MCK - Semiconductors - Oct2019-Full Book-V12-RGB

McKinsey on
Semiconductors
Creating value, pursuing innovation, and
optimizing operations
Number 7, October 2019

McKinsey on Semiconductors is Editorial Board: McKinsey Practice Publications
written by experts and practitioners Ondrej Burkacky, Peter Kenevan,
in McKinsey & Company’s Abhijit Mahindroo Editor in Chief:
Semiconductors Practice along with Lucia Rahilly
other McKinsey colleagues. Editor: Eileen Hannigan
Executive Editors:
To send comments or request Art Direction and Design: Michael T. Borruso,
copies, email us: Leff Communications Bill Javetski,
McKinsey_on_ Semiconductors@ Mark Staples
McKinsey.com. Data Visualization:
Richard Johnson, Copyright © 2019 McKinsey &
Cover image: Jonathon Rivait Company. All rights reserved.
© scanrail/Getty Images
Managing Editors: This publication is not intended to
Heather Byer, Venetia Simcock be used as the basis for trading in
the shares of any company or for
Editorial Production: undertaking any other complex or
Elizabeth Brown, Roger Draper, significant financial transaction
Gwyn Herbein, Pamela Norton, without consulting appropriate
Katya Petriwsky, Charmaine Rice, professional advisers.
John C. Sanchez, Dana Sand,
Sneha Vats, Pooja Yadav, Belinda Yu No part of this publication may be
copied or redistributed in any form
without the prior written consent of
McKinsey & Company.
Table of contents
3 What’s next for semiconductor

profits and value creation? 47 How will changes in the
automotive-component
Semiconductor profits have been market affect semiconductor
strong over the past few years. companies?
Could recent changes within the The rise of domain control units
industry stall their progress? (DCUs) will open new opportunities
for semiconductor companies.
16 Artificial-intelligence hardware:
New opportunities for 50 Right product, right time,
right location: Quantifying the
semiconductor companies semiconductor supply chain
Artificial intelligence is opening Problems along the
the best opportunities for semiconductor supply chain
semiconductor companies in are difficult to diagnose. A new
decades. How can they capture metric can help companies
this value? pinpoint performance issues.
27 Blockchain 2.0: What’s in

store for the two ends— 57 Reducing indirect labor costs at
semiconductor companies
semiconductors (suppliers) and Digital tools could bring new
industrials (consumers)? productivity and efficiency
Ten years after blockchain’s gains to indirect functions. Why
inception, it is presenting new do semiconductor companies
opportunities for both suppliers, hesitate to use them?
such as semiconductor companies,
and consumers, such as industrials.
38 Rethinking car software and

electronics architecture 63 Taking the next leap forward in
semiconductor yield improvement
As the car continues its transition By prioritizing improvements in
from a hardware-driven end-to-end yield, semiconductor
machine to a software-driven companies can better manage
electronics device, the auto cost pressures and sustain higher
industry’s competitive rules are profitability. The path forward
being rewritten. involves advanced analytics.
Introduction
Welcome to the seventh edition of McKinsey on Semiconductors. This publication appears at a time when
our world is being transformed by the growth of artificial intelligence (AI), machine learning, and other
innovative technologies. The pace of change is so fast that leading-edge products today may seem dated
within a year. In this constantly evolving landscape, only one thing is certain: semiconductor companies
will enable some of the most important technological leaps.
Many semiconductor companies are already benefiting from the innovative offerings, with the sector
showing strong and rising profits over the past few years. But there may be challenges ahead, since
companies that want to remain industry leaders must continue to increase their R&D investments. With
costs in labor and other areas rising, some semiconductor companies may have difficulty finding additional
funds for innovation. Moreover, some customers are already designing chips internally, and others may
follow—a trend that could decrease sales. These concerns, and possible solutions, are the focus of the first
article in this issue: “What’s next for semiconductor profits and value creation?”
Other articles discuss recent technological trends that are increasing demand for chips. In “Artificial-
intelligence hardware: New opportunities for semiconductor companies,” the authors explore how the rise
of AI could help players capture more value. Overall, semiconductor growth from AI may be five times greater
than growth from other sources. Opportunity is also the central theme of “Blockchain 2.0: What’s in store for
the two ends—semiconductors (suppliers) and industrials (consumers)?” As this article describes, blockchain’s
role in cryptocurrency and its potential growth as a business application may accelerate demand for chips.
The extent of the change, as well as the timing, is still uncertain, but semiconductor leaders that monitor
developments could have an advantage if blockchain takes off.
This issue also contains two articles that discuss a technology leap that has intrigued consumers: the rise
of autonomous vehicles. “Rethinking car software and electronics architecture” explores how sensors
and other automotive components may evolve, since these changes could influence chip demand. One
specific shift—the rise of domain control units—is described in another article: “How will changes in the
automotive-component market affect semiconductor companies?”
While all of these developments are exciting, semiconductor companies also have to deal with some
stubborn problems that have plagued their businesses for years. Companies will discover a new approach
to eliminating late shipments in “Right product, right time, right location: Quantifying the semiconductor
supply chain.” Similarly, they will learn about strategies for decreasing one of their largest costs in
“Reducing indirect labor costs at semiconductor companies.” The final article, “Taking the next leap forward
in semiconductor yield improvement,” will help companies use analytics and other strategies to optimize
production in both line and die processes.
McKinsey on Semiconductors is designed to help industry executives navigate the road ahead and achieve
continued growth. We hope that you find these articles helpful.
Ondrej Burkacky Peter Kenevan Abhijit Mahindroo

Partner Senior partner Partner
2 McKinsey on Semiconductors Number 7, October 2019

What’s next for
semiconductor profits
and value creation?
Semiconductor profits have been strong over the past few years.
Could recent changes within the industry stall their progress?
by Marc de Jong and Anurag Srivastava
© DuKai photographer/Getty Images
What’s next for semiconductor profits and value creation? 3

Outside Silicon Valley, the semiconductor profit”).¹ With this information, we wanted to
industry’s recent profitability seems unsurprising. answer an even more important question: What can
The general assumption is that these players, like semiconductor players do to ensure that the recent
other tech companies, have long benefited from gains are not a blip but the emergence of a new
the rise of PCs, smartphones, and other devices. industry norm?
But insiders know that the industry’s good fortune
is a relatively recent phenomenon. While software
players were achieving record gains for most of the A decade of change: How value
past two decades, most semiconductor companies creation has evolved within the
achieved limited economic profitability. Overall, semiconductor industry
only microprocessor companies and some fabless Only ten years ago, the semiconductor industry
players could count on consistently strong returns, had mediocre returns. Although many companies
above the cost of capital. were generating value, they lagged far behind
their counterparts in other industries. But a much
Now the semiconductor sector is showing strong different story has unfolded over the past five
and rising profits. What’s more, companies years, with the semiconductor industry reporting
in virtually all subsegments are winning big. record gains. In 2017 alone, it generated $97 billion
McKinsey on Semiconductors 2018
To discover how semiconductor companies in economic profit—more than a threefold increase
What’s next semiconductors
engineered this turnaround, we analyzed trends from the $28 billion captured in 2013 (Exhibit 1).
Exhibit 1 of 10
related to economic profit (see sidebar, “Economic
1
Economic profit equals the net operating profit after tax minus the capital charge (the invested capital, excluding goodwill—the amount of a
purchase that exceeds the value of the assets involved) at previous year-end multiplied by the weighted average cost of capital.
Exhibit 1
From a value-creation perspective, 2017 was a record year for the semiconductor industry.
Economic-profit1 value creation for semiconductor industry² (excluding goodwill), $ billion
Cumulative positive economic profit for companies Cumulative negative economic profit for companies
2010 2011 2012 2013 2014 2015 2016 2017
99
59
52
47
42 40
34
30
–4 –7 –5 –5 –5 –3
–10 –12
Total economic profit, $ billion
38 30 18 28 42 47 54 97
¹ Figures may not sum, because of rounding.

² About 273 companies across all industry subsegments.
Source: Annual reports; Semiconductor CPC database; McKinsey Corporate Performance Center

Economic profit
When we looked at value creation within the semiconductor industry, we deliberately restricted our analysis to economic
profit, which is a periodic measure of value creation. In simplest terms, it is the amount left over after subtracting the cost of
capital from net operating profit. The formula for computing it is as follows:
return on weighted
Economic
profit = invested capital
(ROIC)
x invested
capital – average cost of
capital (WACC)
x invested
capital
We chose to focus on economic profit because this metric comprehensively captures both profitability and the opportunity cost
for the capital deployed. It also allowed us to perform reliable benchmark analyses for companies that followed many different
business models. We only considered operating assets and excluded goodwill and other M&A intangibles. This approach
allowed us to compare operating performance for different companies, regardless of whether their growth occurred organically
or arose from a merger. For the years 2013 through 2017, however, we conducted two analyses: one factored in goodwill, and
one did not (similar to the long-term analysis for 1997 through 2017). We conducted the two analyses to determine if the recent
surge in M&A activity had a significant effect on results.
From 2012 to 2016, the semiconductor

sector ranked tenth out of 59 major
industries for value creation, placing it
in the top 20 percent. That represents
a big jump from the period from 2002
through 2006, when it ranked 18th.

Exhibit 2 of 10
Exhibit 2 Facebook, Google, and Tencent become more

important to the global economy every year.
Value creation has migrated to almost all subsegments. These companies constantly introduce product
or technology upgrades to remain competitive,
Economic-profit value creation across subsegments (excluding
goodwill), cumulative, $ billion and they need chips to enable such advances.
Semiconductor companies have also benefited
Cumulative positive economic profit across subsegments from increased digitization and cloud use across
Cumulative negative economic profit across subsegments other industries, both of which accelerate
chip demand.
1997–2012 –38.7 161.5

In addition to these traditional revenue drivers,
some recent technology innovations, including
the Internet of Things, artificial-intelligence (AI)
2013–17 –0.6 267.5
applications, and blockchain technology, have
created new opportunities for semiconductor
companies to capture value. Advances in
Average annual economic profit (excluding goodwill), $ billion the automotive industry, including vehicle
electrification and the development of self-
driving cars, are also increasing chip demand.
1997–2012 7.9 +575% Such innovations are transforming how much
value semiconductor companies capture from
the technology stack. With AI applications, for
2013–17 53.4 instance, they could potentially achieve a larger
share of total value than they did with PCs and
mobile phones.
From 2012 to 2016, the semiconductor sector Value trends within the semiconductor
ranked tenth out of 59 major industries for value industry
creation, placing it in the top 20 percent. That The rise in economic profit is not the only big shift
represents a big jump from the period from within the semiconductor industry. As we reviewed
2002 through 2006, when it ranked 18th. While the trends, we also found that value distribution
the semiconductor sector still lags far behind has changed. From 1997 to 2012, the cumulative
software, which was second only to biotechnology, positive economic profit across segments was
it now outranks IT services, aerospace and $161.5 billion. But some segments also lost value
defense, chemical, and many other major sectors (Exhibit 2).
for value creation.
Overall, value was highly concentrated in a few
Strong global economic growth since the areas (Exhibit 3). The microprocessor subsegment
2008 recession has propelled the semiconductor generated the most value, followed by fabless.
industry’s revenues, but an even more important Together, they created almost all value in the
factor involves the continued rise of the technology industry, with all other subsegments roughly
sector. Companies such as Alibaba, Amazon, breaking even when their results were totaled.

Exhibit 3 of 10
Exhibit 3
From 1997 through 2012, the microprocessor and fabless subsegments created the most value.
Economic-profit1 value creation by subsegment (excluding goodwill), 1997–2012 cumulative, $ billion
A B C D E F G H I J K
Capital Electronic- IP2 Fabless Micro- Memory Analog Diversified Other Foundry Packaging
equipment design processor IDM3 IDM IDM and
automation assembly
82.4
62.5
38.1
22.4 4.1 42 20.9

1.7 15.2 15.6 4.7 1.1
–0.3 –0.2 –1.0 –2.2

–5.4 –7.6 –8.0
–8.6 –12.6
–47.6 –36.4
13.8 3.8 1.5 57.1 74.8 –9.5 14.2 –20.8 –3.3 8.3 –1.1
Top companies for economic profit, by subsegment
A) AMAT, ASML, KLA-Tencor; B) Synopsys, Cadence, Mentor graphics: C) ARM, Rambus, Spansion; D) Qualcomm, Mediatek, Xilinx; E)
Intel; F) Samsung, Sandisk; G) Linear, Analog, Maxxim; H) TI, ON, NXP; I) Microchip, Powertech, Faraday; J) TSMC; K) Silicon-ware,
Monolithic power.

² Intellectual property.
3
Integrated device manufacturer.
From 1997 to 2012, the cumulative

positive economic profit across segments
was $161.5 billion. But some segments
also lost value.

From 2013 to 2017, by contrast, nearly all Several factors may contribute to the greater
subsegments recorded positive economic distribution of value, including industry
profit (Exhibit 4). Fabless was the best performer consolidation. First, many large conglomerates
during this period, with memory in second have divested their semiconductor units over the
place. The microprocessor subsegment came past ten years to reduce R&D investment and
in third, down from the top ranking it held capital expenditures. Meanwhile, the industry has
from 1997 to 2012. This shift occurred partly also undergone a wave of M&A across segments.
because PCs—key drivers of microprocessor The number of semiconductor companies fell from
demand—have seen much lower growth than 208 in 2012 to 173 by 2017 (Exhibit 5). The fabless
McKinsey on Semiconductors
smartphones and 2018
tablets, which often rely subsegment saw the most consolidation, followed
on chips designed by fabless players. Memory by analog integrated device manufacturers and
Exhibit 4 of 10 players have benefited from less oversupply, diversified integrated device manufacturers, but
resulting in higher average sales prices and the drop in companies is notable in all sectors,
operating margins. including memory.
Exhibit 4
From 2013 through 2017, almost all subsegments demonstrated economic profit.
Economic-profit1 value creation by subsegment (excluding goodwill), 2013–17 cumulative, $ billion
A B C D E F G H I J K
Capital Electronic- IP2 Fabless Micro- Memory Analog Diversified Other Foundry Packaging
equipment design processor IDM3 IDM IDM and
automation assembly
77.6
59.7
39.7
26.9 27.0 29.2
3.3 42 5.3
1.9 13.1 0.9
0 –0.1 –1.6 –0.4 –1.2 –0.2 –1.5
–3.3 –2.7 –3.3 –3.3
23.6 3.3 1.8 76.0 39.3 58.4 12.9 24.3 2.0 25.9 –0.6
Top companies for economic profit, by subsegment
A) AMAT, ASML, Lam Research; B) Synopsys, Cadence, Mentor graphics; C) ARM, Rambus, CEVA; D) Broadcomm, Qualcomm,
Apple; E) Intel; F) Samsung, SK Hynix, Sandisk; G) Analog, Skyworks, Linear; H) TI, Toshiba, NXP; I) Microchip, Nuflare, Fingerprint; J)
TSMC; K) ASE, Silicon-ware.

² Intellectual property.
3
Integrated device manufacturer.

Exhibit 5 of 10
Exhibit 5
Industry consolidation over the past five years has likely contributed to improved profitability.
Economic-profit value creation by number of companies, 2012–17 (all subsegments)
Companies with cumulative positive economic profit Companies with cumulative negative economic profit
208 206 202

187 184
103 114 130 173
119 123
128
105
92
72 68
61
45
2012 2013 2014 2015 2016 2017

A B C D E
Major deals or bankruptcies by subsegment
A) Mindspeed Tech: Acquired by MACOM, Mtekvision Assets: Became private, Transwitch Corp: Filed for bankruptcy; B) International
Rectifier: Acquired by Infineon Tech, Supertex: Acquired by Microchip Technology; C) Altera Corp: Acquired by Intel, IBM Microelectronics:
Acquired by GlobalFoundries, Spansion: Merged with Cypress; D) Actions Semiconductor: Became Private, Anadigics: Acquired by II-VI Inc.,
Fairchild Semicon: Acquired by ON; E) Applied Micro Circuits: Acquired by MACOM
Because there’s greater scale within subsegments, top three ranking. Results for value creation were
companies have more resources to invest similar across most other subsegments, although
in innovation and operating improvements. some notable declines occurred. For instance, in the
Their large size also helps them rebound when microprocessor subsegment, the positive cumulative
downturns occur in this highly cyclical industry, economic profit for the period from 2013 to 2017
since they can take advantage of economies would be reduced from $39.7 billion to $28.3 billion
of scale and rely on more designs than in the past if goodwill is included. Since all semiconductor
for their revenues. If one customer leaves, their subsegments have engaged in M&A to a similar
bottom-line will not see the same hit as a smaller extent, it is not surprising that the relative rankings
player with only a few accounts. Overall, down remained similar.
cycles have been milder and peaks have been
higher within the semiconductor industry over the
past few years (Exhibit 6). Potential challenges: The rise of
in-house chip design
Even when we factored goodwill—the amount of a After five successful years, semiconductor
purchase that exceeds the book value of the assets leaders across the industry have become a bit
involved—into the calculation for the past five years, less optimistic about their prospects. Next to
economic profit remained high. The fabless, memory, global tensions (hitting the semiconductor sector
and microprocessor subsegments retained their significantly, given the international value chains),

Exhibit 6 of 10
Exhibit 6
Within the semiconductor industry, down cycles have been milder and peaks have been higher
in past few years.
Economic-profit value creation by all subsegments (excluding goodwill), $ billion
100 100
80 80
60 60
40 40
20 20
0 0
–20 –20
–40 –40
1997 2000 2010 2017
one trend is generating new questions: the —— Competitive differentiation. By developing

continued rise of in-house chip design at some of proprietary technology, Apple prevents
the semiconductor industry’s largest customers. other companies from replicating its
customer experience.
This shift may be most prominent at Apple. While
the company still relies on external providers for —— Insight into road maps. In-house creation
PC chips, it uses in-house designers to make the gives Apple firsthand insight into processing-
core chips for the iPhone, AppleTV, iWatch, and technology capabilities, allowing it to create
some other offerings. Apple then outsources chip more accurate product road maps and enabling
manufacture to foundries. The company gains superior launch planning for new products.
several advantages by taking this path: For each offering, it can specify how and when
it must update other technology elements to
—— Improved customer experience. Apple wants to complement the processor.
optimize the customer experience and ensure
that it is consistent across devices. While an —— Negotiating leverage. The sheer volume of chips
external provider could create custom designs designed in-house provides a strong negotiating
to meet these goals, an in-house team is position with foundries.
more likely to satisfy the company’s exacting
specifications and possess the necessary While Apple has conducted in-house work for
technical knowledge. many years, the scale, extent, and impact of these

Exhibit 7 of 10
Exhibit 7
Apple has become a large fabless semiconductor company by designing its own chips.
Estimated economic value creation by Apple’s

semiconductor activities (excluding goodwill), Apple’s total unit volume (iPhone, iPad, iPod,
2011–17, $ billion and iWatch), million
3.5 300
3.0 250
2.5
200
2.0
150
1.5
100
1.0
0.5 50
0 0
2011 2013 2015 2017 2011 2013 2015 2017
Source: Alphr; Bloomberg; Business Insider; IHS Markit; S&P Capital IQ
operations might surprise even industry insiders of revenue may further shift away from stand-alone
(Exhibit 7). Apple is now the third largest fabless semiconductor companies.
player in the world, behind Broadcom and Qualcomm
Technologies. If the company were selling chips, its Many technology companies with deep pockets
revenue would be around $15 billion to $20 billion have taken notice of Apple’s success with in-house
annually, in line with Qualcomm Technology’s. And chip design. Several, including large cloud players,
based on current multiples, Apple’s semiconductor are beginning to follow its example by developing
business would be worth $40 billion to $80 billion.² AI chips.⁴ They have already had some significant
These numbers speak volumes about the strength of wins, such as Google’s tensor-processing unit and
Apple’s internal chip operations. Amazon’s Graviton and Inferentia chips, all of which
facilitate cloud computing.⁵ In-house creation
Although shipments of iPhones and iPads appear allows these companies to develop customized
to have peaked, Apple is still expected to expand chips that offer better performance and security.
its semiconductor footprint for iWatches and Costs are also potentially lower, since companies
HomePods. It may also explore internal chip design do not have to pay a designer’s premium. In the
for other products and components, such as those hotly competitive cloud market, these cost savings
that enable power management and graphics.³ If could help companies differentiate themselves
Apple does go down this path, an important source from their rivals.
2
Based on a three- to fourfold revenue multiple of core Qualcomm Technology business (licensing business excluded).
3
Mark Gurman and Ian King, “Apple plans to use its own chips in Macs from 2020, replacing Intel,” Bloomberg, April 2, 2018, bloomberg.com.
4
Richard Waters, “Facebook joins Amazon and Google in AI chip race,” Financial Times, February 18, 2019, ft.com; Argam Atashyan, “Amazon
releases machine learning chips, namely Inferentia and Graviton,” Gizchina Media, November 29, 2018, gizchina.com; Jordan Novet, “Microsoft
is hiring engineers to work on A.I. chip design for its cloud,” CNBC, June 11, 2018, cnbc.com.
⁵ Jordan Novet, “Microsoft is hiring engineers to work on A.I. chip design for its cloud,” CNBC, June 11, 2018, cnbc.com; Tom Simonite, “New at
Amazon: Its own chips for cloud computing,” Wired, November 27, 2018, wired.com.

The development of ARM reference architectures, A look at earnings multiples might suggest
combined with the latest process improvements that investors are even less optimistic about the
at state-of-the-art foundries, could now open industry’s long-term growth prospects. But it’s
the door to other tech companies that want to more likely that they recognize that 2018 profit-
move design in-house—even those without deep ability for semiconductors was significantly higher
pockets. If more companies begin designing chips when it reached the top of the cycle than it had
in-house, the semiconductor industry will confront been in past years. Even as profits inevitably trend
a new type of competitor—and that could have a downward, investors still expect them to reach
long-term impact on demand and profitability. historic levels.
As for the subsegments, fabless now has the

Looking ahead: Investor expectations highest multiple, suggesting that investors
Economic profit is strongly correlated with total think it will remain more profitable and might
returns to shareholders (TRS) across industries. undergo additional M&A, which would increase
Overall, the semiconductor industry’s TRS has its resilience (Exhibit 9). Memory, by contrast,
declined about 10
McKinsey on Semiconductors percent since its peak in late
2018 has a low multiple, even though this subsegment has
2018, partly
What’s next semiconductors because investors are worried that recently generated record profits. Investors may be
Exhibit 8 of 10 the weakening macroeconomic environment could concerned that this segment is more commoditized
affect semiconductor demand (Exhibit 8). and therefore subject to sharper cyclic declines.
Exhibit 8
Total returns to shareholders have been strong within semiconductors.
Industry total returns to shareholders for the largest subsegments, index (Dec 31, 2012 = 100)
450 Diversified integrated

device manufacturer
Fabless
400
Foundry
Memory
350
S&P 500
300
250
200
150
100
50
Q1 Q1 Q1 Q1 Q1 Q1 Q1
2013 2014 2015 2016 2017 2018 2019
Source: Annual reports; S&P Capital IQ; Semiconductor CPC database

Exhibit 9 of 10
Exhibit 9
Performance expectations are highest for the fabless subsegment.

Enterprise value/EBITA¹ ratio, 2018
Fabless Foundry Micro- Capital Diversified Analog Memory

processor equipment IDM² IDM
15
14
12
11 10 10
Implied long-term organic growth at 2018 margins, % CAGR³
3 2 1 –1 1 0 N/A⁴
¹ Earnings before interest, taxes, and amortization; market data as of Dec 31, 2018.
² Integrated device manufacturer.
³ Compound annual growth rate.
⁴ Memory industry reached the peak of the valuation cycle in 2018 with extraordinarily high margins. Margin decline is imminent for the sector and is included in the
sector valuation. At 2018 margins, valuation cannot be reconciled with topline decline alone.
Fabless also has the highest enterprise value competitor for their design needs. Such losses
and is now the largest subsegment by far hurt, but they were often temporary because
(Exhibit 10). Its ability to capture the highest customers often came back to the original
economic profit, strong near-term growth company for future designs. Now if customers
prospects, and potential resiliency all contribute are dissatisfied with a road map, they might
to higher investor expectations. move design capabilities in-house, resulting in
a permanent loss of business.
Potential strategies for
semiconductor companies —— Using M&A in moderation. The semiconductor
Given the current landscape, semiconductor industry is still fragmented in many
companies must accelerate value creation. Four subsegments, and industry consolidation
actions seem essential: still makes sense. The best strategy involves
programmatic M&A, in which companies
—— Creating strong road maps for leading acquire at least one company a year, spending
customers. Semiconductor companies have long an average of 2 to 5 percent of their market
recognized the importance of delivering winning capitalization, with no single deal accounting
road maps for chip design, but the stakes are for more than 30 percent of their market
now higher than ever. In the past, customers that capitalization.⁶ These deals allow players
did not like a proposed road map might go to a
6
Chris Bradley, Martin Hirt, and Sven Smit, Strategy Beyond the Hockey Stick: People, Probabilities, and Big Moves to Beat the Odds, first
edition, Hoboken, NJ: John Wiley & Sons, 2018.

Exhibit 8 of 10
Exhibit 10
Fabless captures the most shareholder value across subsegments.

Enterprise-value distribution across the semiconductor subsegments, % share
100% = $1.5 trillion

Fabless Diversified Memory
25 IDM¹ 7
15
Foundry Analog
13 IDM
5
Microprocessors
17 Others²
7
Capital
equipment
12
Note: Figures may not sum, because of rounding.

¹ Integrated device manufacturer.
² Includes intellectual property, electronic-design automation, packaging and assembly, and other IDM (discrete, opto, microcontroller unit).
Source: S&P Capital IQ; McKinsey analysis
to branch into adjacent areas to strengthen deal could improve any of these areas, it will
their competitive position. Deals that involve help the companies create a more compelling
companies that only offer similar products road map that positions them for future
will not produce as much value. One factor to success. But companies that undertake
consider when contemplating a deal is the value M&A must avoid falling into the trap of paying
that it will bring to customers on measures such too much for goodwill, or else they risk
as price, quality, and performance. If an M&A destroying value.

—— Maintaining price discipline across the cycle. chips, or to exit certain areas altogether instead
The large companies that have emerged from of operating subscale.⁷
deal making have the resources required to
create leading-edge chips. But they will only
win if they focus on smart capacity planning
and maintain relatively stable prices across The semiconductor industry’s recent move to
economic cycles, even if demand slows. value creation is impressive, but companies
cannot assume that the strong profits will continue
—— Preparing for vertical integration among indefinitely. The move to in-house chip design
tech giants. Many large tech players may try among their most important customers could
to acquire small niche companies, especially hit their bottom lines hard. While the pace and
if they have desirable intellectual property, extent of this shift are still unknown, the best
so they can increase their semiconductor companies will begin preparing now by revamping
capabilities. Other large players may choose to their strategies.
license their technologies, rather than buying
7
Benjamin Mayo, “Apple licenses Dialog power management tech, and hires 300 engineers, to develop more custom iPhone chips,” 9to5Mac,
October 10, 2018, 9to5mac.com.
Marc de Jong is a partner in McKinsey’s Amsterdam office, and Anurag Srivastava is an alumnus of the New York office.
Copyright © 2019 McKinsey & Company. All rights reserved.

Artificial-intelligence
hardware: New
opportunities for
semiconductor companies
Artificial intelligence is opening the best opportunities for semi-
conductor companies in decades. How can they capture this value?
by Gaurav Batra, Zach Jacobson, Siddarth Madhav, Andrea Queirolo, and Nick Santhanam
© DuKai photographer/Getty Images

Software has been the star of high tech over the This article begins by reviewing the opportunities that
past few decades, and it’s easy to understand why. they will find across the technology stack, focusing on
With PCs and mobile phones, the game-changing the impact of AI on hardware demand at data centers
innovations that defined this era, the architecture and the edge (computing that occurs with devices,
and software layers of the technology stack enabled such as self-driving cars). It then examines specific
several important advances. In this environment, opportunities within compute, memory, storage, and
semiconductor companies were in a difficult networking. The article also discusses new strategies
position. Although their innovations in chip design that can help semiconductor companies gain an
and fabrication enabled next-generation devices, advantage in the AI market, as well as issues they
they received only a small share of the value coming should consider as they plan their next steps.
from the technology stack—about 20 to 30 percent
with PCs and 10 to 20 percent with mobile.
The AI technology stack will open
But the story for semiconductor companies could many opportunities for semiconductor
be different with the growth of artificial intelligence companies
(AI). Many AI applications have already gained a AI has made significant advances since its
wide following, including virtual assistants that emergence in the 1950s, but some of the most
manage our homes and facial-recognition programs important developments have occurred recently
that track criminals. These diverse solutions, as as developers created sophisticated machine-
well as other emerging AI applications, share one learning (ML) algorithms that can process large
common feature: a reliance on hardware as a core data sets, “learn” from experience, and improve
enabler of innovation, especially for logic and over time. The greatest leaps came in the 2010s
memory functions. because of advances in deep learning (DL),
a type of ML that can process a wider range
What will this development mean for semiconductor of data, requires less data preprocessing by
sales and revenues? And which chips will be most human operators, and often produces more
important to future innovations? To answer these accurate results.
questions, we reviewed current AI solutions and the
technology that enables them. We also examined To understand why AI is opening opportunities for
opportunities for semiconductor companies across semiconductor companies, consider the technology
the entire technology stack. Our analysis revealed stack (Exhibit 1). It consists of nine discrete layers that
three important findings about value creation: allow the two activities that enable AI applications:
training and inference (see sidebar “Training and
—— AI could allow semiconductor companies to inference”). When developers are trying to improve
capture 40 to 50 percent of total value from training and inference, they often encounter
the technology stack, representing the best roadblocks related to the hardware layer, which
opportunity they’ve had in decades. includes storage, memory, logic, and networking. By
providing next-generation accelerator architectures,
—— Storage will experience the highest growth, but semiconductor companies could increase
semiconductor companies will capture the most computational efficiency or facilitate the transfer of
value in compute, memory, and networking. large data sets through memory and storage. For
instance, specialized memory for AI has 4.5 times
—— To avoid mistakes that limited value capture in the more bandwidth than traditional memory, making it
past, semiconductor companies must undertake much better suited to handling the vast stores of big
a new value-creation strategy that focuses on data that AI applications require. This performance
enabling customized, end-to-end solutions for improvement is so great that many customers would
specific industries, or “microverticals.” be more willing to pay the higher price that specialized
memory requires (about $25 per gigabyte, compared
By keeping these beliefs in mind, semiconductor with $8 for standard memory).
leaders can create a new road map for winning in AI.
Artificial-intelligence hardware: New opportunities for semiconductor companies 17

Artificial intelligence semiconductors
Exhibit 1 of 6
Exhibit 1
The technology stack for artificial intelligence (AI) contains nine layers.
Technology Stack Definition Memory

• Electronic data repository for
Services Solution and Integrated solutions that include training data, short-term storage during
use case models, hardware, and other components (eg, processing
voice-recognition systems) • Memory typically consists of DRAM1
Training Data types Data presented to AI systems for analysis
Storage
Platform Methods Techniques for optimizing weights given to model inputs • Electronic repository for long-term
storage of large data sets
Architecture Structured approach to extract features from data (eg, • Storage typically consists of NAND2
convolutional or recurrent neural networks)
Logic
Algorithm A set of rules that gradually modifies the weights given to • Processor optimized to calculate
certain model inputs within the neural network during
neural-network operations, ie,
training to optimize inference
convolution and matrix multiplication
• Logic devices are typically CPU,
Framework Software packages to define architectures and invoke
algorithms on the hardware through the interface GPU, FPGA, and/or ASIC3
Interface Interface systems Systems within framework that determine and facilitate Networking
• Switches, routers, and other
communication pathways between software and
underlying hardware equipment used to link servers
in the cloud and to connect
Hardware Head node Hardware unit that orchestrates and coordinates edge devices
computations among accelerators
Accelerator Silicon chip designed to perform highly parallel operations
required by AI; also enables simultaneous computations
¹ Dynamic random access memory.

² Not AND.
³ CPU = central processing unit, GPU = graphics-processing unit, FPGA = field programmable gate array, ASIC = application-specific integrated circuit.
Source: Expert interviews; literature search
AI will drive a large portion of (Exhibit 3). By 2025, AI-related semiconductors

semiconductor revenues for data could account for almost 20 percent of all demand,
centers and the edge which would translate into about $65 billion in
With hardware serving as a differentiator in AI, revenue. Opportunities will emerge at both data
semiconductor companies will find greater demand centers and the edge. If this growth materializes
for their existing chips, but they could also profit by as expected, semiconductor companies will be
developing novel technologies, such as workload- positioned to capture more value from the AI
specific AI accelerators (Exhibit 2). We created a technology stack than they have obtained with
model to estimate how these AI opportunities would previous innovations—about 40 to 50 percent of
affect revenues and to determine whether AI-related the total.
chips would constitute a significant portion of future
demand (see sidebar “How we estimated value” for
more information on our methodology). AI will drive most growth in storage,
but the best opportunities for value
Our research revealed that AI-related semicon- creation lie in other segments
ductors will see growth of about 18 percent annually We then took our analysis a bit further by looking
over the next few years—five times greater than the at specific opportunities for semiconductor players
rate for semiconductors used in non-AI applications within compute, memory, storage, and networking.

Exhibit 2 of 6
Exhibit 2
Companies will find many opportunities in the artificial intelligence (AI) market, with leaders
already emerging.
Opportunities in existing market Potential new opportunities
• Accelerators for parallel processing, such as GPUs1 • Workload-specific AI accelerators

Compute
and FPGAs2
• High-bandwidth memory • NVM⁴ (as memory device)

Memory
• On-chip memory (SRAM3)
• Potential growth in demand for existing storage systems • AI-optimized storage systems
Storage
as more data are retained • Emerging NVM (as storage device)
• Infrastructure for data centers • Programmable switches

Networking
• High-speed interconnect
¹ Graphics-processing units.
² Field programmable gate arrays.
³ Static random access memory.
⁴ Nonvolatile memory.
Source: McKinsey analysis
For each area, we examined how hardware demand Data-center usage. Most compute growth will stem
is evolving at both data centers and the edge. from higher demand for AI applications at cloud-
We also quantified the growth expected in each computing data centers. At these locations, GPUs
category except networking, where AI-related are now used for almost all training applications. We
opportunities for value capture will be relatively expect that they will soon begin to lose market share
small for semiconductor companies. to ASICs, until the compute market is about evenly
divided between these solutions by 2025. As ASICs
Compute enter the market, GPUs will likely become more
Compute performance relies on central processing customized to meet the demands of DL. In addition
units (CPUs) and accelerators—graphics- to ASICs and GPUs, FPGAs will have a small role in
processing units (GPUs), field programmable gate future AI training, mostly for specialized data-center
arrays (FPGAs), and application-specific integrated applications that must reach the market quickly or
circuits (ASICs). Since each use case has different require customization, such as those for prototyping
compute requirements, the optimal AI hardware new DL applications.
architecture will vary. For instance, route-planning
applications have different needs for processing For inference, CPUs now account for about
speed, hardware interfaces, and other performance 75 percent of the market. They’ll lose ground to ASICs
features than applications for autonomous driving or as DL applications gain traction. Again, we expect to
financial-risk stratification (Exhibit 4). see an almost equal divide in the compute market,
with CPUs accounting for 50 percent of demand in
Overall, demand for compute hardware will increase 2025 and ASICs for 40 percent.
through 2025 (Exhibit 5). After analyzing more than
150 DL use cases, looking at both inference and Edge applications. Most edge training now occurs
training requirements, we were able to identify the on laptops and other personal computers, but more
architectures most likely to gain ground in data devices may begin recording data and playing a role
centers and the edge (Exhibit 6). in on-site training. For instance, drills used during

oil and gas exploration generate data related to a is required—typically dynamic random access
well’s geological characteristics that could be used memory (DRAM)—to store input data, weigh model
to train models. For accelerators, the training market parameters, and perform other functions during
is now evenly divided between CPUs and ASICs. both inference and training. Consider a model
In the future, however, we expect that ASICs built being trained to recognize the image of a cat. All
into systems on chips will account for 70 percent intermediate results in the recognition process—
of demand. FPGAs will represent about 20 percent for example, colors, contours, and textures—need
of demand and will be used for applications that to reside on memory as the model fine-tunes its
require significant customization. algorithms. Given these requirements, AI will create
a strong opportunity for the memory market, with
When it comes to inference, most edge devices now value expected to increase from $6.4 billion in 2017
rely on CPUs or ASICs, with a few applications—such to $12.0 billion in 2025. That said, memory will see
as autonomous cars—requiring GPUs. By 2025, we the lowest annual growth of the three accelerator
expect that ASICs will account for about 70 percent of categories—about 5 to 10 percent—because of
the edge inference market and GPUs 20 percent. efficiencies in algorithm design, such as reduced
bit precision, as well as capacity constraints in the
Memory industry relaxing.
AI applications have high memory-bandwidth
requirements, since computing layers within Most short-term memory growth will result from
deep neural networks must pass input data to increased demand at data centers for the high-
thousands of cores as quickly as possible. Memory bandwidth DRAM required to run AI, ML, and
Training and inference

All artificial-intelligence (AI) The cloud is an ideal location for training have time to send images to the cloud for
applications must be capable of training because it provides access to vast stores processing once it detects an object in
and inference. To understand the of data from multiple servers—and the road, nor do medical applications that
importance of these tasks, consider their the more information an AI application evaluate critically ill patients have leeway
role in helping self-driving cars avoid reviews during training, the better its when interpreting brain scans after a
obstacles. During the training phase, algorithm will become. Further, the hemorrhage. And that makes the edge,
developers present images to the neural cloud can reduce expenses because it or in-device computing, the best choice
network—for instance, those of dogs or allows graphics-processing units (GPUs) for inference.
pedestrians—and perform recognition and other expensive hardware to train
tests. They then refine network multiple AI models. Since training occurs
parameters until the neural network intermittently on each model, capacity is
displays high accuracy in visual detection. not an issue.
After the network has viewed millions
of images and is fully trained, it enables With inference, AI algorithms handle
recognition of dogs and pedestrians less data but must generate responses
during the inference phase. more rapidly. A self-driving car doesn’t

Exhibit 3 of 6
Exhibit 3
Growth for semiconductors related to artificial intelligence (AI) is expected to be five times
greater than growth in the remainder of the market.
AI semiconductor total AI semiconductor total Estimated AI semiconductor total
available market,1 $ billion available market, % available market CAGR,2 2017–25, %
AI
362 18–19
Non-AI
7 11
65 19
93
88
256 295 81
240
17 32
5x
223 224
3–4
2017 2020E 2025E 2017 2020E 2025E Non-AI AI
¹ Total available market includes processors, memory, and storage; excludes discretes, optical, and micro-electrical-mechanical systems.
² Compound annual growth rate.
Source: Bernstein; Cisco Systems; Gartner; IC Insights; IHS Markit; Machina Research; McKinsey analysis
DL algorithms. But over time, the demand for AI that signals their customers are willing to
memory at the edge will increase—for instance, pay for expensive AI hardware in return for
connected cars may need more DRAM. performance gains.¹
Current memory is typically optimized for CPUs, but —— On-chip memory. For a DL compute processor,
developers are now exploring new architectures. storing and accessing data in DRAM or other
Solutions that are attracting more interest include outside memory sources can take 100 times
the following: more time than memory on the same chip. When
Google designed the tensor-processing unit
—— High-bandwidth memory (HBM). This (TPU), an ASIC specialized for AI, it included
technology allows AI applications to process enough memory to store an entire model on
large data sets at maximum speed while the chip.² Start-ups such as Graphcore are also
minimizing power requirements. It allows increasing on-chip memory capacity, taking
DL compute processors to access a three- it to a level about 1,000 times more than what
dimensional stack of memory through a fast is found on a typical GPU, through a novel
connection called through-silicon via (TSV). architecture that maximizes the speed of AI
AI chip leaders such as Google and Nvidia calculations. The cost of on-chip memory is
have adopted HBM as the preferred memory still prohibitive for most applications, and chip
solution, although it costs three times more designers must address this challenge.
than traditional DRAM per gigabyte—a move
1
Liam Tung, “GPU killer: Google reveals just how powerful its TPU2 chip really is,” ZDNet, December 14, 2017, zdnet.com.
2
Kaz Sato, “What makes TPUs fine-tuned for deep learning?,” Google, August 30, 2018, google.com.

Exhibit 4 of 6
Exhibit 4
The optimal compute architecture will vary by use case.

Example use-case analysis of importance
Hardware Backward Small size; Low cost of

Very
interfaces compatibility1 form factor development
Somewhat Processing Processing/ Flexibility Technical Processing/

speed watt capabilities $ spent
Less Training Inference
required
Language translation, GPU2 CPU4
speech understanding ASIC3 ASIC
Face
GPU CPU
detection
FPGA5 ASIC
Financial-risk
stratification GPU CPU
FPGA
Route planning and

optimization GPU CPU
Dynamic
pricing GPU CPU
ASIC
Autonomous
driving ASIC GPU
ASIC
FPGA
¹ Can use interfaces and data from earlier versions of the system.
² Graphics-processing unit.
³ Application-specific integrated circuit.
⁴ Central processing unit.
⁵ Field programmable gate array.
Source: McKinsey analysis
Storage and those depend on whether an application is used

AI applications generate vast volumes of data—about for training or inference. For instance, AI training
80 exabytes per year, which is expected to increase systems must store massive volumes of data as they
to 845 exabytes by 2025. In addition, developers are refine their algorithms, but AI inference systems
now using more data in AI and DL training, which also only store input data that might be useful in future
increases storage requirements. These shifts could training. Overall, demand for storage will be higher
lead to annual growth of 25 to 30 percent from for AI training than inference.
2017 to 2025 for storage—the highest rate of all
segments we examined.³ Manufacturers will One potential disruption in storage is new forms
increase their output of storage accelerators of nonvolatile memory (NVM). New forms of NVM
in response, with pricing dependent on supply have characteristics that fall between traditional
staying in sync with demand. memory, such as DRAM, and traditional storage,
such as NAND flash. They can promise higher
Unlike traditional storage solutions that tend to take density than DRAM, better performance than
a one-size-fits-all approach across different use NAND, and better power consumption than both.
cases, AI solutions must adapt to changing needs— These characteristics will enable new applications
³ When exploring opportunities for semiconductor players in storage, we focused on not AND (NAND). Although demand for hard-disk drives will
also increase, this growth is not driven by semiconductor advances.

Exhibit 5 of 6
Exhibit 5
At both data centers and the edge, demand for training and inference hardware is growing.
Data center, total market, $ billion Edge, total market, $ billion
Inference Training Inference Training

9–10
4–5 4–5
4–4.5
1–1.5
~1
<0.1 <0.1
2017 2025 2017 2025 2017 2025 2017 2025
Source: Expert interviews; McKinsey analysis
and allow NVM to substitute for DRAM and NAND Networking

in others. The market for these forms of NVM is AI applications require many servers during training,
currently small—representing about $1 billion to and the number increases with time. For instance,
$2 billion in revenue over the next two years—but it developers only need one server to build an initial
is projected to account for more than $10 billion in AI model and under 100 to improve its structure.
revenue by 2025. But training with real data—the logical next step—
could require several hundred. Autonomous-driving
The NVM category includes multiple technologies, models require more than 140 servers to reach
all of which differ in memory access time and cost 97 percent accuracy in detecting obstacles.
and are all in various stages. Magnetoresistive
random-access memory (MRAM) has the lowest If the speed of the network connecting servers
latency for read and write, with greater than five- is slow—as is usually the case—it will cause
year data retention and excellent endurance. training bottlenecks. Although most strategies
However, its capacity scaling is limited, making it for improving network speed now involve data-
a costly alternative that may be used for frequently center hardware, developers are investigating
accessed caches rather than a long-term data- other options, including programmable switches
retention solution. Resistive random-access that can route data in different directions. This
memory (ReRAM) could potentially scale vertically, capability will accelerate one of the most important
giving it an advantage in scaling and cost, but it training tasks: the need to resynchronize input
has slower latency and reduced endurance. Phase- weights among multiple servers whenever model
change memory (PCM) fits in between the two, parameters are updated. With programmable
with 3D XPoint being the most well-known switches, resynchronization can occur almost
example. Endurance and error rate will be key instantly, which could increase training speed from
barriers that must be overcome before more two to ten times. The greatest performance gains
widespread adoption. would come with large AI models, which use the
most servers.

Another option to improve networking involves applications within the healthcare and military
using high-speed interconnections in servers. industries. Alternatively, semiconductor companies
This technology can produce a threefold could focus on developing AI hardware that
improvement in performance, but it’s also about enables broad, cross-industry solutions, as Nvidia
35 percent more expensive. does with GPUs.
The path taken will vary by segment. With memory

Semiconductor companies need new and storage players, solutions tend to have the same
strategies for the AI market technology requirements across microverticals. In
It’s clear that opportunities abound, but success compute, by contrast, AI algorithm requirements
isn’t guaranteed for semiconductor players. may vary significantly. An edge accelerator in an
To capture the value they deserve, they’ll need autonomous car must process much different
to focus on end-to-end solutions for specific data from a language-translation application that
industries (also called microvertical solutions), relies on the cloud. Under these circumstances,
ecosystem development, and innovation that companies cannot rely on other players to build
goes far beyond improving compute, memory, and other layers of the stack that will be compatible with
networking technologies. their hardware.
Customers will value end-to-end solutions Active participation in ecosystems is vital

for microverticals that deliver a strong return for success
on investment Semiconductor players will need to create an
AI hardware solutions are only useful if they’re ecosystem of software developers that prefer their
compatible with all other layers of the technology hardware by offering products with wide appeal.
stack, including the solutions and use cases in In return, they’ll have more influence over design
the services layer. Semiconductor companies choices. For instance, developers who prefer a
can take two paths to achieve this goal, and a few certain hardware will use that as a starting point
have already begun doing so. First, they could when building their applications. They’ll then look for
work with partners to develop AI hardware for other components that are compatible with it.
industry-specific use cases, such as oil and gas
exploration, to create an end-to-end solution. For To help draw software developers into their
example, Mythic has developed an ASIC to support ecosystem, semiconductor companies should
edge inference for image- and voice-recognition reduce complexity whenever possible. Since
How we estimated value

We took a bottom-up approach for artificial intelligence (AI). We then the number of devices that were used
to estimate the value at stake for identified the type of logic device they for AI, rather than servers. By combining
semiconductor companies. Consider commonly used and the average sales our insights for data centers and edge
accelerators used for compute functions. price for related accelerators. For edge devices, we could estimate the potential
First, we determined the percent of computing, we conducted a similar value for semiconductor companies
servers in data centers that were used review, but we focused on determining related to compute functions.

Exhibit 6 of 6
Exhibit 6
The preferred architectures for compute are shifting in data centers and the edge.
Data-center architecture, % Edge architecture, %
ASIC1 CPU2 FPGA3 GPU4 Other
Inference Training Inference Training
50 50 50
60
75 70 70
97
40 40
30 50
10 20 20
15 10 10 10 10 10
2017 2025 2017 2025 2017 2025 2017 2025
¹ Application-specific integrated circuit.

² Central processing unit.
³ Field programmable gate array.
⁴ Graphics-processing unit.
Source: Expert interviews; McKinsey analysis
there are now more types of AI hardware than object-detection libraries that can help applications
ever, including new accelerators, players should interpret data from cameras and sensors in self-
offer simple interfaces and software-platform driving cars.
capabilities. For instance, Nvidia provides
developers with Compute Unified Device As preference for certain hardware architectures
Architecture, a parallel-computing platform and builds throughout the developer community,
application programming interface (API) that works semiconductor companies will see their visibility
with multiple programming languages. It allows soar, resulting in better brand recognition. They’ll
software developers to use Compute Unified also see higher adoption rates and greater
Device Architecture–enabled GPUs for general- customer loyalty, resulting in lasting value.
purpose processing. Nvidia also provides software
developers with access to a collection of primitives Only platforms that add real value to end users will
for use in DL applications. The platform has now be able to compete against comprehensive offerings
been deployed across thousands of applications. from large high-tech players, such as Google’s
TensorFlow, an open-source library of ML and DL
Within strategically important industry sectors, Nvidia models and algorithms.⁴ TensorFlow supports
also offers customized software-development kits. Google’s core products, such as Google Translate,
To assist with the development of software for self- and also helps the company solidify its position
driving cars, for instance, Nvidia created DriveWorks, within the AI technology stack, since TensorFlow is
a kit with ready-to-use software tools, including compatible with multiple compute accelerators.
4
An open-source, machine-learning framework for everyone, available at tensorflow.org.

Innovation is paramount, and players must solutions could eliminate. On the technical side,
go up the stack companies should decide if they want to focus
Many hardware companies that want to enable AI on hardware for data centers or the edge.
innovation focus on improving the computation
process. Traditionally, this strategy has involved —— H
ow to play? When bringing a new solution
offering optimized compute accelerators or to market, semiconductor companies should
streamlining paths between compute and data adopt a partnership mind-set, since they
through innovations in memory, storage, and might gain a competitive edge by collaborating
networking. But hardware companies should with established companies within specific
go beyond these steps and seek other forms of industries. They should also determine what
innovation by going up the stack. For example, organizational structure will work best for their
AI-based facial-recognition systems for secure business. In some cases, they might want to
authentication on smartphones were enabled by create groups that focus on certain functions,
specialized software and a 3-D sensor that projects such as R&D, for all industries. Alternatively, they
thousands of invisible dots to capture a geometric could dedicate groups to select microverticals,
map of a user’s face. Because these dots are much allowing them to develop specialized expertise.
easier to process than several millions of pixels from
cameras, these authentication systems work in a —— W
hen to play? Many companies might be
fraction of a second and don’t interfere with the user tempted to jump into the AI market, since the cost
experience. Hardware companies could also think of being a follower is high, particularly with DL
about how sensors or other innovative technologies applications. Further, barriers to entry will rise as
can enable emerging AI use cases. industries adopt specific AI standards and expect
all players to adhere to them. While rapid entry
might be the best approach for some companies,
Semiconductor companies must define others might want to take a more measured
their AI strategy now approach that involves slowly increasing their
Semiconductor companies that take the lead in AI investment in select microverticals over time.
will be more likely to attract and retain customers
and ecosystem partners—and that could prevent
later entrants from attaining a leading position in
the market. With both major technology players The AI and DL revolution gives the semiconductor
and start-ups launching independent efforts in the industry the greatest opportunity to generate value
AI hardware space now, the window of opportunity that it has had in decades. Hardware can be the
for staking a claim will rapidly shrink over the next differentiator that determines whether leading-edge
few years. To establish a strong strategy now, they applications reach the market and grab attention.
should focus on three questions: As AI advances, hardware requirements will shift for
compute, memory, storage, and networking—and
—— W
here to play? The first step to creating a that will translate into different demand patterns.
focused strategy involves identifying the target The best semiconductor companies will understand
industry microverticals and AI use cases. At the these trends and pursue innovations that help take
most basic level, this involves estimating the AI hardware to a new level. In addition to benefiting
size of the opportunity within different verticals, their bottom line, they’ll also be a driving force
as well as the particular pain points that AI behind the AI applications transforming our world.
Gaurav Batra is a partner in McKinsey’s Washington, DC, office; Zach Jacobson and Andrea Queirolo are associate partners
in the New York office; Siddarth Madhav is a partner in the Chicago office; and Nick Santhanam is a senior partner in the
Silicon Valley office.
The authors wish to thank Sanchi Gupte, Jo Kakarwada, Teddy Lee, and Ben Byungchol Yoon for their contributions to this article.

Blockchain 2.0: What’s in
store for the two ends—
semiconductors
(suppliers) and industrials
(consumers)?
Ten years after blockchain’s inception, it is presenting new
opportunities for both suppliers, such as semiconductor companies,
and consumers, such as industrials.
by Gaurav Batra, Rémy Olson, Shilpi Pathak, Nick Santhanam, and Harish Soundararajan
© Imaginima/Getty Images
Blockchain 2.0: What’s in store for the two ends—semiconductors (suppliers) and industrials (consumers)? 27
Blockchain is best known as a sophisticated specific strategies for capturing value. One caveat:
and somewhat mysterious technology that allows all information in this article reflects data available
cryptocurrencies to change hands online without as of December 2018. Cryptocurrency values
assistance from banks or other intermediaries. fluctuate widely, so the numbers reported, including
But in recent years, it has also been promoted as those for market capitalization, may not reflect the
the solution to business issues ranging from fraud most recent data. Blockchain technology and the
management to supply-chain monitoring to identity competitive landscape are also evolving rapidly, and
verification. Despite the hype, however, block- there may have been changes since publication.
chain’s use in business is still largely theoretical. A
few pioneers in retail and other sectors are exploring
blockchain business applications related to supply- Blockchain 1.0: The cryptocurrency era
chain management and other processes, but most It is not surprising that many people conflate
are reluctant to proceed further because of high blockchain with Bitcoin, the first and most dominant
costs, unclear returns, and technical difficulties. cryptocurrency. Until recently, the vast majority
of blockchain applications involved enabling
But we may now be at a transition point between cryptocurrency transactions. Around 2014, however,
Blockchain 1.0 and Blockchain 2.0. In the new era, private companies began investigating the use of
blockchain-enabled cryptocurrency applications will blockchain for other business applications. Since
likely cede their prominence to blockchain business most of these players are still at the pilot stage, it is
applications that can potentially increase efficiency fair to say that blockchain-enabled cryptocurrency
and reduce costs. These applications will be in a has been the focus of the Blockchain 1.0 era.
good position to gain steam, since many large tech
companies may soon begin offering blockchain The emergence of cryptocurrencies
as a service (BaaS). Rather than just providing Bitcoin hit the market in 2009 as an open-source
the hardware layer, as they’ve traditionally done, software application. It was first used in a commer-
these companies will extend their services up the cial transaction in 2010, when two pizzas were
technology stack to blockchain platforms and tools. bought for 10,000 bitcoin (under $10 then, but
As blockchain deployment becomes less complex about $35 million as of December 2018). With no
and expensive, companies that have sat on the central authority or server to verify transactions,
sidelines may now be willing to take the plunge. (See the public was initially skeptical about Bitcoin and
sidebar, “What advantages do blockchain business reluctant to use it. Beginning in 2014, however,
applications offer?”) Bitcoin experienced a meteoric increase in user
base, brand-name recognition, and transaction
Will blockchain business applications continue to volume. Its value is extremely volatile, however, and
grow and finally validate their promise? Industrial it has declined sharply from its late 2017 peak of
companies, which were largely on the sidelines more than $19,000.
during the Blockchain 1.0 era, want an answer to
this question because they could find opportunities The past two years have seen the most growth in
to deploy business applications that improve their blockchain-enabled cryptocurrencies, with the
bottom line. Semiconductor companies are also number increasing from 69 in 2016 to more than
interested in the growth of both blockchain business 1,500 in 2018. Even though Bitcoin’s value has
applications and blockchain-enabled cryptocurrency decreased this year, an influx of initial coin offerings
because this could increase demand for chips. (ICOs) has increased the market capitalization for
cryptocurrencies (Exhibit 1).
Both industrial and semiconductor players will
need a solid understanding of specific blockchain- Many of the additional currencies—also called “alt-
enabled use cases and the market landscape to coins”—were created to address certain gaps or
succeed in the new era. To assist them, this article inefficiencies with Bitcoin, and they are available
reviews the changing market and then focuses on through various networks. Popular altcoins

include Dash, Litecoin, and XRP (offered through Transaction verification
Ripple). Of all the alternative cryptocurrency The method used to verify transactions varies by
networks, Ethereum is most popular. It is an cryptocurrency. With Bitcoin, the first participant,
open-source platform that allows users to build or “miner,” to validate a transaction and add a new
and launch decentralized applications, including block of data to the digital ledger will receive a
cryptocurrencies or digital ledgers. Users must certain number of tokens as a reward. Under this
spend a specific digital currency, Ether, to run model, which is referred to as a proof-of-work
applications on Ethereum. Ether can also serve (PoW) system, miners have an incentive to act
as an alternative to regular money, but its primary quickly. But validating a transaction doesn’t simply
purpose is to facilitate Ethereum operations. involve verifying that Bitcoin has been transferred
from one account to another. Instead, a miner has
Together, the market capitalization of a select set to answer a cryptographic question by correctly
of major cryptocurrencies
McKinsey on Semiconductors 2018 was about $150 billion in identifying an alphanumeric series associated
Blockchain 2.0 December 2018, with Bitcoin and the four leading with the transaction. This activity requires a lot of
Exhibit 1 of 3 altcoins representing about 75 percent of this value. trial and error, making the hash rate—the compute
Bitcoin’s market capitalization of about $60 billion speed at which an operation is completed—
was the highest. extremely important with Bitcoin.
Exhibit 1
The number of active cryptocurrencies and their market capitalization has soared.
Cryptocurrencies active in the market, number Cryptocurrency market capitalization,¹ $ billion
1,500+ 177
~150
392
69 11
33 7
29 2 4
8
2013 2018 2013 2018
¹ This is the market capitalization for a select bundle of cryptocurrencies. Bundle includes: Bitcoin, Dash, Ethereum, Litecoin, Ripple, and several other altcoins. Figures
are as of Dec 11, 2018.
In the beginning, many individuals mined Bitcoin to be ASIC resistant—and that means miners must
as a hobby. But as interest in cryptocurrencies fetch random data and compute randomly selected
grew, the number and size of Bitcoin miners soared, transactions to solve their cryptographic questions.
necessitating more sophisticated hardware and Both activities require frequent access to memory,
more intense computing power. This shift has which ASICs alone won’t provide. Ethereum miners
favored the rise of large mining pools. Many of these, primarily rely on a system that utilizes a GPU in
including AntPool and BTC.COM, are based in combination with memory.
China. The top five mining pools account for 70 to 85
percent of the overall Bitcoin network’s collective
hash rate, or computing power. Blockchain 2.0: Uncertainty about
cryptocurrencies and the emergence of
Hardware for cryptocurrency players business applications
In the early day of cryptocurrency, amateur The Blockchain 2.0 era will likely usher in many
hobbyists relied on central processing units (CPUs) changes. The cryptocurrency market could become
to optimize compute performance. When the more diverse if Bitcoin continues to decrease
Bitcoin network began expanding around 2010, in price, since ICOs may see the situation as an
the graphics-processing unit (GPU) replaced opportunity to stake their claims. Consumers may
the CPU as the accelerator of choice. The ascent also begin demonstrating more interest in other
of GPUs was short lived, however, since many established altcoins. For instance, users may come
companies began designing application-specific to favor Dash or Litecoin for some transactions,
integrated circuits (ASICs) for cryptocurrency since they offer faster transaction speed than
mining to improve hash rates. Bitcoin does. Companies and the general public
are generally becoming more comfortable with
About 50 to 60 percent of companies that cryptocurrency transactions, which could increase
manufacture ASICs for Bitcoin transactions usage rates.¹
are based in the Greater China region (Exhibit
2). (Some of these began creating ASICs for In tandem with these changes, the market for
cryptocurrency mining before Bitcoin entered the blockchain business applications is heating up as
market in 2008, since this was already viewed BaaS simplifies implementation. Demand for these
as a potential growth area.) BitMain Technologies, applications is expected to be strong, and corporate
a China-based company, supplied 70 to 80 percent users could soon outnumber cryptocurrency miners.
of the cryptocurrency ASICs in 2017. Its customers
typically use “crypto rigs”—multiple ASICs Investors are showing continued interest in
working together—to optimize compute speed. blockchain, although funding levels have recently
By conservative estimates, BitMain Technologies declined. Venture-capital funding peaked in 2017
has a gross margin of 65 to 75 percent and an at about $900 million for both cryptocurrency
operating margin of 55 to 65 percent—equivalent and business applications, and it will likely still be
to $3 billion to $4 billion in 2017. That figure is between $600 million and $800 million in 2018. It is
roughly the same as the profit margin for Nvidia, unclear whether 2019 will show continued decline,
which has been in business for 20 years longer. a plateau, or greater investment.
Although most major cryptocurrencies now reward Although it is difficult to make predictions about
miners with high compute speed, some have taken blockchain, since it is a relatively new technology,
steps to prevent large mining pools with crypto rigs we were able to identify several trends in the
from dominating the market. For instance, Ethash, the cryptocurrency and business-application markets
hashing algorithm that Ethereum uses, is designed
Josh Ong, “The branding of cryptocurrency,” Forbes, March 1, 2018, forbes.com.

1

Blockchain 2.0
Exhibit 2 of 3
Exhibit 2
Many companies have designed application-specific integrated chips to mine cryptocurrencies.

A timeline of cryptocurrency chip manufacturers
Company INNOSILICON Technology Bitfury BitMain Technologies CoinBau GmbH Butterfly Labs
HQ Ningbo, China DC, US Beijing, China Dresden, Germany Leawood, KS, US
Recent product T2 Turbo+ 32T Bitfury Tardis Antminer S9-Hydro WolfCAVE XE Monarch
Launch Sept 2018 Oct 2018 Aug 2018 Not available Aug 2014
Hash rate 32 Th/sec 80 Th/sec 18 Th/sec 4.8 Th/sec 725 or 825 Gh/sec
Power efficiency 0.069 J/Gh 0.055 J/Gh 0.096 J/Gh 0.27 W/Gh 0.7 W/Gh
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
Company Ebang Communication Black Arrow Canaan Creative CoinTerra Halong Mining
HQ Hangzhou, China Guangdong, China Beijing, China CA, US Not applicable (online only)
Recent product EBIT E11+ Prospero X36 AvalonMiner 851 TerraMiner IV Dragonmint T16
Launch Oct 2018 Dec 2015 Aug 2018 Jan 2014 Mar 2018
Hash rate 37 Th/sec 2.2 Th/sec 15 Th/sec 1.6 Th/sec 16 Th/sec
Power efficiency 0.055 J/Gh 0.7 W/Gh 0.11 J/Gh 0.6 W/Gh 0.075 J/Gh
Note: Gh = gigahash; J = joule; sec = second; Th = terahash; W = watt.
that could affect demand for this technology. Here may have lingering doubts that are difficult to
is what we found. overcome. But we do expect to see greater usage
rates. In addition, miners will have a greater number
The cryptocurrency market is evolving rapidly, of options from which to choose. Although Bitcoin
but uncertainties remain now represents about 40 to 50 percent of market
Despite the widespread press attention that capitalization for cryptocurrency, other altcoins are
cryptocurrencies receive, their practical value is still becoming more popular. Ethereum, for instance, now
limited. Most people regard them as something of accounts for more than 10 percent of the market
an online Swiss bank account—a haven for activities capitalization. And small ICOs—those beyond the
that can’t be closely tracked by authorities. In many top 20—now represent about 20 percent of market
cases, potential users hold back because they don’t capitalization, up from 5 percent only two years ago.
believe cryptocurrencies are secure. Digital-ledger
technology, the backbone of blockchain, has never Government intervention—particularly the
been hacked, but cryptocurrencies are vulnerable in development of laws and regulations—may strongly
other ways. The most infamous theft occurred in 2014 influence the cryptocurrency market over the next
when someone took 850,000 bitcoin from the Mt. few years. If the current market provides any clues,
Gox exchange by assuming another person’s identity. it is unlikely that a global consensus will emerge.
In the corporate sphere, only about 3,000 companies For instance, some governments allow individuals
now accept Bitcoin transactions. to use cryptocurrency but prohibit banks and
securities companies from doing so. Other countries
Future growth of cryptocurrencies take a much stricter approach by forbidding
It is difficult to predict whether cryptocurrencies ICOs to operate within their borders. If additional
will experience strong growth in Blockchain 2.0, governments adopt this stance, cryptocurrency
since corporate leaders and members of the public uptake could be limited.
Another big question relates to investment. Funding intermediaries and decrease administrative costs
for ICOs usually comes from venture capitalists associated with record keeping. Over the longer
because pension funds and other institutional term, blockchain might be used to improve fraud
investors consider cryptocurrency too risky. (The management, supply-chain monitoring, cross-
majority of ICOs do not yet have customers nor do border payments, identity verification, and the
they generate revenue.) Even though venture-capital protection of copyrights or intellectual property. It
investment in cryptocurrency has increased, the lack could also help companies with smart contracts—
of interest from institutional investors could restrict transactions that execute automatically when
future growth to some extent. certain conditions are met.
Changing algorithms Many companies and organizations are now

Behind the scenes, more subtle changes are supporting the development of blockchain business
occurring in the cryptocurrency market as players applications. The Linux Foundation has created
try to minimize the importance of compute power by Hyperledger, an open-source collaborative effort
developing new algorithms. For instance, Ethereum to develop blockchain technologies for multiple
is considering the replacement of its PoW system industries. Similarly, the company R3 leads a large
with one based on proof of stake (PoS). In a PoS consortium that developed Corda, a blockchain
system, participants are rewarded based on the platform for use in financial services and commerce.
number of coins they have in their digital wallets Corporate investment in blockchain hit $1 billion in
and the length of time they have had these stakes. 2017 and is expected to grow at a compound annual
The participant that rates highest on these factors growth rate of 50 percent through 2021.
is chosen to validate a transaction and receive a
reward. Many other large cryptocurrency networks, Despite these efforts, blockchain business
including Cardano, Dash, and EOS, are also applications arguably remain stuck at the pilot
investigating PoS algorithms. stage, with most companies still attempting to
demonstrate proof of concept (PoC). (The greatest
PoS systems have several advantages. First, they wave of business applications undergoing PoC
help cryptocurrency networks build a trusted occurred from 2016 to 2017; the number at this
network of loyal participants—and this may make stage is now smaller.) Many start-ups that offer
security breaches less common. Second, they level business applications have failed to obtain Series
the playing field for cryptocurrency miners, since C funding—the investment designed to promote
those with the greatest compute power will not growth and scale operations. The emergence of
necessarily be the winners. Players also appreciate competing technologies is the major reason for
that PoS systems are more energy efficient and the lack of progress. For instance, with payments,
allow faster transactions. A shift to PoS systems financial institutions can now use a messaging
could have major implications for semiconductor network that allows for greater transaction speeds
companies that serve cryptocurrency players, since and more transparency than past methods. This
it would shift chip demand in new directions. technology reduces the need for blockchain-based
solutions and discourages incumbents in the
A new look at business applications, but with financial sector from investing in blockchain.
doubts about scalability
Recent McKinsey research has identified more than Much interest in blockchain business applications
90 use cases for blockchain business applications stems from the recent advent of BaaS, which
across industries. Many near-term use cases simplifies the creation of the complex, five-layer
will involve applying blockchain to reduce costs blockchain technology stack (Exhibit 3). Until the past
associated with existing processes, such as the year, enterprise customers had to build individual
exchange of medical records among providers, layers themselves or cobble them together from
insurers, researchers, and patients. In these disparate sources. Among other tasks, they had to
activities, blockchain can remove the need for customize existing digital-ledger fabric platforms

(distributed computing platforms with a base By registering a device on blockchain, the
protocol and configurable functions). They also had company could give it a unique digital identity
to acquire and integrate data, define permissions that could not be altered. The company could
and governance protocols, and code software. Most easily update the digital identity in real time
enterprises simply lacked the funds or in-house to reflect any changes—a service it could not
technology talent to make this happen. perform with physical certificates.
With the emergence of BaaS, the onus of

deployment has moved from customers to providers. To help blockchain applications gain traction at
While BaaS is typically limited to the infrastructure industrial companies, stakeholders must address
layer, some providers also create tools that extend four structural challenges: inertia that prevents
into the data and digital-ledger layers. With access players from collaborating, a lack of standards,
to these offerings, customers can significantly unclear legal and regulatory frameworks, and
reduce the deployment costs of a new blockchain latency issues that make it difficult to verify multiple
system. For instance, they will no longer have to transactions rapidly. For instance, Bitcoin is limited
invest heavily in data or in ledger software and to seven transactions per second, and Ethereum
services to make their fabric platforms operational. can achieve 20 transactions per second. Financial
institutions, such as credit-card companies, can
handle between 24,000 and 56,000 transactions
How industrial companies can create per second.
value in Blockchain 2.0: Core beliefs
Across industries, companies have been exploring Based on our review of the industrial sector, we
blockchain opportunities. Many consumer-facing have identified three core beliefs about the ability
and industrial companies were somewhat late to the of companies to create and capture value during
game because most applications were geared toward Blockchain 2.0.
cryptocurrency or financial transactions during
Blockchain 1.0. But their involvement will increase as Belief 1: The value is in specific use cases that
more blockchain business applications move from the depend on incorruptible record keeping
concept stage to reality. For industrial companies, the Blockchain’s value proposition is clear: it functions
potential use cases span all areas of their operations, as a decentralized, incorruptible database that
and a few have already become reality: allows peers to conduct transactions without
relinquishing control to an intermediary or accepting
—— An industrial company formed a partnership counterparty risk. For industrial companies,
with a technology business that uses such incorruptible record keeping (IRK) can be
blockchain to track the origin of goods and invaluable. For instance, a global wireless-network-
their progress along the supply chain. By equipment company used blockchain to provide
providing greater transparency, the company cybersecurity for various industrial companies that
helped customers understand the quality of its used IoT, including those in utilities, oil and gas,
materials, the supply-chain process, and the and transportation. The IoT devices had tens of
sources of raw ingredients. thousands of nodes, each of which represented a
potential entry point for hackers. With blockchain,
—— A leading manufacturer of Internet of Things the company could track security threats by
(IoT) devices formed a partnership with a assigning each node a unique key that allowed it
blockchain start-up to create “digital passports” to detect unusual behavior or hacker intrusions
for individual IoT devices. The goal was to immediately. In those cases for which IRK is not
improve the expensive and time-consuming essential, industrial companies should consider
process for authentication, which involved using a traditional shared database for transactions,
obtaining physical certificates from authorities. since it is less expensive to maintain.
Blockchain 2.0
Exhibit 3 of 3
Exhibit 3
The blockchain technology stack includes five layers.

Technology layer Core function
Enable user interface and implement business logic (ie, portion of an enterprise
system that determines how data are transformed or calculated, and how they are routed to
Applications people or software)
Are typically domain or industry specific
Ensure interoperability of systems and manage permissions, disaster recovery, and governance
Digital-ledger software
and services
Contain ecosystem of data, such as sales information and shipping records, pulled into
blockchain application
Data
Provide base protocol and configurable functionalities for various services, such as smart
contracting
Digital-ledger fabric
platform
Provide infrastructure for hosting and developing blockchain, and for operating nodes
Include hardware
Development platform
and infrastructure
Source: Asian Venture Capital Journal; VCCEdge; McKinsey analysis
Belief 2: Scalable use cases will involve high preferable to a traditional shared database, for
value, low volume, and collaborative mechanisms instance. Similarly, blockchain applications
The list of potential blockchain applications that that significantly reduce cost by increasing
industrial companies could implement is long. They efficiency are well worth exploring. For instance,
could facilitate smart contracts, provide customers a machinery manufacturer may have a supply
with a clear record of a product’s origin, enhance chain that involves multiple intermediaries. A
logistics and supply chain, improve product quality, blockchain application that could reduce cost
or help satisfy regulatory requirements. But not and complexity during shipping would deliver
every industrial use case with strong potential will enormous value.
survive past the PoC stage. Those that are most
likely to gain traction share three characteristics: —— Low transaction volume. Blockchain
technology still has limited processing
—— High value. Each blockchain application power, which makes it difficult to perform
must deliver significant value to the bottom many transactions simultaneously. Until the
line. If an information breach could cause technology advances, industrial companies
a company to lose millions of dollars, a should apply it to use cases that involve limited
blockchain application might be infinitely transaction volume. For instance, a consumer-

equipment manufacturer could use blockchain can optimize a process, reduce costs, and improve
to track and manage a few SKUs for select end efficiency. These companies will be the most willing
consumers, rather than its entire customer base. to participate in private, permissioned networks in
order to restrict access to sensitive information, such
—— Market mechanisms for ensuring collaboration. as pricing data, to select groups or individuals.
Several blockchain use cases, such as those
for tracking goods through supply chains, will BaaS providers typically offer their platforms for free
require players to share data and participate and then charge customers for each node deployed.
in a common blockchain platform. Initially, few This pricing strategy could help industrial players,
companies may be willing to engage in such since companies generally deploy few nodes during
collaborations. In some specific cases where early implementation. Since industrial companies’
companies have the market power, either financial risks are lower, they may be enticed to
because of their size or position, they will be embark on more blockchain projects, even though
more likely to have other players participate and they are uncertain about the potential returns.
obtain value from blockchain solutions.
How semiconductor players can create

By concentrating on use cases with these value in Blockchain 2.0: Core beliefs
characteristics, industrial companies will prioritize Semiconductor companies have found many
those that are most likely to provide a suitable opportunities in blockchain since its inception.
return on investment. As blockchain technology That will still be the case in the Blockchain 2.0
progresses and the cost of application development era, but we anticipate some important changes as
falls, they may investigate additional use cases. the cryptocurrency sector evolves and business
applications potentially become the primary
Belief 3: Blockchain 2.0 will take off in sources of chip demand. So, what trends must
private, permissioned networks within the semiconductor players understand to succeed? And
industrial ecosystem who will win in this new era, for both cryptocurrency
Unlike cryptocurrency transactions, industrial and blockchain business applications? After
business applications will occur over private analyzing the hardware market, we identified four
blockchains that limit access to invited participants, beliefs about value creation and capture by silicon
rather than over public blockchains. Some of companies during Blockchain 2.0.
these blockchains will have central administrators
to determine which nodes have permission to Belief 1: Value for silicon players will migrate away
access, edit, and validate data. Along with providing from cryptocurrencies (and therefore compute
greater confidentiality, these private, permissioned power) in the near future
networks are the most technically feasible, given that Until blockchain business applications gain traction
blockchain speed decreases and latency increases and demonstrate a positive return on investment—
as more nodes are added. something that is not expected to occur for at least
two to three years—semiconductor companies
For industrial companies, the first private, should continue to focus on cryptocurrency
permissioned blockchains will focus on specific customers. In particular, they should try to optimize
“microverticals”—groups of related tasks—such compute power and minimize power consumption to
as supply-chain management. Within such micro- satisfy the large mining pools that rely on crypto rigs.
verticals, participants are more likely to identify a Recently, BitMain Technologies made an important
common problem that they want to solve through advance in this area by developing a seven-
blockchain and recognize the return on investment. nanometer node miner.
They are also more willing to share implementation
costs, since they can easily see blockchain’s value. A long-term focus on compute power isn’t the
For example, leaders at industrial companies and best strategy, however, since many altcoins are
the vendors that serve them will all benefit if they considering moving from PoW to PoS systems,
What advantages do blockchain business applications offer?
Think of blockchain as a database in real time through the blockchain fails, the information will still be available
shared across a number of participants, network. Companies doing business elsewhere. Another advantage involves
each with a computer. At any moment, with each other must thus store most of the audit trail. Users can go back through
each member of the blockchain holds an their transactional information in digital the blocks of information and easily see
identical copy of the blockchain database, form to take advantage of blockchain. the information previously recorded in the
giving all participants access to the same database, such as the previous owner of
information. All blockchains share three —— A public or private network that a piece of property. And perhaps most
characteristics: enables sharing. Anyone can join or important, blockchain maintains process
leave a public network without express integrity. The database can only be updated
—— A cryptographically secure database. permission. Admission into private when two things happen. First, a user must
When data are read or written, users networks is by invitation only. provide the correct public and private keys.
must provide the correct cryptographic Second, a majority of participants in the
keys—one public (essentially, the Blockchain’s cryptographic keys provide network must verify those credentials. This
address) and one private. Users cannot leading-edge security that goes far reduces the risk that a malicious user will
update the blockchain unless they have beyond that found in a standard distributed gain illicit access to the network and make
the correct keys. ledger. The technology also eliminates the unauthorized updates.
possibility that a single point of failure will
—— A digital log of transactions. emerge since the blockchain database is
Transactional information is available distributed and decentralized. If one node
in which compute power is less important. For companies will need to develop new strategies
blockchain business applications, which could that align with their customers’ priorities. To do so
represent the wave of the future, compute power effectively, they must ask themselves four questions:
is essential but not a differentiator. Instead,
semiconductor companies and other players will win —— In which specific use cases and microverticals
by enabling or providing BaaS. are customers likely to adopt a blockchain
solution at scale?
Belief 2: To win in Blockchain 2.0, semiconductor
companies can’t just understand their —— Which customers or end markets have the
customers—they also have to understand their market position and structure to ensure
customers’ customers that all relevant companies will be willing to
Cryptocurrency ASICs have been in extremely collaborate?
high demand since 2016, because miners began
getting higher rewards for adding the next block. —— How do end customers plan to use blockchain
Most orders come from the top five Bitcoin mining and what aspects of our hardware—for
pools in China, and the demand could increase instance, cost, compute capability, or power
over the next few years. This trend will keep orders consumption—will differentiate the winners
flowing into substrates, ASIC designers, foundries, from the losers?
outsourced assembly and testing companies, and
equipment manufacturers. —— How can we work with (or without) BaaS players,
including those who provide other hardware
With value migrating from cryptocurrencies to components, software integration, or go-to-
blockchain business applications, and with BaaS market capabilities, to enable end-to-end
players gaining market share, semiconductor solutions for customers?

Belief 3: As value migrates away from hardware, and Ebang Communication, are now in strong
semiconductor companies must go ‘up the stack’ positions. But they might not be the long-term
Within the current BaaS technology stack, value winners, despite their first-mover advantage.
predominantly lies within the lowest layer: hardware. The barriers to market entry are low, since new
But over the next several years, as blockchain companies with domain expertise can easily design
business applications start to gain a foothold within ASICs, and some well-known companies are already
large industries, demand will increase for hardware planning to move into the market.
customized for specific use cases or microverticals.
This development will cause value to migrate up the If the new players can differentiate themselves
technology stack from hardware to other layers. based on product performance or price, they may
dethrone the current market leaders. Companies
Given these trends, semiconductor companies with strong end-to-end BaaS offerings may lead
should consider enabling or providing the the pack, while those that continue to focus on
entire BaaS technology stack for specific hardware alone may find themselves sidelined.
microverticals or use cases. After developing a
clear understanding of how customers plan to use
their blockchain chips, semiconductor companies
could then provide platforms and plug-ins that help If blockchain were a tool, it would be a Swiss Army
integrate the layers of the blockchain technology knife that has a blade, a screwdriver, a can opener,
stack, allowing for easier implementation. A and many other attachments—a clever technology
combined offering would meet all customer that enables a diverse set of use cases that go
needs for blockchain, just as TensorFlow does for far beyond cryptocurrency. But like a Swiss Army
machine learning and deep learning. knife, blockchain can be unexpectedly complicated.
Industrial companies must know what networks
This strategy will become even more important as and transactions are most likely to benefit their
the use cases and microverticals start to mature, business. They must also understand which use
since hardware will become a commodity. Those cases have features that are most likely to deliver
semiconductor providers that don’t move “up value at scale—for instance, characteristics that
the stack” will have an increasingly difficult time encourage other participants to join the blockchain
capturing value and thriving. In fact, they could and collaborate. Likewise, semiconductor
find themselves in the same situation they face companies must understand how blockchain is
in the data-center market, where “hyperscalers” being applied in the cryptocurrency market and
have a great deal of control because of their the business sphere and closely follow market
purchasing power. developments in both areas. With blockchain
evolving so rapidly, it can be difficult to keep pace
Belief 4: The semiconductor companies that with change. But those semiconductor companies
were leaders in Blockchain 1.0 are not preordained and industrials that pursue innovation while
to be future winners aggressively enabling blockchain use cases are
Today’s top blockchain hardware providers, likely to reap the greatest rewards.
including BitMain Technologies, Canaan Creative,
Gaurav Batra is a partner in McKinsey’s Washington, DC, office; Rémy Olson is an alumnus of the San Francisco office; Shilpi
Pathak is an alumnus of in the Chicago office; Nick Santhanam is a senior partner in the Silicon Valley office; and Harish
Soundararajan is an alumnus of the Boston office.
The authors wish to thank Jo Kakarwada and Celine Shan for their contributions to this article.
Rethinking car
software and
electronics architecture
As the car continues its transition from a hardware-driven machine to
a software-driven electronics device, the auto industry’s competitive
rules are being rewritten.
by Ondrej Burkacky, Johannes Deichmann, Georg Doll, and Christian Knochenhauer
© Just_Super/Getty Images

The engine was the technology and engineering electronic-system companies are boldly entering
core of the 20th-century automobile. Today, the tech giants’ original feature-and-app turf,
software, large computing power, and advanced and premium automakers are moving into areas
sensors have increasingly stepped into that role; they further down the stack–such as operating systems,
enable most modern innovations, from efficiency to hardware abstractions, and signal processing–
connectivity to autonomous driving to electrification in order to protect the essence of their technical
and new mobility solutions. distinction and differentiation.
However, as the importance of electronics and One consequence of these strategic moves is that
software has grown, so has complexity. Take the the vehicle architecture will become a service-
exploding number of software lines of code (SLOC) oriented architecture (SOA) based on generalized
contained in modern cars as an example. In 2010, computing platforms. Developers will add new
some vehicles had about ten million SLOC; by 2016, connectivity solutions, applications, artificial-
this expanded by a factor of 15, to roughly 150 million intelligence elements, advanced analytics, and
lines. Snowballing complexity is causing significant operating systems. The differentiation will not be in
software-related quality issues, as evidenced by the traditional vehicle hardware anymore but in the
millions of recent vehicle recalls. user-interface and experience elements powered
by software and advanced electronics.
With cars positioned to offer increasing levels
of autonomy, automotive players see the quality Tomorrow’s cars will shift to a platform of new
and security of vehicle software and electronics brand differentiators (Exhibit 2). These will likely
as key requirements to guarantee safety. include infotainment innovations, autonomous-
This means the industry must rethink today’s driving capabilities, and intelligent safety features
approaches to vehicle software and electrical and based on “fail operational” behaviors (for example,
electronic architecture. a system capable of completing its key function
even if part of it fails). Software will move further
down the digital stack to integrate with hardware
Addressing an urgent industry concern in the form of smart sensors. Stacks will become
As the automotive industry transitions from horizontally integrated and gain new layers that
hardware- to software-defined vehicles, the average transition the architecture into an SOA.
software and electronics content per vehicle is
rapidly increasing. Software represents 10 percent Ultimately, the new software and electronic
of overall vehicle content today for a D-segment architecture will come from several game-changing
(large) car (approximately $1,220), and the average trends that drive complexity and interdependencies.
share of software is expected to grow at a compound For example, new smart sensors and applications
annual rate of 11 percent, to reach 30 percent of will create a “data explosion” in the vehicle that
overall vehicle content (around $5,200) in 2030. Not companies need to handle by processing and
surprisingly, companies across the digital automotive analyzing the data efficiently, if they hope to remain
value chain are attempting to capitalize on inno- competitive. A modularized SOA and over-the-air
vations enabled through software and electronics (OTA) updates will become key requirements to
(Exhibit 1). Software companies and other digital- maintain complex software in fleets and enable new
technology companies are leaving their current tier- function-on-demand business models. Infotainment
two and tier-three positions to engage automakers and, to a lesser degree, advanced driver-assistance
as tier-one suppliers. They’re expanding their systems (ADAS) will increasingly become “appified”
participation in the automotive technology stack as more third-party app developers provide vehicle
by moving beyond features and apps into operating content. Digital-security requirements will shift the
systems. At the same time, traditional tier-one focus from a pure access-control strategy to an
Rethinking car software and electronics architecture 39

Rethinking car software
Exhibit 1 of 3
Exhibit 1
Software enables critical automotive innovations.

Software innovation examples
Connectivity Autonomous driving

• Integration of 3rd-party services • Rise of built-in sensors and actuators
• Updates over the air to deploy new features faster • Higher demand for computing power and
• Operation of future cars partly in the cloud communication
• Unlimited need for reliability
Innovation
through
software
Electrification Diverse mobility

• Introduction of new electronics • Shared-mobility services and robo-taxis via app
• Reduction of energy consumption through • Customized driver experience
advanced software algorithms
Source: Automotive Electronics Initiative; Robert N. Charette, “This car runs on code,” IEEE Spectrum, February 2009, spectrum.ieee.org; HAWK; McKinsey analysis
integrated-security concept designed to anticipate, developments are already under way and will hit the
avoid, detect, and defend against cyberattacks. The market in two to three years’ time. This consolidation
advent of highly automated driving (HAD) capabilities is especially likely for stacks related to ADAS and
will require functionality convergence, superior HAD functionality, while more basic vehicle functions
computing power, and a high degree of integration. might keep a higher degree of decentralization.
In the evolution toward autonomous driving,

Exploring ten hypotheses on future virtualization of software functionality and
electrical or electronic architecture abstraction from hardware will become even more
The path forward for both the technology and the imperative. This new approach could materialize
business model is far from fixed. But based on our in several forms. One scenario is a consolidation
extensive research and insights from experts, we of hardware into stacks serving different
developed ten hypotheses regarding tomorrow’s requirements on latency and reliability, such as
automotive electrical or electronic architecture and a high-performance stack supporting HAD and
its implications for the industry. ADAS functionality and a separate, time-driven,
low-latency stack for basic safety features. In
There will be an increasing consolidation of another scenario, the ECU is replaced with one
electronic control units (ECUs) redundant “supercomputer,” while in a third, the
Instead of a multitude of specific ECUs for specific control-unit concept is abandoned altogether in
functionalities (the current “add a feature, add a favor of a smart-node computing network.
box” model), the industry will move to a consolidated
vehicle ECU architecture. The change is driven primarily by three factors:
costs, new market entrants, and demand through
In the first step, most functionality will be centered HAD. Decreasing costs, both for the development
on consolidated domain controllers for the main of features as well as the required computing
vehicle domains that will partially replace functionality hardware, including communication hardware, will
currently running in distributed ECUs. These accelerate the consolidation. So too will new market

Exhibit 2 of 3
Exhibit 2
Architecture will become service oriented, with new factors for differentiation.
Future layered in-vehicle and back-end architecture
Existing layer Modified layer New layer
Combine in-vehicle data with Future factors for brand differentiation:

Cloud platform
environmental data
• Infotainment features requiring “plug and play”
capabilities
Connectivity (back-haul)
• Autonomous capabilities including sensor-fusion
User interface/user experience/ algorithms as a complement to hardware
human–machine interface
• Safety features based on “fail operational”² behavior
Significant increase in number
Applications • Software will move further down stack to
of applications
hardware (smart sensors)
Artificial intelligence/ Analyze data for real-time
advanced analytics decisions and autonomous driving • Stacks become horizontally integrated
Middleware layer/ Abstract applications • New layers will be added to stack

operating system from hardware
Electronic/electrical
hardware1
Closely controlled add-on app
Power
Sensors Actuators and modules due to safety
components
considerations
Vehicle
¹ Including operating system in status quo.

² For example, a system capable of completing its key function even if part of it fails.
entrants into automotive that will likely disrupt the and embedded firmware (including the operating
industry through a software-oriented approach to system) will depend on key nonvehicle functional
vehicle architecture. Increasing demand for HAD requirements instead of being allocated part of a
features and redundancy will also require a higher vehicle functional domain. To allow for separation
degree of consolidation of ECUs. and a service-oriented architecture, the following
four stacks could become the basis for upcoming
Several premium automakers and their suppliers are generations of cars in five to ten years:
already active in ECU consolidation, making early
moves to upgrade their electronic architecture, —— Time-driven stack. In this domain, the controller
although no clear industry archetype has emerged is directly connected to a sensor or actuator
at this point. while the systems have to support hard real-time
requirements and low latency times; resource
The industry will limit the number of stacks used scheduling is time based. This stack includes
with specific hardware systems that reach the highest Automotive
Accompanying the consolidation will be a Safety Integrity Level classes, such as the
normalization of limited stacks that will enable a classical Automotive Open System Architecture
separation of vehicle functions and ECU hardware (AUTOSAR) domain.
that includes increased virtualization. Hardware

—— Event- and time-driven stack. This hybrid stack it possible to reconfigure cars and enable the
combines high-performance safety applications, installation and upgrade of their software. Unlike
for example, by supporting ADAS and HAD today, where middleware within each ECU facilitates
capability. Applications and peripherals are communication across units, in the next vehicle
separated by the operating system, while generation it will link the domain controller to access
applications are scheduled on a time base. functions. Operating on top of ECU hardware in the
Inside an application, scheduling of resources car, the middleware layer will enable abstraction and
can be based on time or priority. The operating virtualization, an SOA, and distributed computing.
environment ensures that safety-critical
applications run on isolated containers with clear Evidence already suggests automotive players
separation from other applications within the car. are moving toward more flexible architectures,
A current example is adaptive AUTOSAR. including an overarching middleware. AUTOSAR’s
adaptive platform, for example, is a dynamic system
—— Event-driven stack. This stack centers on the that includes middleware, support for a complex
infotainment system, which is not safety critical. operating system, and state-of-the-art multicore
The applications are clearly separated from the microprocessors. However, current developments
peripherals, and resources are scheduled using appear restricted to a single ECU.
best-effort or event-based scheduling. The
stack contains visible and highly used functions In the middle term, the number of onboard
that allow the user to interact with the vehicle, sensors will spike significantly
such as Android, Automotive Grade Linux, In the next two to three vehicle generations,
GENIVI, and QNX. automakers will install sensors with similar
functionalities to ensure that sufficient safety-
—— Cloud-based (off-board) stack. The final stack related redundancies exist (Exhibit 3). In the long
covers and coordinates access to car data term, however, the automotive industry will develop
and functions from outside the car. The stack specific sensor solutions to reduce the number
is responsible for communication, as well as of sensors used and their costs. We believe that
safety and security checks of applications a combined solution of radar and camera might
(authentication), and it establishes a defined car be dominant for the next five to eight years. As
interface, including remote diagnostics. autonomous-driving capabilities continue to rise,
the introduction of lidars will be necessary to
ensure redundancy for both object analysis and
Automotive suppliers and technology players have localization. Configurations for SAE International
already begun to specialize in some of these stacks. L4 (high automation) autonomous driving, for
Notable examples are in infotainment (event- example, will likely initially require four to five lidar
driven stack), where companies are developing sensors, including rear-mounted ones for city
communications capabilities such as 3-D and operation and near-360-degree visibility.
augmented navigation. A second example is artificial
intelligence and sensing for high-performance In the long term, we see different possible scenarios
applications, where suppliers are joining with key concerning the number of sensors in vehicles:
automakers to develop computing platforms. further increase, stable numbers, or decrease. Which
scenario will come to pass depends on regulation,
In the time-driven domain, AUTOSAR and JASPAR the technical maturity of solutions, and the ability
are supporting the standardization of these stacks. to use multiple sensors for different use cases.
Regulatory requirements might, for example, enforce
An expanded middleware layer will abstract closer driver monitoring, resulting in an increase
applications from hardware of sensors inside the vehicle. It can be expected
As vehicles continue to evolve into mobile that more consumer-electronics sensors will be
computing platforms, middleware will make used in the automotive interior. Motion sensors and

Exhibit 3 of 3
Exhibit 3
Sensor fusion will provide redundancy for autonomous functions.

Sensor-function ratings
Good Fair Poor

Ultra- Radar + Lidar + Radar +
Camera Radar Lidar sonic lidar camera camera
Object detection
Object classification
Distance estimation
Object-edge precision
Lane tracking
Range of visibility
Functionality in bad weather
Functionality in poor lighting
Cost
Production readiness
Radar and camera most likely combination in next 5–8 years, although solid-state lidar and camera¹ will be dominant in the long term
when proved and integrated into mass-production designs
¹ Comparison with other technologies not yet possible due to low maturity of technology.
health monitoring of measures such as heart rate Sensors will become more intelligent
and drowsiness, as well as face recognition and iris System architectures will require intelligent and
tracking, are just a few of the potential use cases. integrated sensors to manage the massive amounts
However, as an increase or even a stable number of of data needed for highly automated driving. While
sensors would require a higher bill of materials, not high-level functions such as sensor fusion and
only in the sensors themselves but also in the vehicle 3-D positioning will run on centralized computing
network, the incentive to reduce the number of platforms, preprocessing, filtering, and fast reaction
sensors is high. With the arrival of highly automated cycles will most likely reside in the edge or be
or fully automated vehicles, future advanced done directly in the sensor. One estimate puts the
algorithms and machine learning can enhance amount of data an autonomous car will generate
sensor performance and reliability. Combined with every hour at four terabytes. Consequently,
more powerful and capable sensor technologies, a intelligence will move from ECUs into sensors to
decrease of redundant sensors can be expected. conduct basic preprocessing requiring low latency
Sensors used today might become obsolete as their and low computing performance, especially if
functions are overtaken by more capable sensors (for weighting costs for data processing in the sensors
instance, a camera- or lidar-based parking assistant against costs for high-volume data transmission
could replace ultrasound sensors). in the vehicle. Redundancy for driving decisions

in HAD will nevertheless require a convergence likely to be replaced by automotive Ethernet and its
for centralized computing, likely based on extensions, AVB and TSN.
preprocessed data. Intelligent sensors will supervise
their own functionality while redundancy of sensors Going forward, we expect the automotive industry
will increase reliability, availability, and hence safety to also embrace future Ethernet technologies such
of the sensor network. To ensure correct sensor as high-delay bandwidth products (HDBP) and
operation in all conditions, a new class of sensor- 10-gigabit technologies.
cleaning applications—such as deicing capabilities
and those for dust or mud removal—will be required. OEMs will always tightly control data connectivity
for functional safety and HAD but will open
Full power and data-network redundancy will interfaces for third parties to access data
be necessary Central connectivity gateways transmitting and
Safety-critical and other key applications that require receiving safety-critical data will always connect
high reliability will utilize fully redundant circles for directly and exclusively to an OEM back end,
everything that is vital to safe maneuvering, such as available to third parties for data access, except
data transmission and power supply. The introduction where obliged by regulation. In infotainment,
of electric-vehicle technologies, central computers, however, driven by the “appification” of the vehicle,
and power-hungry distributed computing networks emerging open interfaces will allow content and app
will require new redundant power-management providers to deploy content, while OEMs will keep
networks. Fail-operational systems to support the respective standards as tight as possible.
steer-by-wire and other HAD functions will require
redundancy system designs, which is a significant Today’s on-board diagnostics port will be replaced
architectural improvement on today’s fail-safe with connected telematic solutions. Physical
monitoring implementations. maintenance access to the vehicle network will
not be required anymore but can go through the
The ‘automotive Ethernet’ will rise and become OEMs’ back ends. OEMs will provide data ports in
the backbone of the car their vehicle back end for specific use cases such
Today’s vehicle networks are insufficient for the as lost-vehicle tracking or individualized insurance.
requirements of future vehicles. Increased data Aftermarket devices, however, will have less and less
rates and redundancy requirements for HAD, safety access to vehicle internal data networks.
and security in connected environments, and the
need for interindustry standardized protocols Large fleet operators will play a stronger role in
will most likely result in the emergence of the the user experience and will create value for end
automotive Ethernet as a key enabler, especially for customers, for example, by offering different vehicles
the redundant central data bus. Ethernet solutions for different purposes under one subscription (such
will be required to ensure reliable interdomain as weekend or daily commute). This will require
communication and satisfy real-time requirements them to utilize the different OEMs’ back ends and
by adding Ethernet extensions like audio-video start consolidating data across their fleets. Larger
bridging (AVB) and time-sensitive networks (TSN). databases will then allow fleet operators to monetize
Industry players and the OPEN Alliance support consolidated data and analytics not available on the
the adoption of Ethernet technology, and many OEM level.
automakers have already made this leap.
Cars will use the cloud to combine onboard
Traditional networks such as local interconnected information with offboard data
networks and controller area networks will continue Nonsensitive data (that is, data that are not personal
to be used in the vehicle, but only for closed lower- or safety related) will increasingly be processed
level networks, for instance, in the sensor and actor in the cloud to derive additional insights, though
area. Technologies such as FlexRay and MOST are availability to players beyond OEMs will depend on

future regulation and negotiation. As the volumes mission critical, we can expect OEMs to take more
of data grow, data analytics will become critically ownership in this market segment.
important for processing the information and turning
it into actionable insights. The effectiveness of using Vehicles will receive software and feature upgrades
data in such a way to enable autonomous driving as well as security updates for the designed life
and other digital innovations will depend on data span. Regulators will likely enforce software
sharing among multiple players. It’s still unclear how maintenance to ensure the safety integrity of the
this will be done and by whom, but major traditional vehicle designs. The obligation to update and
suppliers and technology players are already maintain software will lead to new business models
building integrated automotive platforms capable of for maintenance and operations of vehicles.
handling this new plethora of data.
Cars will feature updatable components that Assessing the future implications
communicate bidirectionally of vehicle software and electronic
Onboard test systems will allow cars to check architecture
function and integration updates automatically, While the trends affecting the automotive industry
thus enabling life-cycle management and the today are generating major hardware-related
enhancement or unlocking of aftersales features. All uncertainties, the future looks no less disruptive
ECUs will send and receive data to and from sensors for software and electronic architecture. Many
and actuators, retrieving data sets to support strategic moves are possible: automakers could
innovative use cases such as route calculation create industry consortia to standardize vehicle
based on vehicle parameters. architecture, digital giants could introduce onboard
cloud platforms, mobility players could produce their
OTA update capabilities are a prerequisite for own vehicles or develop open-source vehicle stacks
HAD; they will also enable new features, ensure and software functions, and automakers could
cybersecurity, and enable automakers to deploy introduce increasingly sophisticated connected and
features and software quicker. In fact, it’s the OTA autonomous cars.
update capability that is the driver behind many
of the significant changes in vehicle architecture The transition from hardware-centric products
described previously. In addition, this capability also to a software-oriented, service-driven world is
requires an end-to-end security solution across all especially challenging for traditional automotive
layers of the stack outside the vehicle to the ECUs companies. Yet, given the described trends and
in the vehicle. This security solution remains to be changes, there is no choice for anyone in the
designed, and it will be interesting to see how and by industry but to prepare. We see several major
whom this will be done. strategic pushes:
To achieve smartphone-like upgradability, the —— Decouple vehicle and vehicle-function

industry needs to overcome restrictive dealer development cycles. OEMs and tier-one
contracts, regulatory requirements, and security suppliers need to identify how to develop, offer,
and privacy concerns. Here too, a variety of and deploy features largely apart from vehicle-
automotive players have announced plans to development cycles, both from a technical
deploy OTA service offerings, including over-the- and organizational perspective. Given current
air updates for their vehicles. vehicle-development cycles, companies need
to find a way to manage innovations in software.
OEMs will standardize their fleets on OTA Further, they should think about options to
platforms, working closely with technology create retrofitting and upgrade solutions (for
providers in this space. As vehicle connectivity example, computing units) for existing fleets.
and OTA platforms will become increasingly

—— Define the target value add for software and architecture asks for potentially breaking up
electronics development. OEMs must identify the current “vertical” setup and introducing
the differentiating features for which they are new “horizontal” organizational units. Further,
able to establish control points. In addition, it is they need to ramp up dedicated capabilities
crucial to clearly define the target value add for and skills for their own software and electronics
their own software and electronics development development teams.
and to identify areas that become a commodity
or topics that can only be delivered with a —— Design a business model around automotive
supplier or partner. features as a product (especially for automotive
suppliers). To remain competitive and capture
—— Attach a clear price tag to software. Separating a fair share of value in the field of automotive
software from hardware requires OEMs to electronics, it is crucial to analyze which features
rethink their internal processes and mechanisms add real value to the future architecture and
for buying software independently. In addition therefore can be monetized. Subsequently,
to the traditional setup, it is also important to players need to derive new business models for
analyze how an agile approach to software the sale of software and electronics systems,
development can be anchored in procurement be it as a product, a service, or something
processes. Here suppliers (tier one, tier two, and completely new.
tier three) also play a crucial role, as they need
to attach a clear business value to their software
and system offerings to enable them to capture a
larger revenue share. As the new era of automotive software and
electronics begins, it’s drastically changing a
—— Design a specific organizational setup around wide variety of prior industry certainties about
new electronics architecture (including related business models, customer needs, and the nature
back ends). Next to changing internal processes of competition. We are optimistic about the revenue
in order to deliver and sell advanced electronics and profit pools that will be created. But to benefit
and software, automotive players—both OEMs from the shifts, all players in the industry need to
and suppliers—should also consider a different rethink and carefully position (or reposition) their
organizational setup for vehicle-related value propositions in the new environment.
electronics topics. Mainly, the new “layered”
This article was developed in collaboration with the Global Semiconductor Alliance (GSA).
Ondrej Burkacky is a partner in McKinsey’s Munich office, where Georg Doll is a senior expert; Johannes Deichmann is an
associate partner in the Stuttgart office; and Christian Knochenhauer is a senior expert in the Berlin office.
The authors wish to thank Silviu Apostu, Michaela Brandl, and Virginia Herbst for their contributions to this article. They
also wish to thank executives from GSA member companies and others who participated in the interviews and survey that
contributed to this article.

How will changes in
the automotive-component
market affect semi-
conductor companies?
The rise of domain control units (DCUs) will open new opportunities
for semiconductor companies.
Ondrej Burkacky, Johannes Deichmann, and Jan Paul Stein
© Yuhan Liao/Getty Images
How will changes in the automotive-component market affect semiconductor companies? 47

The automotive industry will change more in the sector and adjacent industries. The global market
next decade than it has in the past century. The for software and E/E components is expected
shake-up stems from four mutually reinforcing to grow about 7 percent annually through 2030,
trends that are rapidly gaining traction: autonomous although results will vary by segment. That’s
driving, connected cars, electrification of vehicles, more than double the rate of 3 percent for the
McKinsey on Semiconductors
and shared mobility. All these trends have one automotive sector as a whole (exhibit).
commonmarket
Automotive component enabler: advances in automotive software
Exhibit 1 of 1 and electrical/electronic (E/E) components. As the trends accelerate, automotive systems
will change significantly, especially with respect
These developments are generally good news for to control-unit architecture. Currently, vehicles
semiconductor companies serving the automotive rely on a decentralized architecture in which each
Exhibit
The global market for automotive components is expected to grow about 7 percent annually
through 2030.
Automotive software and electrical/electronic market, $ billion CAGR¹ 2020–30, %
~7 Overall
469
~3
85 Other electronic components2
~10
362 34 Integration, verification, and validation services
76 50 Software3 ~9
25 63 Sensors
238 37 ~8
63 44 81 Power electronics (excluding battery cells)

13 ~15
20 50
30
20
156 ECUs/DCUs4
129
92
~5
2020 2025 2030
3,800
2,755 3,027
Automotive sales,
$ billion ~3
2020 2025 2030
¹ Compound annual growth rate.

² Harnesses, controls, switches, displays.
³ Functions, operating systems, middleware.
⁴ Electronic control units/domain control units.
Source: IHS; McKinsey analysis

individual function, such as parking assistance, In the new automotive age, OEMs will less often
runs on a separate electronic control unit (ECU). follow the traditional sourcing approach in which
These functions are typically “hard coded” in ECU they either rely on tier-one vendors for guidance
hardware that includes embedded software in its or else define specifications and expect suppliers
design and configuration. to deliver on them. Instead, they will depend much
more on tech natives, including semiconductor
Future generations of cars will have a centralized players, for insights about the best technologies
architecture in which a few domain control units and architectures. To succeed, semiconductor
(DCUs) control multiple functions. For instance, companies must have more direct discussions with
one DCU may cover all functions in advanced OEMs about their needs, rather than solely relying
driver-assistance systems, including parking on reports from tier-one suppliers. Without this
assistance and blind-spot detection. DCUs have understanding, they could invest in technologies
less hard coding than ECUs, so software will take that commoditize quickly or get fully translated into
the lead. If an OEM wants to add another function software based on standard DCU hardware.
to a DCU, it can likely add software, rather than
creating new hardware. With this shift, it will no Semiconductor companies must also monitor market
longer be necessary to develop or source hardware trends and place their bets wisely—especially if
and software in tandem. they want to expand from hardware provisioning.
While the changes in E/E architecture offer several
As centralized architecture gains traction, DCUs opportunities to expand into software, many OEMs
will increase their share of the automotive- are still debating their future sourcing strategies.
controller market from about 2 percent to around Their decisions, including those related to whether
40 percent between 2020 and 2030. ECUs will still they should purchase software or create it internally,
be necessary, especially for lower-level functions, could determine the extent of the opportunities
such as pre-processing of sensor data for cameras, available to semiconductor companies.
or for functions where latency is critical. But
ECUs will become increasingly standardized and
commoditized as vehicles transition to software-
defined functions, as will sensors, harnesses, and
other hardware components.
Ondrej Burkacky is a partner in McKinsey’s Munich office, where Jan Paul Stein is a consultant. Johannes Deichmann is an
associate partner in the Stuttgart office.
How will changes in the automotive-component market affect semiconductor companies? 49

Right product, right
time, right location:
Quantifying the semi-
conductor supply chain
Problems along the semiconductor supply chain are difficult
to diagnose. A new metric can help companies pinpoint
performance issues.
by Gaurav Batra, Kristian Nolde, Nick Santhanam, and Rutger Vrijen
© TimeStopper/Getty Images

The semiconductor supply chain stretches performance. With the insights that the RPRTRL
from fabs to back-end factories, with the intricate measurement provides, companies can, for the first
process of chip manufacturing sometimes requiring time, identify all root causes behind performance
four to six months to complete. At the end of issues, develop an improvement plan, and quantify
the line, some of the world’s leading companies their progress.
are waiting for the semiconductors required to
launch their latest innovations. Any delays could
alienate distributors and end customers, placing a On-time delivery—a priority for
semiconductor company on an unofficial blacklist. customers
So why are late shipments so common? When it comes to customer retention, supply-
chain performance matters. That much became
Most players can’t answer this question. clear when we asked managers at six major
Although they’re aware that their supply chains semiconductor customers and distributors to rate
are suboptimal, they generally look at different the factors that influenced their purchase decisions
outcomes in isolation, including the portion of on a scale of one to ten, with ten being the most
on-time deliveries (OTDs), overall cycle times, fill influential. Product specifications, which include
rates, excessive days of inventory, or the number quality and features, ranked first at 9.7, but OTD tied
of orders canceled because of delays. The reasons price for second, at 8.2 (Exhibit 1).
behind their poor performance receive much less
scrutiny, partly because it’s difficult to pinpoint Interview subjects frequently noted that they gave
when and where problems occurred along the preference to semiconductor companies with a
lengthy and complex supply chain. And that means strong OTD record. One said, “For suppliers with
the same mistakes get repeated each time a good delivery performance, we invest more, as
company gets a new order. we feel more comfortable that we can deliver the
products to our own customers.” Another remarked,
A new and comprehensive metric can provide “If a supplier consistently can’t meet delivery dates,
detailed insights into the end-to-end performance we will stock them in reaction to customer orders
of the supply chain. For each order, it asks several but not actively push their sales.”
questions: Were demand forecasts accurate,
allowing companies to deliver the right product Our analysis of one semiconductor company
(RP)? Did execution occur on schedule, allowing revealed the dire consequences of late deliveries.
all tasks to be completed at the right time (RT)? For customers at which the OTD rate was between
Was inventory staged along the supply chain at 0 percent and 40 percent, the semiconductor
the right locations (RL)? This metric—abbreviated company’s revenue dropped 28 percent within one
as RPRTRL—is calculated based on hard data, year. When the semiconductor company’s OTD rate
resulting in an objective assessment of supply-chain was 80 percent or higher, its revenue declined only
A new and comprehensive metric

can provide detailed insights
into the end-to-end performance
of the supply chain.
Right product, right time, right location: Quantifying the semiconductor supply chain 51
McKinsey Semiconductors 2019
Right product semiconductor
Exhibit 1 of 5
Exhibit 1
On-time delivery is an important consideration in buying decisions made by semiconductor

customers and distributors.
Importance to buying decision, score out of 10, n = 6
Product On-time Customer

specifications delivery Price service Terms
9.7
8.2 8.2
5.8
5.3
2 percent over the same period. These findings Most missteps that lead to late deliveries relate
suggest that supply-chain inefficiencies are a major to one of three areas: forecasting, execution, and
cause of customer churn. inventory (Exhibit 2). For instance, if the order lead
time is shorter than the three to four weeks required
What’s behind the low OTD rates? The root causes for back-end processing, a semiconductor company
are as complex as the supply chain itself. When must have sufficient finished-goods inventory to
semiconductor companies receive an order, they meet the target delivery date. But many players
have chips at every stage of the supply chain, with inaccurately forecast future demand and don’t
some undergoing front-end processing, others have enough finished goods in stock when such
in die-bank or back-end processing, and the requests arrive.
remainder sitting in warehouses as finished goods.
Likewise, the lead times for orders may vary, with
some customers expecting quick shipments and A comprehensive metric for assessing
others requesting deliveries along a more relaxed supply-chain performance
timeline. All too often, however, semiconductor The three elements of the RPRTRL metric allow
companies discover that the requested lead time is companies to quantify their performance in
shorter than the cycle time needed to fulfill the order. forecasting, execution, and inventory management

Exhibit 2 of 5
Exhibit 2
The ability to meet target delivery dates depends on order lead time, inventory along the supply
chain, and other factors.
Supply-chain stages Raw Front-end Die Back-end Finished goods and

to fulfill an order material cycle time bank cycle time shipping preparation
Order lead time Shorter than full cycle time Longer than back-end cycle time Shorter than back-end
but less than full cycle time cycle time
Fulfillment Order can be produced from raw Order can be fulfilled from die Order must be fulfilled
requirement materials, if: bank, if: from finished-goods
• front-end planning and execution • sufficient die-bank inventory inventory, since there is
happen in line with anticipated exists at order time no time to produce from
front-end cycle times die bank
• planning and execution happen
• there are no front-end capacity in line with anticipated
constraints back-end cycle times
Possible root cause • Inaccurate front-end planning, • Testing-execution delays • Low or nonexistent
of on-time- including forecasting • Assembly-execution delays inventory, often
delivery gaps • Fabrication-execution delays resulting from poor
• Inaccurate back-end planning, forecasting
• Fabrication-capacity constraints resulting in insufficient dies
planned for assembly
• Insufficient die-bank inventory
(Exhibit 3). Companies must evaluate these methods to determine if they are making decisions
elements for every product ordered, to ensure based on insufficient or flawed information. For
that the overall metric reflects the most up-to- instance, companies may only look at past-
date information. order data to forecast demand, even if they have
other information that provides important clues
Right product about future trends, such as customer financial
If companies can’t predict when products will be statements, the number of web-page views for
needed, it doesn’t matter whether the rest of their certain product parts, and data-sheet downloads
supply chain is efficient. They simply won’t be able for different products on their website. Some
to fulfill orders, or they’ll have excessive inventory companies also encounter problems because
because they make more products than they they use the same forecasting model for all SKUs,
need. The right-product component of RPRTRL which can lead to inaccuracies. If a product has
measures how companies perform in this area by intermittent spikes in demand, it needs a different
calculating the extent of a company’s forecasting model than does a product with low but
bias (the arithmetic mean of a forecasting error) and steady demand.
the magnitude of the forecasting error (the sum of
mistakes on all orders). Right time
The right-time component focuses on how well
Companies that score low on the right-product companies execute orders once they are received—
component will need to reexamine their forecasting basically, it evaluates whether a company is
Exhibit 3 of 5
Exhibit 3
The right product, right time, right location (RPRTRL) metric evaluates the three major
components of supply-chain performance.
Right product Right time

Were demand forecasts Did execution occur on
accurate, allowing schedule, allowing all
companies to deliver tasks to be completed
the right product? at the right time?
The
RPRTRL
metric
Right location
Was inventory staged along
the supply chain at the
right
locations?
completing all tasks, including those related to on progress, so semiconductor companies don’t
fab operations, sorting, assembly, and testing, learn about delays until it’s too late to address
within the expected time frame. The right-time them. In other cases, companies may not use all
score is computed by determining the volume- available vendor capacity or may fail to manage
weighted percentage of individual tasks for which their priorities. As one example, companies might
the actual cycle time was shorter than or equal not accelerate production for “hot lots”—those that
to the planned cycle time, in both back-end and need to enter production quickly because timelines
front-end processing. This calculation of execution will be tight.
performance provides more insights than current
measurement methods, which typically involve Right location
looking at overall cycle times and determining the Are inventory levels sufficient at all locations
percentage of orders with delays. along the supply chain, including die banks and
warehouses for finished goods? Many companies
If companies score low on the right-time can’t answer this question because their current
component, they should review their production- inventory systems haven’t been properly tested
management processes, including those related or implemented. All too often, they just consider
to vendors. For instance, foundry and back-end- average supply and demand, rather than examining
process partners may not provide daily updates the factors that might change these variables.

Companies could gain better insights about In many cases, a right-location analysis will reveal
inventory by calculating their right-location score, that inventory requirements vary significantly by
which measures the percentage of orders for which stage. For instance, one semiconductor company
they had enough inventory to satisfy demand, received many orders for which lead times were
weighted by volume, for each part. All orders are shorter than back-end cycle times. It could only
grouped into buckets based on lead time. For achieve an OTD rate of more than 80 percent (a
instance, a company might receive an order for best-practice figure) when it had enough finished-
which the lead time is shorter than the back-end goods inventory to satisfy projected demand for
cycle time. In this case, the right-location score at least 91 to 120 days (Exhibit 4). When the order
would be determined by calculating whether there lead time was longer than or equal to the back-end
was enough finished-goods inventory for the order. cycle time, the company could draw on its die-
If the lead time was longer than the back-end cycle bank inventory to satisfy the order. In such cases,
time but shorter than the full cycle time, the right- it achieved an OTD rate of more than 80 percent
location score would be based on whether the only when it had enough die-bank inventory to
company had sufficient inventory in the die bank. satisfy projected demand for the next 16 to 30 days.
Unfortunately, the company seldom had die-bank or
Companies may score low on this component if finished-goods inventory at that level.
they stage inventory at the wrong locations or
their finished-goods inventories are too low to fill Calculating RPRTRL scores
the orders that have short lead times. In addition To compute the RPRTRL metric, companies calculate
to delaying OTD, incorrect staging can create a separate scores for each component: right product,
surplus at the finished-goods and die-bank stages, right time, and right location (Exhibit 5). These scores
Exhibit 4 of 5 resulting in higher inventory costs. are then multiplied to determine the total RPRTRL
Exhibit 4
The right-location analysis looks at inventory sufficiency per SKU at key points and its impact
on on-time delivery.
Approximate back-end cycle time Approximate back-end cycle time

100 100
+16 pp
On-time delivery, +38 pp¹ On-time delivery,

% of volume % of volume
Lead time longer

Lead time shorter than or same as
than back-end back-end cycle
cycle time time
0 0
0–15 16–30 31–60 61–90 ≥91 0–15 16–30 31–60 61–90 ≥91
Finished-goods Die-bank
inventory, days inventory, days
¹ Percentage points.
Exhibit 5 of 5
Exhibit 5
The right product, right time, right location (RPRTRL) calculation uses scores from
three areas.
Right product Right time
Forecasting accuracy: evaluates extent Execution quality: measures “pull”
of forecasting bias (arithmetic mean part of supply chain (events that occur
of a forecasting error) and after order is received) and
magnitude of forecasting calculates volume-weighted
error (sum of mistakes Each percentage of orders for
on all orders) component is which actual cycle time
scored on a scale was shorter than or
of 0 to 1. Individual equal to planned
scores are multiplied to cycle time
calculate the total
RPRTRL score
Inventory policy: determines

die-bank/finished-goods sufficiency
by measuring volume-weighted
percentage of orders with
enough inventory at
these locations to
satisfy demand
Right location
score. For the initial computation, companies Calculating an RPRTRL score provides valuable
typically use anywhere from one to two years’ worth insights, but it’s just the first step in any supply-
of data. To measure progress, they should recalculate chain transformation. Companies must then
RPRTRL at monthly or weekly intervals (when they assess the costs and benefits of addressing each
have sufficient data). problem before developing appropriate solutions.
Fore
Since supply-chain issues will vary, companies e
The total RPRTRL score will range from zero to one. In must develop customized strategies for improving
our benchmark analysis of semiconductor companies, forecasting accuracy, execution, and inventory
the best-in-class players had an RPRTRL score in management. Some might get the most benefit
the range of 0.6 to 0.7. The average semiconductor from improved vendor management, for instance,
company scores 0.3. The key question for all semi- while others gain by adopting new predictive data
conductor executives is this: sets that decrease forecasting errors. But in all
cases, the RPRTRL metric will provide a common
Do you know your RPRTRL score? view of the supply chain that helps all groups deploy
a coordinated response. That alone will provide
invaluable assistance.
Gaurav Batra is a partner in McKinsey’s Washington, DC, office; Kristian Nolde is an associate partner in the Vancouver office;
and Nick Santhanam is a senior partner in the Silicon Valley office, where Rutger Vrijen is a partner.

Reducing indirect labor
costs at semiconductor
companies
Digital tools could bring new productivity and efficiency gains
to indirect functions. Why do semiconductor companies hesitate to
use them?
by Koen De Backer, Bo Huang, Matteo Mancini, and Amanda Wang
© gorodenkoff/Getty Images
Reducing indirect labor costs at semiconductor companies 57

When chip components shrink, manufacturing digital solutions for a variety of indirect functions,
and testing costs rise. This adage holds true since they have only applied them to one or two
even though Moore’s law has slowed, since jobs. In that respect, they lag far behind companies
expenses related to semiconductor production in many other sectors that have made more
have increased over the past few years. At every progress in digitizing operations and applying
semiconductor company, cost efficiency is advanced technologies.
now at the top of the agenda, although annual
revenues are solid and have been trending So how should semicos gain a greater
upward. While better margins are one motivator, understanding of their indirect labor? And what
companies also want more funds to invest in digital solutions are likely to produce the best
innovative chips for autonomous vehicles and other results in different functions? Companies might
emerging technologies. Demand for such chips be able to answer these questions through an
could surge as these technologies advance, and analysis that provides transparency into the
companies without leading-edge products will be purpose, end products, and activities (PEA) of
at a disadvantage. indirect employees. With insights from a PEA
analysis, semiconductor companies can recalibrate
In addition to implementing lean programs—a the workload and ensure that employees focus
traditional cost-control approach—many semicos on tasks that truly add value. They can then
are improving labor efficiency by using simple implement appropriate digital solutions for these
digital tools, such as dashboards on mobile tasks, ensuring even greater gains. Semiconductor
phones. They have also adopted more advanced companies that have successfully followed the
digital solutions, such as artificial intelligence PEA approach have reduced their indirect labor
(AI), machine learning, virtual reality, advanced costs by 20 to 30 percent across all functions.
analytics, automation, and 3-D printing. To date,
however, semicos have focused their efforts on
functions directly involved in manufacturing. They An approach for identifying and
have been less aggressive in using digital tools to capturing savings for indirect labor
improve indirect labor costs—those for technicians, At semiconductor fabs, indirect labor typically
engineers, back-office staff, R&D, and other represents a significant proportion of the
functions that support manufacturing but are not cost base. For instance, it accounts for about
involved in the conversion of materials to finished 18 to 20 percent of yearly manufacturing expenses
products. Their hesitation is understandable, since (exhibit). While engineering represents a large
indirect labor costs at semiconductor companies share of these costs, operations management
are much more difficult to quantify than direct and support also account for much spending.
costs, which can be measured based on operator Companies often have trouble estimating the
touch time. potential impact of cost-cutting programs because
many productivity drivers are difficult to quantify,
As digital tools become more sophisticated and particularly within engineering. For instance, a
produce increasingly greater gains, they will take team’s composition—such as the experience level
semiconductor companies further into the age of employees or the number of engineers—can
of Industry 4.0—a period of greater digitization strongly influence its efficiency. Moreover, a lot
in the manufacturing sector. If any companies of productivity information is not available or
resist using these tools, they risk falling behind inaccurately tracked, such as data on a team’s
more aggressive competitors. But even the most return on investment for the products it creates.
ambitious and dedicated semicos may have trouble
expanding their efforts into indirect functions. A PEA analysis can help bring some clarity
They often have limited insight into indirect jobs, to the murky world of indirect costs, both in
including the activities that consume the most manufacturing and R&D. It begins with workshops
time and the areas where productivity lags. Many for indirect managers and frontline staff.
companies also have difficulty selecting the best Participants identify the main purpose and end

Indirect labor
Exhibit 1 of 1
Exhibit
Indirect labor is a key cost driver for semiconductor fabrication plants.
Yearly manufacturing costs for example fabrication plants (fabs), %
Labor
20
100
15
200mm 15
fab 20
10 20
15
300mm 100
20
fab
(highly 25
automated) 20
2 18
10 15
100
Back-end 15
fab 30
10 20
Direct Indirect Materials Process Spare parts Facilities/other Total
Engineering Operations management Operations support
Equipment engineering Test/yield engineering Manufacturing management Procurement Facilities

Process engineering Other engineering Supply chain/planning Quality Other support
Source: Disguised examples from semiconductor companies
products associated with every job description, as Such analyses may not seem new to many
well as the activities that employees perform during industries, since companies across sectors
a typical week and the time spent on each one. already have established methods for identifying
This activity mapping often reveals findings that value drivers. Their analyses may not focus on the
surprise both managers and frontline employees. purpose, end products, or activities of employees,
One executive of a global memory-solution but their overall goal is to gain insight into different
company commented, “PEA is just like a magnetic- functions and reduce costs. In the semiconductor
resonance-imaging scan. Now I finally understand industry, however, such value analyses have been
how my engineers’ time is spent.” Often, a PEA rare, particularly with respect to indirect labor.
analysis will show that employees spend many
hours on activities that are not considered vital to Once companies have baseline metrics and a
their jobs or which do not contribute substantially solid understanding of all job functions, they can
to the creation of a desired end product. identify initiatives to improve efficiency and reduce

workloads. Typically, they will propose more than The recommendations from a PEA analysis will
50 solutions, all of which require funding and differ by indirect function because of the nature of
dedicated employees for implementation. Many of jobs within those functions—for example, technical
these will involve implementing digital solutions, roles, engineering, support services, and R&D.
but there will also be a few simple suggestions that The following sections describe the most relevant
produce good results, such as the elimination of digital solutions for a variety of indirect jobs.
unnecessary meetings or reports. Although every
proposal may sound great on paper, managers
need to conduct a reality check through feasibility Research and development:
assessments. Does an initiative require extensive Increasing productivity
funding or other resources? Will it create a burden Semiconductor R&D budgets are growing by
for the staff responsible for implementation? These about 6 percent annually as companies grapple
questions, and more, must be resolved before with the slowing of Moore’s law and the increased
moving forward. complexity of development processes, including
coding, testing, and verification. Companies now
In addition to assessing feasibility, companies require larger software groups to handle R&D
must quantify the savings for each initiative—this tasks, adding to indirect labor costs. Advanced
includes the number of workload hours eliminated analytics, one of the most popular digital solutions,
for certain tasks as well as cost reductions. They can help tame expenses by identifying the factors
should also determine whether employees can contributing to long development timelines and low
be reassigned or placed into new groups, or if product quality. While many semicos have already
workers lost to attrition must be replaced. The applied advanced analytics, their efforts have
cost-benefit analysis will help them determine tended to focus on streamlining basic engineering
bottom-line impact, prioritize initiatives, and tasks, such as chip design.
monitor progress. Once they have a plan, managers
can assign responsibility for implementation to Consider the example of a semiconductor company
groups or individuals, set timelines, estimate that saw only about 40 percent of its designs
the complexity of implementation, and track the become marketplace winners. To identify the
savings achieved for each initiative. elements of strong products, the company applied
advanced analytics to more than 80 data sets,
Since PEA analyses are conducted across including information on competitors, sales-force
functions, they identify solutions that will benefit records, and market data. It then looked at more
the organization as a whole, rather than those than 500 product features, identifying those that
that help only individual departments. For significantly contributed to value, as well as those
example, a top company that offered electronic- that did not. With this information, it was able to
manufacturing services conducted a PEA analysis channel its product investments more wisely.
across its engineering group. The analysis
revealed that employees spent most of their The company also used advanced analytics to
time completing a yield-management report. The improve its development process. When trying
time that each department spent on this activity to identify the elements of a successful team,
was not significantly high. The burden only became the company considered numerous variables,
apparent when the company totaled results for including tenure and the employee’s record for
employees across the entire engineering group. design wins. It was surprised to discover that
Leaders then created a cross-functional yield- several seemingly insignificant factors strongly
management approach to remedy the problem, influenced the success rate. For instance,
which is expected to reduce the number of teams that had members spread across multiple
engineering hours spent on the report by locations tended to have weaker performance.
52 percent. The insights from these analyses, combined with

the better understanding of product value, helped whatever component they are examining, or wear
the company increase the number of products devices on their wrists that note how far they have
classified as market winners by 10 percent, to walk within a plant to complete their tasks.
improving its total annual revenue by about Such solutions, which may improve technician
$750 million. productivity by up to 45 or 50 percent, are already
familiar to many industries. Within fabs, which have
Another semiconductor company was facing a been slower to embrace digitization, they represent
weak market as its PC sales declined and it lagged a new and untapped opportunity.
far behind its competitors in R&D productivity.
To engineer a turnaround, the company applied One semiconductor company originally had a
advanced analytics to identify productivity- very time-consuming maintenance process
improvement levers. Among other insights, the that involved having technicians make multiple
company discovered that frequent starts and inputs into a computer system, including notes
stops were one of the greatest problems across acknowledging work orders and updating
projects. If a team had to pause for a week or two, equipment status. They often had to leave their
its productivity plummeted. The company also workstations or the plant floor where the
discovered other hidden issues. For instance, machines were located to make these updates.
teams that had more than seven engineers tended To increase maintenance efficiency, the company
to have lower productivity. On the plus side, the implemented a simple digital solution—one familiar
company was also able to identify factors that to companies in other sectors but never before
improved performance, such as having a team in tested in its fab: it created a mobile-phone platform
which members had previously worked together. that allowed technicians to record and track
Once the semiconductor company applied the maintenance activity and machine performance
insights from these analytics, it increased R&D without leaving their work station. When needed,
engineering productivity by 15 to 20 percent. In one they could update aspects of the maintenance
group alone, run-rate savings amounted to order, such as the parts required. Technicians could
$15 million. also access checklists and standard operating
procedures for machine maintenance through the
Semicos that apply advanced analytics may also mobile app. The company was able to reduce
find that many other unexpected factors influence the indirect workload by about 14 percent through
R&D performance. For instance, conventional this initiative.
wisdom says that engineers should focus on one or
two projects. In one analysis, however, productivity
increased when they worked on more projects.¹ Engineering: Introducing more
sophisticated solutions
While some engineering tasks are simple and
Technical fields: Bringing automation straightforward, others require technical judgment
to the fore and customized solutions. The digital solutions
For most manufacturing-support technicians at that can help engineers are therefore more diverse
semicos, daily activities are somewhat repetitive— than those typically applied in other technical
and that means some of the greatest efficiency fields. Robotic process automation (RPA) alone
gains may come from greater automation of might be helpful in some cases, but it will be more
maintenance work flows, or by asking employees powerful if combined with advanced analytics, AI,
to use augmented-reality tools or wearable and machine learning. Although results will vary,
devices that track their movements. For instance, digital solutions can typically reduce engineering
maintenance technicians could use smart costs at semiconductor companies by 30 to
glasses that display the maintenance history of 35 percent.
1
Eoin Leydon, Ernest Liu, and Bill Wiseman, “Moneyball for engineers: What the semiconductor industry can learn from sports,” McKinsey on
Semiconductors, March 2017, McKinsey.com.

One semiconductor company improved its time for each back-office transaction by about
decision-making process for lots that were on 56 percent. Staff members then had more time to
hold—those deferred from further processing— focus on complex tasks.
by applying an RPA solution. For many years,
the company had relied on an IT system that Similarly, a professional-services company
automatically put 15 percent of lots on hold at determined that it could improve recruitment by
the final testing stage. Product-test engineers applying digital solutions. The company received
(PTEs) then reviewed each lot by logging onto more than 250,000 résumés per year, and it wanted
various systems and independently deciding to reduce screening costs and improve its ability to
whether it should proceed. This process accounted identify top candidates. (While automated screening
for about 50 percent of the PTE workload. To is common in many industries, most fabs haven’t
increase efficiency, the company analyzed past taken advantage of it.) After reviewing past résumés,
decisions about lots, including the factors that it created an algorithm that identified the applicants
determined whether they would be rejected. who were most likely to be successful employees,
Based on these insights, the company found that as well as the 50 percent that were unlikely to be
decisions for about 70 percent of lots on hold were hired. When applied to incoming resumes, the
straightforward and could be handled by RPA algorithm picked out the top 5 percent of applicants
solutions in combination with AI algorithms. The and automatically passed them to the next screening
PTEs who previously handled these decisions were stage. The bottom 50 percent were automatically
redirected to yield-improvement tasks or freed rejected. The company is expected to increase hiring
up to make decisions about more complex lots on efficiency by 30 to 50 percent and will also improve
hold. Overall, processing time for lots on hold was its return on investment by 400 to 500 percent.
reduced by 20 percent.
Support functions: Making sense out of Indirect labor is as essential to semiconductor

multiple systems companies as silicon. But many businesses
Employees in support functions must often have little insight into the costs associated with
deal with various IT systems, none of which are the technical fields, engineering roles, support
integrated. Work-flow automation, analytics, services, and R&D jobs that make up this vital
and RPA solutions can typically improve their function. With the continued rise of digital
productivity by 40 to 45 percent. solutions, semiconductor companies can no longer
afford to overlook this area when attempting
Many companies across industries have already to improve efficiency and productivity. If they
applied digital approaches to their support continue to focus only on direct functions during
functions with good results, and semiconductor cost-reduction efforts, they will soon fall behind
companies can expect similar gains. Consider competitors that undertake a more comprehensive
the example of a major bank that had recently approach to improving labor productivity. Although
streamlined its back office. To address work the best digital solutions will vary by company,
backlogs and the risks that might accompany a PEA analysis can be an important first step in
them, the bank worked with process and robotics helping semicos sort through the confusion and
experts to define work flows, identify exceptions, create a path forward, followed by advanced
and establish business rules. It then used RPA to analytics, automation, and more sophisticated
automate about 80 percent of tasks, relieving the digital solutions.
workload pressures and reducing the completion
Koen De Backer is an alumnus of McKinsey’s Singapore office, where Bo Huang is an associate partner and Matteo Mancini
is a partner. Amanda Wang is a consultant in the Shanghai office.

Taking the next leap
forward in semiconductor
yield improvement
By prioritizing improvements in end-to-end yield, semiconductor
companies can better manage cost pressures and sustain higher
profitability. The path forward involves advanced analytics.
by Koen De Backer, RJ Huang, Mantana Lertchaitawee, Matteo Mancini, and Choon Liang Tan
© Monty Rakusen/Getty Images
Taking the next leap forward in semiconductor yield improvement 63

As we progress into the digital era, semi- mind-sets, an insufficient view of data, and
conductor manufacturing competition is isolated efforts, as well as a lack of advanced-
intensifying, with companies looking to make analytics capabilities.
productivity improvements while undertaking a
record level of M&A activity. Front-end fabs and As devices continue to get smaller and more
back-end manufacturers have typically focused sophisticated, the effects of Moore’s law—that is,
transformational improvement efforts on direct and the estimation that the number of transistors in a
indirect labor-cost reduction, overall equipment given chip doubles every two years—will continue
effectiveness and throughput increases, material unabated. Thus in the semiconductor industry,
consumption and cost reductions, and global- the risks to yield due to process variability and
procurement and spending adjustments. Although contaminations are ever increasing, as is the
lean techniques have been the standard method importance of continuously improving design and
of achieving productivity gains, many companies— machine capabilities. In this article, we describe a
particularly back-end manufacturers—have new approach to changing mind-sets, gathering
difficulty sustaining lasting impact.¹ Our experience the right data to inform improvement initiatives,
working in Asia shows that a differentiating factor and achieving sustainable yield increases through
to effectively manage increasing cost pressures systemic improvements. We also offer an overview
and sustain higher profitability is improving end- of the impact that advanced analytics can have on
to-end yield—encompassing both line yield (wafers semiconductor yield and highlight seven capabilities
that are not scrapped) and die yield (dice that pass that semiconductor companies can pursue to inform
wafer probe testing). their efforts.
Yield optimization has long been regarded as one

of the most critical yet difficult to attain goals— Current perspectives on
thus a competitive advantage in semiconductor improving yield
operations. According to the Integrated Circuit The advent of Industry 4.0 tools to improve yield
Engineering Corporation, yield is “the single most across front-end and back-end manufacturers
important factor in overall wafer processing costs,” has been a big topic of discussion. Yet without
as incremental increases in yield significantly even entering that stage of technological maturity,
reduce manufacturing costs.² In this regard, yield most semiconductor companies are still trying
can be viewed as being closely tied to equipment to understand yield data by focusing on excursions,
performance (process capability), operator percentage, or product—or a combination of
capability, and technological design and complexity. the three.
Over the years, advances in fab technology, such
as more efficient air-circulation systems and A focus on percentage involves a bottom-up
better operator capabilities, as well as efforts to approach toward viewing yield percentages, either
lessen direct human contact with the production as an integrated view or by specific process areas.
process through the use of automation, have led This information is typically highly dependent upon
to a decline in particulate problems.³ And yet many the accuracy of the data captured by operators
semiconductor companies struggle to implement and made readily available for engineers through
sustainable yield improvements due to ingrained manufacturing-execution systems.
1
For more, see Koen De Backer, Matteo Mancini, and Aditi Sharma, “Optimizing back-end semiconductor manufacturing through Industry 4.0,”
February 2017, McKinsey.com.
2
“Yield and yield management,” in Cost Effective IC Manufacturing, Scottsdale, AZ: Integrated Circuit Engineering Corporation,1997.
3
Jim Handy, “What’s it like in a semiconductor fab?,” Forbes, December 19, 2011, forbes.com.

Some manufacturers focus on a specific set of end yield. Next, it can use a loss matrix to develop
products or product families, either by highest a holistic view of the company’s greatest sources
volumes or lowest yield performances. Resources of loss; then it can use that data to design more
are then assigned to solve for the root causes of targeted initiatives that will have the biggest impact
specific product problems, as a means of prioritizing on increasing yield—and thus on improving the
the company’s efforts. This approach requires company’s bottom line.
engineering resources from cross-functional teams,
such as equipment, process, product, quality, Align the language and data of engineering
testing, and, of course, yield. and finance
In our experience with semiconductor
Excursion—that is, when a process or piece of manufacturers, there is a consistent disconnect
equipment moves out of preset specifications—can between the engineering and finance functions.
be a significant contributor to yield loss, particularly Engineers focus on and celebrate gains in
if it goes undiscovered until after fabrication. An percentage yield, but they often overlook the
excursion focus can thus be defined as tackling the connection between yield and cost. Indeed, the
highest and most obvious sources of yield loss or celebrated percentage increases may or may
excursion cases identified from past occurrences not lead to any significant impact on the bottom
either in the plant or from customer incidents. line. Furthermore, many engineering and finance
The key is to ensure that the root causes of those functions use different systems to track yield,
yield losses and their potential failure modes are which can result in near-constant misalignment
addressed to avoid a repeat occurrence. between the functions, rendering data less usable
by the lack of agreement about which to use as the
These approaches can enable manufacturers to source of truth.
capture, monitor, and control various forms of yield
losses—but they may leave other opportunities The first step in ensuring that all functions are
on the table. To target the highest impact on aligned in a yield-transformation effort is to speak
profitability, semiconductor companies must first a common language. Then not only can engineers
translate yield loss into actual monetary value (rather and finance personnel understand each other but
than simply volumes or percentages), enabling them the ease of translation and communication can
to more effectively direct resources toward solutions extend vertically through the organizational ladder,
across all products and processes. This approach allowing both ground-level engineers and top-level
goes beyond a yield-loss focus on specific products management to agree on justifications for pursuing
or excursion cases to encompass a more end-to- initiatives and on progress achieved for successful
end view. As a result, semiconductor companies improvement activities.
can more effectively implement systemic process
changes and, particularly given the different cost To overcome divergent sources of truth,
structures for each product, result in significant and semiconductor companies can construct a cost-
as yet unrealized cost savings. of-non-quality (CONQ) baseline that uses cost data
from finance as well as engineering (Exhibit 1). For
example, finance provides data on standard costs,
A new approach to semiconductor standard yields, and yearly volumes per product
yield improvements while engineering provides detailed breakdowns on
To translate yield loss into actual monetary value, the nature (reject category) and source (process)
a semiconductor company must begin by aligning of the defects by product. Merging these two views
the language and data used by engineering and provides a full and accessible view of the cost of
finance to gain a better understanding of end-to- yield losses.

Exhibit 1 of 5
Exhibit 1
Cost-of-non-quality (CONQ) calculation can be broken down into three components: Volume,
standard cost, and yield.
CONQ-calculation breakdown
CONQ total Actual scrap volume Average standard cost Average standard yield
per unit
Example • Chips detected as defective • Cost per chip increased by • Expected die-yield %
• Chips falling into bins allocated expected scrap (yield)
for scrap
Description • Quantity of scrap attributable to • Average cost per chip, including: • % of expected yield loss used
die-yield loss, ie, products — Variable (material) costs for standard chip costing,
discarded during production multiplied by the average
• Scrap quantity measured at — Overhead (including labor) costs standard cost per unit, gives
process output, reported in — Yield adjustment, ie, additional unit the “unyielded” cost, ie, the
enterprise-resource-planning cost due to yield losses real cost at input
system
Not included Utilization loss: Working chips that are unsold and scrapped after 6 months
Rework: Defective chips thrown into bins for reprocessing to ideally produce a good chip
Freight costs: Cost of transportation of wafers from upstream processing
Develop a holistic, data-driven view of what production process, the company was losing almost
needs to improve and where $68 million due to yield losses overall, including
Typically, engineers are dedicated to discrete almost $19 million during electrical testing alone
processes, enabling them to develop deep expertise (Exhibit 2). Engineers can use their technical
in a given area and more effectively serve on the line. knowledge of what happens in particular processes
However, when embarking on a yield transformation, to determine why certain reject codes are high
a semiconductor company must develop a holistic within those processes. By also calculating the
view of the manufacturing process. Therefore addressable amount of loss, this heat-map view
engineering must take a step back to see exactly enables the organization to prioritize what to focus
what parts of the process, and specifically what on and allocate resources to the process areas
reject categories, lead to the greatest amount of most likely to improve profitability.
loss. While some companies already bring a product
focus to yield losses, an overarching view of the entire In our experience, having this view handy is extremely
manufacturing line is usually not top of mind. Thus, useful not only to ensure that everyone has a view of
instead of a singular transformation, what usually what must be addressed and where but also to keep
happens is a lot of the efforts are siloed into individual track of what areas have been covered—and which
processes, products, and even pieces of equipment. ones are still unexplored. The heat map also enables
engineers to take a top-management approach
A loss matrix enables engineering to map process toward the line as a whole, instead of focusing only on
areas (in a heat map) and reject categories against their particular process, and reinforces the view that
yield performance of the manufacturing line all engineers are responsible for managing quality
from start to finish. One manufacturer found that and yield.
across the eight major steps of its semiconductor

Exhibit 2 of 5
Exhibit 2
An example loss matrix illustrates how manufacturers can identify major yield losses by
category to help prioritize improvement efforts.
2018 estimated cost of non-quality, $ million
X Addressable amount Y Targeted savings amount <1 1–2 >2
Reject category/ Electrical Tape Visual Assembly Pick Assembly Assembly

process area testing and reel inspection package 1 Die attach and place package 2 package 3
3.1 1.0 1.2 2.2

Loose dies 3.1 1.0 1.2 - 0.2 2.2 - 0.2
1.0 0.4 0.3 1.0
Contamination/ 1.1 1.1 2.1
1.1 0.3 1.1 2.1 0.4 0.9 - 0.9
foreign material 0.5 0.4 1.0
1.3 3.5 1.0
Bulge/bubble/wrinkles - 1.3 - 3.5 1.0 - - -
0.3 1.5 0.5
0.7
Flux losses 0.7 0.2 - - - - - -
0.3
1.1 1.5
Quality reject - - 1.1 - - 2.5 - -
0.3 0.9
Previous reject:
- 0.2 0.9 0.9 - - - -
from upstream
Broken tile/ 0.4
0.5 0.3 1.0 1.1 0.9 0.5 - -
ceramic crack 0.1
1.5 1.0
Insufficient silicone - - 1.9 - 0.9 - 1.0 -
0.5 0.4
Evaluation 0.6 0.4 1.0 - 0.9 0.5 - -
Contact resistance/ 1.5

1.5 - - - - - - -
no contact 0.4
3.5
Tile-calibration retesting 3.5 - 0.9 - - - - -
1.0
Minor coverage of 1.0
- - - - - - 1.0 -
assembly package 0.3
0.9 0.9
Other 1.5 0.5 0.9 - 1.5 1.3 0.9 -
0.2 0.3
Remaining 2.0 2.0
6.2 1.6 0.5 0.3 1.5 1.6 2.0 2.1
long-tail losses 1.5 0.1
Total loss 18.7 5.8 10.5 7.9 7.3 9.5 4.9 3.2
Implement systemic improvements to identify per-product analysis ensures that action is taken
yield loss only on items that have the biggest impact on yield.
Once the biggest loss areas are identified using
the loss matrix, it is important to ensure the actions As a result, engineers have the detailed insight they
taken to improve the identification of yield loss are need to address the key issues that drive the
sustainable; this starts by isolating the products that particular losses identified by the loss matrix. They
are the biggest contributors to scrap (Exhibit 3). This can also use a product Pareto analysis to identify

Exhibit 3 of 5
Exhibit 3
Product analysis helps manufacturers identify the biggest contributors to overall yield
loss as well as the size gap.
Week:
X top products contribute
Process 1 target Process 2 actual Gap to target Contribution to X% drop of yield (overall gap to
target is x.x%, partly positively
85% 81.3% –10% –10% affected by parts that are better
than target)
Product Yield Gap to target Contribution % of scrap % mix Scrap/mix
Product 1 72% –18% –7% 26% 21% 6.4

Product 2 34% –57% –8% 21% 11% 8
Product 3 43% –48% –7% 17% 9% 7.6
Product 4 68% –22% –6% 11% 9% 6.5
Product 5 84% –6% –5% 10% 11% 5.8
Product 6 12% –78% –6% 10% 6% 8.9
Product 7 88% 8% 5% 9% 11% 5.7
Product 8 73% –17% –5% 8% 8% 6.3
Product 9 90% 10% 5% 8% 10% 5.6
Product 10 91% 10% 5% 7% 9% 5.6
Product 11 68% –5% –5% 7% 6% 6.5
Product 12 86% 5% 5% 7% 7% 5.8
Product 13 83% –8% 5% 6% 7% 5.9
Product 14 87% 6% 5% 6% 7% 5.7
Product 15 87% 6% 5% 6% 7% 5.7
Product 16 99% 14% 6% 6% 11% 5.2
Product 17 93% 13% 5% 6% 7% 5.5
Product 18 92% 12% 5% 6% 7% 5.5
Product 19 94% 14% 5% 6% 8% 5.4
Product 20 89% 9% 5% 6% 6% 5.6
the use cases where addressing an issue will solve which rejects are true and which are false, as
the most significant, far-reaching problems. well as discuss which potential cross-functional
collaborations may help solve the issue. One
Key improvement themes are generally structured manufacturer completed an analysis of four of the
using the traditional “5 Ms” of lean manufacturing— Ms (measurement was not applicable in that case)
machine, man, material, measurement, and method. and sorted out true from false rejects while also
While organizing loss categories along these lines, developing a sound foundation for improvement
semiconductor companies should also analyze initiatives (Exhibit 4).

Exhibit 4 of 5
Exhibit 4
Key improvement themes are identified, evaluated, and structured in close collaboration
with experts.
True rejects False rejects
External Internal
Loss Specifications/
categories material Method (process) Machine Man
Loss 1 • N/A • Reduce mechanical • Target corrective • Reduce contamination

stress and optimize actions to problematic of clean area by
heating profile (regular tools based on golden- banning operators
update; multiple flow analysis from wearing
initiatives) cosmetics
Loss 2 • Feedback to upstream • Create protocol • Perform height

location on wafer- for loss 2 adjustment for
thickness variation troubleshooting machines with high
loss 2 short failure
Loss 3 • Relax • Enhance loss 3 • Offset optimization • Retrain staff to

specifications troubleshooting protocol • Improve machine setup improve execution
• Technical initiatives • Prioritize through (eg, combined jig-holder discipline of the
upstream (etching pattern recognition set, jig maintenance, etc) troubleshooting
process, design) protocol
Loss 4 • Improve • Harder dicing blades

pick-and-place
process
Loss 5 • Move electrical loss • Enhance reel-rework

5 before earlier process, machine to capture new
perform in wafer form at products (jig-set capital
loss 2 tools rework option expenditures required)
Overall • Digital tools to enable efficiency of yield-improvement measures:

— Yield-loss pattern-recognition tool—to enable focused daily yield-loss troubleshooting
— Parametric-analysis tool—to enable effective targeted design revisions and spec relaxation
— Golden-flow tool—to identify problematic tools and support efficient root-cause analysis
One finding from the yield-loss analysis showed that Given their cross-functional nature, the machine-
the manufacturer was experiencing contamination variability initiatives entailed both internal effort and
and wrinkle issues at a particular process point. external involvement. Internally, product, process,
The ensuing problem-solving session identified and test engineers, quality engineering, and R&D
underlying, systemic issues in the manufacturing worked together to run the necessary tests and
process, resulting in four improvement initiatives qualifications to ensure the activity had no negative
relating to both true and false rejects (Exhibit 5). impact on semiconductor quality. Armed with their

Exhibit 5 of 5
Exhibit 5
The idea-generation process starts by brainstorming ways to reduce both true and false
rejects and focusing on addressable issues.
Hypothesis tree for process X yield losses
Developed into initiative
Estimated
improvement
potential, %
Decrease intermachine and interoperator performance

16
variability
Decrease false
Adjust tool recipe settings (calibrations) on machines to
rejects 10
decrease false rejects
Add incoming mapping of missing dies in front-end

manufacturing from upstream site(s) N/A
Decrease
Reject
cost of non- 40
code 1
quality at
process X Match internal operational specs at process X to customer specs
Reject
(no additional further tightening) 5
code 2
Improve lamination-assembly process (shift from standard to

advanced material Y where possible) 9
Decrease true
rejects
Improve baking operational process (root-cause “killer tiles”) in
back-end assembly N/A
Improve front-end manufacturing and delivery process from

upstream site(s) N/A
analysis, engineers could have more meaningful Impact on a yield engineer’s typical
discussions with external vendors about legacy day, with the holistic view of yield
patches to existing equipment and ideas to improve improvements
machine performance. Yield engineering resources are typically spent
supporting or leading improvement activities
The implementation of these four initiatives reduced across both product and process engineering. At
contamination rejects for identified products by one manufacturer, yield engineers’ daily activities
90 percent and wrinkle rejects by 40 percent, and ranged across three main areas—root-cause
in the long term gave valuable insight to engineers problem solving of excursions and other critical
on both collaborating with third parties as well as identified yield losses, cross-functional yield-
ingraining an ownership mind-set. improvement activities and collaborations with other

teams, and operational tracking and reporting of one that will have the biggest forecasted impact to
yield performances across the fab. By applying a the bottom line. Internal problem solving is further
holistic approach toward yield improvements based strengthened with the help of big data analytics
on the steps described above, a typical day in the solutions that proactively highlight commonalities
life of a yield engineer improved in all three realms. or pattern recognition—for example, a particular
tool, process group, or even upstream product
Root-cause problem solving or process that contributes significantly to yield
The majority of yield engineering resources used losses (see sidebar, “The role of advanced analytics
to be spent on yield-loss analyses and low-yield- in semiconductor yield improvement: Converting
threshold troubleshooting, for both mature products data into actions”). Yield solutions can help push
and new product releases from product development, efficiency improvements to the team by providing
including buy-off approvals. Due to the yield-loss proactive, low-yield threshold warnings and
analysis, the manufacturer’s yield engineers could reporting while also improving turnaround time for
shift from a reactive “firefighting” stance on tackling lot releases.
ad hoc requests or manufacturing execution system
triggers to solving for root causes of major excursions Cross-functional yield improvements
or other weekly yield losses on the line. Previously, resources were spread across multiple
projects or initiatives with other engineering teams,
Engineers can now identify key losses as per the with the main task of using analytics to identify the
loss matrix that are unaddressed and start with the
The role of advanced analytics in semiconductor yield improvement: Converting data into actions
As noted by the CEO of advanced- is a big difference between insights from cause identification. This benefit
analytics company Motivo Engineering, traditional quantitative analysis and those is greater when we try to uncover
“Each fab has thousands of process from advanced analytics. Furthermore, root causes of low- and medium-
steps, which, in turn, have thousands of semiconductor manufacturing is in a frequency errors, which are difficult
parameters that can be used in different unique position compared with other to detect using traditional analytics.
combinations. With so many factors industries to reap the benefits of advanced
in play, we see a lot of chip failures or analytics, given the massive amount of —— Improved value-added time for
defects.”¹ Given its complexities, traditional data embedded in fabs’ highly automated engineers. At one organization, for
quantitative analysis wouldn’t help fabs and sensor-laden environment. Fabs can example, data pulling and analysis
uncover all improvement opportunities, benefit from yield analytics through three in line-maintenance activities can
resulting in a lengthy process of root-issue key levers: take up more than triple the time
discovery—and thus massive yield losses. required than if data infrastructure
—— Early defect detection and root- and interface are well designed. This
For that reason, the use of advanced cause identification. Advanced- situation represents an opportunity
analytics offers a new paradigm for analytics tools can help uncover to free up engineers’ time to focus
yield improvement in the semiconductor issues much faster and in much instead on core issues and production
industry. Indeed, the nature of greater detail, leading to faster root- design solutions.
manufacturing complexity means there
For more, see Koen De Backer, Matteo Mancini, and Aditi Sharma, “Optimizing back-end semiconductor manufacturing through Industry 4.0,” February 2017, McKinsey.com.
1

(continued from page 71)
—— Powerful tool for “past learning” and management of the manufacturing —— Golden-flow analysis is a crucial
continuous improvement. Machine- process. These tools and processes analytical capability to determine tool
learning algorithms, a well-organized enable data to be managed and commonality and identify which tools
data lake, and the appropriate tools reported by the engineers so it’s most are performing at optimal levels—and
allow fabs to accumulate learning beneficial to their target audience, be which are not. These data help with
from past experiences and enable they process engineers, managers, or both tool matching and ensuring that
continuous improvement. Whereas the third parties such as customers. production is as high yield and efficient
traditional approach eliminates defects as possible, maximizing throughput
by adjusting multiple parameters, which —— Parametric analysis refers to and optimizing manufacturing flow
helps with the current batch, it fails to testing how product parameters are (see case study “Golden-flow analysis
offer any insight into the root cause of distributed at performance testing in action”).
the problem—meaning it is likely to be and inspections and comparing these
repeated in future batches.² findings to product development’s —— Equipment optimization as an
specification limits. This analysis analytical capability refers to how
Identify core analytics capabilities that ultimately aims to enable the software can perform predictive
can improve yield optimization of specifications—tight analyses to determine potential
Seven core analytics capabilities are enough to ensure good quality but issues before they occur. This
important in yield-management solutions: also reasonable enough to prevent ability is closely linked to predictive
monitoring and reporting, parametric unnecessary over- or under-rejection. maintenance and aims to avoid
analysis, correlation analysis, golden-flow yield loss by tackling predictable
analysis, equipment optimization, pattern —— Correlation analysis finds tool variation and necessary
recognition, and event analysis: correlations between test parameters parameter tuning.
at earlier stages versus final
—— Monitoring and reporting is the most inspections. This assessment aims to —— Pattern recognition is about looking
basic among the capabilities—but also maximize final product performance at the distribution of parameter
one of the most important. This process and help manage end-to-end yield by patterns across wafer maps and
refers to trend charts, histograms, adjusting test parameters depending connecting the findings to equipment,
Pareto analysis, proactive reporting on how they correlate with testing manufacturing trends, and
and notification, and enhanced results, either electrical or visual correlations with process and test
statistical process control, all of test parameters. parameters. With this capability, live
which enable enhanced performance feedback can be given to engineers
“Yield and yield management,” in Cost Effective IC Manufacturing, Scottsdale, AZ: Integrated Circuit Engineering Corporation,1997.
2

Case study
Golden-flow analysis in action
Golden-flow analysis helps identify and configuration, was experiencing an indicatication of a problem until after
bad actors and golden tools in situations uptick in normalized defect density across it got worse. The advanced warning
where
McKinseytrends Semiconductors
are unclear. At one 2019 different layers over a seven-day period of increased defect density allowed
manufacturer,
Taking the next the analysis
leap detected (exhibit). The uptick had not surpassed the manufacturer to take down the
that a specific tool (XYZ-1), which the upper control limit (UCL), so without tool for investigation, repairs, or
Exhibit
was one of three tools in the same class the analysis there would have been no calibration interventions.
Exhibit
Commonality analysis helps to identify low-performing and golden tools in situations

where trends are unclear.
Normalized litho-defect density at Tool 1–Tool 2 by lithography tool, number per cm²
XYZ-1 XYZ-2 XYZ-3
1.1
1.0
Upper control limit
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7
Days of week, #

(continued from page 72)
so tools and process parameters can highly automated and sensor-laden collaboration and action ( see case study
be adjusted to reduce yield loss systems in fabs, data quality is usually “Feedback loop finds costs savings”).
(see case study “Using analytics to a challenge in implementing analytics
reduce losses”). software or using data for analysis; for Partnerships with technology and
example, different product families have analytics vendors. As our colleagues
—— Event analysis entails studying different data formats and complex have noted, many analytics and
production events, such as production processes. The important machine-learning vendors believe that
maintenance and supply changes, step is to get individuals with a strong semiconductor companies prefer to
to discover their effect on yield. technical knowledge of data and database develop solutions in-house,³ which
Identifying root causes for quality optimization to create the right data discourages them from building strong
shifts or parametric surges can infrastructure to enable scale-up of relationships with other semiconductor
be done by tying them to the analytics solutions. companies. In reality, active partnerships
occurrence of various events on the with analytics vendors will help
manufacturing floor. Right organization setup to take data increase the speed of building analytics
insights to fast action and feedback loop. capabilities for fabs. Given the fast-
Undertake key enablers to Converting data and insights into actions changing environment and highly
overcome typical challenges in is among the most critical steps—and specialized capability in analytics,
implementing yield analytics challenges—to capture benefits from ongoing collaboration and partnership
Well-organized data integration and analytics. In particular to yield, issues will help semiconductor companies stay
interface. Data pull and cleaning (that is, always cross sites and require end-to-end on the cutting edge and employ solutions
the creation of a data lake) are important collaboration to get breakthrough results. that enhance in-house capability.
steps in deploying analytics. Despite The key to success is to have effective
the richness of data gathered through yield tracking and a platform to enable
Ondrej Burkacky, Mark Patel, Nicholas Sergeant, and Christopher Thomas, “Reimagining fabs: Advanced analytics in semiconductor manufacturing,” March 2017, McKinsey.com.
3
impact of recommended improvements. Armed with Teams can effectively link decisions from customer
end-to-end traceability of yield losses from front requirements (either by R&D or business units),
end to back end, yield teams benefit from a more down to bottom-line impact on front-end and
granular view of bottom-line impact, reducing the back-end expected yield losses, to identify systemic
analytical resources needed and allowing for more root causes cutting across processes, reject
insights to be shared with the cross-functional team, categories, or products. This capability helps yield
including R&D, business-unit sales and marketing engineers be more precise in identifying which
teams, and front- and back-end managers. teams (product or process engineers) are needed
and helps to prioritize the initiatives in which they

Case study
Using analytics to reduce losses
One manufacturer developed a false- The algorithm provides a daily automated estimation and monitoring on a monthly
reject estimator analytics tool for final report of false rejects at tool and part basis. This approach reduced losses from
inspection equipment to help the fab number (product) levels, enabling a material wastes and customer quality
detect and estimate sizes of false rejects focused effort to tackle problems in a issues while enhancing overall capacity
based on a pattern-recognition algorithm. timely manner by comparing with manual (for example, dice output per day).
Case study
Feedback loop finds cost savings
One semiconductor player operating end-to-end yield monitoring and speed this yield PMO has delivered 10 percent
across regions in Asia and America up the feedback loop. Along with the yield improvement and identified and
set up a cross-site yield project- development of four analytical tools and implemented a $12 million cost-savings
management office (PMO) to facilitate a performance-management dashboard, opportunity within six months.
ought to invest most of their time. From an efficiency the value chain to collaborate on more data and to
improvement and workload-reduction perspective, push initiatives to be more fact based and prioritize
teams can better rationalize meeting participation. resources to maximize profitability.
Yield engineers are further empowered with data Yield-performance tracking and reporting
to highlight potential opportunities to implement For both mature and new unreleased products,
more yield gains by aligning or relaxing internal yield engineers have shifted from daily or weekly
specifications, without affecting customer demand yield-percentage monitoring to more continuous
or satisfaction. Transparency enables teams across monitoring thanks to the capabilities of the loss

matrix. Performance baselines and improvements —— Develop a holistic, data-driven view of what
can be tracked and reported either in the form of needs to improve and where. Work on yield
the loss matrix or with the help of analytical yield can often be siloed due to how manufacturing
solutions. Teams can now visualize the distribution organizations are structured. Using the loss
of key forecasted shifts in yield losses as measured matrix and analytical solutions—where costs
by monetary impact, which helps prioritize the next can be easily viewed by processes, reject codes,
wave of improvement initiatives. Reporting is more or products—allows engineers and managers
mutually exclusive and collectively exhaustive than to gain a better view of the health of the entire
previously limited reporting by process and integral manufacturing process, from R&D through
yield percentages. wafer fabrication and die packaging, to push
improvement efforts to the right areas. This view
also gives engineers and managers a chance to
track what areas they are already tackling, as
For semiconductor companies, the successes of well as what areas have yet to be explored.
effective yield improvement lead not only to increased
profitability but also to better organizational health as —— Implement systemic improvements. Yield
a whole. Our experience points to three central key improvements should address excursion cases—
pillars that make yield transformations successful: but more important, they should also tackle the
baseline yield. By setting up discussions where
—— Align the language and data of engineering and engineers can explore historic causes of yield
finance. Looking at yield percentages only loss, new levers can be discovered that will
provides one view of the situation; engineering increase overall yield performance for a certain
and finance alike must align on using the cost of product or process. There can also be situations
poor quality as the method for understanding where certain losses are tolerated simply
and guiding the direction of the company’s because they have historically been seen as
yield improvement efforts. Collaboration on the acceptable. Focusing on standout issues of yield
creation of a CONQ calculation can ensure that loss, as well as working to continuously improve
improvement initiatives are based on a viable the baseline yield percentage as a whole, leads
foundation of data and collaboration. to more sustainable yield improvement.
Koen De Backer is an alumnus of McKinsey’s Singapore office, where Matteo Mancini is a partner. RJ Huang is a consultant
in the Manila office, Mantana Lertchaitawee is a consultant in the Bangkok office, and Choon Liang Tan is an alumnus of the
Kuala Lumpur office.

This McKinsey Practice Publication meets the Forest
Stewardship Council® (FSC®) chain-of-custody
standards. The paper used in this publication is certified
as being produced in an environmentally responsible,
socially beneficial, and economically viable way.
Printed in the United States of America
November 2019
Designed by Global Editorial Services
Copyright © McKinsey & Company
McKinsey.com

MCK - Semiconductors - Oct2019-Full Book-V12-RGB

Uploaded by

Copyright:

Available Formats

MCK - Semiconductors - Oct2019-Full Book-V12-RGB

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MCK - Semiconductors - Oct2019-Full Book-V12-RGB

Uploaded by

Copyright:

Available Formats

What are the main topics covered in the publication?

What are the main topics covered in the publication?

What are some of the opportunities discussed for semiconductor companies?

What are some of the opportunities discussed for semiconductor companies?

McKinsey on

Number 7, October 2019

3 What’s next for semiconductor

27 Blockchain 2.0: What’s in

38 Rethinking car software and

Ondrej Burkacky Peter Kenevan Abhijit Mahindroo

2 McKinsey on Semiconductors Number 7, October 2019

by Marc de Jong and Anurag Srivastava

© DuKai photographer/Getty Images

What’s next for semiconductor profits and value creation? 3

2010 2011 2012 2013 2014 2015 2016 2017

Total economic profit, $ billion

¹ Figures may not sum, because of rounding.

4 McKinsey on Semiconductors Number 7, October 2019

From 2012 to 2016, the semiconductor

What’s next for semiconductor profits and value creation? 5

Exhibit 2 Facebook, Google, and Tencent become more

1997–2012 –38.7 161.5

6 McKinsey on Semiconductors Number 7, October 2019

22.4 4.1 42 20.9

–0.3 –0.2 –1.0 –2.2

Top companies for economic profit, by subsegment

¹ Figures may not sum, because of rounding.

From 1997 to 2012, the cumulative

What’s next for semiconductor profits and value creation? 7

26.9 27.0 29.2

Total economic profit, $ billion

Top companies for economic profit, by subsegment

¹ Figures may not sum, because of rounding.

8 McKinsey on Semiconductors Number 7, October 2019

208 206 202

2012 2013 2014 2015 2016 2017

Major deals or bankruptcies by subsegment

What’s next for semiconductor profits and value creation? 9

one trend is generating new questions: the —— Competitive differentiation. By developing

10 McKinsey on Semiconductors Number 7, October 2019

Estimated economic value creation by Apple’s

Source: Alphr; Bloomberg; Business Insider; IHS Markit; S&P Capital IQ

What’s next for semiconductor profits and value creation? 11

As for the subsegments, fabless now has the

Total returns to shareholders have been strong within semiconductors.

450 Diversified integrated

Source: Annual reports; S&P Capital IQ; Semiconductor CPC database

12 McKinsey on Semiconductors Number 7, October 2019

Performance expectations are highest for the fabless subsegment.

Fabless Foundry Micro- Capital Diversified Analog Memory

Implied long-term organic growth at 2018 margins, % CAGR³

What’s next for semiconductor profits and value creation? 13

Fabless captures the most shareholder value across subsegments.

100% = $1.5 trillion

Note: Figures may not sum, because of rounding.

14 McKinsey on Semiconductors Number 7, October 2019

Copyright © 2019 McKinsey & Company. All rights reserved.

What’s next for semiconductor profits and value creation? 15

© DuKai photographer/Getty Images