Quantum Computing Is a Mirage a Practical Machine Might Never Be Possible P

QUANTUM COMPUTING IS A MIRAGE A PRACTICAL MACHINE MIGHT NEVER BE POSSIBLE P. 24

FOR THE TECHNOLOGY INSIDER | 03.19

THE INTERNET OF COWS HOW LoRa WIRELESS TECH MAKES EXCLUSIVE IoT EASY P. 07 SPECIAL CHINA’S BID FOR GRID SUPREMACY DELIVERY UPPING THE NIMBLE MACHINES WITH ANTE TO DYNAMIC LEGS WILL GO 1 MEGAVOLT WHERE NO ROBOTS HAVE P. 36 GONE BEFORE P. 30 HERE’S HACKING’S STATE OF THE ART BRANCH- PREDICTOR ATTACKS ARE ON THE RISE P. 42 starting at AWGs $4,980 per channel 2.4 GSa/s, 16 bit, 750 MHz 4, 8 and more channels <50 ns trigger latency Typical Applications Semiconductor testing, quantum computing, phased-array radar design & test, Boost your lidar, spectroscopy, NMR lab's performance Impedance starting at Analyzers $11,270 DC to 5 MHz, 1 mΩ to 1 TΩ 0.05% basic accuracy Compensation Advisor and Confidence Indicators Typical Applications High-Q dielectrics, capacitive sensors, supercapacitors, PV materials, component characterization

Lock-in starting at Amplifiers $6,210 Up to 600 MHz Scope, FFT, FRA, Sweeper, Imaging tool Optional: AWG, PID, PLL, Boxcar, Counter, AM & FM Typical Applications AFM, LVP, CARS, SRS, SNOM, graphene, optical PLL, THz, pump-probe, RFID, MEMS, NEMS, gyros, NDT, MRFM LabOne® Software All instruments are equipped with LabOne®, the Zurich Instruments control software, providing a wealth of features, efficient workflows and ease of use. Access your instruments from any web browser or integrate it into your LabVIEW, MAT- LAB, Python, C, and .NET programs.

Intl. +41 44 515 0410 USA 855-500-0056 (Toll Free) [email protected] www.zhinst.com

Zurich Let's discuss your application Instruments Start the conversation today CONTENTS_03.19

WALK THIS WAY A new generation of legged robots is learning to move with greater agility and efficiency. by jonathan hurst Page 30

24 THE CASE AGAINST QUANTUM COMPUTING A skeptical physicist explains why he’s not expecting practical quantum computers to arrive anytime soon. By Mikhail Dyakonov

36 A GRID AS BIG AS CHINA A massive expansion leads to the first ultrahigh- voltage AC-DC power grid. By Peter Fairley

42 HOW THE SPECTRE AND MELTDOWN HACKS REALLY WORKED An in-depth look at these dangerous microprocessor vulnerabilities. By Nael Abu-Ghazaleh, Dmitry Ponomarev & Dmitry Evtyushkin

07 NEWS 16 RESOURCES 21 OPINION THE INSTITUTE 01–16

On the cover and this page Photograph for IEEE Spectrum by Dan Saelinger

IEEE SPECTRUM (ISSN 0018-9235) is published monthly by The Institute of Electrical and Electronics Engineers, Inc. All rights reserved. © 2019 by The Institute of Electrical and Electronics Engineers, Inc., 3 Park Avenue, New York, NY 10016-5997, U.S.A. Volume No. 56, Issue No. 3. The editorial content of IEEE Spectrum magazine does not represent official positions of the IEEE or its organizational units. Canadian Post International Publications Mail (Canadian Distribution) Sales Agreement No. 40013087. Return undeliverable Canadian addresses to: Circulation Department, IEEE Spectrum, Box 1051, Fort Erie, ON L2A 6C7. Cable address: ITRIPLEE. Fax: +1 212 419 7570. INTERNET: [email protected]. ANNUAL SUBSCRIPTIONS: IEEE Members: $21.40 included in dues. Libraries/institutions: $399. POSTMASTER: Please send address changes to IEEE Spectrum, c/o Coding Department, IEEE Service Center, 445 Hoes Lane, Box 1331, Piscataway, NJ 08855. Periodicals postage paid at New York, NY, and additional mailing offices. Canadian GST #125634188. Printed at 120 Donnelley Dr., Glasgow, KY 42141-1060, U.S.A. IEEE Spectrum circulation is audited by BPA Worldwide. IEEE Spectrum is a member of the Association of Business Information & Media Companies, the Association of Magazine Media, and Association Media & Publishing. IEEE prohibits discrimination, harassment, and bullying. For more information, visit http://www.ieee.org/web/aboutus/whatis/policies/p9-26.html.

SPECTRUM.IEEE.ORG | MAR 2019 | 01 BACK STORY_ use a nanoreinforced THE INSTITUTE JOINS EPOXY? IEEE SPECTRUM

key benefits of f you turn to the back of this issue, you’ll find something nanosilica ﬁlled new. IEEE’s member-news publication, The Institute, will epoxy EP30NS now be integrated into IEEE Spectrum’s print magazine, once a quarter. For the past several decades, The Institute has been a stand-alone newspaper that was most recently Abrasion resistant bundled with issues of Spectrum and distributed four per ASTM D466-14 times a year. We’ll I also be migrating The Institute’s Optically clear online content Refractive index: 1.56 onto Spectrum’s website over the course of 2019. NASA low outgassing Why the change? per ASTM E595 There are of course some cost savings and environmental benefits Dimensionally stable to be had by combining Hardness: 80-90 Shore D the distribution of these two very different member Low CTE publications. But, 30-35 x 10-6 in/in/°C even more important, we’re hoping to raise awareness of the fine work The Institute’s staff does covering the global activities of IEEE and the fascinating work of its members around the world. learn more While Spectrum is an authoritative and independent about EP30NS journalistic source of news and information about new and emerging high-impact technologies, The Institute pulls back the curtain on the vital role IEEE members are playing in these exciting ventures. Think of it as IEEE’s “member clubhouse,” a place to catch up on what’s happening around IEEE. The Institute shines a spotlight on the best and the brightest of IEEE all year long, in print and online. In this issue of The Institute, you’ll find IEEE President José Moura’s first President’s Column in the opening pages, as well Scan to watch as an Editor’s Note from Kathy Pretz, discussing some of the improvements you can look forward to this year. Let us know what you think of the new format, at [email protected]. ■

154 Hobart St., Hackensack, NJ 07601 USA 03.19 +1.201.343.8983 • mainmasterbond.com www.masterbond.com New Advanced Optical Transceiver Test Suite

Complementing its 400GE test features, VeEX’s RXT-6400 Advanced Optical Transceiver test suite allows users to quickly test, qualify, and validate the performance of optical modules. New high speed optical modules that operate with PAM4 modulation, can easily be tested with this new test suite before they are used in production environments such as data centers, core and transport networks.

Product Highlights

• Voltage Sweep: Pre-FEC BER validation over operational voltage range (±5%) to verify optical module integrity • Pre-FEC BER: BER testing on a per-lane basis to verify the integrity of all electrical and optical lanes before FEC is applied to the PAM4 signal. Aggregate and Per Lane bit error count and BER measurements qualify the optical module performance under pre-FEC conditions • Frequency Offset Sweep: Pre-FEC BER testing on up to five frequency offset values. Pulling range can be varied between ±100 ppm from nominal clock frequency • Voltage/Temperature/BER monitoring: Voltage, temperature, and Pre-FEC BER are monitored and displayed for the duration of the test. A histogram function clearly displays all three measurements for easy correlation and tracking of any abnormal changes throughout the test • Pass/Fail Thresholds: Configurable threshold values for PASS/FAIL indication are available for transmit and receive optical power levels and pre-FEC BER • I2C Read/Write and Pre-Emphasis settings: For expert users, individual I2C registers can be read and programmed with an intuitive read/write user interface. All register settings can be saved for future referencing. Per-Lane PAM4 Pre-Emphasis settings are configurable. Pre-taps, Post-taps, and attenuation settings allow for PAM4 signal conditioning on the host side to help verify and stress transceiver tolerance and performance. In Auto and Manual modes, the PAM4 Pre-Emphasis settings are set to default values

Website: www.veexinc.com Email: [email protected] 400GE Solutions Tel: +1.510.651.0500 Visit us at OFC® 2019 - Booth #1816 The Verification Experts Fax: +1.510.651.0505 CONTRIBUTORS_

EDITOR IN CHIEF Susan Hassler, [email protected] ADVERTISING PRODUCTION MANAGER EXECUTIVE EDITOR Glenn Zorpette, [email protected] Felicia Spagnoli, [email protected] Nael Abu-Ghazaleh EDITORIAL DIRECTOR, DIGITAL SENIOR ADVERTISING PRODUCTION COORDINATOR Harry Goldstein, [email protected] Nicole Evans Gyimah, [email protected] Abu-Ghazaleh is chair of the computer engineering MANAGING EDITOR Elizabeth A. Bretz, [email protected] EDITORIAL ADVISORY BOARD, IEEE SPECTRUM department at the University of California, SENIOR ART DIRECTOR Susan Hassler, Chair; David C. Brock, Sudhir Dixit, Riverside. With Dmitry Ponomarev of the State Mark Montgomery, [email protected] Limor Fried, Robert Hebner, Grant Jacoby, Leah Jamieson, SENIOR EDITORS University of New York at Binghamton and Dmitry Deepa Kundur, Norberto Lerendegui, Steve Mann, Stephen Cass (Resources), [email protected] Evtyushkin of the College of William and Mary, Allison Marsh, Jacob Østergaard, Umit Ozguner, John Rogers, Erico Guizzo (Digital), [email protected] he has been studying side-channel cybersecurity Jonathan Rothberg, Umar Saif, Takao Someya, Jean Kumagai, [email protected] Maurizio Vecchione, Yu Zheng, Kun Zhou, Edward Zyszkowski attacks. In this issue, they explain the Spectre and Samuel K. Moore, [email protected] Meltdown vulnerabilities [p. 42]. Abu-Ghazaleh Tekla S. Perry, [email protected] EDITORIAL ADVISORY BOARD, THE INSTITUTE says to expect such threats to continue. “In general, Philip E. Ross, [email protected] Kathy Pretz, Chair; John Baillieul, Philip Chen, attacks on computer architecture are here to stay.” David Schneider, [email protected] Michael B. Forster, Shashank Gaur, Susan Hassler, DEPUTY ART DIRECTOR Brandon Palacio, [email protected] Hulya Kirkici, Cecilia Metra, San Murugesan, Mirela Sechi PHOTOGRAPHY DIRECTOR Randi Klett, [email protected] Annoni Notare, Joel Trussell, Hon K. Tsang, Chonggang Wang ASSOCIATE ART DIRECTOR Erik Vrielink, [email protected] MANAGING DIRECTOR, PUBLICATIONS SENIOR ASSOCIATE EDITOR Michael B. Forster Peter Fairley Eliza Strickland, [email protected] IEEE Spectrum contributing editor Fairley traveled NEWS MANAGER Amy Nordrum, [email protected] EDITORIAL CORRESPONDENCE IEEE Spectrum, 3 Park Ave., 17th Floor, to China to report on the massive ultrahigh-voltage ASSOCIATE EDITORS Willie D. Jones (Digital), [email protected] New York, NY 10016-5997 power grid that the country is building [p. 36]. At Michael Koziol, [email protected] TEL: +1 212 419 7555 FAX: +1 212 419 7570 one point, Fairley—a self-described “grid junkie”— SENIOR COPY EDITOR Joseph N. Levine, [email protected] BUREAU Palo Alto, Calif.; Tekla S. Perry +1 650 752 6661 was thrilled to sit down with eight Ph.D. engineers COPY EDITOR Michele Kogon, [email protected] DIRECTOR, BUSINESS DEVELOPMENT, EDITORIAL RESEARCHER Alan Gardner, [email protected] and economists to discuss the world’s largest power MEDIA & ADVERTISING Mark David, [email protected] system. Negotiating political sensitivities was tricky. ADMINISTRATIVE ASSISTANT ADVERTISING INQUIRIES Naylor Association Solutions, “In China, the intersection of technology and policy Ramona L. Foster, [email protected] CONTRIBUTING EDITORS Evan Ackerman, Mark Anderson, Erik Henson +1 352 333 3443, [email protected] is an uncomfortable topic,” Fairley says. “I sensed Robert N. Charette, Peter Fairley, Tam Harbert, Mark Harris, REPRINT SALES +1 212 221 9595, ext. 319 a chill on public discourse.” David Kushner, Robert W. Lucky, Prachi Patel, Morgen E. Peck, REPRINT PERMISSION / LIBRARIES Articles may be Richard Stevenson, Lawrence Ulrich, Paul Wallich photocopied for private use of patrons. A per-copy fee must EDITOR IN CHIEF, THE INSTITUTE be paid to the Copyright Clearance Center, 29 Congress Kathy Pretz, [email protected] St., Salem, MA 01970. For other copying or republication, Jonathan Hurst ASSISTANT EDITOR, THE INSTITUTE contact Managing Editor, IEEE Spectrum. Joanna Goodrich, [email protected] Hurst is cofounder and chief technology officer COPYRIGHTS AND TRADEMARKS IEEE Spectrum is a of Agility Robotics and an associate professor DIRECTOR, PERIODICALS PRODUCTION SERVICES Peter Tuohy registered trademark owned by The Institute of Electrical and Electronics Engineers Inc. Responsibility for the in robotics at Oregon State University. In this EDITORIAL & WEB PRODUCTION MANAGER Roy Carubia SENIOR ELECTRONIC LAYOUT SPECIALIST Bonnie Nani substance of articles rests upon the authors, not IEEE, issue, he writes about his quest to understand the PRODUCT MANAGER, DIGITAL Shannan Dunlap its organizational units, or its members. Articles do not underlying principles of legged locomotion and WEB PRODUCTION COORDINATOR Jacqueline L. Parker represent official positions of IEEE. Readers may post to build robots that walk like animals [p. 30]. His MULTIMEDIA PRODUCTION SPECIALIST Michael Spector comments online; comments may be excerpted for latest creation is Digit, a humanoid that may one ADVERTISING PRODUCTION +1 732 562 6334 publication. IEEE reserves the right to reject any advertising. day walk up to your doorstep to make a deli very. “When robots can go anywhere humans go, the world will be very different,” Hurst says.

IEEE BOARD OF DIRECTORS CHIEF MARKETING OFFICER Karen L. Hawkins Julien Mailland PRESIDENT & CEO José M.F. Moura, [email protected] +1 732 562 3964, [email protected] +1 732 562 3928 FAX: +1 732 465 6444 CORPORATE ACTIVITIES Donna Hourican Mailland is an assistant professor of telecommu PRESIDENT-ELECT Toshio Fukuda +1 732 562 6330, [email protected] TREASURER Joseph V. Lillie SECRETARY Kathleen A. Kramer nications at Indiana University. He and Kevin MEMBER & GEOGRAPHIC ACTIVITIES Cecelia Jankowski Driscoll, at the University of Virginia, are the PAST PRESIDENT James A. Jefferies VICE PRESIDENTS +1 732 562 5504, [email protected] authors of Minitel: Welcome to the Internet Witold M. Kinsner, Educational Activities; Hulya Kirkici, STANDARDS ACTIVITIES Konstantinos Karachalios (MIT Press, 2017), and they wrote about the history Publication Services & Products; Francis B. Grosz Jr., Member +1 732 562 3820, [email protected] of the Minitel network for IEEE Spectrum. In this & Geographic Activities; K.J. “Ray” Liu, Technical Activities; EDUCATIONAL ACTIVITIES Jamie Moesch Robert S. Fish, President, Standards Association; Thomas M. +1 732 562 5514, [email protected] issue, they describe how to convert obsolete Coughlin, President, IEEE-USA CHIEF FINANCIAL OFFICER & Minitel terminals for modern use [p. 16]. “Minitel is DIVISION DIRECTORS ACTING CHIEF HUMAN RESOURCES OFFICER Renuka P. Jindal (I); David B. Durocher (II); Sergio Benedetto a great way to start conversations about openness Thomas R. Siegert +1 732 562 6843, [email protected] in digital-network design,” says Mailland. (III); John P. Verboncoeur (IV); John W. Walz (V); Manuel Castro (VI); Bruno Meyer (VII); Elizabeth L. “Liz” Burd (VIII); Alejandro TECHNICAL ACTIVITIES Mary Ward-Callan “Alex” Acero (IX); Ljiljana Trajkovic (X) +1 732 562 3850, [email protected] REGION DIRECTORS MANAGING DIRECTOR, IEEE-USA Chris Brantley Babak Dastgheib-Beheshti (1); Wolfram Bettermann (2); +1 202 530 8349, [email protected] Gregg L. Vaughn (3); David Alan Koehler (4); Robert C. Dan Saelinger Shapiro (5); Keith A. Moore (6); Maike Luiken (7); IEEE PUBLICATION SERVICES & PRODUCTS BOARD Saelinger, a Portland, Ore.–based still-life Magdalena Salazar-Palma (8); Teófilo J. Ramos (9); Akinori Nishihara (10) Hulya Kirkici, Chair; Derek Abbott, John Baillieul, Sergio photographer, is accustomed to working with DIRECTOR EMERITUS Theodore W. Hissey Benedetto, Ian V. “Vaughan” Clarkson, Eddie Custovic, Samir inanimate objects, so shooting robots Cassie M. El-Ghazaly, Ron B. Goldfarb, Larry Hall, Ekram Hossain, and Digit for this month’s cover article, “Walk IEEE STAFF W. Clem Karl, Om P. Malik, Aleksander Mastilovic, Carmen This Way” [p. 30], provided an intriguing EXECUTIVE DIRECTOR & COO Stephen Welby S. Menoni, Paolo Montuschi, Lloyd A. “Pete” Morley, Sorel +1 732 562 5400, [email protected] Reisman, Gianluca Setti, Gaurav Sharma, Maria Elena Valcher, challenge. Saelinger says he enjoyed depicting CHIEF INFORMATION OFFICER Cherif Amirat John Vig, Steve Yurkovich, Bin Zhou, Reza Zoughi scenarios in which “robots can become part of +1 732 562 6017, [email protected] IEEE OPERATIONS CENTER our mundane, everyday world. If a robot could GENERAL COUNSEL & CHIEF COMPLIANCE OFFICER Jack S. Bailey, +1 212 705 8964, [email protected] 445 Hoes Lane, Box 1331 go to the fridge and whip up some dinner for PUBLICATIONS Michael B. Forster Piscataway, NJ 08854-1331 U.S.A. us, I’d be all for it.” +1 732 562 3998, [email protected] Tel: +1 732 981 0060 Fax: +1 732 981 1721

04 | MAR 2019 | SPECTRUM.IEEE.ORG This spiral antenna was optimized with EM simulation.

Visualization of the log-scaled norm of the electrical field on the slotted surface and far- field radiation pattern of a spiral slot antenna.

Wireless communication, sensing, positioning, and tracking. All of these technologies can take advantage of the spiral slot antenna’s consistent radiation pattern and impedance over a large bandwidth. To optimize spiral slot antenna designs for particular applications, engineers can turn to EM analysis software that calculates S-parameters and far-field patterns. The COMSOL Multiphysics® software is used for simulating designs, devices, and processes in all fields of engineering, manufacturing, and scientific research. See how you can apply it to designing spiral slot antennas.

comsol.blog/spiral-antennas SPECTRAL LINES_ 03.19

Workers assemble the enclosure for the IBM Q System One quantum computer, which was shown at the Consumer Electronics Show in Las Vegas in January.

The committee was to provide “an independent assessment of the feasibility and implications of creating a functional quantum computer capable of addressing real-world problems....” It was to estimate “the time and resources required, and how to assess the probability of success.” The experts who took up the challenge included John Martinis of the University of California, Santa Barbara, who heads Google’s quantum-hardware efforts; David Awschalom of the University of Chicago, who formerly directed the Cen- ter for Spintronics and Quantum Com- putation at UCSB; and Umesh Vazirani of the University of California, Berkeley, who codirects the Berkeley Quantum Quantum Computing’s Prospects Information and Computation Center. How long will it be before we see practical To their credit, in their report, released quantum-computing machines? in December, they didn’t sugarcoat the difficulties. Quite the opposite. The committee concluded that “it ur romance with new technologies always seems to follow the is highly unexpected” that anyone will be able to same trajectory: We are by turns mesmerized and adoring, dis- build a quantum computer that could compromise appointed and disheartened, and end up settling for less than we public-key cryptosystems (a task that quantum com- originally imagined. In 1954, Texas Instruments touted its new puters are, in theory, especially suitable for tackling) transistors as bringing “electronic ‘brains’ approaching the human in the coming decade. And while less-capable “noisy brain in scope and reliability” much closer to reality. In 2000, U.S. intermediate-scale quantum computers” will be built president Bill Clinton declared that the Human Genome Project within that time frame, “there are at present no known O would lead to a world in which “our children’s children will know algorithms/applications that could make effective use the term cancer only as a constellation of stars.” And so it is now with of this class of machine,” the committee says. quantum computing. ¶ The popular press is awash with articles touting its Okay, if not a decade, then how long? The commit- promise. Tech giants are pouring huge amounts of money into building pro- tee was not prepared to commit itself to any estimate. totypes. You get the distinct impression that the computer industry is on the Authors of a commentary in the January issue of the verge of an imminent quantum revolution. ¶ But not everyone believes that Proceedings of IEEE devoted to quantum computing quantum computing is going to solve real-world problems in anything like were similarly reticent to make concrete predictions. the time frame that some proponents of the technology want us to believe. So the answer is: Nobody really knows. Indeed, many of the researchers involved acknowledge the hype has got- The people working in this area are nevertheless ten out of control, cautioning that quantum computing may take decades thrilled by recent progress they’ve made on proof-of- to mature. ¶ Theoretical physicist Mikhail Dyakonov, a researcher for many concept devices and by the promise of this research. years at Ioffe Institute, in Saint Petersburg, Russia, and now at the University They no doubt consider the technical hurdles to be of Montpellier, in France, is even more skeptical. In “The Case Against Quan- much more tractable than Dyakonov concludes. So tum Computing,” on p. 24 of this issue, he lays out his view that practical don’t be surprised when you see their perspectives general-purpose quantum computers will not be built anytime in the fore- appear in Spectrum, too. seeable future. ¶ As you might expect, his essay ruffled some feathers after —David A. Schneider & Susan Hassler it was published online. But as it turns out, while his article was being pre- The editorial content of IEEE Spectrum does not represent official positions pared, a committee assembled by the U.S. National Academies of Sciences, of the IEEE or its organizational units. For additional commentary on the National Academies’ recent study on the feasibility of quantum computing, go

Engineering, and Medicine had been grappling with the very same question. to https://spectrum.ieee.org/quantum0319 RESEARCH IBM

06 | MAR 2019 | SPECTRUM.IEEE.ORG 51: NUMBER OF COUNTRIES WHERE OPERATORS HAVE DEPLOYED AT LEAST ONE LoRa NETWORK

Cattle may be at home COW PATROL: A company called on the range, but mod- Moovement has sold more than LoRa’s BID TO 5,000 LoRa-based ear tags to ern ranchers need to be able ranchers in Australia. to find their wayward cattle, RULE THE IoT and inefficiencies in tracking cost the cattle industry around US $4.8 billion a year. At a recent conference about connected The standard for low-power, devices in Amsterdam, Jan Willem Smeenk of the Dutch com- wide-area coverage is pany Sodaq and Thomas Telkamp of the startup Lacuna Space gaining momentum talked about connecting cattle into a future Internet of bovines. Smeenk’s company builds solar-powered ear tags that alert ranchers to the whereabouts and well-being of their cattle. The tags work up to distances of about 5 kilometers via LoRa, one of the leading low-power, long-range standards for the Internet of Things (IoT). Telkamp’s company is launching LoRa gateways on satellites, which could one day help ranchers track down

MOOVEMENT cattle that stray beyond their owners’ local coverage areas.

SPECTRUM.IEEE.ORG | MAR 2019 | 07 BEE WELL: When placed in a hive, Hostabee’s connected device measures temperature and humidity every hour.

large amounts of data or many devices per gateway (though Smeenk says he tells ranchers to expect each gateway to handle 2,000 cattle). Meanwhile, NB-IoT and LTE-M make use of spectrum already assigned to carriers for mobile telephony. That means they are a little easier for operators to integrate, and offer slightly higher data rates. Because NB-IoT and LTE-M share bands with 4G, they can penetrate buildings in urban areas better than LoRa and Sigfox. But NB-IoT has a harder time than LoRa when it comes to handling devices that move between towers. So NB-IoT makes more sense “What makes LoRa interesting for us is the only one where you can build out a for devices that stay in one spot, such is you can have two approaches: Use network from scratch without relying on as water-meter readers. a public network, or deploy your own any third-party providers,” says Alasdair But for all the talk of standards, scal- network,” says Smeenk. For now, Sodaq Allan, a technology researcher and con- ing up the Internet of Things is still a is mainly selling its cattle trackers in sultant in Exeter, England. work in progress. Software developer Australia to ranches that lie beyond the LoRa sends data in short bursts or mes- Maxime Mularz’s startup, Hostabee, reach of conventional mobile networks. sages by modulating the spread of data uses LoRa to track the temperature In January, the industry alliance “chirps.” Those messages are similar in and humidity inside 1,000 beehives in behind LoRa declared that 100 network size to those sent via SMS on cellular places such as Austria and the United operators around the world were using networks, but they cost less energy to States. The company’s devices must be its standard, putting it a smidge ahead send and receive. The LoRa patent dates capable of interacting with networks of competing IoT standards, which have to 2008, although it didn’t appear in a in different countries. Using LoRa is about 60 network implementations each. commercial wide-area network stan- an advantage because so many net- With LoRa service in place, companies dard until 2015. works have already deployed it, and have started using it to follow the move- Sigfox, the original low-power stan- beekeepers who work outside of a cov- ments of people with dementia, connect dard, began development in 2009 erage area can build their own compat- with pollution monitors, and track when and came as an integrated hardware ible gateways and either manage the recycling bins need to be changed. and network package. Then in 2016, data locally or send it up to the cloud The intellectual property of LoRa 3GPP, the industry collaboration that via a backhaul operator. Still, Mularz chips belongs to Semtech, which has defined standards for mobile ser- says, “we spend a lot of time working licenses it to other chipmakers, too. vice from 2G onward, established the with network operators to adapt to their Users can set up their own gateway, Narrowband Internet of Things (NB-IoT) back ends.” either by building their own hardware standard and the LTE-M standard. Both Some operators, including the French around a LoRa-compatible chip, or by are designed to work with cheap chips telecom giant Orange, have been testing buying one of a growing number of pre- and to extend battery life by limiting ways to make it possible for LoRa devices fabricated options, some based on Rasp- data rates. to roam across national borders. If they berry Pi or Arduino. LoRa and Sigfox both use unlicensed succeed, beekeepers, some of whom “The main differentiator for LoRa over spectrum, which makes them subject rent hives in different countries, may the other competing low-power, low- to interference and puts a cap on their soon waggle for joy. —Lucas Laursen bandwidth, and long-range protocol stan- duty cycles. That means users cannot ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/ dards for the Internet of Things is that it use these options for anything requiring loraiot0319 HOSTABEE

08 | MAR 2019 | SPECTRUM.IEEE.ORG guidance (the kind that counts wheel SELF-DRIVING TECH revolutions to measure distance). These sensors will also allow it to plan two- NEWS dimensional paths where no lane mark- FOR NON-CARS ings exist. The test-track phase is now half done, and sometime in the first half of From mall shuttles to monster trucks, Perrone this year, the roboticized truck will go off Robotics is ready to debut its self-driving tech to its final phase—field trials, probably in either A ustralia or Arizona, where open- pit mines are common and self-driving technology is particularly welcome. Like early mammals scuttering between the legs of tyran- The appeal of monster robotrucks nosaurs, a lot of little companies are trying to weave involves simple labor economics. It’s around—and maybe even outlast—the big boys of s elf-driving getting ever harder to find truck drivers technology. for two-shift-a-day operations in faraway One such example is Perrone Robotics, a small Virginia company that places. “Say you’re a company in Canada, has developed a self-driving package that it says can be quickly adapted hauling logs out of the woods,” Hofert to any vehicle. This Swiss Army knife of an AI can give smarts to an says. “That’s where the trees are, in the existing car, shuttle bus, or truck—even the gargantuan trucks used in middle of nowhere, and so you have to mining. Tiny shuttles and behemoth trucks sell in small numbers, and pay a lot to get people to go up there.” equipping them to drive themselves is beneath the dignity of major play- Closer to home, Perrone will also ers, like Alphabet’s Waymo and General Motors’ Cruise Automation. launch a pilot shuttle service in March “What we’re doing, certainly Waymo and GM Cruise could do, but they in the shopping area of Crozet. Though are focused on their own agenda. This is our niche, and we are going the driverless shuttle is fully certified by where we can add real value,” says David Hofert, the chief marketing Virginia’s road-safety regulators, for now officer at Perrone Robotics. it will hew to downtown Crozet’s speed The company predates today’s craze for self-driving technology. limit of 40 kilometers per hour (25 miles It had self-driving prototypes running around its headquarters in per hour). That’s fast enough to com- Crozet, Va., in the early 1990s, and it was among the original partici- plete a circuit every 10 minutes or so. pants in the DARPA Grand Challenges that sparked the robocar revo- Ann Mallek, who represents the city’s lution. When IEEE Spectrum last wrote about Perrone, in 2015, it was district in the Albemarle County gov- showing off a package that could be quickly dropped into an existing car, using external actuators to turn the steering wheel and depress the accel- erator and brake pedals. “Could we do your car? Absolutely,” Hofert says. “But it’s kind of expensive for a one-off. Our drop-in costs about [US] $150,000, and you have to add sensors, so altogether, it’s a couple hundred thousand.” More promising are one-off jobs like outfitting the Liebherr T 282 C, a ver- itable truckosaurus for the mining industry. This beast works off-road, in wide-open spaces, where it doesn’t have to avoid obstacles such as pedestrians and small dogs. Even so, the truck will be equipped

PERRONE ROBOTICS PERRONE with an array of radars, lidars, and ste- reo cameras, as well as GPS and inertial

SUPER SIZE: The Liebherr T 282 C can haul up to 363 metric tons of coal or debris.

SPECTRUM.IEEE.ORG | MAR 2019 | 09 A CHIP DESIGN WITH 40 GPUs Engineers plan to use “silicon interconnect fabric” to put 40 GPUs on a single wafer

Back in the 1980s, p arallel-computing pioneer Gene Amdahl hatched a plan to speed mainframe computing: a silicon-wafer-size processor. By keeping most of the movement of data on the processor itself, computing would be faster and more ernment, says the county, MICRO TRANSIT: Perrone energy efficient. With US $230 million from ven- Robotics’s self-driving shuttle can ture capitalists, the most ever at the time, Amdahl the local transit company, seat up to six people. and Perrone are splitting the founded Trilogy Systems to make his vision a real- roughly $200,000 cost of the trundling around malls ity. This first commercial attempt at “wafer-scale pilot. If the shuttle pilot goes and campuses. Navya, a integration” was such a disaster that it reportedly well, she says, the partners French company, has a introduced the verb “to crater” into the financial will launch a second project boxy shuttle that follows press lexicon. Engineers at the University of Illinois involving “a larger transit a set path at the pace of a Urbana-Champaign and at the University of Cali- vehicle to bring commuters trot and brakes for frolick- fornia, Los Angeles, think it’s time for another go. from Crozet to the central ing deer and other obstacles; At the IEEE International Symposium on High- shopping area on Sundays.” it’s now running students Performance Computer Architecture in February, Showing that the larger bus around the campus of the Illinois electrical and computer engineering associ- can be retrofitted in the University of Michigan, in ate professor Rakesh Kumar and his collaborators same way as the small shut- Ann Arbor. However, each made the case for a wafer-scale computer consisting tle should ease the process of one costs about $290,000, of as many as 40 GPUs. Simulations of this multi- obtaining approval for the sec- and the maintenance runs processor monster improved the rate of calcula- ond round of tests, she adds. in the tens of thousands of tions nearly 19-fold and the combination of energy Elsewhere, there are dollars per year. EasyMile, consumption and signal delay more than 140-fold. other roboshuttles already also from France, has put “The big problem we are trying to solve is the com- its EZ10 shuttle bus on munication overhead between computational units,” the roads in Helsinki and Kumar explains. Supercomputers routinely spread other cities, although applications over hundreds of GPUs that live on sep- “What we’re only in slow lanes. arate printed circuit boards and communicate over doing, certainly Perrone is betting long-haul data links. These links soak up energy and it’s got a niche inside are slow compared with the interconnects inside Waymo and GM this niche—it alone can the chips themselves. What’s more, because of the Cruise could take someone else’s mismatch between the mechanical properties of shuttles and give them chips and those of printed circuit boards, proces- do, but they are eyes, that they might sors must be kept in packages that severely limit focused on their see, and actuators, the number of inputs and outputs a chip can use. that they might steer. So getting data from one GPU to another entails own agenda.” —Philip E. Ross “an incredible amount of overhead,” says Kumar. What’s needed are connections between GPU mod- — David Hofert, ↗ POST YOUR COMMENTS at https://

Perrone Robotics spectrum.ieee.org/perrone0319 ules that consume very little energy and are as fast ROBOTICS PERRONE

10 | MAR 2019 | SPECTRUM.IEEE.ORG Bandwidth (gigabytes/second) ing a defect. If your chip is the size of a dinner plate, it’s almost guaranteed to have a system-killing flaw somewhere on it. So it makes more sense to start with normal-size GPU chips that have already passed quality tests and find a technology to better connect them. The team believes they have that in a technology called silicon interconnect fabric (Si-IF). Si-IF replaces the circuit board with silicon, so there is no mechanical mismatch between the chip and the board and therefore no need for a chip package. The Si-IF wafer is patterned with one or more lay- ers of 2-micrometer-wide copper interconnects spaced as little as 4 µm apart. That’s a spacing comparable to that of the top Energy (femtojoules/byte) level of interconnects on a chip. In the spots where the GPUs are meant to plug in, the silicon wafer is patterned with short copper pillars spaced about 5 µm apart. The GPU is aligned above these, pressed down, and heated. This well-established process, called thermal compression bonding, causes the copper pillars to fuse to the GPU’s copper interconnects. The combination of narrow interconnects and tight spacing means you can squeeze at least 25 times as many inputs and outputs on a chip, according to the Illinois and UCLA researchers. Kumar and his collaborators had to take a number of constraints into account in designing the wafer-scale GPU, including how much heat could be removed from the wafer, how the Latency (picoseconds) GPUs could most quickly communicate with one another, and how to deliver power across the entire wafer. Power turned out to be one of the more limiting constraints. At a chip’s standard 1-volt supply, the Si-IF wafer’s wiring would consume a full 2 kilowatts. So Kumar’s team boosted the voltage supply to 48 volts, reducing the amount of current needed and therefore the power lost. That solution required spreading voltage regulators and signal-conditioning capacitors around the wafer, taking up space that might have gone to more GPU modules. Still, in one design they were still able to squeeze in 41 GPUs. They tested a simulation of this design and found it sped both computation and the movement of data while consuming less On-Chip Short-haul Long-haulShort-haul Long-haul energy than 40 standard GPU servers would have. Medium-haul Medium-haul The Si-IF wafer-scale GPU “overcomes the problems that Conventional Wafer-scale integration integration the early wafer-scale work was not able to solve,” says Robert W. Horst of Horst Technology Consulting, in San Jose, Calif. BETTER, FASTER, CHEAPER: Connection bandwidth, latency, and More than two decades ago, at Tandem Computers, Horst was energy generally become worse the further you go from the chip in a computer. Wafer-scale integration using silicon interconnect fabric involved in creating the only wafer-scale product ever commer- improves all three metrics, so that they’re closer to those of the GPU chip. cialized, a memory system used in place of fast hard drives in stock exchanges. He expects that cooling will be one of the most and as plentiful as the interconnects on the chips. Such speedy challenging aspects. “If you put that much logic in that close connections would integrate those 40 GPUs to the point where proximity, the power dissipation can be pretty high,” he says. they act as one giant GPU. From the perspective of the program- Kumar says the team has started work on building a wafer- mer, “the whole things looks like the same GPU,” says Kumar. scale prototype processor system, but he would not give fur- One solution is to use standard chipmaking techniques to build ther details. —Samuel K. Moore all 40 GPUs on the same silicon wafer and add interconnects A version of this article appears in our Tech Talk blog. between them. But that’s the philosophy that killed Amdahl’s

SOURCE: RAKESH KUMAR/UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN AT ILLINOIS OF KUMAR/UNIVERSITY RAKESH SOURCE: attempt in the 1980s. Larger chips have a higher chance of hav- ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/40gpu0319

NEWS

SPECTRUM.IEEE.ORG | MAR 2019 | 11 NEWS

neered the floating solar concept in THAILAND EXPERIMENTS 2006, already has projects under way in Thailand. It opened a floating solar-equipment-manufacturing facil- WITH FLOATING SOLAR ity there last year that customizes a product for the country, where a major PLANTS requirement is low cost. “Pontoons are the more expensive part of floating solar, The state-run utility aims to install the tech at eight and Ciel & Terre has a monopoly on this,” dams over the next two decades says Jordan Macknick, the lead energy- water-land analyst at the United States’ National Renewable Energy Laboratory. But plastics are SCG’s specialty, and Solar farms take up its ability to produce pontoons in-house land, which is espe- should keep costs down. In fact, the com- cially precious in areas with pany has teamed up with EGAT to con- dense populations—and duct R&D for floating buoy materials at those areas need renewable a pilot plant in the Tha Thung Na Dam, power most. To solve this according to the Bangkok Post. paradox, populous countries Thailand today gets 12 percent of its short on land have started energy from renewables and hydro- to turn to solar farms that power. The government aims to increase float on reservoirs and dams. the country’s renewables share to 37 per- Of the world’s 1.1 gigawatts cent by 2036, with solar providing 17 GW. of floating solar capacity, Half of that, or 6 percent of the country’s 450 megawatts exists in China, Japan, SOLAR FLOAT: This 1-megawatt installation in total power, could in principle come India, and South Korea. a Thai reservoir owned by SCG Chemicals was from floating solar farms, according to the company’s first floating solar farm. Thailand now wants in, and an the Thai law firm Pugnatorius. unlikely player is vying to get a big share Floating solar is part of the petro- Floating solar is attractive in Thailand, of that market. Thailand’s Siam Cement chemical producer’s goals to diversify and the country’s several hydropower Group (SCG), one of Southeast Asia’s its business, Kettong says. SCG makes plants offer ideal staging waters. “Hydro- largest building-material companies, the mounting platforms for its floating power and floating solar make a lot of has developed floating solar modules solar farms from durable and recycla- sense together because the grid infra- that it will build, install, and maintain. ble high-density polyethylene. The pon- structure is already built,” Macknick As the only large Thai company mak- toons should last 50 years, twice the says. Hydropower can balance out ing floating panels, it hopes to land con- expected life-span of the encapsulated, solar’s variable output, he adds, and tracts to build at least some of the 1 GW water-resistant silicon solar panels that “floating systems on reservoirs could of floating solar capacity that state-run SCG buys from various suppliers. Each reduce evaporation from those reser- Electricity Generating Authority of Thai- module generates 100 kilowatts, and voirs, which means more fuel for hydro- land (EGAT) plans to install across eight they can be easily bolted together. Their power systems.” dams over the next two decades. design is also more streamlined, he says: Solar farms are also easier to set afloat EGAT announced in January that instal- A 1-MW floating plant takes 10 percent on water than to install on land, which lation of the first project, a 45-MW farm less space than competitors’ plants. can require clearing, excavation, and at the Sirindhorn Dam on the country’s The company has already installed grading, says Kettong. The two remain- eastern border, would begin in April, with 5 MW of floating solar projects in east- ing challenges facing floating solar in the four others following soon after. But the ern and southern Thailand, mostly at country are the complicated permit pro- utility still hasn’t picked a supplier. “Right factories. And the company is starting cess (which can take three to six months) now, EGAT is in a competitive bidding to sell to clients in other Southeast Asian and building credibility. “This is a new process, and we’re trying to be one of the countries. product, and a new concept, for Thai winners,” says Chatree Kettong, energy The French company Ciel & Terre companies,” he says. —Prachi Patel development manager at SCG Chemicals could give the Thai firm some stiff ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/

Co., a subsidiary of Siam Cement Group. competition. The firm, which pio- floatingsolar0319 CHEMICALS SCG

12 | MAR 2019 | SPECTRUM.IEEE.ORG OP2 Courtesy of The World’s ROBOTIS Best ROBOTS GUIDE Is Here! ROBOTS.IEEE.ORG

IEEE Spectrum’s new ROBOTS site features more than 200 robots from around the world.

• Spin, swipe and tap to make • Rate robots and check • Play Faceoff, an interactive robots move. their ranking. question game. • Read up-to-date robotics • View photography, videos news. and technical specs. Check out Robots.ieee.org on your desktop, tablet, or phone now! 14 | MAR 2019 | SPECTRUM.IEEE.ORG PHOTOGRAPH BY Wolfgang Rattay/Reuters CHARGING UP

TWO NAGGING questions continue to come up in discussions about ditching fossil- fueled automotive propulsion for electric vehicles: First, how are the increasing numbers of people in urban areas, with no garages in which to charge their cars overnight, supposed to rely on electric power? And second, if a significant proportion of the driving public purchases EVs, will the grid be able to deliver all the electric power needed to keep us moving? This fast-charging facility in Duisburg, Germany, is being touted as the answer to both questions. Its 180-square- meter roof is covered with solar panels, and it has 210 kilowatt-hours of battery storage to keep it from overtaxing the local grid when vehicles are plugged into its four DC charging stations. The 150 kW delivered at each bay is enough power to boost a car’s batteries to an 80 percent state of charge in about 10 minutes.

THETHE BIG BIG PICTUREPICTURE newsNEWS

SPECTRUM.IEEE.ORG | MAR 2019 | 15 FROM TRASH TO TREASURE TURN A MINITEL TERMINAL INTO A LINUX TERMINAL OR VIDEOTEX DISPLAY

NE OF US (MAILLAND) GREW UP IN PARIS IN THE 1980S, SURROUNDED BY ADVERTISEMENTS FOR RACY O “pink” chat rooms, accessible through terminals connected to France’s Minitel network. They were a lucrative part of the wider Minitel economy, which also let you send messages, check bank balances, and read news. By 2000, as the Internet displaced the Minitel network, the billboards started being replaced by piles of terminals abandoned by trash RESOURCES_HANDS ON cans. In 2010, while researching Minitel law and policy for a project that became the first English-language academic book on Minitel, I’d collected over 15 terminals of various models, when I met an American home brewer (Driscoll) who wanted to play with them. • We quickly completed our first project: turning a terminal into a Twitter client. Then we turned it into a webcam client; then, into a videotex slideshow display. We found a lot of help along the way because we weren’t the only people unwill- ing to let these stylish terminals end up as e-waste. Indeed, there’s now a vibrant Minitel hacking scene. You don’t need to be in France to try any of these upcyling projects either. Minitel was an open platform, and a number of failed attempts to replicate Minitel’s success means there are localized terminals to be found, with ones for Ireland, South Africa, and the United States, among others. • However, a lot of the guidance we relied upon is in French, so here we’d like to offer an English-language introduction to Minitel hacking. • The first thing you’ll need is a terminal. French terminals are plentiful on eBay, and despite their age, very sturdy. A reasonable offer shouldn’t top US $23 (shipping is the issue, thanks to relatively heavy cathode ray tubes. Expect that to cost at least $90 to North America). We recommend getting the “1B” model, the sturdiest and most versatile. Steer away from

Magis, M2, or M12, each of which present problems for the home brewer. Your terminal must have a five-pin DIN connector in the back. USA LAB, RESEARCH MINITEL

16 | MAR 2019 | SPECTRUM.IEEE.ORG If you’re in North America, or anywhere else that uses 110-volt AC power, you’ll need a step-up transformer to feed your terminal with 220-V AC power (unless you’ve managed to obtain a native U.S. version, of course!). If all goes well, you’ll be able to turn it on, and be able to type on the keyboard and see characters echoed to the screen. Our next step was to connect the Minitel 1B to a laptop running Linux. With the help of Google Translate, Driscoll found instructions to build a simple circuit for connecting a Minitel’s DIN serial port to a standard RS-232 port, posted by Frédéric Leperchois and Furrtek. This particular circuit—which shifts signal voltage levels appropriately between the DIN and RS-232 ends—has a long history in French hacker culture, appearing in 1980s computer magazines and “download kits,” enabling PC owners to use their Minitels as modems. We’ve reproduced the schematic here with English labels. We laid out the circuit and sent a stream of data to the Minitel from the laptop. (Comput- ers with RS-232 ports are becoming harder to find, so you may have to use an RS-232– to–USB converter cable). On our Linux system, the Minitel was connected to the port mount- VIVE LE MINITEL: Using a voltage-level conversation about the role of openness in the converter [schematic and top left] many Minitel design of digital networks. ed at /dev/ttyS0. We used the agetty utility to terminals can be connected to PCs, such as create a console interface set to 4,800 baud, this one running a Twitter interface [top right]. In recent years, the Minitel revival scene 7 data bits, and 1 even parity bit. We set the has seen an influx of participants, ranging port speed on the Minitel by hitting P and the from old-timers to teenage geeks. Legend- FNCT button together, followed by 4. actly how Minitel graphics were encoded and ary pink site 3615 SM has now been port ed Instead of nice alphanumeric characters, displayed. Driscoll’s rudimentary French be- to the Internet and was recently demoed however, we got strange symbols on the gan to include technical jargon such as des during an exhibit in New York City. Christian screen. The problem was that Minitel was ex- caractères semi-graphiques. Using the specs, Quest has ported other services, including pecting not ASCII but videotex. (Older British we built a simple Arduino Uno–based video- Minitel’s electronic phone-book service, and Irish readers will be familiar with videotex tex server. We connected the serial commu- complete with up-to-date information ex- from its use by the Ceefax and Aertel television nication pins on the Arduino to the Minitel’s tracted via France Telecom’s public API. information systems and the BBC Micro.) DIN port and used the SoftwareSerial library Github lists dozens of repositories tagged We had to switch the terminal into the to write videotex. The Minitel serial port was “minitel” including those by Eric Sérandour. ASCII-derived téléinformatique mode, a designed to interconnect with peripherals A French engineering school has even in- little-known feature of the Minitel 1B. After using a logical-high threshold level of 2.5 V or troduced Minitel as part of its curriculum, we pressed FNCT+T followed by A, the famil- more and a logical-low threshold of 0.4 V or assigning students the task of “giving a iar Linux prompt appeared. Within a few days, less. The Minitel detected Arduino’s 5-V sig- second life to Minitel.” The Minitel commu- we had put together a Linux system monitor, nals without a problem, but we needed a nity is not shy about sharing tips and code. an ASCII art webcam, and a Twitter client. logic-level converter for the Arduino to reli- We hope you join us and reanimate these We were hooked. Thanks to Frederic ably detect the Minitel’s 2.5-V signals. With terminals by unleashing your creativity. Cambus, maintainer of a large repository of the help of an old 16-bit teletext authoring —KEVIN DRISCOLL & JULIEN MAILLAND resources, we got our hands on the official tool, we created Minitel screens and anima- POST YOUR COMMENTS at https://spectrum.ieee.org/

MINITEL RESEARCH LAB, USA LAB, RESEARCH MINITEL technical specifications, which spell out ex- tions. We’ve exhibited this system to provoke minitel0319

SPECTRUM.IEEE.ORG | MAR 2019 | 17 RESOURCES_GEEK LIFE

erty away for free. I started to really get inter- THE FACES OF ested in that, and meeting these people, and OPEN SOURCE seeing who they really were,” says Adams. Capturing a good portrait is very differ- THE HUMANS ent from taking a simple snap: The objec- BEHIND THE tive is not simply to record a likeness, but to evoke something of the spirit that animates CODE the subject’s features. A good connection between the photographer and subject is more important than the quality of the cam- era gear. This is where Adams’s background gives him the edge: He was in on the ground 1 floor of the commercial Internet at a number ne thing we never forget at of companies in the 1990s and 2000s, be- O IEEE Spectrum is that technol- fore jumping careers into professional pho- ogy doesn’t automatically con- tography about 10 years ago. dense out of the laws of physics, like dew “Because I’m fairly technically literate, I find on leaves. Every bit of tech that exists, ways to get people comfortable with the fact exists because someone worked to make that I understand what they’ve done and its it so. But it’s a harsh irony of engineering importance. I think that puts people in the that the better designed something is, the right zone,” says Adams. more inconspicuous its creator becomes. Adams typically invites subjects to his This is especially true for software “down Silicon Valley studio, but he’s traveled the stack”—the protocols, servers, operat- around the United States to conferences ing systems, and other infrastructure upon and o ther events in his efforts to fill his roster. which every app depends. Is there a common strand in the wide range Peter Adams is trying make some of those of personalities he has photographed? “This creators more visible—literally—with his online isn’t a mercenary group of people, meaning Faces of Open Source project. Adams takes they’re not developing software for money. striking photographs of contributors to the There’s a much larger force that’s driving open source movement, from Unix designer these people to do what they do. Obviously, Ken Thompson to more recent players such as some people are employed by companies 5 information-security expert Yan Zhu. to work on the software, but others are Adams embarked on his photographic not. Others are independent consultants mission in 2014. “Consumers don’t have that choose to make a living, working and much awareness that underneath the shiny developing open source software that the wrappers that the commercial companies world can use. I think that says something like Apple and Google are making is a foun- about somebody who’s willing to devote, in dation of open source software. I felt it was some cases, their entire professional life to, an incredible story that trillions of dollars say, a network. It’s a pretty amazing thing.” of economic value could be created on the —STEPHEN CASS work of a relatively small group of people who essentially gave their intellectual prop- POST YOUR COMMENTS at https://spectrum.ieee.org/faces0319

Gallery Guide: 1. Yan Zhu, infosec contributor to projects including SecureDrop and Tor. 2. Ken Thompson, cocreator of Unix. 3. Vint Cerf, cocreator of TCP/IP. 4. Mary Ann Horton, creator of uuencode. 5. Keila Banks, evangelist. 6. Eric Raymond, author. 7. Limor Fried, open source hardware developer. 8. Larry Wall, creator of Perl. 9. John Osterhout, creator of TcL/Tk. 10. Sanjay Ghemawat, cocreator of TensorFlow. 11. Jessica McKellar, Python contributor. 12. Junio Hamano, maintainer of Git. 9

18 | MAR 2019 | SPECTRUM.IEEE.ORG 2 3 4

6 7 8

10 11 12

PHOTOGRAPHS BY Peter Adams SPECTRUM.IEEE.ORG | MAR 2019 | 19 RESOURCES_AT WORK

In hopes of increasing the supply of data DEFINING DATA SCIENTISTS scientists—and clarifying the definition— A CERTIFICATION PROGRAM the Open Group, an IT industry consor- tium that offers certification, announced AIMS TO HELP in January a program for data scientists. Open Group certification requires an ap- plicant to complete projects that demonstrate knowledge of statistics, machine learning, AI, and business communications, explained Martin Fleming, chief ana- lytics officer for IBM. There are three levels: certified data scientist, master certified data scientist, and distinguished certified data scientist. Applicants can be reviewed by their peers through the Open Group or within their own organizations, providing that their companies receive Open Group accreditation to do so. IBM, Fleming said, is the first company to be accredited, but he expects other companies will soon follow. Offering certification, Fleming said, will “help make our organization an attractive place to come and work because it will give our workers a credential that is valuable to them. It will also help them improve their skills, which are valuable to the organization, and provide clar- ity around kinds of skills that are required for HEAP CONNECTIVITY AND STORAGE TECHNOL the profession.” IBM currently employs about ogies have encouraged companies to gather as much 15,000 tech workers that it defines as data sci- data as possible, whether tracking user behavior on a entists, and expects that number to grow faster website, how a plane’s engines are performing, or how than its overall workforce, Fleming indicated. electricity flows across a grid. But all that data has IBM also announced that it will estab- no value in itself. Someone has to coax useful knowl- lish a 24-month data science apprentice- edge from it. Consequently, data scientists are in ship program, allowing less-experienced demand. • But while a lot of companies want to hire people, including those without four-year data scientists, there is little agreement about what a data scientist actu- college degrees, to build skills in the field Cally is. In a Hacker News discussion about my recent View From the Valley and receive Open Group certification. The blog post on the demand for data scientists, one commenter wrote: “ ‘Data apprenticeship program is certified by the scientist’ is a title recently thrown around a lot for positions that used to U.S. Department of Labor. That took a little be called ‘data analyst,’ with no strong [machine learning] or [software doing, Fleming pointed out, because it engineering] ability required.” • So can we say anything more specific? required the Department of Labor, for the Job-search firm Indeed has sketched out a picture of a data scientist as first time, to include “data scientist” as a rec- a technologist with a degree “in computer science, statistics, or a quan- ognized occupation. titative social science, along with some training in statistical modeling, “We tend to think of data scientists as machine learning, and programming.” Another Hacker News commenter superstars,” Fleming said, “but it is a large stated: “I would have thought [data science involves] a serious, sustained community. You need folks at junior level and study of statistics—starting with a strong base knowledge of the mathe at superstar level.” —TEKLA S. PERRY matics of probability and building from there. But based on the résumés A version of this article appears in our View I’ve seen, that doesn’t seem to be the common opinion.” A fuzzy defi- • From the Valley blog. nition doesn’t make filling data science jobs—or helping a data scientist

figure out if a job offer is right for them—any easier. POST YOUR COMMENTS at https://spectrum.ieee.org/data0319 ISTOCKPHOTO

20 | MAR 2019 | SPECTRUM.IEEE.ORG REFLECTIONS_BYINTERNET OF EVERYTHING_BY ROBERT LUCKY STACEY HIGGINBOTHAM OPINIONOPINION

own sensors and autonomous system would take over. Remote-control operations require fast, reliable cellular coverage. That’s why, in November, Ericsson and Einride, a Swedish electric-autonomous-vehicle company, built a test track in Sweden to see how remote operations would work in practice. Einride provided the trucks, called T-pods, while Ericsson, in collaboration with Scandinavian telco Telia, supplied the 5G network. In ongoing trials, the T-pods are making deliveries between two DB Schenker warehouses in Sweden—and a human has to take over by remote control in roughly 10 percent of the tests. These tests show that even though Einride trucks can drive themselves most of the time, it’s best to have a person to easily step in when the AI encounters something it can’t handle. But how many trucks can one person safely monitor? If each person can monitor only one truck, many of the costs associated with eliminating human drivers fall by the wayside. Åsa Tamsons, a senior vice president at Ericsson, has suggested a network operations center AUTONOMOUS TRUCKS (NOC), with alerts notifying monitoring humans which trucks need assistance. NEED PEOPLE It could be like a cross between an air- traffic control tower and a traditional telecommunications NOC. WE ALL KNOW how a self-driving vehicle is supposed to act, thanks Robert Falck, CEO of Einride, says that to companies and startups like Waymo and Uber: These cars and the company wants to reach a 10:1 ratio trucks should be outfitted with enough sensors and computing of vehicles to people. In Falck’s mind, the power to put all humans in the passenger seat. • But that rosy vision combination of humans and autonomous ignores current laws and practices. For example, while California has approved cars will continue for quite some time. autonomous-vehicle testing without a safety driver since February 2018, those “It’s about experience and understand- vehicles must still have a remote operator who can step in to pilot the car in an ing how even if you have an autonomous, emergency. It’s not sexy, but the big secret of autonomous cars is that there’s still reliable system, you need a human in a human behind many of them. • Consider commercial vehicles, like semitrailer the mix,” says Falck. “Then you start trucks. Trucking is a relatively high-paying job that’s approachable for a diversity of to understand that remote drivers and people, and these drivers fear self-driving versions. This fear over massive unem- autonomous vehicles will be working ployment is a big part of the discussion around such truck fleets. • But that’s not the together even if we reach Level 5 auton- future that’s developing. Instead of creating truly autonomous fleets, the industry omy.” Level 5 autonomy is when a vehi- is looking at remotely operated autonomous vehicles. Companies like autonomous cle can be considered to be driving itself. truck company Starsky Robotics and telecom-equipment manufacturer Ericsson are Except apparently, we’ll still need a per- creating a future of semi-autonomous semitrucks. • In December 2018, Starsky filed son behind the wheel—even if they’re not, a plan with the U.S. National Highway Traffic Safety Administration to use remote- you know, physically behind the wheel. n controlled trucks for deliveries. A human would remotely drive the truck from a ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/ distribution center to the highway. However, once the truck is on the highway, its autonomoustrucks0319

ILLUSTRATION BY Dan Page SPECTRUM.IEEE.ORG | MAR 2019 | 21 NUMBERS DON’T LIE_BY VACLAV SMIL OPINION

many boxes over distances 400 times ELECTRIC CONTAINER as long at speeds three to four times as fast as the pioneering electric ship can handle. SHIPS ARE A HARD SAIL What would it take to make an electric ship that can carry 18,000 TEUs? JUST ABOUT EVERYTHING you wear or use around the house In a 31-day trip, today’s efficient diesel once sat in steel boxes on ships whose diesel engines propel vessel burns 4,650 metric tons of fuel them from Asia, emitting particulates and carbon dioxide. (bunker or diesel), each ton packing Surely, you would think, we can do better. • After all, we’ve 42 gigajoules. That’s an energy density had electric locomotives for more than a century and high-speed electric of about 11,700 watt-hours per kilogram, trains for more than half a century, and recently we have been expanding versus 300 Wh/kg for today’s lithium- the global fleet of electric cars. Why not get electric container ships? Actu- ion batteries, a nearly 40-fold difference. ally, the first one should begin to operate this year: the Yara Birkeland, The total fuel demand for the trip is built by Marin Teknikk, in Norway, is not only the world’s first electric- about 195 terajoules, or 54 gigawatt-hours. powered, zero-emissions container ship but also the first autonomous Large diesels (and those in the ships are the commercial vessel. • But don’t write off giant diesel-powered container largest we have) are about 50 percent effi- ships and their critical role in a globalized economy just yet. Here is a cient, hence their useful propulsive energy back-of-the-envelope calculation that explains why. • Containers come demand is about 27 GWh. To match that in different sizes, but most are the standard twenty-foot equivalent demand, large electric motors operating units (TEU)—rectangular prisms 6.1 meters (20 feet) long and 2.4 meters at 90 percent efficiency would need about wide. The first small container ships of the 1960s carried mere hundreds 30 GWh of electricity. of TEUs; now Maersk’s Triple-E class ships load 18,000 TEUs, and OOCL Load the ship with today’s best com- Hong Kong holds the record, at 21,413. At the “super slow steaming,” mercial Li-ion batteries (300 Wh/kg) and fuel-saving speed of 16 knots, these ships can make the journey from still it would have to carry about 100,000 Hong Kong to Hamburg in 31 days. • Now look at the Yara Birkeland. metric tons of them to go nonstop from It will carry just 120 TEU, its service speed will be 6 knots, its longest Asia to Europe in 31 days. Those batter- intended operation will be 30 nautical miles—between Herøya and ies alone would take up about 40 percent Larvik, in Norway—and its batteries will deliver 7 to 9 megawatt hours. of maximum cargo capacity, an eco- Today’s state-of-the-art diesel container vessels thus carry 150 times as nomically ruinous proposition, never mind the difficulties involved in charging and operating the ship. And even if we push batteries to an energy density of 500 Wh/kg sooner than might be expected, an 18,000-TEU vessel would still need nearly 60,000 metric tons of them for a long intercontinental voyage at a relatively slow speed. The conclusion is obvious. To have an electric ship whose batteries and motors weighed no more than the fuel (about 5,000 metric tons) and the diesel engine (about 2,000 metric tons) in today’s large container vessels, we would need batteries with an energy density more than 10 times as high as today’s best Li-ion units. That’s a tall order indeed: In the past 70 years the energy density of the best commercial batteries hasn’t even quadrupled. n

↗ POST YOUR COMMENTS at https://spectrum.ieee.org/ electricships0319

22 | MAR 2019 | SPECTRUM.IEEE.ORG ILLUSTRATION BY Anders Wenngren REFLECTIONS_BY ROBERT W. LUCKY OPINION

Canada, called DeepStack, which has trounced human opponents in Texas hold ’em. The researchers write that it “plays using ‘intuition’ honed through deep learning to reassess its strategy with each decision.” To me, having little poker experience, this was a revelation: Contrary to my naive belief that winning at poker was based on psyching out opponents, poker is really a game of strategy. And the computer has learned a better strategy than we have discovered on our own. A couple of years ago, another poker program, Libratus, from Carnegie Mellon University, in Pittsburgh, bested human champions. Libratus uses a technique, unhelpfully called Monte Carlo Counter- factual Regret Minimization, which is a clever way to prune an enormous decision tree and to choose among the multi- tude of possible pathways. This technique has been shown to lead to a Nash equilib- rium strategy in which neither player can gain by changing strategy (assuming fixed CAN MACHINE LEARNING strategies by the others). In other words, this results in a tie among equally profi- cient and knowledgeable participants. TEACH US ANYTHING? This strategy interests and puzzles me. It seems to me that the computer is play- THE BREATHLESS HEADLINE caught my eye: “Computer Shows ing not to lose. I would assume that the Human Intuition—AI Breakthrough!” (or words to that effect). human players are playing to win, and I was intrigued but skeptical. Reading further, I learned that in so doing, lose. This seems counter- a computer program, AlphaZero, developed by a team at intuitive. I recall too well all the times DeepMind, in London, had beaten other champion chess-playing pro- I have yelled at my TV when my sports grams, as well as (of course) humans. That wasn’t the interesting news, teams have been in the lead and start as we take that kind of dominance for granted these days. What fasci- playing not to lose, and then do lose. I nated me was how the program had been constructed. Instead of being suppose those teams’ don’t-lose strate- tuned by expert players, AlphaZero initially knew nothing more than gies are not so optimal as those of the the rules of chess. It learned how to play, and to win, by playing against poker-playing program. itself. Soon it got so good it could beat everyone and everything. • But, I A deeper understanding of all this wondered, isn’t this what humans have been doing for centuries—learning eludes me. However, I’m left with a resi- by playing chess against ourselves? What, if anything, has the computer due of envy of those engineers work- learned so quickly that we haven’t in all those years? Unfortunately, the ing on these game-playing programs. neural network isn’t telling us. It appears, for instance, that it sacrifices What a great privilege! I know that it isn’t pieces to gain position at a greater frequency than humans would usu- all fun and games (literally), and that ally attempt. Whatever, I am still intrigued, but also still skeptical reading there is a great deal of nitty-gritty as headlines about the software showing human intuition, which is defined as well as creative thought involved. None- “immediate apprehension or cognition without reasoning.” Is AlphaZero theless, I am indeed intrigued. This is evoking human intuition, or is this superhuman intuition? Or should it great stuff. n not be called intuition at all? Similar claims are being made about a • ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/ new poker-playing program from a team at the University of Alberta, in reflections0319

ILLUSTRATION BY Edmon DeHaro SPECTRUM.IEEE.ORG | MAR 2019 | 23 When will useful quantum computers be constructed? Not in the foreseeable future, this physicist argues. Here’s why

The Case Against QUANTUM COMPUTING By Mikhail Dyakonov

ILLUSTRATIONS BY CHRISTIAN GRALINGEN

24 | MAR 2019 | SPECTRUM.IEEE.ORG

QUANTUM COMPUTING IS ALL THE RAGE. It seems like hardly a day goes by without some news outlet describing the extraordinary things this technology promises. Most commentators forget, or just gloss over, the fact that people have been working on quantum computing for decades—and without any practical results to show for it. ¶ IBM notes that quantum computers could “provide breakthroughs in many disciplines, including materials and drug discovery, the optimization of complex systems, and artificial intelligence.”M icrosoft assures us that quantum computers will “forever alter our economic, industrial, academic, and societal landscape.” And journalists repeatedly warn that quantum computers may soon break the encryption that protects the Internet. It has gotten to the point where many researchers in various fields of physics feel obliged to justify whatever work they are doing by claiming that it has some relevance to quantum computing.

Meanwhile, government research agencies, academic departments (many of them funded by government agencies), and corporate laboratories are spending billions of dollars a year developing quantum computers. On Wall Street, Morgan Stanley and other financial giants expect quantum computing to mature soon and are keen to figure out how this technology can help them. It’s become something of a self- perpetuating arms race, with many organizations seemingly staying in the race if only to avoid being left behind. Some of the world’s top technical tal- ent, at places like Google, IBM, and Microsoft, are working hard, and with lavish resources in state-of-the-art laboratories, to realize their vision of a quantum-computing future. In light of all this, it’s natural to wonder: When will useful quantum computers be constructed? The most optimistic experts estimate it will take 5 to 10 years. More cautious ones predict 20 to 30 years. (Similar predictions have been voiced, by the way, for the last 20 years.) I belong to you want to make a simulation of nature, you’d a tiny minority that answers, “Not in the foreseeable future.” Hav- better make it quantum mechanical, and by ing spent decades conducting research in quantum and condensed- golly it’s a wonderful problem, because it matter physics, I’ve developed my very pessimistic view. It’s based doesn’t look so easy,” he opined. A few years on an understanding of the gargantuan technical challenges that later, University of Oxford physicist David would have to be overcome to ever make quantum computing work. Deutsch formally described a general-purpose quantum computer, a quantum analogue of the universal Turing machine. The idea of quantum computing first appeared nearly 40 years ago, The subject did not attract much atten- in 1980, when the Russian-born mathematician Yuri Manin, who now tion, though, until 1994, when mathemati- works at the Max Planck Institute for Mathematics, in Bonn, first put for- cian Peter Shor (then at Bell Laboratories and ward the notion, albeit in a rather vague form. The concept really got on now at MIT) proposed an algorithm for an the map, though, the following year, when physicist Richard Feynman, ideal quantum computer that would allow at the California Institute of Technology, independently proposed it. very large numbers to be factored much faster Realizing that computer simulations of quantum systems become than could be done on a conventional com- impossible to carry out when the system under scrutiny gets too computer. This outstanding theoretical result plicated, Feynman advanced the idea that the computer itself should triggered an explosion of interest in quan- operate in the quantum mode: “Nature isn’t classical, dammit, and if tum computing. Many thousands of research

26 | MAR 2019 | SPECTRUM.IEEE.ORG When will useful quantum (called quantum amplitudes), which, being complex numbers, have computers be constructed? real parts and imaginary parts. Those complex numbers, α and β, each have a certain magnitude, and according to the rules of quan- Optimistic More cautious tum mechanics, their squared magnitudes must add up to 1. experts’ experts’ That’s because those two squared magnitudes correspond to the prediction prediction probabilities for the spin of the electron to be in the basic states ↑ and Years from now ↓ when you measure it. And because those are the only outcomes possible, the two associated probabilities must add up to 1. For example, if

0 5 10 15 20 25 30 35 40 45 50 the probability of finding the electron in the↑ state is 0.6 (60 percent), then the probability of finding it in the↓ state must be 0.4 (40 percent)—nothing else would make sense. papers, mostly theoretical, have since been In contrast to a classical bit, which can only be in one of its two published on the subject, and they continue basic states, a qubit can be in any of a continuum of possible states, as to come out at an increasing rate [see graph, defined by the values of the quantum amplitudesα and β. This prop- “Up and Away”]. erty is often described by the rather mystical and intimidating state- The basic idea of quantum computing is to ment that a qubit can exist simultaneously in both of its ↑ and ↓ states. store and process information in a way that Yes, quantum mechanics often defies intuition. But this concept is very different from what is done in conven- shouldn’t be couched in such perplexing language. Instead, think of tional computers, which are based on classi- a vector positioned in the x-y plane and canted at 45 degrees to the cal physics. Boiling down the many details, x-axis. Somebody might say that this vector simultaneously points in it’s fair to say that conventional computers both the x- and y-directions. That statement is true in some sense, but operate by manipulating a large number of it’s not really a useful description. Describing a qubit as being simulta- tiny transistors working essentially as on-off neously in both ↑ and ↓ states is, in my view, similarly unhelpful. And switches, which change state between cycles yet, it’s become almost de rigueur for journalists to describe it as such. of the computer’s clock. In a system with two qubits, there are 22 or 4 basic states, which can The state of the classical computer at the be written (↑↑), (↑↓), (↓↑), and (↓↓). Naturally enough, the two qubits can start of any given clock cycle can therefore be described by a quantum-mechanical wave function that involves be described by a long sequence of bits cor- four complex numbers. In the general case of N qubits, the state of responding physically to the states of individ- the system is described by 2N complex numbers, which are restricted ual transistors. With N transistors, there are by the condition that their squared magnitudes must all add up to 1. 2N possible states for the computer to be in. While a conventional computer with N bits at any given moment Computation on such a machine fundamen- must be in one of its 2N possible states, the state of a quantum com- tally consists of switching some of its tran- puter with N qubits is described by the values of the 2N quantum ampli- sistors between their “on” and “off” states, tudes, which are continuous parameters (ones that can take on any according to a prescribed program. value, not just a 0 or a 1). This is the origin of the supposed power of In quantum computing, the classical the quantum computer, but it is also the reason for its great fragility two-state circuit element (the transistor) is and vulnerability. replaced by a quantum element called a quan- How is information processed in such a machine? That’s done by tum bit, or qubit. Like the conventional bit, it applying certain kinds of transformations—dubbed “quantum gates”— also has two basic states. Although a variety that change these parameters in a precise and controlled manner. of physical objects could reasonably serve as Experts estimate that the number of qubits needed for a useful quantum bits, the simplest thing to use is the quantum computer, one that could compete with your laptop in electron’s internal angular momentum, or solving certain kinds of interesting problems, is between 1,000 and spin, which has the peculiar quantum prop- 100,000. So the number of continuous parameters describing the erty of having only two possible projections state of such a useful quantum computer at any given moment must on any coordinate axis: +1/2 or –1/2 (in units of be at least 21,000, which is to say about 10300. That’s a very big number the Planck constant). For whatever the chosen indeed. How big? It is much, much greater than the number of sub- axis, you can denote the two basic quantum atomic particles in the observable universe. states of the electron’s spin as ↑ and ↓. To repeat: A useful quantum computer Here’s where things get weird. With the needs to process a set of continuous param- quantum bit, those two states aren’t the only 1,000–100,000 eters that is larger than the number of sub- ones possible. That’s because the spin state THE NUMBER OF atomic particles in the observable universe. of an electron is described by a quantum- LOGICAL QUBITS At this point in a description of a possi- NEEDED FOR A mechanical wave function. And that func- USEFUL QUANTUM ble future technology, a hardheaded engi- tion involves two complex numbers, α and β COMPUTER neer loses interest. But let’s continue. In

SPECTRUM.IEEE.ORG | MAR 2019 | 27 NUMBER OF PAPERS PER YEAR 1,500

The literature is any real-world computer, you have that have been reported were extremely diffi- light on studies to consider the effects of errors. In a cult to conduct, though, and must command conventional computer, those arise 1,000respect and admiration. describing when one or more transistors are The goal of such proof-of-principle experi- experiments switched off when they are supposed ments is to show the possibility of carrying with actual to be switched on, or vice versa. This out basic quantum operations and to dem- hardware. unwanted occurrence can be dealt onstrate some elements of the quantum algo- with using relatively simple error- rithms500 that have been devised. The number of Those reported correction methods, which make qubits used for them is below 10, usually from were extremely use of some level of redundancy built 3 to 5. Apparently, going from 5 qubits to 50 difficult to into the hardware. (the goal set by the ARDA Experts Panel for YEAR In contrast, it’s absolutely unimag- the year 2012) presents experimental difficul- conduct and 0 inable how to keep errors under ties1996 that are hard 2001 to overcome. 2006 Most probably2011 2016 must command control for the 10300 continuous they are related to the simple fact that 25 = 32, respect and parameters that must be processed while 250 = 1,125,899,906,842,624. admiration. by a useful quantum computer. Yet By contrast, the theory of quantum com- quantum-computing theorists have puting does not appear to meet any substan- succeeded in convincing the general tial difficulties in dealing with millions of public that this is feasible. Indeed, qubits. In studies of error rates, for example, they claim that something called 1,500variousNUMBER noise OF models are being considered. PAPERS the threshold theorem proves it can It hasPER been YEAR proved (under certain assump- be done. They point out that once tions) that errors generated by “local” noise the error per qubit per quantum can be corrected by carefully designed and gate is below a certain value, indef- 1,000very ingenious methods, involving, among initely long quantum computation other tricks, massive parallelism, with many becomes possible, at a cost of substan- thousands of gates applied simultaneously to tially increasing the number of qubits different pairs of qubits and many thousands needed. With those extra qubits, they argue, you can handle errors by of measurements done simultaneously, too. forming logical qubits using multiple physical qubits. 500A decade and a half ago, the ARDA Experts How many physical qubits would be required for each logical qubit? Panel noted that “it has been established, under No one really knows, but estimates typically range from about 1,000 to certain assumptions, that if a threshold pre- 100,000. So the upshot is that a useful quantum computer now needs cision per gate operation could be achieved, a million or more qubits. And the number of continuous parameters quantum error correction would allow a quanYEAR - 0 defining the state of this hypothetical quantum-computing machine— tum1996 computer 2001 to compute 2006 indefinitely.”2011 Here, 2016 which was already more than astronomical with 1,000 qubits—now the key words are “under certain assumptions.” becomes even more ludicrous. That panel of distinguished experts did not, Even without considering these impossibly large numbers, it’s sobering that no one has yet figured out how to combine many physical qubits into a smaller number of logical qubits that can compute UP AND AWAY: The number of research reports published each year that are about quantum something useful. And it’s not like this hasn’t long been a key goal. computing and algorithms has grown enormously In the early 2000s, at the request of the Advanced Research and since the notion of quantum computing first Development Activity (ARDA—a funding agency of the U.S. intelligence became popular in the 1990s. community that is now part of Intelligence Advanced Research Proj- ects Activity), a team of distinguished experts in quantum informa- 1,500 tion established a road map for quantum computing. It had a goal for NUMBER OF 2012 that “requires on the order of 50 physical qubits” and “exercises PAPERS PER YEAR multiple logical qubits through the full range of operations required 1,000 for fault-tolerant [quantum computation] in order to perform a simple instance of a relevant quantum algorithm….” It’s now 2019, and that ability has still not been demonstrated. 500

The huge amount of scholarly literature that’s been generated YEAR about quantum computing is notably light on studies describing 0 1996 2001 2006 2011 2016 experiments with actual hardware. The relatively few experiments 2018 DECEMBER SCIENCES, OF ACADEMY NATIONAL U.S. SOURCE:

28 | MAR 2019 | SPECTRUM.IEEE.ORG however, address the question of whether Grover developed at Bell Laboratories in 1996, and other special- these assumptions could ever be satisfied. ized applications that are suitable for quantum computers. I argue that they can’t. In the physical world, On the hardware front, advanced research is under way, with a continuous quantities (be they voltages or the 49-qubit chip (Intel), a 50-qubit chip (IBM), and a 72-qubit chip (Google) parameters defining quantum-mechanical having recently been fabricated and studied. The eventual outcome wave functions) can be neither measured of this activity is not entirely clear, especially because these compa- nor manipulated exactly. That is, no continu- nies have not revealed the details of their work. ously variable quantity can be made to have While I believe that such experimental research is beneficial and an exact value, including zero. To a mathe- may lead to a better understanding of complicated quantum systems, matician, this might sound absurd, but this I’m skeptical that these efforts will ever result in a practical quantum is the unquestionable reality of the world we computer. Such a computer would have to be able to manipulate—on live in, as any engineer knows. a microscopic level and with enormous precision—a physical system Sure, discrete quantities, like the number characterized by an unimaginably huge set of parameters, each of of students in a classroom or the number of which can take on a continuous range of values. Could we ever learn transistors in the “on” state, can be known to control the more than 10300 continuously variable parameters defin- exactly. Not so for quantities that vary contin- ing the quantum state of such a system? uously. And this fact accounts for the great difference between a conventional digital computer and the There is a tremendous gap between hypothetical quantum computer. the rudimentary experiments that have Indeed, all of the assumptions that theorists make about the been carried out and the extremely preparation of qubits into a given developed quantum-computing theory. state, the operation of the quantum gates, the reliability of the My answer is simple. No, never. measurements, and so forth, cannot be ful- I believe that, appearances to the contrary, the quantum-computing filledexactly . They can only be approached fervor is nearing its end. That’s because a few decades is the maximum with some limited precision. So, the real lifetime of any big bubble in technology or science. After a certain period, question is: What precision is required? With too many unfulfilled promises have been made, and anyone who has what exactitude must, say, the square root been following the topic starts to get annoyed by further announce- of 2 (an irrational number that enters into ments of impending breakthroughs. What’s more, by that time all the many of the relevant quantum operations) tenured faculty positions in the field are already occupied. The probe experimentally realized? Should it be ponents have grown older and less zealous, while the younger gen- approximated as 1.41 or as 1.41421356237? eration seeks something completely new and more likely to succeed. Or is even more precision needed? There All these problems, as well as a few others I’ve not mentioned here, are no clear answers to these and similar raise serious doubts about the future of quantum computing. There crucial questions. is a tremendous gap between the rudimentary but very hard experiments that have been carried out with a few qubits and the extremely

developed quantum-computing theory, which relies on manipulat- While various strategies for building quan- ing thousands to millions of qubits to calculate anything useful. That tum computers are now being explored, an gap is not likely to be closed anytime soon. approach that many people consider the To my mind, quantum-computing researchers should still heed most promising, initially undertaken by the an admonition that IBM physicist Rolf Landauer made decades ago Canadian company D-Wave Systems and now when the field heated up for the first time. He urged proponents of being pursued by IBM, Google, Microsoft, quantum computing to include in their publications a disclaimer and others, is based on using quantum sys- along these lines: “This scheme, like all other schemes for quantum tems of interconnected Josephson junctions computation, relies on speculative technology, does not in its current cooled to very low temperatures (down to form take into account all possible sources of noise, unreliability and about 10 millikelvins). manufacturing error, and probably will not work.” n The ultimate goal is to create a universal quantum computer, one that can beat conventional computers in factoring large Continue the conversation: Readers have been commenting about this numbers using Shor’s algorithm, perform- article since it was first published online in N ovember 2018. You can add your c omments ing database searches by a similarly famous on this topic at https://spectrum.ieee.org/quantumskeptic0319 See also this month’s Spectral Lines, “Quantum Computing’s Prospects,” p. 6.

SOURCE: U.S. NATIONAL ACADEMY OF SCIENCES, DECEMBER 2018 DECEMBER SCIENCES, OF ACADEMY NATIONAL U.S. SOURCE: quantum-computing algorithm that Lov

SPECTRUM.IEEE.ORG | MAR 2019 | 29

Walk This Way To be useful around people, robots need to learn how to move like we do By Jonathan Hurst Photography by Dan Saelinger

SPECTRUM.IEEE.ORG | MAR 2019 | 31 REAL-LIFE DROIDS: Agility Robotics designed its legged robots Cassie [left] and Digit to move in a more dynamic fashion than regular robots do. robots have walked on legs for decades. Today’s most advanced humanoid robots can tramp along flat and inclined surfaces, climb up and down stairs, and slog through rough terrain. Some can even jump. But despite the progress, legged robots still can’t begin to match the agility, efficiency, and robustness of humans and animals. • Existing walking robots hog power and spend too much time in the shop. All too often, they fail, they fall, and they break. For the robotic helpers we’ve long dreamed of to become a reality, these machines will have to learn to walk as we do. • We must build robots with legs because our world is designed for legs. We step through narrow spaces, we navigate around obstacles, we go up and down steps. Robots on wheels or tracks can’t easily move around the spaces we’ve optimized for our own bodies.

Indeed, many humanoids feature legs that look similar to disabilities more mobility. They’ll finally bring the robots ours, with hips, knees, ankles, and feet. But the similarities imagined in science fiction into our daily lives. usually end there: If you compare, for example, the forces these robots exert on the ground with those a human exerts, you’ll discover that they are often quite different. Most humanoids, descended from early industrial robot arms, control their Some birds run better than they can fly, if they can fly at all. limbs to follow specified trajectories as accurately and rigidly Ostriches, turkeys, guinea fowl, and quail can’t soar like a as possible. However, legged locomotion does not require hawk, but they are fast on their feet. In collaboration with position control so much as force control, with plenty of pli- Monica Daley from the University of London’s Royal Veteri- ability and elasticity—known in robotics as compliance—to nary College, my colleagues and I have spent countless hours allow for unexpected contacts. observing birds walking and running in the laboratory. We A number of research groups have been trying to build robots want to understand how these animals move so nimbly and that are less stiff and that can move in a more dynamic, human- efficiently—most of these feathered machines are powered like fashion. The most famous such robot is perhaps Atlas, from only by seeds! Boston Dynamics, a humanoid that can run on hard and soft In one experiment, a guinea fowl is running down a track terrain, leap over logs, and even do backflips. But again, when when it steps on a pit concealed by a sheet of tissue paper. The we compare the motion of even the most sophisticated robots bird didn’t know it was going to step into a pothole about half with what animals can achieve, the machines fall short. a leg deep; yet the animal doesn’t miss a beat, its leg stretch- What are we missing? Technology isn’t the biggest hurdle: ing to adjust to the drop as it darts past the obstacle. What’s Motors are powerful enough, materials are strong enough, happening here is pretty remarkable: The bird’s brain doesn’t and computers are fast enough. Rather, the limiting factor have to sense and react to the disturbance because its legs appears to be our basic understanding of how legged loco- can handle it on their own. motion works. This offers an important insight for robot designers: If you At Oregon State University’s Dynamic Robotics Laboratory, first build your robot and only then try to program it to be I lead a group of researchers seeking to identify the underlying agile, you’re likely doomed to fail. As with the guinea fowl, a principles of legged locomotion and to apply our discoveries robot’s agility will derive in great part from inherent mechan- to robots. I’m also the cofounder and chief technology offi- ical properties of its body, or what roboticists call passive cer of Agility Robotics, a startup based in Albany, Ore., that’s dynamics. And it’s been neglected in most legged-robot proj- exploring commercial uses for robotic legged mobility. In 2017, ects. By thoughtfully engineering the passive dynamics of we unveiled Cassie, a bipedal platform we have sold to several your robot in parallel with software control as an integrated research groups. Soon we will have a new robot ready to step system, you increase your chances of making a robot that out into the world: Digit has legs similar to Cassie’s, but it also approaches the performance of an animal. features perception sensors and a pair of arms, which it will Now, it’s worth pointing out that although animals serve as our use for stability and, in the future, manipulation. inspiration, we don’t replicate the shape of a bird’s foot or the Through both the lab and the company, we’re working arrangement of muscles and bones in a human leg. Instead, we toward a future in which robots will be able to go anywhere want to capture the physics of animal locomotion and abstract people go. I believe dynamic-legged robots will one day help from it a mathematical model that we can understand, test take care of elderly and infirm people in their homes, assist in computer simulations, and then implement in real robots. with lifesaving efforts in fires and earthquakes, and deliver Because we are using metal and electronics rather than bones packages to our front doors. Robotic legs will also enable exo- and brains to build these robots, they may look quite different skeletons and powered prosthetic limbs to give people with from an animal while still capturing the same physics.

PHOTOGRAPHS BY Dan Saelinger SPECTRUM.IEEE.ORG | MAR 2019 | 33 SPECIAL DELIVERY: Digit is learning to walk on different types of terrain and also climb stairs, so it may one day deliver packages directly to your front door.

One of the simplest mathematical models consists of a point part of the model, handling ground impacts and storing mass (representing the upper body) attached to a pair of ideal mechanical energy. springs, which represent the legs. This model, known as a At first, ATRIAS could barely stand, and we supported it with spring-mass model, is a simplification, of course; it resembles an overhead tether. But as we refined its controller, which kept a stick figure and doesn’t take into account that legs have track of the body’s speed and tilt, the robot took its first steps joints, or that the feet don’t touch the ground at discrete and soon was pacing through the lab. Next ATRIAS learned points. Still, the spring-mass model can produce surprising to recover from disturbances, which in one experiment con- results: In simulations, it can generate nearly every walking sisted of my students throwing dodgeballs at it. We also took and running gait observed in people and animals. ATRIAS to the university’s football field, accelerated the robot to a top speed of 7.6 kilometers per hour (4.7 miles per hour), and then quickly brought it to a stop—in the end zone, of course. To understand better what the robot was doing, imagine To test the spring-mass model in a machine, my group at you’re blindfolded and wearing stilts, with your upper body Oregon State—in collaboration with Hartmut Geyer from wrapped up in a carpet, so you can’t use your arms for balanc- Carnegie Mellon University and Jessy Grizzle from the Univer- ing. All you can do is keep stepping, and that’s what ATRIAS sity of Michigan—developed ATRIAS, a bipedal robot whose did. It was even able to handle obstacles, such as a pile of ply- name is an acronym reflecting our main premise: Assume wood placed on its path. The Robot Is A Sphere. The idea was that the passive dynam- While agility was important, it was also critical that ATRIAS ics of the robot could come as close as possible to reproduc- economize on power. We confirmed it did when we measured ing those of the point mass with springy legs. a parameter known as cost of transport (COT). It’s defined as We built each leg out of lightweight carbon-fiber rods, the ratio of power consumption to weight times velocity, and arranged in a parallelogram-shaped structure known as a it’s used to compare the energy efficiency of how animals and four-bar linkage. This structure minimized the mass of the machines move. The lower the COT, the better. A person walk- legs and their associated inertia, approximating the spring- ing, for example, has a COT of 0.2, while conventional human- mass model. We also equipped the upper portion of the legs oid robots have much higher values—ranging between 2 and 3, with fiberglass springs, which physically embody the “spring” according to some estimates. Our experiments showed that

34 | MAR 2019 | SPECTRUM.IEEE.ORG PHOTOGRAPHS BY Dan Saelinger ATRIAS, in walking mode, had a COT of 1.13, demonstrating we set a high bar: We wanted Cassie to be able to run through the efficiency benefits of dynamic robots. Indeed, a few small a forest, handling rough terrain and operating for hours on lithium-polymer batteries—the type used in radio-controlled a battery, without a safety tether. cars—could keep ATRIAS operating for about an hour. Cassie builds on the same concepts developed for ATRIAS, but We also measured the forces the robot was exerting on the we decided to give it brand-new legs. We’d used two motors to ground. We placed ATRIAS, which weighs in at 72.5 kilograms, power the four-bar linkage on each leg of ATRIAS. This arrange- or about the same as a person, on a force plate, an instru- ment minimized their mass, but there was a trade-off: During ment often used in sports medicine to assess a person’s gait part of the step cycle, one motor acted as a brake on the other, by gauging ground-reaction forces. While the robot walked, costing significant energy unnecessarily. For Cassie, we stud- we recorded the force data. Then we replaced ATRIAS with ied alternative leg configurations to eliminate this effect. The one of my students, and recorded his steps. When we plot- new design allows the motors to be smaller, and thus makes the ted the ground-reaction forces over time, the two data sets robot even more efficient than ATRIAS was. had the exact same shape. As far as we know, it is the most It’s important to note that Cassie’s leg configuration was a realistic robotic implementation of human gait dynamics result of this analysis. The fact that the leg resembles that of that has ever been done. an ostrich or other theropod may be a sign that we are on the The results confirmed that a simple actuated spring-mass right track, but it was never our goal to create a robot that, with system is something that can be built into a robot; enable a few well-placed feathers, could fit right in with a mob of emus. many of the traits we seek such as efficiency, robustness, Each of Cassie’s legs has five axes of motion—or degrees and agility; and get at the core of what legged locomotion is. of freedom in robotic parlance—each one driven by a motor. It was time to build our next robot. The hips have three degrees of freedom, similar to our own, allowing the leg to swing in any direction. Two other motors power joints in the knee and foot. Cassie has additional degrees of freedom in its shin and ankle; these are passive, not con- Cassie is a dynamic walker like ATRIAS. We optimized every trolled by motors but rather attached to springs, which help aspect of its design with the goal of building a rugged and the robot move through complex terrain that flat-footed capable robot that Agility Robotics could commercialize. And humanoids can’t handle. | CONTINUED ON PAGE 50

SPECTRUM.IEEE.ORG | MAR 2019 | 35 BIG PICTURE: This Beijing dispatch center controls most of China’s ultrahigh-voltage lines and monitors renewable energy use.

STATE GRID CORP. OF CHINA

A Grid as Big as China

POWER SHIFT: This station in Zhejiang province imports hydropower from Sichuan province as direct current and converts it to alternating current.

XU YU/XINHUA/REDUX

36 | MAR 2019 | SPECTRUM.IEEE.ORG NIMBY: Coal plants in Inner Mongolia feed this station near Shanghai, reducing the megacity’s air pollution.

IMAGINECHINA/AP

Building the world’s first ultrahigh-voltage AC-DC power grid is hard. Operating it is harder By Peter Fairley

SUPERSIZED: Pushing UHV technology to 1,100 kilovolts requires upscaled components like this 800-metric-ton transformer.

ABB

SPECTRUM.IEEE.ORG | MAR 2019 | 37 CRUSHING IT: China’s newest UHV line from Xinjiang to Anhui has set world records for transmission distance, power, and voltage.

Wind rips across an isolated utility station in northwestern China’s desolate Gansu Corridor. More than 2,000 years ago, Silk Road traders from Central Asia and Europe crossed this arid, narrow plain, threading between forbidding mountains to the south and the Gobi Desert to the north, bearing precious cargo bound for Imperial Beijing. Today the corridor carries a distinctly modern commodity: gigawatts of electricity destined for the megacities of eastern China. One waypoint on that journey is this ultrahigh-voltage (UHV) converter station outside the city of Jiuquan, in Gansu province.

Electricity from the region’s wind tur- that grid operators have built over the that the new long-haul DC lines don’t bines, solar farms, and coal-fired power past decade. In the northwestern region destabilize China’s regional AC grids. plants arrives at the station as alternat- of Xinjiang, China recently switched on For example, if the 8-gigawatt DC line ing current. Two dozen 500-metric-ton its largest UHV link: a 1,100-kV DC cir- from Gansu were to unexpectedly go off transformers feed the AC into a cavern- cuit that cost over 40.7 billion yuan. The line, the power shock could cause wide- ous hall, where AC-DC converter circuits new line’s taller transmission towers and spread blackouts in Hunan and beyond. hang from the 28-meter-high ceiling, beefier wires parallel the Gansu–Hunan To minimize the threat, the State Grid emitting a penetrating, incessant buzz. line through the Gansu Corridor, before Corp. of China, a state-owned company Within each circuit, solid-state switches diverting to Anhui province in the east. that runs most of China’s transmission known as thyristors chew up the AC and The result of all this effort is an emerg- and distribution grids, intentionally spit it out as DC flowing at 800 kilovolts. ing nationwide supergrid that will inter- limits the line’s throughput to no more From here, the transmission line tra- connect China’s six regional grids and than 4.5 GW. In practice, the line has car- verses three more provinces before ter- rectify the huge geographic mismatch ried less than one-quarter of its design minating at a sister station in Hunan between where China produces its clean- capacity on average. That’s one reason province, more than 2,300 kilometers est power (in the north and west) and why over one-third of Gansu province’s away. There, the DC is converted back where power is consumed (in the densely theoretical wind output and one-fifth of to AC, to be fed onto the regional power populated east). By using higher voltages its solar potential went unused in 2017. grid. Since it opened in mid-2017, the of direct current, which flows through Other UHV lines in neighboring regions 26.2 billion yuan (US $3.9 billion) Gansu– conductors more uniformly than does have similarly operated below capacity. Hunan transmission line has moved alternating current, the new trans- And eastern provinces don’t have suf- about 24 terawatt-hours. mission lines dramatically reduce the ficient incentive to import the cleaner The sheer scale of the new line and the amount of power that’s lost along the way. power that the UHV lines offer. advanced grid technology that’s been But even as China celebrates the com- The ultimate solution to both issues, developed to support it dwarf anything pletion of more than 30,000 km of UHV according to State Grid engineers, is to going on in pretty much any other coun- lines, power engineers are struggling double down on UHV. They argue that try. And yet, here in China, it’s just one of to master the resulting hybrid AC-DC the country must move far more energy

22 such ultrahigh-voltage megaprojects transmission system. They must ensure via UHV DC to maximize the use of renew- IMAGES VCG/GETTY

38 | MAR 2019 | SPECTRUM.IEEE.ORG able energy while slashing reliance on coal. Over the next decade, Liu delivered. China’s nuclear power capacity and con- State Grid is also building a world-leading He put some 2,000 State Grid engi- straints on local coal power due to air- set of ultrahigh-voltage AC lines, to help neers on the project and funded more quality concerns. The new UHV grid is eastern China’s regional AC grids absorb than 300 professors and 1,000 grad- also helping the country lead the global the output from those massive lines. uate students at Chinese universities transition to renewable generation, mov- “The UHV AC power grid is like a deep- to conduct power-grid-related R&D. ing 161.5 terawatt-hours of hydro, wind, water port, and the UHV DC is like a State Grid expanded and refocused its and solar energy in 2017 alone. 10,000-ton ship. Only the deep-water port research centers to attack specific UHV ABB, Siemens, and other international can support the 10,000-ton ship,” says issues, including how to safely handle power-technology companies have been Qin Xiaohui, vice director of power system the higher electromagnetic fields and instrumental in developing and validat- planning with State Grid’s China Electric the more potent impulses during switching key components of the Chinese UHV Power Research Institute, in Beijing. ing and faults. grid. But State Grid has insisted on shar- Meanwhile, power authorities every- In January 2009, State Grid ener- ing the intellectual property for the tech- where are watching. Gregory Reed, a gized its first UHV demonstration line— nologies developed at its behest. DC transmission expert who runs the a 650-km, 1,000-kV UHV AC transmission In a 2014 interview, Executive Vice University of Pittsburgh’s Center for line that linked the North China and President Liu Zehong described one Energy, says China’s UHV grid puts it far Central China regional grids. Ten years tense episode in 2006 when State Grid ahead of the rest of the world. “They’re on, State Grid has completed 19 of 30 asked international suppliers to help investing significantly, and they’ve gone proposed UHV lines. develop 6-inch-diameter thyristors capa- right to the highest levels of technology That aggressive build-out has helped ble of handling more current than 5-inch capability from day one. There’s no com- fast-growing urban centers such as thyristors could. The suppliers initially parison anywhere else in the world. It’s Shanghai stave off power shortages balked, said Liu, but ultimately relented like we’re all still pedaling our bicycles, despite delays in the expansion of because of State Grid’s “determined atti- while the Formula 1 race car goes flying by.”

China’s UHV movement was born of a limo ride. It was late 2004, and Liu China’s Hybrid AC-DC Grids Ultrahigh-voltage DC lines move coal-fired and renewable Zhenya, then president of State Grid, was generation thousands of kilometers to China’s megacities. sharing a car with Kai Ma, minister of UHV AC helps distribute the imported electricity. the National Development and Reform Commission (NDRC), the powerful state body that regulates China’s growth and major investments. As Chinese policy Northwest China’s wind expert Yi-chong Xu describes in her 2017 and solar power book Sinews of Power (Oxford University plants are the Harbin Press), Kai complained of the crippling world’s largest.

power shortages of the day. Liu blamed Urumqi

“weak and fragmented” grids, ones ill- Jiuquan equipped to exchange bulk power. And Shenyang

he proposed a bold solution: massive Beijing cross-country power lines utilizing the most advanced UHV technologies. Within a year, Kai’s NDRC had approved an ambitious and compre-

hensive plan that embraced Liu’s vision. Shanghai Wuhan Chongqing It combined UHV DC lines, which excel Lhasa Chengdu at moving bulk power from one spot to Greater another over long distances, and a UHV Changsha Shanghai’s Coal UHV DC energy AC backbone to reliably distribute that Hydro UHV DC under construction imports power to consumers. State Grid would Nuclear may triple in UHV AC Guangzhou lead the engineering and ensure that Solar the coming UHV AC domestic suppliers would manufacture Wind under construction Hong Kong decade. 90 percent of the UHV equipment, thus building up a new high-tech export sec- China Southern Power Grid split its tor for China. AC grid in two to make it more stable.

ILLUSTRATION BY Erik Vrielink SPECTRUM.IEEE.ORG | MAR 2019 | 39 tude” and the “huge market opportunities” of the Chinese market. Two years later, Chinese firms were manufacturing the resulting 6-inch switches.

For all of State Grid’s progress, its UHV deployment remains uneven and incomplete. China could end up with just half of the 89,000 km of UHV lines that its plans called for by 2020 and none of the anticipated UHV links to Kazakhstan, Mongolia, and Russia. Many proposed projects—particularly for the UHV AC backbone—have failed to gain the NDRC’s blessing. As a result, many areas still have no UHV AC lines, and both types of UHV are delivering well below expectations. What has blocked full implementation is an intense debate over the future of UHV. Some Chinese grid experts question the hundreds of billions of yuan spent on UHV projects and what they see as State Grid’s monopolization of consume what’s known as reactive power— on the AC grid that feeds it. “The UHV grid engineering and manufacturing. found in AC systems in which the current DC line is actually acting as an amplifier. Provincial officials have chafed at the and voltage are out of phase. (By contrast, A small AC disturbance in the receiving centralization of grid planning and oper- active, or real, power is the power that’s end can become a large AC disturbance in ation that UHV requires. actually consumed by the grid’s loads; its the sending-end grid,” says Zhang. Some experts have also criticized Liu’s current and voltage waves are aligned.) To minimize the risk of multiple con- ultimate goal for the UHV AC backbone— By consuming reactive power, the UHV verter failures and cascading blackouts, linking up and synchronizing China’s converters tend to pull down the voltage engineers for State Grid’s East China regional grids—as far too risky. Han of surrounding AC lines, so converter sta- regional grid have deployed a fiber- Yingduo , a member of the prestigious tions have equipment to supply reactive optic control network that automatically Chinese Academy of Engineering and power and prop up the AC voltage. rebalances supply and demand. If nec- a professor at Tsinghua University, in But if an AC line’s voltage sags, nearby essary, it can boost line voltage within Beijing, has warned that unifying C hina’s converters will consume even more reac- 200 milliseconds of a voltage drop, using grid would make it far more vulnerable to tive power, pulling voltage down further. a set of fault responses that have been cascading blackouts, like the one in 2003 A voltage sag can also disrupt the thy- built into the East China grid’s AC-DC that knocked out power in the northeast- ristors’ ability to switch from one cur- converters. As soon as the fiber-optic ern United States and Canada. rent path to another, a process known network flags an outage on a UHV DC Because no other country has ever as commutation. A severe commutation line, the converters pull up to 10 percent built a hybrid UHV AC-DC grid, State failure will cause the converter to shut more power over the remaining DC lines Grid engineers are having to feel their down, deepening the AC voltage drop to keep the grid operational. The optical way along. In a traditional lower-voltage and starting a potentially destructive control scheme can also restore balance network, the grid operator typically feedback loop that could end in a black- by releasing power from pumped hydro reserves emergency power to cover the out. “Successive DC commutation fail- plants, which store energy by pushing sudden loss of the grid’s largest asset. ures will trigger a chain reaction,” says water uphill. And it can trigger small That may mean keeping a gigawatt or two Qin, the system planning expert at State controlled blackouts, shutting off some of extra power generation at the ready. Grid’s Beijing research institute. distribution feeders to reduce demand Now add multiple UHV lines to your The resulting blackout could travel far while sparing hospitals and other essen- network, each carrying 8 to 12 GW, and and fast, notes Zhang Fang, a system oper- tial loads. your requirements for reserve power ator in State Grid’s National Electric Power These measures have enabled a trio of rise dramatically. Dispatching and Control Center, in Beijing. UHV DC lines that deliver hydropower Maintaining the ideal voltage on a When a UHV DC circuit goes off line unex- from the Southwest China grid to oper- UHV grid is also enormously challeng- pectedly, it creates a power surge hun- ate continuously at their combined

ing. Thyristor-based UHV converters dreds or thousands of kilometers away, 21.6-GW design capacity. The result is FAIRLEY PETER

40 | MAR 2019 | SPECTRUM.IEEE.ORG Controllers may reroute power from a It has led the creation of nine UHV province with low electricity demand standards through the International to another where demand is higher. Or Electrotechnical Commission and they may steer electricity to one of State the IEEE—a move that researchers at Grid’s 21 pumped hydro plants, which Argonne National Laboratory, in Illinois, collectively can soak up 19 GW. warned would help Chinese suppliers In theory, Chinese law has long “crowd others out of the global market.” required grid operators to prioritize State Grid is already working on its first renewable energy. But in practice, each international UHV DC project: a pair of province has its own plans and priorities, 800-kV lines to move power from Brazil’s which tend to favor electricity generated Belo Monte megadam. But subsequent locally. For instance, in Zhejiang province, UHV sales have been slow to material south of Shanghai, significant opposition ize. That may be because most coun to importing electricity has hampered tries do not yet need, or cannot afford, the operation of an 8-GW UHV DC line a 1,000-kV AC or DC line. MODERN IMPORTS: A trio of UHV DC lines traces the Silk Road from Ningxia province, according to ana Undaunted, former State Grid chair in China’s Gansu province. lysts at Bloomberg New Energy Finance. man Liu is now crusading to build trans On the windy, sunny day when I vis continental and intercontinental UHV ited Gansu’s DC converter station last grids. The same technology that went year, its UHV line was carrying just 3 GW into building the 1,100-kV line from of its 8-GW capacity. That was the cumu Xinjiang to Anhui could efficiently move lative output from several renewable power up to 5,000 kilometers. “If we an electrical trifecta: Greater S hanghai, plants. But the province also has an addi just turn that line around to point west, China’s most densely urbanized and tional 15 GW of solar and wind that’s we are getting close to Europe. So the industrialized region, gets more clean connected to the new line but not yet technology is available,” says Magnus power; the Yangtze River Delta’s mega authorized to feed power into it. Callavik, general manager of ABB dams spill less excess water during flood Change is coming. Two months after Sifang Power System, a Beijing-based season; and State Grid earns more rev my visit, power companies in coastal joint venture between Swiss power- enue from its UHV investment. Even Jiangsu province struck a deal to buy engineering giant ABB and China’s so, Shanghai still runs short of power power from Gansu’s largest wind farm Sifang Automation. for several weeks each summer, forc via another UHV DC line. And last Callavik says he is convinced that ing State Grid to pay big customers to November, State Grid began building continental-scale UHV DC will happen, idle their factories. Keeping pace with a UHV DC line from Qinghai province sooner or later. In a world that must growth may require tripling Shanghai’s to move even more of Gansu’s renew decarbonize, figuring out how to bal electricity imports within a decade. able generation. Meanwhile, the NDRC ance variable energy supplies such as is stoking demand by mandating mini solar and wind generation with regional At the national control center, in mum rates of renewable energy use by loads is a growing concern. “Transmis Beijing, mounting pressure to push each region. sion is a very cost-efficient way of doing more clean power through State Grid’s State Grid’s long-term goal to inter that,” says Callavik. UHV lines is hard to miss. The main connect its regional grids should also In China the question is how quickly screen displays the status of the AC and reduce curtailment, experts say. Zhang State Grid will overcome the technical DC trunk lines, providing a real-time Ning, an authority on renewables inte and political obstacles that are holding view of the entire system. Dominating gration at Tsinghua University, points back UHV’s carbon-slashing potential. If the left wall are warning lights tracking out that the Southwest grid’s hydro the country continues to rely heavily on renewable energy curtailment in each of power can balance the fluctuations in coal power, importing that power over 25 provinces—and who should be fixing it. the Northwest’s wind and solar out thousands of kilometers will help clear Green lights mean that all of the poten put. “If we interconnect the West, cur the air in China’s eastern megacities. tial solar and wind power is being used. tailment of wind power there can be But the country’s carbon footprint will Blue, yellow, and orange lights indicate reduced from more than 20 percent remain unchanged, and the benefits for renewable energy waste, which State to 5 percent,” he estimates, and both the global climate will be nil. Mobiliz Grid’s provincial, regional, or national regions’ use of coal can also be cut. ing gigawatts of renewable power over a controllers, respectively, must try to stop. UHV grid, on the other hand, promises a “We are determined to consume the Even as State Grid irons out the kinks real change, for China and the world. n renewable energy to the maximum in its UHV grids, the company is push ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/ extent. That’s our job,” says Zhang. ing its equipment and expertise abroad. chinauhvgrid0319

SPECTRUM.IEEE.ORG | MAR 2019 | 41 42 | MAR 2019 | SPECTRUM.IEEE.ORG An in-depth look at these dangerous exploitations of How the microprocessor vulnerabilities SPECTRE and MELTDOWN HACKS Really Worked

WE’RE USED TO THINKING of computer processors as orderly machines that proceed from one simple instruction to the next with complete regularity. But the truth is, that for decades now, they’ve been doing their tasks out of order and just guessing at what should come next. They’re very good at it, of course. So good in fact, that this ability, called speculative execution, has underpinned much of the improvement in computing power during the last 25 years or so. But on 3 January 2018, the world learned that this trick, which had done so much for modern computing, was now one of its greatest vulnerabilities. Throughout 2017, researchers at Cyberus Tech- nology, Google Project Zero, Graz University of Technology, Rambus, University of Adelaide, and University of Pennsylvania, as well as independent researchers such as cryptographer Paul Kocher, separately worked out attacks that took advantage of speculative execution. Our By Nael Abu- own group had discovered Ghazaleh, Dmitry the original vulnerability Ponomarev behind one of these attacks & Dmitry back in 2016, but we did not Evtyushkin put all the pieces together.

SPECTRUM.IEEE.ORG | MAR 2019 | 43 These types of attacks, called Meltdown and Spectre, were cution: When one instruction is done using a stage, the next no ordinary bugs. At the time it was discovered, Meltdown instruction is free to use it. could hack all Intel x86 microprocessors and IBM Power pro- Since the 1990s, microprocessors have relied on two tricks cessors, as well as some ARM-based processors. Spectre and to speed up the pipeline process: out-of-order execution and its many variations added Advanced Micro Devices (AMD) speculation. If two instructions are independent of each other— processors to that list. In other words, nearly the whole world that is, the output of one does not affect the input of another— of computing was vulnerable. they can be reordered and their result will still be correct. And because speculative execution is largely baked into pro- That’s helpful, because it allows the processor to keep work- cessor hardware, fixing these vulnerabilities has been no easy ing if an instruction stalls in the pipeline. For example, if an job. Doing so without causing computing speeds to grind into instruction requires data that is out in DRAM main memory low gear has made it even harder. In fact, a year on, the job is rather than in the cache memory located in the CPU itself, far from over. Security patches were needed not just from the it might take a few hundred clock cycles to get that data. processor makers but from those further down the supply chain, Instead of waiting, the processor can move another instruc- such as Apple, Dell, Linux, and Microsoft. The first computers tion through the pipeline. powered by chips that are intentionally designed to be resis- The second trick is speculation. To understand it, start with tant to even some of these vulnerabilities arrived only recently. the fact that some instructions necessarily lead to a change Spectre and Meltdown are the result of the difference in which instructions come next. Consider a program con- between what software is supposed to do and the processor’s taining an “if” statement: It checks for a condition, and if the microarchitecture—the details of how it actually does those condition is true, the processor jumps to a different location things. These two classes of hacks have uncovered a way for in the program. This is an example of a conditional-branch information to leak out through that difference. And there’s instruction, but there are other instructions that also lead every reason to believe that more ways will be uncovered. to changes in the flow of instructions. We helped find two, Branchscope and SpectreRSB, last year. Now consider what happens when such a branch instruc- If we’re going to keep the pace of computing improvements tion enters a pipeline. It’s a situation that leads to a conun- going without sacrificing security, we’re going to have to drum. When the instruction arrives at the beginning of the understand how these hardware vulnerabilities happen. pipeline, we do not know its outcome until it has progressed And that starts with understanding Spectre and Meltdown. fairly deep into the pipeline. And without knowing this outcome, we cannot fetch the next instruction. A simple but naive IN MODERN COMPUTING SYSTEMS, software programs solution is to prevent new instructions from entering the pipe- written in human-understandable languages such as C++ are line until the branch instruction reaches a point at which we compiled into assembly-language instructions—fundamen- know where the next instruction will come from. Many clock tal operations that the computer processor can execute. To cycles are wasted in this process, because pipelines typically speed execution, modern processors use an approach called have 15 to 25 stages. Even worse, branch instructions come pipelining. Like an assembly line, the pipeline is a series of up quite often, accounting for upwards of 20 percent of all stages, each of which is a step needed to complete an instruc- the instructions in many programs. tion. Some typical stages for an Intel x86 processor include To avoid the high performance cost of stalling the pipeline, those that bring in the instruction from memory and decode modern processors use an architectural unit called a branch it to understand what the instruction means. Pipelining basi- predictor to guess where the next instruction, after a branch, cally brings parallelism down to the level of instruction exe- will come from. The purpose of this predictor is to speculate

HOW A PIPELINE Branches in pipelines SPEEDS COMPUTING What if the order of USING SPECULATIVE Waiting execution depends on EXECUTION instructions the results of one of the instructions? Here, Pipelines 1 whether to execute 2 instruction B or instruction By breaking up The pipeline execution into 3 C next depends on the stages, a pipeline 4 outcome of A. Waiting for completes A to finish wastes time. instructions at a higher rate than if each instruction had Completed to finish before the instructions next began. The program wastes three clock cycles. Time Time

44 | MAR 2019 | SPECTRUM.IEEE.ORG about a couple of key points. First, will a conditional branch it to wait, provided that the condition is eventually resolved be taken, causing the program to veer off to a different sec- and any results from bad guesses are, effectively, undone. tion of the program, or will it continue on the existing path? It’s this type of intra-instruction speculation that’s behind And second, if the branch is taken, where will the program go— all variants of the Meltdown bug, including its arguably more what will be the next instruction? Armed with these predic- dangerous version, Foreshadow. tions, the processor pipeline can be kept full. The insight that enables speculation attacks is this: During Because the instruction execution is based on a predic- misspeculation, no change occurs that a program can directly tion, it is being executed “speculatively”: If the prediction observe. In other words, there’s no program you could is correct, performance improves substantially. But if the write that would simply display any data generated dur- prediction proves incorrect, the processor must be able to ing speculative execution. However, the fact that specula- undo the effects of any speculatively executed instructions tion is occurring leaves traces by affecting how long it takes relatively quickly. instructions to execute. And, unfortunately, it’s now clear The design of the branch predictor has been robustly that we can detect these timing signals and extract secret researched in the computer-architecture community for data from them. many years. Modern predictors use the history of execution within a program as the basis for their results. This scheme WHAT IS THIS TIMING INFORMATION, and how does a achieves accuracies in excess of 95 percent on many differ- hacker get hold of it? To understand that, you need to grasp ent kinds of programs, leading to dramatic performance the concept of side channels. A side channel is an unintended improvements, compared with a microprocessor that does not pathway that leaks information from one entity to another speculate. Misspeculation, however, is possible. And unfor- (usually both are software programs), typically through a tunately, it’s misspeculation that the Spectre attacks exploit. shared resource such as a hard drive or memory. Another form of speculation that has led to problems is As an example of a side-channel attack, consider a device speculation within a single instruction in the pipeline. That’s that is programmed to listen to the sound emanating from a pretty abstruse concept, so let’s unpack it. Suppose that a printer and then uses that sound to deduce what is being an instruction requires permission to execute. For instance, printed. The sound, in this case, is a side channel. an instruction could direct the computer to write a chunk of In microprocessors, any shared hardware resource can, data to the portion of memory reserved for the core of the in principle, be used as a side channel that leaks informa- operating system. You wouldn’t want that to happen, unless tion from a victim program to an attacker program. In a it was sanctioned by the operating system itself, or you’d risk commonly used side-channel attack, the shared resource is crashing the computer. Prior to the discovery of Meltdown the CPU’s cache. The cache is a relatively small, fast-access and Spectre, the conventional wisdom was that it is okay to memory on the processor chip used to store the data most start executing the instruction speculatively even before the frequently needed by a program. When a program accesses processor has reached the point of checking whether or not memory, the processor first checks the cache; if the data is the instruction has permission to do its work. there (a cache hit), it is recovered quickly. If the data is not In the end, if the permission is not satisfied—in our exam- in the cache (a miss), the processor has to wait until it is ple, the operating system has not sanctioned this attempt to fetched from main memory, which can take several hundred fiddle with its memory—the results are thrown out and the clock cycles. But once the data arrives from main memory, program indicates an error. In general, the processor may it’s added to the cache, which may require tossing out some speculate around any part of an instruction that could cause other data to make room. The cache is divided into segments

Speculative execution Misspeculation To save time, a processor If the prediction predicts what the next Waiting Speculative was wrong, the instruction will be. The instructions execution results of the instructions are executed speculation are speculatively and become 1 thrown out. visible to the program 2 The pipeline only if the prediction 3 was correct. 4

The processor doesn’t know if the prediction was Spectre and Meltdown take correct until instruction Completed advantage of speculatively A reaches the end of the instructions executed instructions during pipeline. misspeculation.

Time Time

SPECTRUM.IEEE.ORG | MAR 2019 | 45 called cache sets, and each location in main memory has a simple, a single instruction often consists of multiple operations corresponding set in the cache. This organization makes it that can depend on one another. For example, memory-read easy to check if something is in the cache without having to operations are often dependent on the instruction satisfying search the whole thing. the permissions associated with the memory address being Cache-based attacks had been extensively researched even read. An application usually has permission to read only from before Spectre and Meltdown appeared on the scene. Although memory that’s been assigned to it, not from memory allocated the attacker cannot directly read the victim’s data—even when to, say, the operating system or some other user’s program. that data sits in a shared resource like the cache—the attacker Logically, we should check the permissions before allowing can get information about the memory addresses accessed the read to proceed, which is what some microprocessors do, by the victim. These addresses may depend on sensitive data, notably those from AMD. However, provided the final result is allowing a clever attacker to recover this secret data. correct, CPU designers assumed that they were free to specu- How does the attacker do this? There are several possible latively execute these operations out of order. Therefore, Intel ways. One variation, called Flush and Reload, begins with the microprocessors read the memory location before checking attacker removing shared data from the cache using the “flush” permissions, but only “commit” the instruction—making the instruction. The attacker then waits for the victim to access results visible to the program—when the permissions are sat- that data. Because it’s no longer in the cache, any data the vic- isfied. But because the secret data has been retrieved specula- tim requests must be brought in from main memory. Later, the tively, it can be discovered using a side channel, making Intel attacker accesses the shared data while timing how long this processors vulnerable to this attack. takes. A cache hit—meaning the data is back in the cache—indi- The Foreshadow attack is a variation of the Meltdown vul- cates that the victim accessed the data. A cache miss indicates nerability. This attack affects Intel microprocessors because that the data has not been accessed. So, simply by measuring of a weakness that Intel refers to as L1 Terminal Fault (L1TF). how long it took to access data, the attacker can determine While the original Meltdown attack relied on a delay in check- which cache sets were accessed by the victim. It takes a bit of ing permissions, Foreshadow relies on speculation that occurs algorithmic magic, but this knowledge of which cache sets during a stage of the pipeline called address translation. were accessed and which were not can lead to the discovery of Software views the computer’s memory and storage assets encryption keys and other secrets. as a single contiguous stretch of virtual memory all its own. But physically, these assets are divided up and shared among MELTDOWN, SPECTRE, AND THEIR VARIANTS all fol- different programs and processes. Address translation turns low the same pattern. First, they trigger speculation to exe- a virtual memory address into a physical memory address. cute code desired by the attacker. This code reads secret data Specialized circuits on the microprocessor help with the without permission. Then, the attacks communicate the secret virtual-to-physical memory-address translation, but it can be using Flush and Reload or a similar side channel. That last part slow, requiring multiple memory lookups. To speed things is well understood and similar in all of the attack variations. up, Intel microprocessors allow speculation during the trans- Thus, the attacks differ only in the first component, which is lation process, allowing a program to speculatively read the how to trigger and exploit speculation. contents of a part of the cache called L1 regardless of who owns that data. The attacker can do this, and then disclose the Meltdown Attacks data using the side-channel approach we already described. Meltdown attacks exploit speculation within a single instruc- In some ways Foreshadow is more dangerous than Meltdown, tion. Although assembly-language instructions are typically in other ways it is less. Unlike Meltdown, Foreshadow can

MELTDOWN Logically, reading from memory should work like this: But to save time, Intel processors Read from memory Meltdown Does the application have permission to execute these read from this address? steps out of order.

Yes No Does the application have permission to read from this address?

Read from No Raise an Yes Raise an memory “exception” “exception”

Raise an “exception” By masking exceptions and using a side channel that Report results Report results times how long it takes to read from memory, Meltdown can steal secret data.

46 | MAR 2019 | SPECTRUM.IEEE.ORG read the contents only of the L1 cache, because of the spe- SPECTRE VARIANT 1 cifics of Intel’s implementation of its processor architecture. Spectre v1 can be summed up in the following piece of code: However, Foreshadow can read any contents in L1—not just If (x<256){ data addressable by the program. secret=array1[x] y=array2[secret] } Spectre Attacks Spectre attacks manipulate the branch-prediction system. This system has three parts: the branch-direction predic- 1. The code is run several times with x less Take the branch? tor, the branch-target predictor, and the return stack buffer. than 256. This primes the branch predictor to Yes No expect x to be less than 256 the next time. The branch-direction predictor predicts whether a conditional branch, such as one used to implement an “if” statement in a programming language, will be taken or not taken. Cache memory 0 To do this, it tracks the previous behavior of similar branches. 0 For example, it may mean that if a branch is taken twice in a 0 0 2. The attacker empties the row, future predictions will say it should be taken. 0 processor’s cache using flush 0 The branch-target predictor predicts the target memory 0 instructions, so that any data read address of what are called indirect branches. In a conditional 0 by the program must be brought in 0 from main memory. branch, the address of the next instruction is spelled out, but 0 0 for an indirect branch that address has to be computed first. 0 The system that predicts these results is a cache structure called the branch-target buffer. Essentially, it keeps track of Memory the last computed target of the indirect branches and uses address 3. The attacker runs the code with Secret = 3 x these to predict where the next indirect branch should lead to. x set to a value greater than 256. The The return stack buffer is used to predict the target of a processor begins to “speculatively” “return” instruction. When a subroutine is “called” during a execute the rest of the code as if x were less than 256. The memory 0 256 program, the return instruction makes the program resume address at x contains secret data the work at the point from which the subroutine was called. Try- attacker wants. The software assigns ing to predict the right point to return to based only on prior this data to the variable secret. 100 3 0 2 return addresses won’t work, because the same function may 0 1 be called from many different locations in the code. Instead, the system uses the return stack buffer, a piece of memory Memory address on the processor, that keeps the return addresses of func- Brought in 4. Next, the code 3 x uses the value of tions as they are called. It then uses these addresses when from main memory secret as a memory a return is encountered in the subroutine’s code. because the address. It reads the Each of these three structures can be exploited in two differ- cache is empty 0 256 data at that address, bringing that data ent ways. First, the predictor can be deliberately mistrained. into the cache from In this case, the attacker executes seemingly innocent code main memory y = 100 Secret = 3 designed to befuddle the system. Later, the attacker deliber- 0 2 ately executes a branch that will misspeculate, causing the 0 1 program to jump to a piece of code chosen by the attacker, called a gadget. The gadget then sets about stealing data. 5. Eventually, the processor y Secret Take the branch? A second manner of Spectre attack is termed direct injec- realizes that it should not have taken the branch. It does not Yes No tion. It turns out that under some conditions the three pre- make the values secret and y dictors are shared among different programs. What this visible to the program. means is that the attacking program can fill the predictor structures with carefully chosen bad data as it executes. Cache memory When an unwitting victim executes their program either 0 at the same time as the attacker or afterward, the victim 0 6. The attacker accesses each 0 address in the cache systematically. will wind up using the predictor state that was filled in by 0 Because the memory address equal the attacker and unwittingly set off a gadget. This second 0 0 to the secret is the only one whose attack is particularly worrisome because it allows a victim 0 data has already been brought in 0 program to be attacked from a different program. Such a from main memory, it takes less 0 time to access than all the others, threat is especially damaging to cloud-service providers 100 0 thereby revealing the secret. because they cannot then guarantee that their client data 0 is protected. Time SOME SPECULATIVE EXECUTIONS VULNERABILITIES have started thinking about how to design future CPUs that Vulnerability What’s exploited? Disclosed keep speculation without compromising security. Spectre Variant 1 Branch-direction predictor 1/2018 One defense, called kernel page-table isolation (KPTI), is now built into Linux and other operating systems. Recall that each Spectre Variant 2 Branch-target buffer 1/2018 application views the computer’s memory and storage assets Branchscope Branch-direction predictor 3/2018 as a single contiguous stretch of virtual memory all its own. But Spectre Variant 4 Speculative store bypass 5/2018 physically, these assets are divided up and shared among differ- Spectre RSB/ Return stack buffer 6/2018 ent programs and processes. The page table is essentially the Ret2Spec operating system’s map, telling it which parts of a virtual mem- Lazy FPU Lazy restore of floating-point 6/2018 unit registers ory address correspond to which physical memory addresses. TLBleed Translation-lookaside buffer 6/2018 The kernel page table is responsible for doing this for the core of the operating system. KPTI and similar systems defend NetSpectre Branch-direction predictor 7/2018 against Meltdown by making secret data in memory, such as Meltdown Page-permission check 01/2018 the OS, inaccessible when a user’s program (and potentially an Foreshadow/L1TF Address translation 08/2018 attacker’s program) is running. It does this by removing the for- bidden parts from the page table. That way, even speculatively executed code cannot access the data. However, this solution THE SPECTRE AND MELTDOWN vulnerabilities pre- means extra work for the operating system to map these pages sented a conundrum to the computing industry because when it executes and unmap them afterward. the vulnerability originates in hardware. In some cases the Another class of defenses gives programmers a set of tools to best we can do for existing systems—which make up the bulk limit dangerous speculation. For example, Google’s Retpoline of installed servers and PCs—is to try to rewrite software to patch rewrites the kind of branches that are vulnerable to attempt to limit the damage. But these solutions are ad hoc, Spectre Variant 2, so that it forces speculation to target a benign, incomplete, and often result in a big hit to computer perfor- empty gadget. Programmers can also add an assembly-language mance. At the same time, researchers and CPU designers instruction that limits Spectre v1, by restricting speculative

ROBOTIC END-EFFECTORS

Measure all six components of force and torque in a compact, rugged sensor.

Interface Structure high-strength alloy provides IP60, High-Speed Electronics IP65, and IP68 environmental interfaces for Ethernet, PROFINET, protection as needed EtherNet/IP, Analog, USB, CAN EtherCAT, Wireless, and more Sensing Beams and Flexures designed for high stiffness and over- Silicon Strain Gages load protection without compromising provide high noise immunity, resolution accuracy, and high factor-of-safety, standard on all F/T models

Engineered for high-performance and maximum stiffness, with the highest resolution and accuracy available, it’s the ultimate force/torque sensor. Only from ATI.

www.ati-ia.com 919.772.0115

48 | MAR 2019 | SPECTRUM.IEEE.ORG memory reads that follow conditional branches. Conveniently, puter performance, making it difficult to eliminate such vul- this instruction is already present in the processor architecture nerabilities in current system designs. It is likely that new CPU and is used to enforce the correct ordering between memory designs will evolve to retain speculation, while preventing operations originating on different processor cores. the type of side-channel leakage that enables these attacks. As the processor designers, Intel and AMD had to go deeper Nevertheless, future computer-system designers, including than a regular software patch. Their fixes update the processor’s those designing processor chips, must be aware of the secu- microcode. Microcode is a layer of instructions that fits between rity implications of their decisions, and no longer optimize the assembly language of regular software and the processor’s only for performance, size, and power. n actual circuitry. Microcode adds flexibility to the set of instructions a processor can execute. It also makes it simpler to design ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/spectremeltdown0319 a CPU because when using microcode, complex instructions are translated to multiple simpler instructions that are easier to execute in a pipeline. New Version! Basically, Intel and AMD adjusted their microcode to change the behavior of some assembly-language instructions in ways that limit speculation. For example, Intel engineers added options that interfere with some of the attacks by allowing the operating system to empty the branch-predictor structures in certain circumstances. A different class of solutions attempts to interfere with the attacker’s ability to transmit the data out using side channels. For example, MIT’s DAWG technology securely divides up the processor cache so that different programs don’t share any of its resources. Most ambi- tiously, there are proposals for new processor architectures that would introduce structures on the CPU that are dedicated to speculation and sep- arate from the processor’s cache and other hardware. This way, any operations that are executed speculatively but are not eventually committed are never visible. If the speculation result is confirmed, the speculative data is sent to the processor’s main structures. Speculation vulnerabilities have lain dormant in processors for over 20 years, and they remained, so far as anyone knows, unexploited. Their discovery Over 75 New Features & Apps in Origin 2019! For a FREE 60-day has substantially shaken industry and Over 500,000 registered users worldwide in: evaluation, go to highlighted how cybersecurity is not ◾ 6,000+ Companies including 20+ Fortune Global 500 OriginLab.Com/demo only a problem for software systems ◾ 6,500+ Colleges & Universities and enter code: 8547 but for hardware as well. Since the ini- ◾ 3,000+ Government Agencies & Research Labs tial discovery, around a dozen variants of Spectre and Meltdown have been revealed, and it is likely that there are 25+ years serving the scientific & engineering community more. Spectre and Meltdown are, after all, side effects of core design principles that we have relied on to improve com-

SPECTRUM.IEEE.ORG | MAR 2019 | 49 WALK THIS WAY CONTINUED FROM PAGE 35

Cassie’s new legs required a more sophisticated low-level controller than ATRIAS did. With ATRIAS, extending a leg was achieved simply by applying equal and opposite torques with the two motors. With Cassie’s leg, moving the foot in a specific direction requires calculating different torques for each motor. To do that, the controller needs to take into account the inertia of the legs as well as the dynamics of the motors and gearboxes. ENHANCE YOUR IEEE True, the control problem became more complex, but this MEMBERSHIP BY JOINING THE method allows for better performance and a much broader Power & Energy Society range of behaviors. Cassie can achieve a walking speed of 5 km/h using one of our initial controllers. The power draw Help shape the future ranges from 100 watts (standing) to about 300 W (walking), of the industry and give and the lithium-ion battery allows for about 5 hours of con- your career a boost tinuous operation. The new legs also allow Cassie to steer in Join the engaged members of PES a way that ATRIAS couldn’t. And thanks to a powered foot who are advancing innovation and joint, it can stand in place without having to constantly move deepening their expertise in important its feet the way ATRIAS does. areas such as: Cassie also weighs just 31 kg (68 pounds), half as much as • Renewable energy ATRIAS. You can load two Cassies into the trunk of a car in • Bulk energy storage less than a minute. And it’s far more robust: Its body parts are • Distributed energy resources made of aluminum and carbon fiber, and a protective shell • Smart grids made of thermoformed acrylic-polyvinyl chloride, a strong More Power to the Future™ Learn more: ieee-pes.org/members plastic, protects it from collisions and falls. Cassie isn’t running through forests just yet. But we’ve taken it outdoors, without safety tethers, and it has walked on dirt, grass, and leaf-strewn paths. We’re now learning how to integrate the robot’s dynamic behaviors with motion planning, which allows it to climb stairs, for example. We’re also working on another feature that could make a robot like Cassie much more useful: arms.

Digit is a direct descendant of Cassie. It has similar legs, but we added a torso and a pair of arms. The arms are designed to help with the robot’s mobility and balance, as they swing in coordination with the gait. They’ll also allow Digit to catch itself when it falls, and reorient its body to get back up. There’s something else that Digit has that Cassie didn’t: integrated perception. We added numerous sensors to the robot, including a lidar atop the torso. These sensors will help gather data to allow the robot to navigate a world full of obstacles such as cluttered rooms and stairs, and to rely on the underlying stable dynamics only when handling unexpected situations and sensing errors. Digit and the legged robots that will follow still have a long way to go. But we’re convinced they will change the world. Their impact could be as big as the automobile in terms of changes to lifestyles, and even traffic patterns and aspects of city layouts, where these robots promise to transform logistics and package delivery. In the not-too-distant future, as vehicles gain autonomy, carmakers and ride-sharing companies like Lyft and Uber

50 | MAR 2019 | SPECTRUM.IEEE.ORG will own large fleets of vehicles that transport people, with peak traffic during rush hour, just like today. ENABLING YOUR INNOVATION But late at night and in the FROM CONCEPT TO PERFORMANCE middle of the day, what will these automated vehicles do? Instead of simply being idle, they could trans- SOFTWARE THAT LIVES UP TO port your packages from automated warehouses THE POWER OF YOUR IDEAS to your home. However, these delivery vehicles are stuck at the curb: With- out a human, getting the package to your doorstep is a major challenge. That’s where legged robots come in. Riding along in these FIRST STEPS: ATRIAS’s legs vehicles, they’ll cover those don’t look like a person’s, but the robot was the first machine to last few meters. Although demonstrate humanlike wheels and wings may walking-gait dynamics. serve some of these roles, in a world designed for bipeds, no mobility platform could be as versatile as a legged robot. Delivery robots will be part of an increasingly more automated logistics system that goes from manufacturers and resellers directly to your door. This system will reduce shipping costs until it’s significantly less expensive to have items delivered to your home than to buy them from a large, well- lit, heated, human-accessible warehouse. Big-box stores that today sell what are essentially commodities will become COMPETITIVE ADVANTAGES superfluous. People will still enjoy in-store shopping for cer- • New! “Smart Workspace” tain goods, of course. But for the week’s groceries, delivery • Search-based design simulation robots will save you time and money. • Program your own applications with API Legged locomotion will also help bring robots into our homes • Precise field calculations using our proprietary and businesses. Robots that can navigate stairs and cluttered BEM and FEM solvers environments, while interacting safely with humans at eye • Intuitive and easy-to-use interface level and at human scale, will allow us to age gracefully in our own homes. They’ll help carry things around and serve as tele- “For our purposes, ‘ELECTRO’ is actually the presence devices, allowing family members and friends to use preferred software tool for every-day engineering. the robot to talk to people remotely and keep them company. It is important to have software that is very quick Legged robots will also go where it’s dangerous for humans and easy to learn and this product is more than to go. They will parachute into forest fires to gather data in capable of handling the large models we require.” real time, rush into burning buildings to check for occupants, or enter disaster zones, like the Fukushima Daiichi nuclear Dr. Beriz Bakija, Siemens AG’s Energy Sector, Germany power plant, to explore hard-to-access areas. They will conduct regular inspections of the internal spaces of hydroelectric ASK FOR AN ONLINE DEMONSTRATION dams and explore abandoned mines, so that we won’t have to. A lot of challenges need to be solved to get us to that future. But I’m convinced the robotics community can make this technology practical. It’ll be one small step for a robot, one giant leap for mankind. n [email protected] / (204) 632-5636

OREGON STATE UNIVERSITY STATE OREGON ↗ POST YOUR COMMENTS at https://spectrum.ieee.org/robotlegs0319 integratedsoft.com Technische Universität Wien, generally referred to as TU Wien, is located in the heart of Europe, a place where one can experience cultural diversity and international life. Research, teaching and learning have been conducted here in the service of progress for 200 years. TU Wien is amongst Europe’s most successful universities of technology and, with over 30,000 students and a staff of about 4,600, is Austria’s largest scientific and engineering research and education institution. At the Faculty of Electrical Engineering and Information Technology the position for a full-time indefinite-term University Professor for the specialist field of “Autonomous Systems” with contractual employment is expected to be filled as of 01.01.2020. This is a professorship regulated by § 98 of the 2002 Universities Act (UG) and a Foundation Professorship supported in the first three years by the B&C Privatstiftung as well as the company TTTech. The Faculty of Electrical Engineering and Information Technology, one of the eight faculties at TU Wien, has an excellent international reputation and plays an active role in national and international research. The current research areas of the faculty are: Photonics, Micro- and Nanoelec- tronics, Telecommunications, Systems and Control Engineering as well One of the as Energy Technology and Energy Systems. The position as professor for Autonomous Systems has been allocated in TU Wien’s development plan to the focal areas of research on Information and Communication most influential Technology. The future incumbent’s field of work should be Autonomous Systems, with a focus on one or more of the following: reference resources • Design, realization and scientific evaluation of safety-critical autonomous and embodied semi-autonomous mobile systems, e.g. for engineers autonomous robots, airborne, waterborne or ground vehicles, min- iature robots and their self-organisation. It is expected that candi- around the world. dates are willing to focus their research on ground vehicles and/or autonomous robots. For over 100 years, Proceedings of the IEEE • Cognition-inspired and machine learning based methods for autono- has been the leading journal for engineers mous systems. looking for in-depth tutorial, survey, and • Novel methods for the system integration of safety-critical autonomous systems as well as their interaction with their environment review coverage of the technical and infrastructure. developments that shape our world. • Distributed, collaborating self-learning autonomous systems. Offering practical, fully referenced articles, This chair is expected to build up an interdisciplinary research group at Proceedings of the IEEE serves as a bridge the Institute of Computer Technologies that will study and develop solu- to help readers understand important tions for challenges of real autonomous systems both theoretically and technologies in the areas of electrical through experimental research. engineering and computer science. The duties of a university professor at TU Wien include, in addition to research, teaching activities in the engineering fields of the Bachelor, Master and PhD programmes of the Faculty of Electrical Engineering and Information Technology as well as collaboration in management of the institute and the faculty. To learn more and start More information concerning this position is available at your subscription today, visit https://etit.tuwien.ac.at/fileadmin/t/etit/Dokumente/Sonstiges/Aus- ieee.org/proceedings-subscribe schreibung/Par_98_Professur_Autonome_Systeme_EN_13_12_2018_ Ute_Koch.pdf The application package should be sent before April 30, 2019 to the Dean by email [email protected]

52 | MAR 2019 | SPECTRUM.IEEE.ORG in 4to6weeks. published author Become a • • • • • • • journal thatallowsyouto: Access isamultidisciplinaryIEEE products toaglobalaudience. recent developments,methods,ornew for authorswhowanttoquicklyannounce Published onlineonly, Access IEEE the IEEE. of worthy journal that’s open access a multidisciplinary Access... IEEE ieeeaccess.ieee.org Learn moreat: only $1,750perarticle Publishwithoutapagelimitfor Connectwithreadersthroughcommenting citation dataforeachpublishedarticle multimediaandtrackusage Integrate in oneoftheSpecialSections contributing totrending,interdisciplinarytopics Establishyourselfasanindustrypioneerby Web ofScience(andhasanImpactFactor) Accessisincludedin a keyfactorwhyIEEE Expectarapidyetrigorouspeerreview— neatly intraditionaljournals thatdonotfit Submitmultidisciplinaryarticles IEEE IEEE Reachmillionsofglobalusersthroughthe Xplore ® digital librarywithfreeaccesstoall digital

is ideal

Best New Journal in STM 2015

17-PUB-013 3/17 and inleadingresearch effortsinArtificial Intelligence,including educationprograms, pursuing innovativeandinterdisciplinary (GUET). SAIisseekingthecandidateswhohaveaninterestin Intelligence (SAI) at Guilin University of Electronic Technology Applicants areinvitedforthefacultyinSchoolofArtificial Opportunities is supported by the Ministry of Industry and Information of Industry is supported by the Ministry technology inChina,GUETisakeypublicuniversitywhich As oneofthefourfamousuniversitiesfocusingonelectronic About GUET package willbeoffered. all countriesandofnationalities.Agloballycompetitive This isaworldwidesearch, opentoqualifiedcandidatesfrom records andpotentialsforresearch, teachingandleadership. and ComputerEngineering,orrelatedfieldswithdemonstrated degrees inComputerScience,ElectricalEngineering, and AutonomousRobotics.ApplicantsshouldpossessPh.D. Computer Vision,NaturalLanguageProcessing,BigData, but notlimitedtoMachineLearning,ImageProcessing, We toyourjoiningus! lookforward Ms. Tung; Cellphone: +86-13217735600. phone: +86-13737731244.Department ofHumanResources: we couldprovide:SchoolofArtificial Intelligence:Mr. Mo;Cell Please donothesitatetocontact us ifthereisanyassistance Contacts Send [email protected]’s subjectof"ApplyforAI". Way toApply sceneries. states, celebrities at homeand abroad, have attracted by Guilin’s travelers byCNNin2013.Numeroustourists,includingheadsof Lijiang Rivermakesthelistinworld’s 15bestriversfor best”, ranking2ndamongstChina’s top10scenicspots.Guilin enjoys ahighprestigeas“Eastorwest,Guilin’s landscape is to theuniquenessofwondersmountainsandrivers,Guilin ofmorethan2,000years,due Boasting along-standinghistory About Guilin and 40,600studentsatalllevels. bachelor programs.GUETownsover2,900facultyandstaff, 1 postdoctoralresearch center, 17masterprogramsand70 Founded in1960,theUniversitynowhas4doctoralprograms, Technology andGuangxiZhuangAutonomousRegion. Full-Time FacultyPositionsinArtificialIntelligence Guilin UniversityofElectronicTechnology Location: Guilin,Guangxi,China School ofArtificialIntelligence Closing Date:unlimited SPECTRUM.IEEE.ORG

MAR 2019

53 Tenure Track Position in Mechatronics Engineering Micron Technology, Inc., is seeking the below positions for The Department of Mechanical its semiconductor R&D facility in Boise, ID; its manufacturing and Industrial Engineering facility in Manassas, VA; and its sales and design facilities in (https://www.ryerson.ca/mie/) Folsom and Milpitas, CA; and design facilities in Austin, TX in the Faculty of Engineering & and Longmont, CO. Architectural Science at Ryerson University invites applications for a The following Micron subsidiaries are also seeking positions: tenure-track position at the rank of Assistant Professor in Mechatronics Micron Semiconductor Products, Inc., at its headquarters Engineering, effective July 1, 2019, subject to final budgetary approval. in Boise, ID; and sales facilities in Meridian, ID and Folsom, Candidates must hold a Ph.D. degree (or be near completion) in Milpitas, and Irvine, CA. Mechatronics Engineering, Mechanical Engineering, Electrical Engineering or a related field. Postdoctoral experience is an asset. Electrical, Electronics, Communications, Chemical, The selected candidate must be eligible to register as a professional Industrial, Mechanical, Materials, Computer System engineer (P.Eng.) in the province of Ontario by the date of appointment. Analysts, and Software Engineering; Physics, Materials Candidates must have a demonstrated commitment to uphold the values Science, Engineering Manager and other related of Equity, Diversity, and Inclusion in teaching, research and service. Engineering occupations. Marketing, Sales, Logisticians, This includes the ability to foster creative and collaborative intellectual Finance, Accounting, and other related business positions. inquiry by bringing diverse knowledge, experiences and perspectives to learning activities and research projects. Candidates must demonstrate Please submit your resume online: well-developed research and teaching abilities in Mechatronics http://www.micron.com/jobs Engineering, with expertise in robotics, controls, autonomous systems, Resume and/or cover letter must reflect each requirement or or other relevant areas. Candidates must hold a strong research profile it will be rejected. Upon hire, all applicants will be subject to (e.g., evidence of an emerging scholarly record, ability to establish and drug testing/screening and background checks. maintain an independent, externally funded research program), evidence of high-quality inclusive teaching and student training, and a capacity Note: Some of these positions may require domestic and for collegial service. international travel for brief business purposes. Please For the full job posting, including qualifications and how to apply, please read the full job description when applying online for such refer to the Ryerson Career Opportunities website: https://www.ryerson. requirements. ca/jobs/ EOE

Institute Director Position The University of Maine is seeking applications for the position of Director of the Frontier Institute for Research in Sensor Technologies (FIRST). This is a tenured appointment reporting to the Vice President for Research and Dean of the Graduate School. FIRST, consisting of an interdisciplinary team of scientists and a world- class set of resources, is dedicated to research and development of cutting-edge, commercially viable sensor systems for biomedical, environmental, energy, defense, information technology, and other applications. Information about the position and application process is available at https:// umaine.hiretouch.com/

ILLUMINATE EDUCATE ENGAGE ENERGIZE

Washington State University - The Department of Biological Systems Engineering seeks a 12-month, The IEEE Foundation is leading a special campaign 75% tenure-track position in the rank of Assistant to raise awareness, create partnerships, and generate financial Professor in Food Engineering. The successful resources needed to combat global challenges. candidate should have Ph.D. in Engineering and demonstrated expertise in sensing technology Our goal is to raise $30 million by 2020. that can be applied for food safety and quality assessment. To apply: visit http://www.wsujobs. com (Position #128473). Screening of application materials will begin April 01, 2019 and will continue DONATE NOW until the position is filled. Contact Dr. Shyam ieeefoundation.org Sablani, Search Committee Chair, at ssablani@wsu. edu, with questions about this position.

54 | MAR 2019 | SPECTRUM.IEEE.ORG Technology insight on demand on IEEE.tv Internet television gets a mobile makeover

A mobile version of IEEE.tv is now available for convenient viewing. Plus a new app for IEEE.tv can also be found in your app store. Bring an entire network of technology insight with you:

• Convenient access to generations of industry leaders. • See the inner-workings of the newest innovations. • Find the trends that are shaping the future.

IEEE Members receive exclusive access to award-winning programs that bring them face-to-face with the what, who, and how of technology today.

Tune in to where technology lives www.ieee.tv

11-MEMB-0574b_IEEEtv_Spectrum_7x10_float_Final.indd 1 1/24/12 1:53 PM PAST FORWARD_BY ALLISON MARSH

In 1856, W.H. Burnap purchased the patent rights to a “magnetic-electric machine” from Walter Kidder, SHOCK YOUR who had purchased them the prior year from the actual inventor, Ari Davis. The machine was designed to give shocks and thus treat the ailing. Burnap went on to manufacture and market, with great success, “Davis & WAY TO GOOD Kidder’s Patent Magneto-Electric Machine for Nervous Diseases.” Included in the long list of maladies and HEALTH afflictions it treated were ennui and female hysteria, as well as toothache and just about everything else. Legitimate doctors and dubious quacks could carry this compact, convenient, and durable device to a patient’s home and apply the panacea of electricity. ■

↗ For more on this magnetoelectric machine, see https://spectrum.ieee.org/pastforward0319 IMAGES LIBRARY/GETTY PICTURE SOCIETY & SCIENCE

56 | MAR 2019 | SPECTRUM.IEEE.ORG Some things don’t make sense.

Some things do. IEEE Member Group 10-Year Level Term Life Insurance Plan — It just makes sense.

10-Year Level Term Life Insurance Plan has affordable group rates, portable coverage, premium and volume IEEE acts as an advocate to ensure member benefits remain valuable. As a result the IEEE Member Group

Wantdiscounts, to join an accelerated thousands death of your benefit fellow and IEEEbenefit members? amounts up to $2 million.* Apply online today!

Learn more about the IEEE Member Group 10-Year Level Term Life Insurance Plan. Visit IEEEinsurance.com/10YTL.**

*The Group 10-Year Level Term Life Insurance Plan is available only for residents of the U.S. Program Administered by Mercer Health & Benefits Administration LLC (except territories), Puerto Rico and Canada (except Quebec). This plan is underwritten by New York Life Insurance Company, 51 Madison Ave., New York, NY 10010 on Policy Form GMR. In CA d/b/a Mercer Health & Benefits This plan is administered by Mercer Consumer, a service of Mercer Health & Benefits Administration Insurance Services LLC LLC. This coverage is available to residents of Canada (except Quebec). Mercer (Canada) Limited, represented by its employees Nicole Swift and Suzanne Dominico, acts as a broker with respect AR Insurance License #100102691 to residents of Canada. CA Insurance License #0G39709 **For information on features, costs, eligibility, renewability, limitations and exclusions. 85475 (3/19) Copyright 2019 Mercer LLC. All rights reserved.

85475 IEEE 10-Y LTL Spectrum Funny Ad_2019.indd 1 1/30/19 10:46 AM 85475 (3/19), Spectrum IEEE 10Y LTL Trim Size: 7.875" x 10.5" Live Area: 7" x 10" Bleed: 8.125" x 10.75" Colors 4C - CMYK Stock: N/A MERCER Title: na-mls-machine-learn-blue-7.87X10.75 Creative ServicesProject Manager This advertisementprepared by: Client Name: The Mathworks [email protected] Size: 7.875”x10.75” Natick, MA01760 Elizabeth Putnam 1 AppleHillDrive REQ #:110718D 508-647-0589 MathWorks Page 4/c LEARNING MACHINE MATLAB mathworks.com/machinelearning With MATLAB and puttheminto production. learning to buildpredictive models regression,deep and classification, SPEAKS Cosmos Communications Cosmos Communications ® you can use clustering, 38686a2 K Y M C 11.9.18 133 ej 1 4

1Q2 Q1

@2019 The MathWorks, Inc.