Best is a
protocols need to happen? senior engineer the type of asymmetric cryptography problems that
Scott Best: Something with 400 logical qubits that at chip-design you’re trying to solve with this new cryptography
can actually complete Shor’s algorithm—that kind company Rambus. have no obvious shortcuts to them.

Photo-illustration by Stuart Bradford JULY 2024 [Link] 21

   SEMICONDUCTORS

The Path to
a 1-Trillion-
Transistor GPU
AI’S BOOM DEMANDS NEW CHIP TECHNOLOGY

B y M a r k L i u & H .- S . P h i l i p W o n g Illustration by Optics Lab

22 [Link] JULY 2024

   SEMICONDUCTORS

In 1997
technology. This last contribution to the generative
AI revolution has received less than its fair share of
credit, despite its ubiquity.
Over the last three decades, the major milestones
in AI were all enabled by the leading-edge semicon-
the IBM Deep Blue supercomputer defeated world ductor technology of the time and would have been
chess champion Garry Kasparov. It was a ground- impossible without it. Deep Blue was implemented
breaking demonstration of supercomputer technol- with a mix of 0.6- and 0.35-micrometer-node
ogy and a first glimpse into how high-performance chip-manufacturing technology. The deep neural
computing might one day overtake human-level network that won the ImageNet competition, kicking
intelligence. In the 10 years that followed, we began off the current era of machine learning, was imple-
to use artificial intelligence for many practical tasks, mented with 40-nanometer technology. AlphaGo
such as facial recognition, language translation, and conquered the game of Go using 28-nm technology,
recommending movies and merchandise. and the initial version of ChatGPT was trained on
Fast-forward another decade and a half and arti- computers built with 5-nm technology. The most
ficial intelligence has advanced to the point where recent incarnation of ChatGPT is powered by servers
it can “synthesize knowledge.” Generative AI, such using even more advanced 4-nm technology. Each
as ChatGPT and Stable Diffusion, can compose layer of the computer systems involved, from soft-
poems, create artwork, diagnose disease, write sum- ware and algorithms down to the architecture, circuit
mary reports and computer code, and even design Advances in design, and device technology, acts as a multiplier for
semiconductor
integrated circuits that rival those made by humans. the performance of AI. But it’s fair to say that the
technology
Tremendous opportunities lie ahead for artificial [above the time- foundational transistor technology is what has
intelligence to become a digital assistant to all line],including enabled the advancement of the layers above.
human endeavors. ChatGPT is a good example of new materials, If the AI revolution is to continue at its current
how AI has democratized the use of high-perfor- advances in pace, it’s going to need even more from the semi-
lithography,
mance computing, providing benefits to every indi- conductor industry. Within a decade, it will need a
new types of
vidual in society. transistors, 1-trillion-transistor GPU—that is, a GPU with 10
All those marvelous AI applications have been and advanced times as many devices as is typical today.
due to three factors: innovations in efficient packaging,
machine-learning algorithms, the availability of have driven the Relentless Growth in AI Model Sizes
massive amounts of data on which to train neural development of The computation and memory access required for
more-capable AI
networks, and progress in energy-efficient comput- systems [below
AI training have increased by orders of magnitude
ing through the advancement of semiconductor the timeline]. in the past five years. Training GPT-3, for example,

 Transistor  Integrated
 Packaging fan-out
 Materials
 System-on-
 Lithography
integrated-chips
 AI  193-nanometer
immersion  3D-stacked DRAM  Chiplets

 High-k/  Chip-on-wafer-
metal gate on-substrate
 Copper  Strained
interconnects silicon  FinFET  EUV  Nanosheet

1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022 2024

 MNIST  ImageNet  AlphaGo  GPT-3

database
 AlexNet  DALL-E
 IBM Deep Blue  Watson  Transformers  GPT-4

 Long short-  AlphaFold  AlphaCode

term memory
 BERT  ChatGPT

24 [Link] JULY 2024

 Vertical connection density in 3D chips has increased at
roughly the same rate as the number of transistors in a GPU.
1,000,000
Projected
600,000
Bonds per square millimeter

400,000

200,000

100,000

60,000
40,000

20,000

10,000
2021 2023 2025 2027 2029 2031 2033
Year
Source: Taiwan Semiconductor Manufacturing Co.

10,000
Transistor count per processor, billions

Transistors per processor are Projected

increasing 1.73 times every two years.

1,000

100

10

1
2008 2010 2012 2014 2016 2018 2020 2022 2024 2026 2028 2030 2032 2034
Year
Source: Taiwan Semiconductor Manufacturing Co.

Largely thanks to advances in semiconductor technology, a measure called energy-

efficient performance is on track to triple every two years.
100
Energy-efficient performance
(1/femtojoules∙picoseconds)

Projected
1

0.01

0.0001

0.000001

2005 2010 2015 2020 2025 2030 2035 2040

Year
Source: TSMC. Energy-efficient performance (EEP) = Throughput/watt x throughput. Data are based on GPU data.

JULY 2024 [Link] 25

   SEMICONDUCTORS

requires the equivalent of more than 5 billion billion square millimeters, what’s called the reticle limit. But
operations per second of computation for an entire we can now extend the size of the integrated system
day (that’s 5,000 petaflops-days), and 3 trillion bytes beyond lithography’s reticle limit. By attaching sev-
(3 terabytes) of memory capacity. eral chips onto a larger interposer—a piece of silicon
Both the computing power and the memory into which interconnects are built—we can integrate
access needed for new generative AI applications a system that contains a much larger number of
continue to grow rapidly. We now need to answer a devices than what is possible on a single chip. For
pressing question: How can semiconductor tech- example, Taiwan Semiconductor Manufacturing Co.’s
nology keep pace? chip-on-wafer-on-substrate (CoWoS) technology
can accommodate up to six reticle fields’ worth of
From Integrated Devices to Integrated Chiplets compute chips, along with a dozen high-band-
Since the invention of the integrated circuit, semi- width-memory (HBM) chips.
conductor technology has been about scaling down HBMs are an example of the other key semicon-
in feature size so that we can cram more transistors ductor technology that is increasingly important for
into a thumbnail-size chip. Today, integration has AI: the ability to integrate systems by stacking chips
risen one level higher; we are going beyond 2D scal- atop one another, what we at TSMC call sys-
ing into 3D system integration. We are now putting tem-on-integrated-chips (SoIC). An HBM consists
together many chips into a tightly integrated, mas- of a stack of vertically interconnected chips of
sively interconnected system. This is a paradigm DRAM atop a control logic IC. It uses vertical inter-
shift in semiconductor-technology integration. connects called through-silicon-vias to get signals
In the era of AI, the capability of a system is through each chip and solder bumps to form the
directly proportional to the number of transistors connections between the memory chips. Today,
integrated into that system. One of the main limita- high-performance GPUs use HBM extensively.
tions is that lithographic chipmaking tools have been Going forward, 3D SoIC technology can provide
designed to make ICs of no more than about 800 an alternative to the conventional HBM technology
of today, delivering far denser vertical interconnec-
tion between the stacked chips. Recent advances
have shown HBM test structures with 12 layers of
chips stacked using hybrid bonding, a cop-
How Nvidia Uses
per-to-copper connection with a higher density
Advanced Packaging than the solder bumps in use now can provide.
CoWoS, TSMC’s chip-on-wafer-on-silicon Bonded at low temperature on top of a larger base
advanced packaging technology, has already logic chip, this memory system has a total thick-
been deployed in products. Examples include ness of just 600 μm.
the Nvidia Ampere and Hopper GPUs. Each With a high-performance computing system com-
posed of a large number of dies running large AI
consists of one GPU die with six
models, high-speed wired communication may
high-bandwidth memory cubes all on a quickly limit the computation speed. Today, optical
silicon interposer. The compute GPU die is interconnects are already being used to connect
about as large as chipmaking tools will server racks in data centers. We will soon need optical
currently allow. Ampere has 54 billion interfaces based on silicon photonics that are pack-
transistors, and Hopper has 80 billion. The aged together with GPUs and CPUs. This will allow
transition from 7-nm technology to the the scaling up of energy- and area-efficient band-
denser 4-nm technology made it possible to widths for direct, optical GPU-to-GPU communica-
pack 50 percent more transistors on essen- tion, such that hundreds of servers can behave as a
single giant GPU with a unified memory. Because of
tially the same area. Ampere and Hopper are
the demand from AI applications, silicon photonics
the workhorses for today’s large language will become one of the semiconductor industry’s
model training. It takes tens of thousands of most important enabling technologies.
these processors to train ChatGPT.

A
Toward a Trillion-Transistor GPU
s noted already, typical GPU chips used
for AI training have already reached the
reticle field limit. And their transistor
count is about 100 billion devices. The
continuation of the trend of increasing transistor
count will require multiple chips, interconnected with
NVIDIA

2.5D or 3D integration, to perform the computation.

The integration of multiple chips, by CoWoS, SoIC,

26 [Link] JULY 2024

 or related advanced packaging technologies, allows
for a much larger total transistor count per system How AMD Uses 3D Technology
than can be squeezed into a single chip. We forecast
that within a decade a multichiplet GPU will have The AMD MI300A accelerated processor unit
more than 1 trillion transistors. leverages not just CoWoS but also TSMC’s 3D
We’ll need to link all these chiplets together in a technology, silicon-on-integrated-circuits
3D stack, but fortunately, industry has been able to (SoIC). The MI300A combines GPU and CPU
rapidly scale down the pitch of vertical intercon- cores designed to handle the largest AI
nects, increasing the density of connections. And workloads. The GPU performs the intensive
there is plenty of room for more. We see no reason matrix multiplication operations for AI, while
why the interconnect density can’t grow by an order
the CPU controls the operations of the entire
of magnitude, and even beyond.
system, and the high-bandwidth memories
Energy-Efficient Performance Trend for GPUs are unified to serve both. The 9 compute dies
So, how do all these innovative hardware technol- built with 5-nm technology are stacked on top
ogies contribute to the performance of a system? of 4 base dies of 6-nm technology, which are
We can see the trend already in server GPUs if dedicated to cache and I/O traffic. The base
we look at the steady improvement in a metric dies and memories sit atop silicon interposers.
called energy-efficient performance. EEP is a com- The compute part of the processor is com-
bined measure of the energy efficiency and speed posed of 150 billion transistors.
of a system. Over the past 15 years, the semicon-
ductor industry has increased energy-efficient
performance about threefold every two years. We
believe this trend will continue at historical rates.
It will be driven by innovations from many sources,
including new materials, device and integration
technology, extreme ultraviolet lithography, circuit
design, system architecture design, and the co-op-
timization of all these technology elements, among
other things.
In particular, the EEP increase will be enabled
by the advanced packaging technologies we’ve been
discussing here. Additionally, concepts such as sys-
tem-technology co-optimization, where the differ- has already been embraced by most of today’s tech-
ent functional parts of a GPU are separated onto nology companies and electronic design automation
their own chiplets and built using the best perform- companies.
ing and most economical technologies for each, will

I
become increasingly critical. The Future Beyond the Tunnel
n the era of artificial intelligence, semiconductor
A Mead-Conway Moment for 3D ICs technology is a key enabler for new AI capabili-
In 1978, Carver Mead, a professor at the California ties and applications. A new GPU is no longer
Institute of Technology, and Lynn Conway at Xerox restricted by the standard sizes and form factors
PARC invented a computer-aided design method for of the past. New semiconductor technology is no
integrated circuits. They used a set of design rules to longer limited to scaling down the next-generation
describe chip scaling so that engineers could easily transistors on a two-dimensional plane. An inte-
design very-large-scale integration circuits without grated AI system can be composed of as many ener-
much knowledge of process technology. gy-efficient transistors as is practical, an efficient
That same sort of capability is needed for 3D chip system architecture for specialized compute work-
design. Today, designers need to know chip design, loads, and an optimized relationship between soft-
system-architecture design, and hardware and soft- ware and hardware.
ware optimization. Manufacturers need to know For the past 50 years, semiconductor-technology
chip technology, 3D IC technology, and advanced development has felt like walking inside a tunnel.
packaging technology. As we did in 1978, we again The road ahead was clear, as there was a well-de-
need a common language to describe these technol- fined path. And everyone knew what needed to be
ogies in a way that electronic design tools under- done: shrink the transistor.
stand. Such a hardware description language gives Now, we have reached the end of the tunnel. From
designers a free hand to work on a 3D IC system here, semiconductor technology will get harder to
design regardless of the underlying technology. It’s develop. Yet, beyond the tunnel, many more possi-
AMD

on the way: An open-source standard, called 3Dblox, bilities lie ahead. 

JULY 2024 [Link] 27

 Undulators and laser: A series
of magnets with opposite
polarity cause the electrons
to undulate, bunch up, and emit
light. The light interacts
with the electrons to amplify a
particular wavelength, resulting
in a free-electron laser.

Injector linac:
This portion of the
accelerator gives
new electrons a kick
before they enter
the magnetic path of
the accelerator.
2nd arc: After
generating laser
light, electrons
bend back around
toward the
main linac.
Electron gun: Particles
are injected into the
accelerator to begin
their journey.

28 [Link] JULY 2024

1st arc: Accelerated
electrons make their first
bend and head toward the
light-emitting undulators.

Beam dump:
Decelerated
electrons are
magnetically
diverted into a
block of material
and stopped.

Deceleration (main linac): The

superconductors sap energy from the
Acceleration (main linac): returning electrons that have already
Superconducting RF circuits made a complete loop, using the energy
boost the speed of the newly to accelerate the new electrons from
injected electrons. the electron gun.

Is the Future of
Moore’s Law
in a Particle Accelerator?
WIGGLING ELECTRONS COULD TURBOCHARGE EUV LITHOGRAPHY
By John Boyd

Illustration by [Link]/KEK JULY 2024 [Link] 29

 A
As Intel, Samsung, TSMC, and Japan’s upcoming
advanced foundry Rapidus each make their separate
preparations to cram more and more transistors into
every square millimeter of silicon, one thing they all
have in common is that the extreme ultraviolet
(EUV) lithography technology underpinning their
efforts is extremely complex, extremely expensive,
and extremely costly to operate. A prime reason is
that the source of this system’s 13.5-nanometer light
is the precise and costly process of blasting flying
droplets of molten tin with the most powerful com-
mercial lasers on the planet.
But an unconventional alternative is in the works.
A group of researchers at the High Energy Acceler-
ator Research Organization, known as KEK, in them. Chipmakers had been getting by with work-
Tsukuba, Japan, is betting EUV lithography might around after workaround for the then most advanced
be cheaper, quicker, and more efficient if it harnesses system, lithography using 193-nm light. Moving to
the power of a particle accelerator. a much shorter, 13.5-nm wavelength was a revolu-
Even before the first EUV machines had been tion that would collapse the number of steps needed
installed in fabs, researchers saw possibilities for in chipmaking and allow Moore’s Law to continue
EUV lithography using a powerful light source well into the next decade.
called a free-electron laser (FEL), which is gener- The chief cause of the continual delays was a
ated by a particle accelerator. However, not just any light source that was too dim. The technology that
particle accelerator will do, say the scientists at ultimately delivered a bright enough source of EUV
KEK. They claim the best candidate for EUV lithog- light is called laser-produced plasma, or EUV-LPP.
raphy incorporates the particle-accelerator version It employs a carbon dioxide laser to blast molten
of regenerative braking. Known as an droplets of tin into plasma thousands of times per
energy-recovery linear accelerator, it could enable second. The plasma emits a spectrum of photonic
a free-electron laser to economically generate tens energy, and specialized optics then capture the nec-
of kilowatts of EUV power. This is more than essary 13.5-nm wavelength from the spectrum and
enough to drive not one but many next-generation guide it through a sequence of mirrors. Subse-
lithography machines simultaneously, pushing quently, the EUV light is reflected off a patterned
down the cost of advanced chipmaking. mask and then projected onto a silicon wafer.
“The FEL beam’s extreme power, its narrow It all adds up to a highly complex process. And
spectral width, and other features make it suitable although it starts off with kilowatt-consuming
as an application for future lithography,” Norio lasers, the amount of EUV light that is reflected onto
Nakamura, researcher in advanced light sources at the wafer is just several watts. The dimmer the light,
KEK, told me on a visit to the facility. the longer it takes to reliably expose a pattern on
the silicon. Without enough photons carrying the
Today’s EUV systems are made by a single manu- pattern, EUV would be uneconomically slow. And
facturer, ASML, headquartered in Veldhoven, pushing too hard for speed can lead to costly errors.
Netherlands. When ASML introduced the first gen- When the machines were first introduced, the
eration of these US $100-million-plus precision power level was enough to process about 100 wafers
KEK

machines in 2016, the industry was desperate for per hour. Since then, ASML has managed to steadily

30 [Link] JULY 2024

 But ultimately, operating costs come down to elec-
tricity [Link] Benson, recently retired
senior research scientist at the Thomas Jefferson
National Accelerator Facility, in Virginia, estimates
that the wall-plug efficiency of the whole EUV-LPP
system might be less than 0.1 percent. Free-electron
lasers, like the one KEK is developing, could be as
much as 10 to 100 times as efficient, he says.

The system KEK is developing generates light by

boosting electrons to relativistic speeds and then
deviating their motion in a particular way.
The process starts, Nakamura explains, when an
electron gun injects a beam of electrons into a
meters-long cryogenically cooled tube. Inside this
tube, superconductors deliver radio-frequency (RF)
signals that drive the electrons along faster and
faster. The electrons then make a 180-degree turn
and enter a structure called an undulator, a series of
oppositely oriented magnets. (The KEK system cur-
rently has two.) The undulators force the speeding
electrons to follow a sinusoidal path, and this
motion causes the electrons to emit light.
What happens next is a phenomenon called
self-amplified spontaneous emissions, or SASE. The
hike the output to about 200 wafers per hour for the The experimental light interacts with the electrons, slowing some and
present series of machines. compact energy- speeding up others, so they gather into “micro-
recovery linear
ASML’s current light sources are rated at 500 bunches,” peaks in density that occur periodically
accelerator at
watts. But for the even finer patterning needed in KEK recaptures along the undulator’s path. The now-structured
the future, Nakamura says it could take 1 kilowatt much of the electron beam amplifies only the light that’s in phase
or more. ASML says it has a road map to develop a energy from with the period of these microbunches, generating
1,000-W light source. But it could be difficult to electrons on a a coherent beam of laser light.
return journey
achieve, says Nakamura, who formerly led the beam It’s at this point that KEK’s compact energy-
to speed up a new
dynamics and magnet group at KEK and came out set of electrons. recovery linac (cERL), diverges from lasers driven
of retirement to work on the EUV project. by conventional linear accelerators. Ordinarily, the
Difficult but not necessarily impossible. Dou- spent beam of electrons is disposed of by diverting
bling the source power is “very challenging,” agrees the particles into what is called a beam dump. But
Ahmed Hassanein who leads the Center for Mate- in the cERL, the electrons first loop back into the
rials Under Extreme Environment, at Purdue Uni- RF accelerator. This beam is now in the opposite
versity, in Indiana. But he points out that ASML has phase to newly injected electrons that are just start-
achieved similarly difficult targets in the past using ing their journey. The result is that the spent elec-
an integrated approach of improving and optimizing trons transfer much of their energy to the new beam,
the light source and other components, and he isn’t boosting its energy. Once the original electrons have
ruling out a repeat. had some of their energy drained away like this, they
But brightness isn’t the only issue ASML faces are diverted into a beam dump.
with laser-produced plasma sources. “There are a “The acceleration energy in the linac is recov-
number of challenging issues in upgrading to higher ered, and the dumped beam power is drastically
EUV power,” says Hassanein. He rattles off several, reduced compared to [that of] an ordinary linac,”
including “contamination, wavelength purity, and Nakamura explains to me while scientists in another
the performance of the mirror-collection system.” room operate the laser. Reusing the electrons’
High operating costs are another problem. These energy means that for the same amount of electricity
systems consume some 600 liters of hydrogen gas the system sends more current through the accel-
per minute, most of which goes into keeping tin and erator and can fire the laser more frequently, he says.
other contaminants from getting onto the optics and Other experts agree. The energy-recovery linear
wafers. (Recycling, however, could reduce this figure.) accelerator’s improved efficiency can lower costs,

JULY 2024 [Link] 31

 In a free-electron laser, accelerated electrons are subject to alternating magnetic fields,
causing them to undulate and emit electromagnetic radiation. The radiation bunches up the
electrons, leading to their amplifying only a specific wavelength, creating a laser beam.

Electron
Undulator magnet
cloud from
particle N S
accelerator N S
S N N S Alternating magnets cause
S N N S
the electrons to wiggle.

S N N S
S N N S
N S S N N S
N S S N
S N N N S N
S N S N
S S S N
N S S N
Electron N S S N
cloud
emits
N S S
photons. N S
N
Photons cause
electrons to bunch Laser light
together and emit
more photons.

In an ordinary linear Free-electron laser Recirculation loop

accelerator, injected electrons
gain energy from an RF field in
a superconducting accelerator. Accelerating beam
The electrons then enter a
free-electron laser and are
immediately disposed of in a RF field
beam dump. In an energy-recovery
linear accelerator, the electrons
circle back into the RF field
and lend their energy to newly Decelerating beam
injected electrons before Recirculating beam Accelerated beam
exiting to a beam dump.

Injector Beam dump

Main superconducting linac
Injection beam Decelerated beam

“which is a major concern of using EUV laser- storage-ring-based synchrotrons—huge circular

produced plasma,” says Hassanein. accelerators that keep a beam of electrons moving
with a constant kinetic energy. So, the KEK
The KEK compact energy-recovery linear acceler- researchers went in search of a more appropriate
ator was initially constructed between 2011 and 2013 application. After talking with Japanese tech com-
with the aim of demonstrating its potential as a syn- panies, including Toshiba, which had a flash-
chrotron radiation source for researchers working memory-chip division at the time, Nakamura and
for the institution’s physics and materials-science his team conducted an initial study that confirmed
divisions. But researchers were dissatisfied with that a kilowatt-class light source was possible with
the planned system, which had a lower perfor- a compact energy-recovery linear accelerator. And
mance target than could be achieved by some so, the EUV free-electron-laser project was born.

32 [Link] JULY 2024 Illustrations by Chris Philpot

In 2019 and 2020, the researchers modified the electron gun, and the undulator. Engineers will also
existing experimental accelerator to start the jour- have to develop good procedural techniques to
ney to EUV light. ensure, for instance, that the electron beam does not
The system is housed in an all-concrete room to degrade or falter during operations.
protect researchers from the intense electromag- And to ensure their approach is cost effective
netic radiation produced during operation. The enough to grab the attention of chipmakers, the
room is some 60 meters long and 20 meters wide researchers will need to create a system that can
with much of the space taken up by a bewildering reliably transport more than 1 kW of EUV power
tangle of complex equipment, pipes, and cables that simultaneously to multiple lithography machines.
snakes along both sides of its length in the form of The researchers already have a conceptual design
an elongated racetrack. for an arrangement of special mirrors that would
The accelerator is not yet able to generate EUV convey the EUV light to multiple exposure tools
wavelengths. With an electron beam energy of without significant loss of power or damage to
17 megaelectron volts, the researchers have been able to the mirrors.
generate SASE emissions in bursts of 20-micrometer
infrared light. Early test results were published in the It’s too early in the development of EUV free-
Japanese Journal of Applied Physics in April 2023. The electron lasers for rapidly expanding chipmakers to
next step, which is underway, is to generate much Lasers pay it much attention. But the KEK team is not alone
greater laser power in continuous-wave mode. based on in chasing the technology. The venture-backed
To be sure, 20 micrometers is a far cry from energy- startup xLight, in Palo Alto, Calif. is also among
13.5 nanometers. And there are already types of par- recovery those chasing it. The company, which is packed
ticle accelerators that produce synchrotron radia- with particle-accelerator veterans from the Stan-
tion of even shorter wavelengths than EUV. But
linear ford Linear Accelerator and elsewhere, recently
lasers based on energy-recovery linear accelerators
accelerators inked an R&D deal with Fermi National Accelerator
could generate significantly more EUV power due could Laboratory, in Illinois, to develop superconducting
to their inherent efficiency, the KEK researchers generate cavities and cryomodule technology. Requests for
claim. In synchrotron radiation sources, light inten- significantly comment from xLight went unanswered, but in
sity increases proportionally to the number of more EUV January, the company took part in the 8th Work-
injected electrons. By comparison, in free-electron power due shop EUV-FEL in Tokyo, and former CEO Erik
laser systems, light intensity increases roughly with to their Hosler gave a presentation on the technology.
the square of the number of injected electrons, inherent Significantly, ASML considered turning to par-
resulting in much more brightness and power. ticle accelerators a decade ago and again more
For an energy-recovery linear accelerator to
efficiency. recently when it compared the progress of free-
reach the EUV range, it will require equipment electron laser technology to the laser-produced
upgrades beyond what KEK currently has room for. plasma road map. But company executives decided
So, the researchers are now making the case for LPP presented fewer risks.
constructing a new prototype system that can pro- And, indeed, it is a risky road. Independent views
duce the needed 800 MeV. on KEK’s project emphasize that reliability and
In 2021, before severe inflation affected econo- funding will be the biggest challenges the research-
mies around the globe, the KEK team estimated the ers face going forward. “The R&D road map will
construction cost (excluding land) for a new system involve numerous demanding stages in order to
at 40 billion yen ($260 million) for a system that develop a reliable, mature system,” says Purdue’s
delivers 10 kW of EUV and supplies multiple lithog- Hassanein. “This will require serious investment
raphy machines. Annual running costs were judged and take considerable time.”
to be about 4 billion yen. So even taking recent infla- “The machine design must be extremely robust,
tion into account, “the estimated costs per exposure with redundancy built in,” adds retired research
tool in our setup are still rather low compared to the scientist Benson. The design must also ensure that
estimated costs” for today’s laser-produced plasma components are not damaged from radiation or laser
source, says Nakamura. light.” And this must be accomplished “without
There are plenty of technical challenges to work compromising performance, which must be good
out before such a system can achieve the high levels enough to ensure decent wall-plug efficiency.”
of performance and stability of operations demanded More importantly, Benson warns that without a
by semiconductor manufacturers, admits Nakamura. forthcoming commitment to invest in the technol-
The team will have to develop new editions of key ogy, “development of EUV-FELs might not come in
components such as the superconducting cavity, the time to help the semiconductor industry.” 

JULY 2024 [Link] 33

 A solar panel broken
down yields silver-laced
silicon, glass, copper,
a junction box and
an aluminum frame.

SPENT
BUT
34 [Link] JULY 2024
NOT
TRASHED
An Italian startup recovers valuable
materials from old solar panels
BY EMILY WALTZ
PHOTOGRAPHY AND ADDITIONAL REPORTING
BY LUIGI AVANTAGGIATO

JULY 2024 [Link] 35

 INSIDE
a shipping container in an industrial area of
Venice, the Italian startup 9-Tech is taking a
crack at a looming global problem: how to
responsibly recycle the 54 million to 160 million
tonnes of solar modules that are expected to
reach the end of their productive lives by 2050.
Recovering the materials won’t be easy. Solar
panels are built to withstand any environment
on Earth for 20 to 30 years, and even after sit-
ting in the sun for three decades, the hardware
is difficult to dismantle. In fact, most recycling
facilities trash the silicon, silver, and copper—
the most valuable but least accessible materials
in old solar panels—and recover only the alu-
minum frames and glass panes.
The need for recycling will only grow as the
world increasingly deploys solar power. More
than 1.2 terawatts of solar power has already
been deployed globally. Solar panels are cur-
rently being distributed at a rate of more than
400 gigawatts per year, and the rate is expected
to increase to a whopping 3 TW per year by
2030, according to a literature analysis by
researchers at the National Renewable Energy
Laboratory (NREL).
In an attempt to stop a mountain of photo-
voltaic garbage from accumulating, research-
ers are pursuing better recycling methods. The
most advanced methods proposed so far can
recover at least 90 percent of the copper, silver,
silicon, glass, and aluminum in a crystalline
silicon PV module. But these processes are
expensive and often involve toxic chemicals.
No recycling method has proven to be as cheap
as landfilling, and very few operate on an
industrial scale, says Garvin Heath, principal
environmental engineer at NREL, who man-
ages a group of international experts assigned
by the International Energy Agency to analyze
PV sustainability.
Above: The startup 9-Tech operates its
The founders of 9-Tech say they have a pilot plant out of a modified shipping
better way. Their process is a noisy one involv- container housed at the Green Propulsion
ing a combustion furnace, an ultrasound bath, Laboratory in the industrial port of
and mechanical sorting, the vibrations of which Marghera in Venice.
shake the floor of the modest freight container
where they have been testing their operation Right: After the frame, glass, and
for nearly two years. The company uses no toxic junction box are removed from a PV panel,
chemicals, releases no pollutants into the envi- the inner, bendable layers of silicon,
ronment, and recovers up to 90 percent of the polymers, and metal conductors remain.
materials in a solar panel, says Francesco Workers cut the inner layers into large
sections in preparation for the oven.
Miserocchi, chief technology officer at 9-Tech.

36 [Link] JULY 2024

JULY 2024 [Link] 37
 Above: Bits of silicon and glass are
separated from the rest of the panel.

Right: The inner layers of a PV panel

are fed into a continuous combustion
furnace, which heats the materials to
over 400 °C, vaporizing the polymers.

Far right: A mechanical roller

separates the copper grid after
the PV materials exit the furnace.

T
HE COMPANY TAILORS its process to laminated array in sheets of tempered glass,
crystalline silicon solar panels, which frame the whole thing in aluminum, seal the
make up 97 percent of the global PV edges, and attach a junction box on the back.
market. The panels typically consist of an When it’s time to recycle a panel, one of the
array of silicon wafers doped with boron and most challenging steps is removing the poly-
phosphorus, and topped with an antireflective mers, which stick to everything. “It’s not just
coating of silicon nitride. Silver conductors the edges or a couple of dots of glue. It’s an
are screen printed onto the wafer surface, and entire surface—several square feet—of poly-
copper conductors are soldered onto the array mer,” says Heath. The polymer can be burned
in a grid pattern. To protect the materials from off, but this releases carbon monoxide, hydro-
moisture and damage, manufacturers laminate fluoric acid, and other harmful pollutants. Sep-
the entire array in adhesive polymers—usually arating the silver conductors also proves
ethylene-vinyl acetate. Then they encase the challenging because they’re applied in a very

38 [Link] JULY 2024

 Above: After the materials exit
the oven, mechanical sieves
separate the glass and silicon.

thin layer–about 10 to 20 micrometers–that is wafers, and metal conductors. Without the frame
strongly attached to the silicon. Removing them or glass, the sandwich layers bend, shattering the
typically involves toxic reagents such as hydro- fragile silicon into small pieces. Workers crack
fluoric acid, nitric acid, or sodium hydroxide. the tempered glass and then send all the mate-
The team at 9-Tech addresses these chal- rials, which are mostly still in place because of
lenges in two ways. They recover the silver the polymers, into a continuous combustion
FAR RIGHT, BOTTOM: 9-TECH

using ultrasound rather than toxic chemicals, furnace. Heated to over 400 °C, the polymers
and although they burn the polymers, they cap- vaporize, and a filter captures the pollutants. The
ture the pollutants emitted. system also captures the heat from the furnace
The process at the company’s pilot plant and reuses it for energy efficiency.
starts with workers manually removing the alu- As the remaining material exits the fur-
minum frame, junction box, and tempered glass. nace, a roller mechanically strips out the
This leaves a sandwich of polymers, silicon copper. A series of sieves sort the broken bits

JULY 2024 [Link] 39

 Left: Workers immerse silver-
laced silicon pieces in a
bath of organic acid and treat
it with ultrasound to loosen
the bonds between the elements.

Right: Silver emerges from the

silicon bath in a fine dust.

Far right: High-purity copper,

glass, and silicon are recovered from
9-Tech’s PV-panel recycling process.

of glass and silicon based on thickness. The can recover 90 percent of the silver, 95 percent
silicon pieces, still laced with silver, are of the silicon, and 99 percent or more of the
immersed in a bath of organic acid and treated copper, aluminum, and glass from a PV module.
with ultrasound to loosen the bonds between What’s more, the material is considered highly
the elements. The ultrasound works by prop- pure, which increases the types of applications
agating sound waves into the acid bath, result- for which it can be reused.
ing in alternating high- and low-pressure The startup’s recycling process is more
cycles. If the waves are intense enough, they expensive than existing methods that recover
create cavitation bubbles that mechanically only aluminum and glass. But the extraction
interact with the material, causing the silver of high-purity silicon, silver, and copper
to dislodge from the silicon, says Pietro- should offset the extra cost, Miserocchi says.
giovanni Cerchier, CEO at 9-Tech. Plus, it’s more efficient than mining for virgin
Finally, workers remove the silicon frag- elements. You can extract about 500 grams of
ments from the ultrasound bath with a mesh silver from a tonne of solar panels, but only
net. This leaves a fine silver dust in the solution, 165 grams of silver from a tonne of ore, he
9-TECH (2)

which can be recovered by filtration or centri- says. “A photovoltaic panel at the end of its
fuge. All told, Cerchier says, 9-Tech’s pilot plant life still has a lot to give,” says Miserocchi. “It

40 [Link] JULY 2024

 efforts globally, including thermal, chemical,
mechanical, and optical approaches. The most
common method involves grinding the silicon,
metal, and polymer layers into small pieces,
separating them by density, and recovering the
silicon and metal with a thermal or chemical
process. Other processes include laser irradi-
ation, high-voltage pulses, optical sorting,
pyrolysis, chemical solvents, etching, and
delaminating with a hot knife.
Driving this innovation, in part, are regula-
tions adopted by the European Union in 2012.
The rules require all PV-panel manufacturers
in the EU to run take-back or recycling pro-
grams, or partner with other recycling schemes.
As a result, Germany, which has the most solar
power capacity in Europe, has one of the largest
PV recycling systems in the world.
But recycling is a high-volume business, and
aside from catastrophic weather events that
wipe out solar power stations, spent solar mod-
ules reach recyclers at a relative trickle. And
then there’s the challenge of finding a second
life for the materials after they are recovered—
a supply chain that’s not well developed.
First Solar, a global PV manufacturer based
in Tempe, Ariz., addressed both of these issues
on a large scale by building an in-house recy-
cling program with seven facilities in five coun-
tries. The company makes cadmium telluride
thin-film solar panels that buyers can purchase
with the recycling price built in. At the end of
the panels’ lives, buyers send them back to First
Solar for recycling into new products. The semi-
conductor material can be recycled up to 41
times, giving it a life-span of more than 1,200
years, according to the company. But the glass
isn’t pure enough to be reused in solar modules,
so the company plans to supply it to float-glass
manufacturers for use in windows and doors.
The challenges with recycling have inspired
can be considered a small mine of precious researchers to rethink the way crystalline solar
elements.” panels are made. For example, some manufac-
turers are trying to reduce or eliminate the

T
HE 9-TECH TEAM will know more difficult-to-recover silver, replacing it with
about the profitability of their method other conductive metals. And a team at NREL
after they build a larger demonstration demonstrated in February a way to eliminate
facility over the next 18 months. That plant, polymers in PV panels by laser welding the
to be located in the same industrial district glass panes instead, which may do a better job
of Venice as the shipping container, will be sealing out moisture. That technique may lend
able to handle up to 800 solar modules a day. itself to perovskite solar modules, a promising
Their pilot plant processes only about seven technology that is particularly susceptible to
modules a day. moisture and corrosion.
The company’s approach is one of many “Recycling shouldn’t be the only strategy,”
recycling methods for crystalline silicon PV says Heath. People should consider alterna-
panels in development. A comprehensive tive ways to repair or reuse solar panels to
review published in April in the Journal of extend their lives before resorting to recycling,
Cleaner Production identified dozens of other he says. 

JULY 2024 [Link] 41

BY ILLUSTRATIONS BY
GWE ND O LY N G REG
RA K MAB LY

THE
SHRINK
IN YOUR
POCKET

JULY 2024 [Link] 43

Apps track mood swings by passively collecting smartphone data

 NEARLY EVERY DAY since she was a child, Alex Leow,
a psychiatrist and computer scientist at the University
of Illinois Chicago, has played the piano. Some days she
plays well, and other days her tempo lags and her fingers
hit the wrong keys. Over the years, she noticed a pattern:
How well she plays depends on her mood. A bad mood
or lack of sleep almost always leads to sluggish,
mistake-prone music.
In 2015, Leow realized that a similar pat- The software then generates feed- people live with a mental illness, includ-
tern might be true for typing. She won- back for users, such as a graph displaying ing 40 million with bipolar disorder.
dered if she could help people with hourly keyboard activity. Researchers These apps differ from most of the
psychiatric conditions track their moods get access to the donated data from more than 10,000 mental-health and
by collecting data about their typing style users’ phones, which they use to develop mood apps available, which typically ask
from their phones. She decided to turn and test machine learning algorithms users to actively log how they’re feeling,
her idea into an app. that interpret data for clinical use. One help users connect to providers, or
After conducting a pilot study, in 2018 of the things Leow’s team has observed: encourage mindfulness. The popular
Leow launched BiAffect, a research app When people are manic—a state of being apps Daylio and Moodnotes, for example,
that aims to understand mood-related overly excited that accompanies bipolar require journaling or rating symptoms.
symptoms of bipolar disorder through disorder—they type “ferociously fast,” This approach requires more of the
keyboard dynamics and sensor data from says Leow. user’s time and may make these apps less
users’ smartphones. Now in use by more BiAffect is one of the few men- appealing for long-term use. A 2019 study
than 2,700 people who have volunteered tal-health apps that take a passive found that among 22 mood-tracking
their data to the project, the app tracks approach to collecting data from a apps, the median user-retention rate was
typing speed and accuracy by swapping phone to make inferences about users’ just 6.1 percent at 30 days of use.
the phone’s onscreen keyboard with its mental states. (Leow suspects that fewer But despite years of research on pas-
own nearly identical one. than a dozen are currently available to sive mental-health apps, their success is
consumers.) These apps run in the back- far from guaranteed. App developers are
ground on smartphones, collecting dif- trying to avoid the pitfalls of previous
ferent sets of data not only on typing but smartphone psychiatry startups, some
also on the user’s movements, screen of which oversold their capabilities
time, call and text frequency, and GPS before validating their technologies. For
location to monitor social activity and example, Mindstrong was an early
sleep patterns. If an app detects an startup with an app that tracked taps,
abrupt change in behavior, indicating a swipes, and keystrokes to identify digital
potentially hazardous shift in mental biomarkers of cognitive function. The
state, it could be set up to alert the user, company raised US $160 million in fund-
a caretaker, or a physician. ing from investors, including $100 mil-
Such apps can’t legally claim to treat lion in 2020 alone, and went bankrupt in
or diagnose disease, at least in the United February 2023.
States. Nevertheless, many researchers Mindstrong may have folded because
and people with mental illness have been the company was operating on a different
using them as tools to track signs of timeline from the research, according to
depression, schizophrenia, anxiety, and an analysis by the health-care news web-
bipolar disorder. “There’s tremendous, site Stat. The slow, methodical pace of
immediate clinical value in helping science did not match the startup’s need
Alex Leow noticed that mood people feel better today by integrating to return profits to its investors quickly,
influenced her style of piano
these signals into mental-health care,” the report found. Mindstrong also strug-
playing, and hypothesized that
says John Torous, director of digital psy- gled to figure out the marketplace and
BIAFFECT

it might affect phone-typing

style too. She created the app chiatry at Beth Israel Deaconess Medical find enough customers willing to pay for
BiAffect to test her theory. Center, in Boston. Globally, one in 8 the service. “We were first out of the

44 [Link] JULY 2024

 to see objective data on their patients in
between office visits, or for people tran-
sitioning from inpatient to outpatient
settings. These apps are “providing a
service that doesn’t exist,” says Colin
Depp, a clinical psychologist and profes-
sor at the University of California, San
Diego. Providers can’t observe their
patients around the clock, he says, but
smartphone data can help close the gap.
Depp and his team have developed
an app that uses GPS data and micro-
phone-based sensing to determine the
frequency of conversations and make
inferences about a person’s social inter-
actions and isolation. The app also
tracks “location entropy,” a metric of
how much a user moves around outside
of routine locations. When someone is
depressed and mostly stays home, loca-
tion entropy decreases.
Depp’s team initially developed the
app, called CBT2go, as a way to test the
effectiveness of cognitive behavioral
therapy in between therapy sessions. The
app can now intervene in real time with
people experiencing depressive or psy-
chotic symptoms. This feature helps
people identify when they feel lonely or
agitated so they can apply coping skills
they’ve learned in therapy. “When people
walk out of the therapist’s office or log
off, then they kind of forget all that,”
Depp says.
Another passive mental-health-app
developer, Ellipsis Health in San Fran-
blocks trying to figure this out,” says depression. Between 30 and 50 percent cisco, uses software that takes voice
Thomas Insel, a psychiatrist who of people with bipolar disorder will samples collected during telehealth
cofounded Mindstrong. attempt suicide at least once in their lives. calls to gauge a person’s level of depres-
Now that the field has completed a Catching early signs of a mood swing can sion, anxiety, and stress symptoms. For
“hype cycle,” Torous says, app developers enable people to take countermeasures each set of symptoms, deep-learning
are focused on conducting the research or seek help before things get bad. models analyze the person’s words,
needed to prove their apps can actually But detecting those changes early is rhythms, and inflections to generate a
help people. “We’re beginning to put the hard, especially for people with mental score. The scores indicate the severity
burden of proof more on those develop- illness. Observations by other people, of the person’s mental distress, and are
ers and startups, as well as academic such as family members, can be subjec- based on the same scales used in stan-
teams,” he says. Passive mental-health tive, and doctor and counselor sessions dard clinical evaluations, says Michael
apps need to prove they can reliably parse are too infrequent. Aratow, cofounder and chief medical
the data they’re collecting, while also That’s where apps come in. Algo- officer at Ellipsis.
addressing serious privacy concerns. rithms can be trained to spot subtle devi- Aratow says the software works for
ations from a person’s normal routine people of all demographics, without
A crucial component of that might indicate a change in mood—an needing to first capture baseline mea-
managing psychiatric ill- objective measure based on data, like a sures of an individual’s voice and speech
ness is tracking changes in diabetic tracking blood sugar. “The ability patterns. “We’ve trained the models in
mental states that can to think objectively about my own think- the most difficult use cases,” he says. The
lead to more severe epi- ing is really key,” says retired U.S. major company offers its platform, including an
sodes of the disease. Bipolar disorder, for general Gregg Martin, who has bipolar app for collecting the voice data, through
example, causes intense swings in mood, disorder and is an advisor for BiAffect. health-care providers, health systems,
from extreme highs during periods of The data from passive sensing apps and employers; it’s not directly available
mania to extreme lows during periods of could also be useful to doctors who want to consumers.

JULY 2024 [Link] 45

 In the case of BiAffect, the app can be also track suicide risk in menstruating trials with control groups and testing
downloaded for free by the public. Leow women and cognition in people with the technology in different patient pop-
and her team are using the app as a multiple sclerosis. ulations, he says.
research tool in clinical trials sponsored BiAffect’s software tracks behaviors How the data is collected can make a
by the U.S. National Institutes of Health. like hitting the backspace key fre- big difference in the quality of the
These studies aim to validate whether quently, which suggests more errors, research. For example, the rate of sam-
the app can reliably monitor mood dis- and an increase in typing “@” symbols pling—how often a data point is col-
orders, and determine whether it could and hashtags, which suggest more lected—matters and must be calibrated
social media use. The app combines this for the behavior being studied. What’s
typing data with information from the more, data pulled from real-world envi-
phone’s accelerometer to determine ronments tends to be “dirty,” with inac-
how the user is oriented and moving— curacies collected by faulty sensors or
for example, whether the user is likely inconsistencies in how phone sensors
lying down in bed—which yields more initially process data. It takes more work
clues about mood. to make sense of this data, says Casey
The makers of BiAffect and Ellipsis Bennett, an assistant professor and chair
Health don’t claim their apps can treat of health informatics at DePaul Univer-
or diagnose disease. If app developers sity, in Chicago, who uses BiAffect data
want to make those claims and sell their in his research.
product in the United States, they would One approach to addressing errors
first have to get regulatory approval is to integrate multiple sources of data
from the U.S. Food and Drug Adminis- to fill in the gaps—like combining
tration. Getting that approval requires accelerometer and typing data. In
rigorous and large-scale clinical trials another approach, the BiAffect team is
that most app makers don’t have the working to correlate real-world infor-
resources to conduct. mation with cleaner lab data collected
in a controlled environment where
The sensing techniques researchers can more easily tell when
upon which passive apps errors are introduced.
rely—measuring typing Who participates in the studies mat-
dynamics, movement, ters too. If participants are limited to a
voice acoustics, and the particular geographic area or demo-
like—are well established. But the algo- graphic, it’s unclear whether the results
rithms used to analyze the data collected can be applied to the broader population.
by the sensors are still being honed and For example, a night-shift worker will
validated. That process will require con- have different activity patterns from
siderably more high-quality research those with nine-to-five jobs, and a city
among real patient populations. dweller may have a different lifestyle
For example, clinical studies that from residents of rural areas.
include control or placebo groups are After the research is done, app devel-
crucial and have been lacking in the opers must figure out a way to integrate
past. Without control groups, compa- their products into real-world medical
nies can say their technology is effective contexts. One looming question is when
“compared to nothing,” says Torous at and how to intervene when a change in
Beth Israel. mood is detected. These apps should
Torous and his team aim to build always be used in concert with a profes-
software that is backed by this kind of sional and not as a replacement for one,
quality evidence. With participants’ says Torous. Otherwise, the app’s assess-
consent, their app, called mindLAMP, ments could be dangerous and distress-
Compared with a healthy user passively collects data from their screen ing to users, he says.
[top], a person experiencing time and their phone’s GPS and accel-
symptoms of bipolar disorder erometer for research use. It’s also cus- No matter how well these
[middle] or depression [bottom]
tomizable for different diseases, passive mood-tracking
may use their phone more
than usual and late at night. including schizophrenia and bipolar apps work, gaining trust
BiAffect measures phone disorder. “It’s a great starting point. But from potential users may
usage and orientation to help to bring it into the medical context, be the biggest stumbling
track those symptoms. there’s a lot of important steps that block. Mood tracking could easily feel
BIAFFECT

we’re now in the middle of,” says Torous. like surveillance. That’s particularly
Those steps include conducting clinical true for people with bipolar or psy-

46 [Link] JULY 2024

MOOD factor prevented her from using the app

SENSORS
enough to find out.
Beyond users’ perception, maintain-
ing true digital privacy is crucial. “Digital
footprints are pretty sticky these days,”
Seven metrics says Katie Shilton, an associate professor
apps use to make at the University of Maryland focused on
social-data science. It’s important to be
inferences about KEYBOARD DYNAM-
ICS: Typing speed and transparent about who has access to per-
your mood accuracy indicate a lot
about a person’s mood.
sonal information and what they can do
with it, she says.
For example, people
who are manic often “Once a diagnosis is established, once
type extremely fast. you are labeled as something, that can
affect algorithms in other places in your
life,” Shilton says. She cites the misuse
of personal data in the Cambridge Ana-
lytica scandal, in which the consulting
firm collected information from Face-
book to target political advertising. With-
out strong privacy policies, companies
producing mental-health apps could
similarly sell user data—and they may be
ACCELEROMETER: CALLS AND TEXTS: GPS LOCATION: particularly motivated to do so if an app
This sensor tracks how The frequency of texts Travel habits offer is free to use.
the user is oriented and and phone calls clues about mood. For
moving. Lying in bed example, a person
Conversations about regulating men-
signifies a person’s
would suggest a social isolation or experiencing tal-health apps have been ongoing for
different mood than activity, which depression may spend over a decade, but a Wild West–style lack
going for a run. indicates mood. more time at home. of regulation persists in the United
States, says Bennett of DePaul University.
For example, there aren’t yet protections
in place to keep insurance companies or
employers from penalizing users based
on data collected. “If there aren’t legal
protections, somebody is going to take
this technology and use it for nefarious
purposes,” he says.
Some of these concerns may be medi-
MIC AND VOICE: SLEEP: Changes in SCREEN TIME: An ated by confining all the analysis to a
Mood can affect how a sleep patterns indicate increase in the amount user’s phone, rather than collecting data
person speaks. mood shifts. Insomnia, of time a person
Microphone-based a common symptom of
in a central repository. But decisions
spends on a phone can
sensing tracks the bipolar disorder, can be a sign of depressive about privacy policies and data structures
rhythm and inflection trigger or worsen mood symptoms and can are still up to individual app developers.
of a person’s voice. disturbances. interfere with sleep. Leow and the BiAffect team are cur-
rently working on a new internal version
of their app that incorporates natu-
ral-language processing and generative
AI extensions to analyze users’ speech.
The team is considering commercializ-
ing this new version in the future, but
only following extensive work with
chotic disorders, where paranoia is part health as a person with schizophrenia. industry partners to ensure strict pri-
of the illness. But when she tested one passive sensing vacy safeguards are in place. “I really see
Keris Myrick, a mental-health advo- app, she opted to use a dummy phone. “I this as something that people could
cate, says she finds passive mental-health didn’t feel safe with an app company eventually use,” Leow says. But she
apps “both cool and creepy.” Myrick, having access to all of that information acknowledges that researchers’ goals
who is vice president of partnerships and on my personal phone,” Myrick says. don’t always align with the desires of the
innovation at the mental-health-advo- While she was curious to see if her sub- people who might use these tools. “It is
cacy organization Inseparable, has used jective experience matched the app’s so important to think about what the
a range of apps to support her mental objective measurements, the creepiness users actually want.” 

JULY 2024 [Link] 47

 HISTORY IN AN OBJECT BY ALLISON MARSH

The Wearable
Computer
as Bling
In 1993, well before Google
Glass debuted, the
artist Lisa Krohn designed
a prototype wearable
computer that looked like
no other. The Cyberdesk
was an experiment in
augmented reality, fusing
fashion with function
to extend the user’s senses.
The four circles along the
breastbone are a four-key
keyboard with a large
trackball at the top center.
A small microphone lies
against the throat, and an
earpiece hooks into the
left ear. Krohn imagined the
yellow tube in front of the
right eye as a retinal scan
display that would project a
laser beam directly onto
the back of the eye, creating
a screen centered in the
user’s field of view. Krohn
never built a working
version of the Cyberdesk.
Rather, she viewed it
as “strategic foresight,
speculative technology,
SAN FRANCISCO MUSEUM OF MODERN ART

predictive design, or
design fiction.” And it’s yet
another case where art,
like science fiction, has the
uncanny ability to
The Cyberdesk was anticipate the future.
a futuristic wearable
computer that looked
beyond the beige FOR MORE ON LISA
and boxy design of KROHN’S CYBERDESK,
traditional computers. see [Link]/
pastforward-jul2024

48 [Link] JULY 2024

 BILLS
DON’T
STOP ON
DISABILITY.
WHY
SHOULD
YOUR
INCOME?
Help protect your
livelihood and lifestyle.

1-800-493-IEEE (4333)
To learn more*, visit [Link]/Lift DISABILITY INSURANCE

Group Disability Income Insurance is available only for residents of the U.S. (except VT and territories), Puerto Rico and Canada (except Quebec).
Underwritten by New York Life Insurance Company, 51 Madison Ave., New York, NY 10010 on Policy Form GMR. It is available to residents of
Canada (except Quebec). *For information on features, costs, eligibility, renewability, limitations and exclusions visit [Link]/Lift.
Program Administered by AMBA Administrators, Inc.
In CA d/b/a Association Member Benefits & Insurance Agency
CA Insurance License #0I96562 | AR Insurance License #100114462
103118 (7/24) 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.

[Link]/ai

© The MathWorks, Inc.