Académique Documents
Professionnel Documents
Culture Documents
Equity Research
Semiconductor Industry Primer 2014
Semiconductors
Source for cover image of Intel Broadwell wafer: Intel (copyright Intel, reproduced with permission)
TABLE OF CONTENTS
Map Of The Semiconductor Industry .............................................................................................................................................................. 5
Semiconductor Companies....................................................................................................................................................................... 5
Types Of Semiconductors .........................................................................................................................................................................6
Semiconductor End Markets .................................................................................................................................................................... 7
Semiconductor Industry Dynamics ................................................................................................................................................................ 8
Growth...................................................................................................................................................................................................... 8
The Semiconductor Cycle ....................................................................................................................................................................... 11
Pricing ..................................................................................................................................................................................................... 14
Capacity And Capacity Investment ........................................................................................................................................................ 17
The Rise Of The Foundries .................................................................................................................................................................... 24
Semiconductor Inventory And The Electronics Supply Chain ............................................................................................................ 29
Semiconductor Segments...............................................................................................................................................................................35
Analog .................................................................................................................................................................................................... 38
Logic ....................................................................................................................................................................................................... 39
Memory ................................................................................................................................................................................................... 41
Discrete Components ............................................................................................................................................................................ 42
Sub-Segments Of Interest ..................................................................................................................................................................... 46
Microprocessor Technology Topics ................................................................................................................................................54
Applications Processors ..................................................................................................................................................................56
Discrete Graphics ........................................................................................................................................................................... 63
Memory............................................................................................................................................................................................65
Analog .............................................................................................................................................................................................. 81
PLDs ................................................................................................................................................................................................ 85
Selected Technology Topics ........................................................................................................................................................................... 91
Semiconductor Wafers And Chips ......................................................................................................................................................... 91
Manufacturing Transitions -- Line Widths And Wafer Size ................................................................................................................ 92
Calculating The Number of Circuits Of A Wafer (Die Per Wafer) ....................................................................................................... 96
Transistors -- What They Are And Some Technical Terms................................................................................................................... 97
Lithography -- Multi-patterning And EUV ......................................................................................................................................... 100
Appendix A: Semiconductor Companies..................................................................................................................................................... 102
Appendix B: Glossary ................................................................................................................................................................................... 105
Semiconductors
Companies with
shares between 1%
and 2%, AMD,
Infineon, SanDisk,
NXP, MediaTek,
Freescale, Sony,
Marvell, Nvidia, 12%
Intel, 15%
Samsung
Electronics, 10%
Qualcomm, 5%
SK Hynix, 4%
Micron Technology,
4%
Toshiba, 4%
Others, 35%
Texas Instruments,
3%
Renesas Electronics,
3%
Broadcom, 3%
STMicroelectronics,
3%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Applications, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
The semiconductor industry as a whole is somewhat fragmented. The top nine companies make up more than
half of total semiconductor sales, but the remaining half of the market is served by a large number of
companies, each with market share of 2% or less. However, the industry contains several quite distinct
segments, of which many semiconductor companies focus on just one or two. As a result, there are a number of
major segments and sub-segments in which there are just two or three primary competitors. We discuss this in
more detail further on in this report.
Intel is the largest semiconductor company, accounting for about 15% of total semiconductor sales. The
bulk of Intels business is related to the computer end market. Intel is the market leader in
microprocessors, the chips that act as the brains of the computers.
Samsung is the second-largest semiconductor company, with about 10% of total semiconductor industry
sales in 2013. Semiconductors are just 16% of Samsungs total sales. Samsung is also a major producer of
wireless handsets and flat panel displays. Samsung is the worlds largest semiconductor memory
company. Samsungs Memory subgroup, which includes mainly dynamic random access memory
(DRAM) and NAND flash memory, accounts for 60-70% of Samsungs semiconductor revenue, with the
rest of Samsungs semiconductor revenue coming from a variety of logic and other chips from Samsungs
LSI group. Samsungs semiconductor division also offers foundry services in which it manufactures chips
for other companies. Apple currently uses Samsung and TSMC as foundry partners for the processors it
uses in its iPhone and iPad products.
Qualcomm, a producer of semiconductor chips for wireless handsets, is the worlds largest fabless
company. A fabless company is a company that does not own its own manufacturing facilities, i.e.,
fabrication facilities (or fabs). Qualcomms main chip business is designing and selling communications
chips and processors for smartphones. In addition to selling chips for wireless handsets and other mobile
Semiconductors
devices such as tablets, Qualcomm also derives about one-third of its total revenue and about two-thirds of
its profit from royalties paid by wireless handset makers.
Texas Instruments, often referred to as TI, is the worlds largest analog semiconductor company. Analog
chips are chips that deal with continuous electrical signals or with electrical power, rather than digital
signals (i.e., signals that represent either a 1, or a 0). Digital chips are used in the heart of many modern
electronics to do calculations. Analog chips are the interface between the real world and the digital heart of
an electronics system. About 60% of TIs revenue is from its analog division, and an additional 20% is
from TIs embedded processing division, which TI considers to be complementary to analog. TIs
embedded processing products include microcontrollers (small thinking chips) and digital signal
processors (thinking chips that are optimized to do the calculations related to analyzing electrical signals).
Types Of Semiconductors
Exhibit 2. Types Of Semiconductors (2013)
Discretes, 6%
Optoelectronics &
Sensors, 12%
IC, 82%
Source: Semiconductor Industry Association and Wells Fargo Securities, LLC
Semiconductors can be divided into three broad categories (see Exhibit 2):
Integrated circuits (ICs). ICs are semiconductor devices (chips) on which an entire electrical circuit is
created. More than 80% of all semiconductor sales are of integrated circuits. ICs are used for most electronics
applications in which semiconductors are needed. The price of an integrated circuit can range from tens of
cents to several thousand dollars, with the average price of an integrated circuit currently running at about
$1.20-1.30. Microprocessors are one category of IC that are quite expensive, with Intels average
microprocessor price being more than $100 and the price of Intels more expensive server chips stretching
above $1,000. Memory chips typically have a price of a dollar to a few dollars. Analog chips, while generally
being highly profitable, tend to have low average selling prices (ASP), of the order of $0.30-0.50. Most of our
discussion in this report is centered on ICs since they account for the bulk of the industry, and most of the
major semiconductor companies are primarily IC companies.
Discrete semiconductors. Discrete semiconductors are single semiconductor devices (as opposed to
integrated circuits, which are made up of several devices all connected together on the same chip). Discretes
are used in many electronic applications, but one important use of discrete devices is in managing electric
power. Prices for discrete semiconductor devices range typically from a few cents to a dollar or more, with the
average price for a discrete being about $0.05. Discretes account for approximately 6% of total semiconductor
sales. Discrete semiconductors have been slowly declining as a percent of total semiconductor sales over the
years, in part, we think, because integrated circuits have been incorporating some of the functionality of
discrete devices.
Optoelectronics and sensors. Semiconductors that can be used to generate light (e.g., for displays, for
traffic lights, to drive communications signals along optical fibers) or sense light (e.g., in digital cameras) fall
into the category of optoelectronics. Generally the technology used to make optoelectronic devices is different
from that needed to make integrated circuits or what we have defined as discrete semiconductor devices, and
so there is, for the most part, a different set of companies that participates in this segment. Sensors are used to
sense temperature, pressure, acceleration (e.g., to activate the airbag in a car), and other things. As with
optoelectronics, this is a fairly small segment of semiconductors. A number of analog IC companies have
become interested in sensor technology since the output of sensors is typically handled by analog chips. Solar
cells are a potentially huge application for semiconductors, but are generally considered a different market
from optoelectronics and sensors. The revenue number we have used to calculate the 12% of semiconductors
does not include solar cell revenue, even though from an engineering point of view solar cells are, in fact,
optoelectronic devices. We do not discuss optoelectronics, sensors, or solar cells further in this report.
Semiconductor End Markets
Exhibit 3 shows semiconductor sales by end market.
Exhibit 3. Semiconductor End Markets (2013)
Data Processing
Compute+Storage,
39%
Industrial/Medical/
Other, 9%
Wireless
Communications,
24%
Military/Aerospace,
1%
Automotive, 8%
Wired
Communications,
6%
Consumer, 12%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Applications, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
Gartner estimates that computing (data processing--compute plus storage) accounts for roughly 39% of total
semiconductor sales. Computing tends to require leading-edge semiconductor technology, especially in
microprocessors and memory (DRAM). Nearly 300 million personal computers (PCs) shipped worldwide in
2013, and about 200 million tablets, we estimate. Server systems, while having lower volume (on the order of
10 million server systems shipped worldwide in 2013) have very high semiconductor content, including many
chips with high prices and high gross margin. For example, Intels Data Center Group (chips for servers and
other computer infrastructure systems) contributes about a quarter of Intels total revenue and nearly a third
of Intels operating profit. Tablets and convertible devices continue to grow as a source of semiconductor
demand. In 2013, Wireless communications (mostly wireless handsets) accounted for 24% of semiconductor
consumption and are a high-volume segment, as about 1.8 billion wireless handsets were shipped worldwide in
2013 (up 4% yr/yr). We think that as handsets become more sophisticated with the rise of smartphones, there
may be good opportunities for increasing semiconductor content per handset. However, we view the overall
handset market as being somewhat saturated, with a worldwide installed base of perhaps close to 4.6 billion
handsets.
Semiconductors
Consumer electronics accounted for 12% of total semiconductor consumption in 2013. This segment contains a
broad range of product such as camcorders, set top boxes, DVD players, TV, cameras, and video game
consoles. We view this as one of the slower growing segments of all the semiconductor end markets. As new
consumer devices crop up, older products fade.
Increasing semiconductor content is a driver for semiconductor growth in the automotive end market (about
8% of semiconductor consumption). Semiconductor applications in cars include sensors (e.g., every airbag has
a micromechanical semiconductor sensor that triggers when rapid deceleration occurs during a crash),
microcontrollers (e.g., antilock brakes), wireless (e.g., the signal that tells the car to lock its doors), electronic
power (e.g., the drivers that lock and unlock the doors), and displays (e.g., dashboard lighting). We believe the
growing popularity of on-board navigation systems and TV/DVD systems will provide ongoing momentum for
increasing content in automobiles. Electric cars also have substantial semiconductor content associated with
the handling of the electric power.
$10,000,000
$1,000,000
$100,000
Dec-77
Dec-78
Dec-79
Dec-80
Dec-81
Dec-82
Dec-83
Dec-84
Dec-85
Dec-86
Dec-87
Dec-88
Dec-89
Dec-90
Dec-91
Dec-92
Dec-93
Dec-94
Dec-95
Dec-96
Dec-97
Dec-98
Dec-99
Dec-00
Dec-01
Dec-02
Dec-03
Dec-04
Dec-05
Dec-06
Dec-07
Dec-08
Dec-09
Dec-10
Dec-11
Dec-12
Dec-13
Dec-14
$100,000,000
Monthly Sales
12/85-12/00 15%/yr
Semiconductor sales grew at a rate of about 15% per year from 1985 to 2000. Although there is, we think, a
widely held belief that there was excessive buying of technology-related goods in 1999 and 2000, the graph
does not show semiconductor sales much above the longer term trend line in 1999 or 2000. From 2001
through today (2014), revenue growth for the semiconductor industry has been far lower, following a trend of
about 4% per year growth. As we discuss below, we believe that the true underlying long-term sales growth
rate for semiconductors may be between these two numbers, of the order of about 10% per year.
32,000,000
30,000,000
28,000,000
26,000,000
24,000,000
22,000,000
20,000,000
18,000,000
16,000,000
14,000,000
12,000,000
10,000,000
8,000,000
6,000,000
4,000,000
2,000,000
-
Dec-77
Dec-78
Dec-79
Dec-80
Dec-81
Dec-82
Dec-83
Dec-84
Dec-85
Dec-86
Dec-87
Dec-88
Dec-89
Dec-90
Dec-91
Dec-92
Dec-93
Dec-94
Dec-95
Dec-96
Dec-97
Dec-98
Dec-99
Dec-00
Dec-01
Dec-02
Dec-03
Dec-04
Dec-05
Dec-06
Dec-07
Dec-08
Dec-09
Dec-10
Dec-11
Dec-12
Dec-13
Dec-14
Exhibit 5 shows:
Semiconductor sales recovered steadily from 2002 through 2006 following the downturn of 2000-01, with
sales reaching year-2000 highs in H2 2004.
The years 2007 and 2008 were ones of modest growth for semiconductors.
Semiconductor demand dropped sharply at the end of 2008 and early 2009 with the global
macroeconomic issues that emerged during this period.
The industry recovered as 2009 progressed and managed to end 2010 with record volume.
Global semiconductor sales were flat year over year in 2011 and down 3% in 2012, even though the global
economy grew 4% in 2011 and 3% in 2012.
Semiconductor sales rebounded in 2013, growing 5% year over year, compared to global growth of about
3%. However, the growth in 2013 was largely a reversal of the decline in 2012.
As electronics continue to penetrate further into essentially every aspect of human activity, we think it is likely
that semiconductor growth will in general pace above global GDP growth. From 2011 through 2013,
semiconductor industry growth in total for the three years was 2%, far below the 10% total expansion of global
GDP. We think this may indicate that there is some pent-up demand for a range of electronics goods that
could be realized as the global economy continues to recover. For the period January-September 2014 (the
most recent data available at the time of writing this report), global semiconductor sales rose 10% year over
year.
Exhibit 6 shows that an interesting picture emerges when we look at unit shipments, as opposed to sales in
dollars. In discussing unit trends, we look at integrated circuit (IC) unit numbers, rather that total
semiconductor unit numbers since otherwise the very low-priced discrete chips would distort the trends.
Discrete semiconductors have an average selling price (ASP) of about $0.05, versus the IC ASP, which is
currently close to $1.30. Therefore, discrete semiconductors account for a far larger percentage of unit
shipments than their proportion of economic value to the semiconductor industry.
Semiconductors
Exhibit 6. Worldwide Integrated Circuit (IC) Unit Shipments
10,000,000
1,000,000
100,000
Dec-78
Dec-79
Dec-80
Dec-81
Dec-82
Dec-83
Dec-84
Dec-85
Dec-86
Dec-87
Dec-88
Dec-89
Dec-90
Dec-91
Dec-92
Dec-93
Dec-94
Dec-95
Dec-96
Dec-97
Dec-98
Dec-99
Dec-00
Dec-01
Dec-02
Dec-03
Dec-04
Dec-05
Dec-06
Dec-07
Dec-08
Dec-09
Dec-10
Dec-11
Dec-12
Dec-13
Dec-14
100,000,000
1/85-12/00 10%/yr
IC unit shipment growth of 10% per year from 1985 to 2000 was 6 percentage points less than sales
growth in dollar terms, which implies pricing expansion during this period.
In contrast to semiconductor sales in dollar terms, while unit shipments did drop sharply in 2001, the
recovery from 2002 to 2007 pulled unit shipment levels back to the longer term trend line extrapolated
from 1985 to 2000. Following the economic downturn of 2008/2009, unit shipments move back up in
2010 to the long-term trend line we have shown in our graph. Hence, for the 25-year period from 1985
through 2010, there was been a consistent trend of solid unit growth of 10% per year for semiconductor
units. Admittedly, the unit numbers have dropped below our trend line in the past 3-4 years, though, as
discussed above, we believe that this is a sign that some amount of pent-up demand has been created,
rather than it being a reflection of the underlying trend of global chip growth suddenly slowing after 2010.
The discrepancy between sales growth and unit growth from 2000 to 2010 was the result of falling ASPs.
As we discuss further on, IC ASP today is close to where it was in 1990. The decline from 2000 to 2010
roughly matched the expansion from 1990 to 2000. The 20+ year trend does not show any obvious ASP
decline.
We conclude that there is good reason to expect that the semiconductor industry should continue to achieve IC
unit growth of 10% per year, with potentially some amount of catch up in the near term from the softness of
2011-13. We expect this unit growth to be driven by a combination of solid unit growth in the major
semiconductor end markets, as well as increasing semiconductor content in a number of key markets such as
wireless handsets and automotive electronics. As discussed below, we think that over the next few years, IC
ASP could well be flat to up, which, combined with unit growth of 10% per year or better, could result in overall
semiconductor industry sales growth in the 10-12% per year range over the next several years.
10
Seasonality
In the past, PC sales showed a clear seasonal pattern, with H2 of each year typically being significantly stronger
than H1. March quarter shipments are often down sharply from those of the preceding December quarter. For
semiconductors as a whole, June quarter shipments tend to be slightly up from March quarter shipment levels,
though in some segments such as for, example, PC-related chips, June sales can be flat to down from March.
September quarter sales are usually up sharply from June quarter sales, and there is often another increase in
sales from the September quarter to the December quarter, driven by the seasonal pattern for electronics
products such as consumer PCs, wireless handsets, and other consumer electronics. Since chip purchases and
system builds occur before the systems are sold, though some chip companies see demand peaking at some
point in the September-November time frame, and so some chip companies typically see muted sequential
growth or even slight sequential declines in the December quarter.
The seasonality of wireless handset sales is similar to that of PCs, with some minor differences. The June
quarter is typically stronger than the March quarter (whereas for PCs, the June quarter is similar to or weaker
than the March quarter in most years). As with PCs, the September quarter is stronger than the June quarter
for wireless handsets, the December quarter is stronger than the September quarter, and the March quarter is
down from the preceding December quarter. The monthly global semiconductor shipment data released by the
Semiconductor Industry Association (SIA) typically shows a month-over-month decline in the month of
December.
The market for communications infrastructure chips often shows softness in the September quarter, with
September sales being flat or down sequentially compared to June. This is to some extent associated with a
drop in demand in Europe, perhaps associated in part with summer vacations in Europe.
Some segments of the industrial end markets show strength in the beginning of the year, with the March
quarter being up sequentially over the December quarter.
There is a delay between the purchase of semiconductor components and the sale of the systems (such as PCs
and wireless handsets) in which the semiconductors are used. We assume that the offset between
semiconductor sales and systems sales is in aggregate about 1-2 months. This results in slight differences
between semiconductor seasonality and end-market seasonality.
The Semiconductor Cycle
In the past, the semiconductor cycle has been supply driven, not demand driven (see Exhibit 7). On the supply
side, we believe the capital-intensive nature of semiconductor manufacturing and the time lag between
investing in new capacity and actually being able to use the new capacity is responsible for this pattern:
A new state-of-the-art fab can cost $6 billion or more to construct and equip today.
The building construction for a new fab might take six months to a year. Moving in the manufacturing
equipment and getting it ready to process semiconductors (i.e., qualification of the equipment) takes
several additional months.
On the demand side, in our view, the major semiconductor end markets have not shown any obvious cyclical
behavior, apart from seasonal patterns over the course of each year and swings associated with macroeconomic
downturns and recoveries.
In theory, when capacity is tight, semiconductor pricing should rise causing an acceleration of semiconductor
growth in dollar terms. Tight capacity also causes end customers and distributors to increase days of
semiconductor inventory held. It takes time to make the decision to invest in more capacity and then to bring
that capacity online. When the additional capacity comes online (and in particular, if excess capacity is
created), pricing moderates, growth decelerates, and end customers decrease days of semiconductor inventory
held.
11
Semiconductors
Yr/Yr Growth
40%
20%
0%
(20%)
(40%)
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
Dec-01
Dec-00
Dec-99
Dec-98
Dec-97
Dec-96
Dec-95
Dec-94
Dec-93
Dec-92
Dec-91
Dec-90
Dec-89
Dec-88
Dec-87
Dec-86
Dec-85
Dec-84
(60%)
Source: Semiconductor Industry Association and Wells Fargo Securities, LLC estimates
Exhibit 7 shows the semiconductor cycle. The solid line represents yr/yr sales growth in dollars and the crosses
show yr/yr unit growth.
The period from 1985 to 1999 shows two complete cycles. The cycle shows up in the yr/yr growth in dollars.
Although unit growth does fluctuate, it does not show the same clear cyclical behavior as sales growth in
dollars.
We believe that the period from 2000 to mid-2003 is unusual and not part of the cyclical pattern because of
the impact of unusual softness in worldwide economies.
In the 1985-1999 period, the length of a full cycle was about seven years. This was made up of about 3.5 years
of strong growth in dollar sales ranging from 20% to 40% per year, and about 3.5 years of weaker growth (or
even declines) in the negative 20% to positive 20% range.
The cycle is driven by dollars sales growth being higher than unit shipment growth (the continuous line is, for
the most part, above the crosses in the graph). This implies pricing expansion. In the negative phase of the
cycle, from 1989 to 1991, sales growth was comparable with unit shipment growth; pricing had a neutral
impact. On the other hand, in the downturn from 1995 to 1998, dollars sales growth was lower than unit
shipment growth (continuous line below the crosses), implying a decline in pricing.
We interpret unit shipment growth as an indicator of overall demand. Unit shipment growth, and by
implication, demand, does not show any obvious cyclical behavior. This is consistent with the absence of
cyclical behavior in PC shipments and any of the other major semiconductor end markets.
In the past, the cyclical pattern for semiconductors was driven by pricing. Given no clear cyclical pattern in
demand, we conclude that semiconductor pricing patterns are far more a function of availability of supply
(capacity utilization) than a result of fluctuating demand. In the next subsection we discuss pricing and in the
following subsection we look at capacity.
In H2 2003 it began to look (at the time) as if the industry was returning to normal cyclical behavior. Units had
been growing steadily in a more or less normal seasonal pattern since early 2002 and in H2 2003, yr/yr sales
growth in dollars overtook unit growth; pricing was expanding. We had expected the industry to go into a three
and a half-year expansion phase, driven by pricing expansion. However, in H2 2004, semiconductor unit
growth started to decline on a month-over-month basis, the result of an inventory correction stemming from
too much chip buying in H1 2004. A mid-cycle correction in the yr/yr growth pattern for unit shipments is
not unusual; it can be seen that there was a dip in unit growth in early 1987 and in 1994, in the middle of the
high-growth phases of the previous two cycles. What was unusual this time, though, was the crossover of sales
12
growth and unit growth following this correction. Pricing declined and yr/yr pricing comparisons turned
negative around mid-2005 and remained negative in the next three years through 2008.
Then, in H2 2008, the global economic downturn drove down semiconductor unit demand sharply. This was
not a semiconductor-specific event but instead, a broader issue affecting all segments of the economy. Growth
picked up sharply in H2 2009 before returning to more normal levels in 2010. So far in 2014, growth has been
moderate, below the 10% growth trend.
The trends of the past 12 years raise the question as to whether there has been any change in the cyclical
behavior of the semiconductor industry. Some things have not changed:
Semiconductor manufacturing remains capital-intensive, although the transition to 300mm has resulted
in a small step down in investment percent over the past few years.
As an example of the costs associated with these new facilities, Intels Fab 42 (for 14nm chips on 300mm
wafers) is expected to cost more than $5 billion to construct. In addition, with the move to ever larger
wafer sizes, the minimum reasonable size of a fab has soared 10-100x over the past 15 years.
Capacity in general has to be brought on in fairly large chunks, especially when a new facility has to be
constructed. Despite this, it still takes a fair amount of time to actually bring up new capacity. It can take
1-2 years to build, equip, and bring up a fab from scratch. Just to expand production in an existing facility
by moving in more production equipment typically takes a minimum of six months or so.
However, some things have changed, and those changes could reasonably be expected to dampen (improve)
the cyclical behavior of the semiconductor industry in the future:
More and more semiconductor companies are fabless, having their manufacturing done by third parties,
that is, the semiconductor foundries. This decouples the capital intensity from the companies selling the
chips. Tightness in capacity leading to chip shortages is still a mechanism that can drive pricing upward.
However, the fabless chip companies are buffered from problems of excess capacity. It is someone elses
problem. Therefore, excess capacity at the foundries does not immediately lead to a collapse of chip
pricing in an effort by the owners of the capacity to fill up their excess capacity.
It appears that in recent years, semiconductor manufacturers have become far more profit-conscious and
cautious about the risk of creating excess capacity. In the decade of the 1990s, we believe the top priority
at many major semiconductor companies was to drive sales growth. We believe that in recent years there
has been far more focus among chip company management teams on improving profit margin. This
change in priority in the industry has led to a lot more focus on capital efficiency and a more rapid
response to cutting back on capital spending when business conditions have weakened.
Our graph shows that since the year 2000, there has been no multiyear sustained period in which pricing
expansion (sales growth tracking above unit growth) has driven the upward leg of a semiconductor-specific
cycle. It appears as if the cyclical component of the semiconductor industry has disappeared,
with no obvious sign of a semiconductor specific cycle (apart from economic cycles that have
affected all industry) over the past 14 years. We believe that there will be more consistent growth and
less cyclical behavior in the semiconductor industry in the future than there was in the past. This is a positive
development, in our view, as it should allow semiconductor companies to be more efficient in their use of
capital and make it easier to plan for growth.
13
Semiconductors
Pricing
Exhibit 8 shows ICs currently have an ASP of around $1.25, while discrete semiconductors are much less
expensive, with an average price of about $0.05. Prices for individual ICs range from tens of cents to hundreds
of dollars. Prices for discrete semiconductor products range from cents to tens of cents.
Exhibit 8. Total Integrated Circuit Average Selling Price
$3.50
$3.00
$2.50
$2.00
$1.50
$1.00
$0.50
$0.00
The very big discrepancy between discrete and IC ASP means that discretes have a far bigger impact on overall
semiconductor unit shipments than their true economic value warrants. This is why we normally analyze and
discuss IC unit shipment trends rather than total semiconductor shipment trends (although the two do tend to
track). Even within ICs there is a broad range of prices and therefore, commentary on even IC unit shipments
can sometimes be misleading. With IC ASPs currently a little above $1, Intels desktop and notebook
microprocessors have an average selling price of about $130. Therefore, Intel, the worlds largest
semiconductor company, has a disproportionately small impact on worldwide semiconductor unit shipments
than some other companies that are far smaller than Intel in terms of total sales.
Blended IC ASPs:
Soared from 1990 to 1995, to as high as about $3.00, driven by high dynamic random access memory
(DRAM) prices;
Dropped from 1995 to 1997 (DRAM price correction), but then rose again from 1997 to 2000, to close to
$2.00;
Rose in 2004, only to fall back in 2005 and drift down from 2005 through 2008.
From 2008 through the present (late 2014), IC ASPs have remained fairly stable, running in the $1.201.50 range.
14
Some factors that have affected semiconductor ASP at various points in the past include memory price
changes, a dip in semiconductor demand, and the rise of microprocessors.
Memory prices drove ASP up from 1990 to 1995. DRAM had an anomalous period from 1991 to 1995, in
which prices deviated from the long-term trend, in the positive direction. Our numbers show that from 1990 to
1995, DRAM ASP rose to nearly $9 from below $4. Jumping forward to Exhibit 50, it can be seen that price per
bit (discussed in some detail further on) for DRAM has generally followed a decline of 35% per year, but price
per bit remained flat from 1991 to 1995. This helped drive overall IC ASPs upward from 1991 to 1995 (high
growth in memory bits shipped with no price-per-bit decline). When memory prices suddenly corrected back
to the longer term trend (a 35% per year price-per-bit decline), this led to a sharp IC ASP decline from 1995 to
1997. By 1999, memory ASP was back to about $3.50, close to where it had been in 1990. The spike in IC ASPs
from 1990 to 1995 can largely be explained (as far as memory played a role) as a memory pricing, rather than a
memory mix, effect.
Increasing microprocessor mix drove ASP up from 1990 to 2000. Although memory pricing had a
huge positive impact on IC ASPs in the first half of the 1990s, the effect was reversed in the second half and so,
over the course of the whole decade, pricing for memory was neutral to IC ASPs. The one thing we can identify
that had a clear and permanent impact on driving up IC ASPs from 1990 to 2000 is microprocessors. From
1990 to 2000, the rise of the PC led to the emergence of microprocessors as a major semiconductor category.
Microprocessors have much higher prices than most other chips, on the order of $40.00-120.00 today ($200+
in the 1990-2000 time frame) for high-volume desktop and notebook microprocessors. This had a positive
influence on IC ASPs from 1990 to 2000. Microprocessors are just one sub-segment of overall logic and PCmicroprocessors are a sub-segment of what the Semiconductor Industry Association (SIA) classifies as
microprocessors. From 1990 to 1999, logic ASPs rose to about $2.80 from $1.30. In Exhibit 9 it can be seen
that logic did grow some as a percentage of the total IC dollar mix from 1990 to 2000, but really, the big
change was in logic ASPs. Within logic, microprocessor ASPs soared, as triple-digit PC microprocessor prices
grew to dominate the microprocessor category, driving up microprocessors (MPU) as a percentage of the total
logic mix (see Exhibit 10). From all this we can see that although on the surface it appears that the contribution
of logic to rising IC ASPs through the 1990s was a pricing effect, it was in reality a mix effect within logic. As
high-priced microprocessors grew to represent a greater percent of the overall logic segment, they drastically
increased the pricing impact that logic had on overall IC ASPs.
A sharp downturn (bursting of the tech bubble) hit pricing in 2001. In 2001, semiconductor
demand plunged by more than 20% yr/yr, producing excess capacity, which then caused prices to decline (see
Exhibit 11).
Rising handset mix diluted the positive effect of microprocessors from 2003 to 2008. Even
though the semiconductor industry did benefit from the recovery of 2003-08, with the number of unit
shipments rising in this period and capacity utilization running at a comfortable 85-95%, IC ASPs continued to
trend downward slowly, to about $1.30 in 2008 from about $1.50 in 2001. We think that this trend was in part
due to slight mix shifts. For example, wireless handsets were an important end market, driving chip growth
from 2001 to 2008. There are some chips in handsets that sell for several dollars, though this is nothing like
the PC-driven microprocessor dynamic of the 1990-2000 period, with microprocessors selling for several
hundred dollars becoming a larger part of the mix. Exhibit 10 shows that microprocessor sales dropped as a
percentage of total logic sales, to 28% by 2009 from 35% in 2001.
15
Semiconductors
ASP stability from 2009 through 2013. The sharp downturn toward the end of 2008 did not have the
same large negative impact on IC ASPs as the downturn of 2001 had. Blended IC ASP has been relatively stable
over the past four years, despite rapidly recovering demand through 2009 and 2010, followed by the inventory
correction in the second half of 2011 and broad-based economic softness in 2011 and 2012. We think this is
probably a result of more focus on management of capacity and capacity investment in the semiconductor
industry, and also, the increasing proportion of chips sold by fabless companies.
Over the next few years, we think that some important drivers for semiconductor growth include smartphones,
tablets, and servers, all of which should help improve mix, leading to a blended IC ASP trend that we think
could be flat to up for some extended period of time. On the other hand, further into the future, the
proliferation of lower cost electronics for Internet-of-things applications could pressure blended IC ASP.
Smartphones continue to grow as a percent of wireless handsets overall. Gartner data suggest that in 2013,
smartphones accounted for about 54% of total handsets shipped worldwide, up from 39% in 2012.
Smartphones have an applications processor, while lower end wireless handsets do not. The memory
content of a smartphone (especially NAND flash) is generally higher than in a standard wireless handset.
Tablets contain applications processors and flash memory chips, both of which typically have prices that
are a fair amount above the IC average.
Servers have a particularly rich mix of chips. A high percentage of servers have two or more
microprocessors, with ASPs in the $600+ area. Servers also tend to have a large amount of high-density
DRAM.
Memory, 27%
Logic, 57%
Analog, 16%
2000
Memory,
28%
Analog,
17%
Logic,
55%
1990
Memory,
30%
Analog,
19%
Logic,
51%
Source for all charts: Semiconductor Industry Association and Wells Fargo Securities, LLC
16
2013
Gen. MOS
Logic &c.,
59%
1999
General
MOS Logic,
31%
DSP, 2%
MCU, 10%
MPU, 28%
Other, 1%
MPR, 14%
1990
MCU, 19%
MPU, 36%
MPR, 14%
Other, 17%
MCU, 19%
MPU, 11%
Source for all charts: Semiconductor Industry Association and Wells Fargo Securities, LLC estimates
17
Semiconductors
Exhibit 11. Worldwide Semiconductor Capacity (millions of square inches) And Capacity Utilization
120.0
3,000.0
100.0
2,500.0
80.0
2,000.0
60.0
1,500.0
40.0
1,000.0
20.0
500.0
0.0
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Gartner, Inc.: Forecast: Semiconductor Wafer Fab Capacity, Worldwide, 2Q14 Update, David Christensen et alia
Exhibit 11 shows worldwide semiconductor capacity and capacity utilization. We consider capacity utilization
of 90% and higher to be a healthy level. The capacity data of Exhibit 11 are described in terms of millions of
square inches. Since the transition to different wafer sizes often takes place over the span of years, at any given
point in time the semiconductor industry as a whole is doing its manufacturing on more than one wafer size.
Currently, most manufacturing is done on 200mm (8 inch) and 300mm (12 inch) wafers, but there still
remains some manufacturing on 6-inch and even some 5-, 4-, and 3-inch wafer processing. To normalize all
these different wafer sizes, capacity numbers (and other things associated with wafers, for example, prices
charged per wafer by foundries) are often couched in wafer equivalent numbers. A 300mm wafer has 2.25x
the area of a 200mm wafer, and hence, 2.25 more chips can be fabricated on such a wafer. Therefore, a
300mm wafer would count as 2.25 equivalent 200mm wafers when adding up wafer capacity.
The data in Exhibit 11 are an aggregate of utilization rates across the semiconductor industry. However, the
sensitivity of pricing to capacity utilization is different for different segments of the industry. The most
sensitive segment is the memory segment since memory is a price-elastic commodity. However, although it is
possible to get qualitative commentary on capacity utilization from the memory makers and some analysts
track fab capacity availability and expansion at memory makers, so many factors appear to play into memory
pricing that we question how fruitful it is to try to quantitatively predict future memory prices from capacity
analysis.
Low capacity utilization does often result in chip pricing pressure. Semiconductor ASPs fell from 1997 to 1998,
as did capacity utilization, and the entire semiconductor industry was hit by a big downturn in 2001, which
resulted in falling capacity utilization and falling pricing. However, in the most recent downturn toward the
end of 2008, although memory prices plunged in H2 2008, prices of other chips remained fairly stable.
Memory prices rebounded in 2009, even as capacity utilization hit a low in the March 2009 quarter. We think
that this was partly the result of aggressive action taken by the various participants of the electronics supply
chain to cut production and inventory. This was effective in preventing problems of excess inventory.
18
90%
85%
80%
75%
70%
65%
60%
55%
50%
Mar-92
Sep-92
Mar-93
Sep-93
Mar-94
Sep-94
Mar-95
Sep-95
Mar-96
Sep-96
Mar-97
Sep-97
Mar-98
Sep-98
Mar-99
Sep-99
Mar-00
Sep-00
Mar-01
Sep-01
Mar-02
Sep-02
Mar-03
Sep-03
Mar-04
Sep-04
Mar-05
Sep-05
Mar-06
Sep-06
Mar-07
Sep-07
Mar-08
Sep-08
Mar-09
Sep-09
Mar-10
Sep-10
Mar-11
Sep-11
Mar-12
Sep-12
Mar-13
Sep-13
Mar-14
95%
Source: Federal Reserve--Industrial Production/Capacity Utilization Release and Wells Fargo Securities, LLC
The U.S. Federal Reserve releases, on a monthly basis, capacity utilization data for semiconductor and other
electronic component manufacturing in the United States. We have plotted this information in Exhibit 12. The
U.S. data are interesting because they are presented monthly rather than quarterly. They generally reflect
similar trends with the Gartner numbers. However, a substantial amount of semiconductor manufacturing is
done outside the United States. Also the U.S. data include many other types of manufacturing besides
semiconductor wafers, such as printed circuit boards, capacitors and resistors, electronic coils and
transformers, electronic connectors, and other electronic components.
140%
7,000,000
120%
6,000,000
100%
5,000,000
80%
4,000,000
60%
3,000,000
40%
2,000,000
20%
1,000,000
0%
TSMC+UMC Capacity
19
Semiconductors
The Taiwanese semiconductor foundries, TSMC and UMC, release quarterly numbers from which their
capacity utilization can be calculated. Capacity utilization for both companies is plotted in Exhibit 13. The
Taiwanese foundries are the only major semiconductor companies of which we are aware that routinely report
specific capacity utilization information when they release earnings. Foundry capacity utilization tends to show
much larger swings than overall worldwide capacity utilization. This makes sense for two reasons:
Some chip companies do their own manufacturing, as well as use foundries for making similar chips.
During downturns, companies that have a choice between manufacturing in their own facilities, versus
using foundries inevitably opt to maximize the use of their own fabs.
Companies that do their own manufacturing, i.e., independent device manufacturers (IDM), tend to
balance the risk of creating excess inventory against the costs of underused fabs. Fabless companies do not
have to concern themselves with holding up utilization at their foundries, and therefore, would be
expected to cut wafer orders more aggressively in downturns than IDMs.
2010
$5.20
$0.95
2011
$10.8
$2.9
2012
11.0
$1.9
2013
$10.7
$1.24
2014E
$10.5-$11.5
$2.8-$3.2
2015E
$3.6-$4.0
Micron3 (Aug FYE)
Samsung
$6.60 $7.80 $6.90 $3.50
$11.0
$11.6 $13.4 ~$11.3
~$13.8
TSMC
$2.50 $2.60 $1.89 $2.67
$5.90
$7.3
$8.3
$9.7
~$9.6
~$10
UMC
$1.00 $0.90 $0.35 $0.50
$1.80
$1.6
$2.0
$1.1
$1.3
Hynix
$5.10 $1.90 ~$0.8 $3.00
$3.3
$3.7
~$3.2
~$3.8
Globalfoundries4
$5.4
$9-10
1 Date last affirmed
2Intels 2013 capital spending included about $1.3 billion for 450mm spending; we think this could be about the same for 2014.
3Increase in Microns spending in 2014 is driven in part by Microns acquisition of Elpida.
4Globalfoundries has indicated that it has $9-10B capital expenditure planned in 2014-15.
Source: Company reports and Wells Fargo Securities, LLC
One way to estimate future capacity increases is to look at the capital expenditure (capex) of the semiconductor
manufacturing companies. Exhibit 14 lists some of the top semiconductor companies in capital spending
planned for 2014/2015.
In recent years the larger chip manufacturers, Intel, Samsung, and TSMC, typically have run
(semiconductor) capex in the $8-14 billion per year range.
TSMC began aggressively ramping up its capex in 2010 and has continued to drive up capex through 2013.
In 2014 and 2015 it appears that TSMCs capex might be plateauing at close to the 2013 level.
Although Intel has noted that difficulties with lithography (imaging) for the most advanced technologies
may well result in a need for higher levels of capital spending in the near term. On the other hand, while
Intels capital spending more than doubled in 2011 from 2010 levels, Intel has been tracking on a plateau
of close to $11 billion in capital spending in each of the past four years, 2011-14.
20
Date1
Oct 14, 2014
Aug 5, 2014
Oct
Oct
Oct
Oct
Oct
30,
16,
29,
23,
20,
2014
2014
2014
2014
2014
30%
25%
20%
15%
10%
5%
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
Dec-01
Dec-00
Dec-99
Dec-98
Dec-97
Dec-96
Dec-95
Dec-94
Dec-93
Dec-92
Dec-91
0%
Dec-90
35%
Source: Semiconductor Industry Association, SEMI, and Wells Fargo Securities, LLC estimates and calculations
A more specific indicator for how semiconductor capacity might change in the near future (and to get a feel as
to whether capacity utilization will remain tight, driving up pricing) is the purchases of semiconductor
equipment. Exhibit 15 shows worldwide semiconductor manufacturing equipment sales as a percent of
worldwide IC sales (basically a measure of the rate of reinvestment). The IC manufacturers account for the
bulk of semiconductor equipment purchases, which is why we use IC sales, rather than total semiconductor
sales to normalize the semiconductor equipment numbers.
In principle, spending below the equilibrium ratio implies a reduction in capacity and potentially, higher
capacity utilization, while spending above the equilibrium point implies an increase in capacity and potentially,
a fall in capacity utilization (if the capacity is growing faster than demand. However, the equilibrium ratio has
shifted over the years, driven by a number of major secular trends:
(1)
The ratio of semiconductor equipment spending as a fraction of IC sales rose from 1990 to 2000. In fact,
although we have not plotted data prior to 1990 on the graph (for lack of a consistent data set), the ratio
of equipment spending to IC sales rose through the first three decades of the life of the semiconductor
industry, from 1970 to 2000. Some of this was due to the fact that semiconductor equipment makers
continually increased their contribution to the semiconductor manufacturing process. One example of
this is automation. In the late 1980s, almost all operators had to manually load wafers onto machines,
and wafers had to be carried from one machine to the other by operators. By the end of the 1990s,
almost all semiconductor equipment had precision robotics that automatically loaded wafers into the
machines, and some fabs had installed fab-wide automation, in which wafers could be mechanically
moved from one machine to the next. Robotic handling is expensive. The price of some machines
increased as much as tenfold, to millions of dollars from hundreds of thousands of dollars. This
increased the capital intensity of the semiconductor business. The cost of constructing and equipping a
semiconductor fab rose to hundreds of millions of dollars in the 1990s from tens of millions, to the
current level of several billion dollars for a leading-edge semiconductor fab. However, it does not follow
that the profitability dropped or that the absolute cost of making semiconductor chips rose through
time. For example, automation results in higher semiconductor equipment cost, but reduced labor cost.
21
Semiconductors
(2) An opposing trend that should in principle decrease the ratio of equipment purchases to chip sales is the
increase in wafer size. Over time, semiconductor manufacturers have increased the size of the wafers used
to make semiconductor chips. In the early 1990s, manufacturers were shifting to six-inch wafers
(diameter) from five-inch. In the mid-1990s, the transition to eight-inch (200mm) wafers from six-inch
wafers took place. Over the past few years, manufacturers have been moving to 300mm (12 inch) wafers
from 200mm (8 inch) wafers. The increase in the cost of a machine that can deal with a larger wafer size is
in principle less than the increase in amount of wafer area. However, up to the year 2000, the increasing
functionality of the machinery had a much bigger impact than the increasing wafer size, and so the ratio of
semiconductor equipment spending to semiconductor sales rose. In recent years, semiconductor
manufacturing techniques have reached a relatively mature phase (although line-width transitions occur
on the same regular schedule that they have historically, as we discuss in more detail in the technology
section). We believe that the current 300mm transition is resulting in reduced capital costs. A 300mm
wafer has more than 2x the area of a 200mm wafer, and so, more than twice the number of chips is made
on a 300mm wafer than on a 200mm wafer.
(3) Semiconductor manufacturing has become more concentrated in a relatively small number of large
players. Semiconductor foundries have grown faster than the chip industry overall as more companies
become fabless, depending on foundries for their chip manufacturing. In addition, we think that over the
past decade or so, the chip industry as a whole has shifted its priorities to focus more on profitability and
capital efficiency than on raw growth. We expect that in the future, these factors could well lead to more
pricing pressure on semiconductor equipment, as well as less duplication of equipment purchases, leading
to more stable and higher capacity utilization in the semiconductor industry.
(4) In recent years a number of chip companies have suggested that transitions to more advanced
technologies are becoming increasingly difficult to make, which might increase the capital intensity of the
semiconductor industry. In particular, as we discuss in more detail in the Selected Technology Topics
section of this report, difficulties with transitioning to a different color of light (shorter wavelength EUV)
in the cameras used to image semiconductor patterns has led to the need for multiple patterning. Multiple
patterning involves the use of the most expensive pieces of semiconductor equipment, the lithography
steppers, two or three times instead of once to transfer a single pattern layer onto a semiconductor wafer.
From Exhibit 15 it can be seen that in the early 1990s, semiconductor equipment spending was, on average,
running at about 15% of IC sales. This number rose to more than 20% (with great volatility) in the late 1990s.
In response to the downturn of 2001, the chip industry held down capital investment through 2002 and 2003,
though the investment ratio rose in 2004 as demand strengthened. We think that many investors are
under the mistaken impression that technology difficulties have led to the rise of capital
intensity of the semiconductor industry in recent years and will continue to rise in the future.
We think that the data show a long-term trend in the ratio of semiconductor equipment
spending to IC sales that has been sloping downward from 2000 through today (2014). We
believe this trend toward a lower percentage of capital reinvestment will continue through the
next several years.
We think it is interesting, though, that the chip industry began to drive down the ratio of manufacturing
equipment purchases to semiconductor sales, beginning in mid-2007, and with the downturn of 2009, the
ratio fell to 7-8% in H1 2009, lower than at any other point in the preceding 19 years. While the ratio
rebounded from mid-2009 to mid-2010, we think the huge and rapid drop in spending in H1 2009 shows the
heightened sensitivity that chip companies have developed to the risks of excess capacity and excess inventory.
The very low ratio of reinvestment seen in 2008 and in H1 2009 is almost certainly far below the maintenance
capex level, i.e., the level of capital spending needed for replacement of obsolete or aging equipment and other
maintenance functions. As shown in the preceding Exhibit 11, worldwide chip capacity started to fall in
December 2008 and dropped substantially in H1 2009. Since it takes time to install and qualify new
equipment (on the order of weeks to months), capacity changes in any given quarter are a reflection of
equipment purchases in previous quarters. We look at the fact that chip equipment investment levels were very
low in H1 2009 as part of the reason for the declines in chip capacity through H2 2009 and into H1 2010.
Similarly, rising, but still modest spending in 2010, led to worldwide chip capacity holding steady through
most of 2010, but rising slowly toward the end of 2010 and into 2011.
22
Bookings
Billings
Feb13
Dec13
Apr12
Jun11
Oct09
Aug10
Feb08
Dec08
Jun06
Apr07
Oct04
0.00
Aug05
$0
Feb03
0.20
Dec03
$500
Jun01
0.40
Apr02
$1,000
Oct99
0.60
Aug00
$1,500
Feb98
0.80
Dec98
$2,000
Jun96
1.00
Apr97
$2,500
Oct94
1.20
Aug95
$3,000
Feb93
1.40
Dec93
$3,500
Jun91
1.60
Apr92
$4,000
Book/Bill
NorthAmericanChipEquipment($MM)
Book:Bill
Source: Semiconductor Equipment and Materials Institute and Wells Fargo Securities, LLC
Exhibit 16 shows North American semiconductor equipment bookings and billing (shipment) data as released
by the Semiconductor Equipment and Materials Institute (SEMI). Our interest from a semiconductor point of
view is more in the worldwide semiconductor equipment purchases, and we have used worldwide data for
Exhibit 15 in calculating the ratio of semiconductor equipment purchases to IC sales. However, the North
American numbers track the worldwide numbers closely as far as overall trend goes. The North American data
are issued by SEMI some weeks ahead of the worldwide data, and we believe that it is the more widely tracked
and discussed information and so, we have shown it here.
As shown in Exhibit 4, by 2004, semiconductor revenue had recovered to reach the peak first hit in 2000 and
then continued to rise from 2004 through 2007. In 2010, semiconductor revenue passed its prior 2007 peak
and was more than 20% higher than during the year-2000 peak, and chip sales are on track to hit a new peak
in 2014. However, Exhibit 16 shows that in 2014, North American semiconductor equipment purchases are
running at less than half their year 2000 levels. Worldwide semiconductor equipment data show a similar
qualitative trend, with worldwide semiconductor equipment sales in 2014 being about a third below year 2000
levels. The quantitative difference in the U.S. versus global semiconductor equipment trend arises in part
because some important segments of semiconductor equipment with strong secular growth characteristics, in
particular, lithography (machines that project the circuit patterns), are served by non-U.S. companies (e.g.,
ASML in the case of lithography).
Exhibit 16 also shows how aggressively the chip industry responded to the downturn of 2008, by cutting
semiconductor equipment purchases in early 2009 down to levels comparable to what was last seen in the
early 1990s. Similar caution can be seen in the sharp fall of semiconductor bookings and semiconductor
equipment book-to-bill in mid-2011 and mid-2012, in response to ongoing softness in semiconductor endmarket demand.
23
Semiconductors
The Rise Of The Foundries
The semiconductor foundries are companies that manufacture (process) semiconductor wafers for fabless chip
companies. The trend to outsource manufacturing is growing steadily. In 2002, foundries accounted for 9-11%
of all semiconductor processing (in terms of wafer-equivalent starts), a percentage that had risen to about
14-15% by H1 2008. Foundry wafer production grew at a 21% per year rate from 2002 through mid-2008,
while overall semiconductor wafer production grew 11%. By 2013, foundries were responsible for about 20%30% of all worldwide semiconductor production.
The trade group Semiconductor International Capacity Statistics (SICAS) gathered from information provided
by the member semiconductor companies of SICAS. Unfortunately, through 2011, a number of the regular
contributors of data to SICAS decided to stop providing numbers, with the result being that the group stopped
publishing its worldwide semiconductor capacity information.
The foundry model achieves the following:
The foundries, by aggregating the business of several fabless companies, can achieve the scale that is
required for the large investments in fabs and technology development.
The foundries take on substantial capital risk, but in return, do not have product development and product
competition risk.
MagnaChip, 1%
WIN, 1%
Other, 9%
TowerJazz, 1%
Dongbu, 1%
Huahong Grace, 2%
TSMC, 46%
Vanguard, 2%
Powerchip**, 3%
SMIC*, 5%
Samsung, 9%
UMC, 9%
GlobalFoundries***,
10%
24
The worlds largest foundry, Taiwan Semiconductor Manufacturing Corporation (TSMC) was formed in 1986.
Today the company accounts for almost half of the IC foundry market. Since the foundry business is capitalintensive, we think size is helpful in maintaining leading-edge technology. Globalfoundries is the secondlargest foundry, with about a 10% market share.
GlobalFoundries was formed in 2009 by AMD teaming up with an investor (ATIC, the investment wing of the
Abu Dhabi Government), to spin out its manufacturing operations into a joint venture. Since AMD owned a
stake in GlobalFoundries and through 2009, AMD was essentially its sole customer, GlobalFoundries did not
really look like a true foundry in a business sense. However, at the beginning of 2010, Chartered
Semiconductor was acquired and folded into GlobalFoundries to create a foundry of size that is comparable
with UMCs. In March 2012, AMD announced that it had divested its remaining stake in GlobalFoundries. In
October 2014 IBM announced an agreement in which it plans to transfer its semiconductor manufacturing
operations to GlobalFoundries (and pay GlobalFoundries $1.5 billion to do this!) If this deal is successfully
closed, the size of GlobalFoundries should increase by more than the 1% foundry revenue share that we have
shown in our pie chart for IBM because GlobalFoundries is to take over the manufacturing of IBMs fairly
substantial internal chip needs, which are not reflected in IBMs foundry revenue.
Samsung is a leading manufacturer of memory chips and also does some chip foundry work. In recent years,
Samsungs foundry business has grown dramatically, driven by the fact that Samsung manufactured all of
Apples processor chips (designed by Apple) for Apples iPhone and iPad products through most of 2013. In
2013, Samsung accounted for about 9% of world foundry revenue, down from 11% in 2012. However, we
believe Apple moved a large portion of its processor business away from Samsung to TSMC in late 2013,
hurting Samsungs global foundry share, which we think might fall further in 2014. However, we think that
Samsung might regain some of the Apple foundry business in 2015.
Intel has begun to offer foundry services. Intel has indicated that it does not wish to be an all-purpose foundry
for a broad range of chip companies, and specifically, that Intel has no intention of providing foundry services
for companies that it considers its competitors. We expect that in the near term, Intels foundry volume will
likely be very small. However, we believe that Intel is very interested in competing for Apples foundry
business. If Intel does win Apple or any other substantial foundry customer, this could give Intel meaningful
worldwide foundry market share. Altera has already announced that it will be using Intel to manufacture its
most advanced high-end products, but we do not expect Altera to begin ramping its foundry business at Intel
in any volume until 2016, and even then, at just one portion of Alteras total foundry needs, the Altera business
would still represent only a small percent of the global foundry market.
TSMC and UMC, since they are Taiwanese companies, report sales monthly, which provides a good monitor of
chips produced for the fabless semiconductor companies. Exhibit 18 shows monthly sales for these two
companies combined, while Exhibit 19 shows year-over-year growth. The foundry segment has higher growth,
but more volatility than the overall semiconductor industry.
25
Semiconductors
Revenue(US$MM)
$2,200
$2,000
$1,800
$1,600
Pre-UMC
Consolidation
$1,400
$1,200
$1,000
$800
$600
$400
$200
Dec93
Mar94
Jun94
Sep94
Dec94
Mar95
Jun95
Sep95
Dec95
Mar96
Jun96
Sep96
Dec96
Mar97
Jun97
Sep97
Dec97
Mar98
Jun98
Sep98
Dec98
Mar99
Jun99
Sep99
Nov99
Feb00
May00
Aug00
Nov00
Feb01
May01
Aug01
Nov01
Feb02
May02
Aug02
Nov02
Feb03
May03
Aug03
Nov03
Feb04
May04
Aug04
Nov04
Feb05
May05
Jul05
Oct05
Jan06
Apr06
Jul06
Oct06
Jan07
Apr07
Jul07
Oct07
Jan08
Apr08
Jul08
Oct08
Jan09
Apr09
Jul09
Oct09
Jan10
Apr10
Jul10
Oct10
Jan11
Apr11
Jun11
Sep11
Dec11
Mar12
Jun12
Sep12
Dec12
Mar13
Jun13
Sep13
Dec13
Mar14
Jun14
Sep14
$0
Yr/Yr%Change
40%
20%
0%
(20%)
(40%)
(60%)
Dec00
Mar01
May01
Aug01
Nov01
Feb02
May02
Aug02
Nov02
Feb03
May03
Aug03
Nov03
Feb04
May04
Aug04
Nov04
Feb05
May05
Aug05
Nov05
Feb06
May06
Aug06
Oct06
Jan07
Apr07
Jul07
Oct07
Jan08
Apr08
Jul08
Oct08
Jan09
Apr09
Jul09
Oct09
Jan10
Apr10
Jul10
Oct10
Jan11
Apr11
Jul11
Oct11
Jan12
Apr12
Jul12
Sep12
Dec12
Mar13
Jun13
Sep13
Dec13
Mar14
Jun14
Sep14
(80%)
Foundry sales are not, in fact, equivalent to semiconductor sales, but they are roughly equivalent to
semiconductor cost of goods sold (COGS) since the fabless companies report the cost of the wafers they buy
from the foundries on their cost of goods line in their income statements (though cost of goods contains other
elements, too, including the packaging and test costs of the chips once the foundries have delivered the
wafers). A wafer typically takes about 13 weeks to process, and this can be the bulk of the cost of a
semiconductor chip. The packaging and testing of the chips takes a week or two. Some fabless semiconductor
companies, such as Altera, keep the bulk of their inventory in wafer form (this is referred to as being in die
bank since the chips that are cut from the wafers are called dice, die in singular).
26
$2,000
$1,800
$1,600
$1,400
$1,200
$1,000
$800
$600
Exhibit 20 shows average wafer pricing at each foundry (the price at which the foundries sell the fully
processed wafers to the fabless companies). The average wafer price is above $1,000 for an 8-inch equivalent
wafer (i.e., 12-inch wafers have more than double this price since they have more than double the area). We
believe, however, that prices vary by a fair amount, depending on what technology node the wafer is processed.
Exhibit 21 shows TSMCs wafer sales by technology. About 60% of TSMCs sales are for the most advanced
technologies, 40nm, 28nm, and 20nm line widths (an explanation of line widths is provided further on in this
report). Many fabless companies do not need cutting-edge technology, and so they buy wafers fabricated with
somewhat older (and less expensive) technology nodes.
27
Semiconductors
Exhibit 21. TSMC Wafer Sales By Technology
100%
90%
80%
70%
0.50um+
0.25/0.35um
60%
0.15/0.18um
0.11/0.13um
50%
90nm
40%
65nm
40nm
30%
28nm
20nm
20%
10%
28
Q3 14
Q1 14
Q3 13
Q1 13
Q3 12
Q1 12
Q3 11
Q1 11
Q3 10
Q1 10
Q3 09
Q1 09
Q3 08
Q1 08
Q3 07
Q1 07
Q3 06
Q1 06
Q3 05
Q1 05
Q3 04
Q1 04
Q3 03
Q1 03
Q3 02
Q1 02
Q3 01
Q1 01
0%
It is important for semiconductor companies to carry some inventory and for there to be adequate inventory in
the supply chain (e.g., at distributors, contract manufacturers, etc.), so that demand for product can be met in
an efficient manner. If inventory levels are too low, that can lead to lost business. On the other hand, investors
can be very sensitive to the risk of excess inventory building up.
Exhibit 22 illustrates places in the electronics supply chain where excess semiconductor inventory can build
up.
Exhibit 22. Semiconductor Supply Chain
Foundry
Semiconductor
Distributor
Contract Manufacturer
Wireless Communications
OEM
Communications
Infrastructure OEM
Electronics Resellers,
Distributors, and Retailers
End Customer
Excess inventory can affect sales and profitability of semiconductor companies in several ways:
Excess inventory at chip companies can lead to inventory write-downs. If a chip company is
holding too much inventory and the inventory gets obsolete or the value of the inventory drops excessively
(e.g., for a commodity-like memory), the company may have to take inventory write-downs that drive up
cost of goods in a given quarter. Memory companies have this risk, and microprocessor companies appear
to sometimes get affected, too. There is a class of companies, like analog companies and PLD companies,
that has products with long life cycles and fairly stable pricing. Such companies have less risk of needing
inventory write-downs. However, there have been times in the past, for example, during the downturn of
2000-01, when even companies that we would view as having a lower risk of inventory write-downs (e.g.,
PLD companies like Altera and Xilinx) did take inventory charges.
29
Semiconductors
Excess inventory at distributors can sometimes affect chip revenue. Excess inventory at
distributors can lead to cutbacks in orders from distributors as they work to reduce their inventory levels.
Inventory at distributors is of particular importance to analog and PLD companies that sell a high
percentage (sometimes as much as 90%) of their product through distribution.
o
Companies that recognize revenue on sell-in to distribution can have their sales affected by
distributors driving down their inventory levels. The converse is true, too. When inventory levels at
distributors are lean at the end of a downturn with business gathering momentum, companies that
recognize revenue on sell-in to distribution get the double benefit of increasing end-customer
demand, as well as inventory builds in distribution.
Companies that recognize revenue on sell-through from distribution are not at risk of a revenue hit.
For such companies, inventory at distribution can look similar to internal inventory at the company
from a financial point of view. In such cases, excess inventory at distribution creates the similar
problems to excess internal inventory. Altera, for example, often discusses its inventory targets in
terms of internal inventory + inventory at distribution.
Excess inventory at systems manufacturers can affect chip revenue. Systems manufacturers
(makers of electronics goods like PCs, wireless handsets, etc.) include contract manufacturers (companies
that make things for other companies), as well as original equipment manufacturers, companies that both
make and sell their own products. If there is excess chip inventory at systems manufacturers, this can lead
to the systems manufacturers slowing their chip purchases while they work down their excess inventory,
resulting in a drop in revenue for the chip companies.
Excess systems inventory at systems resellers, systems distributors, and retailers can affect
chip revenue. Sometimes lack of demand can lead to a buildup of inventory in electronics systems that
have already been fully manufactured. This can be a particular risk for products sold to consumers in the
second half of the year. If there is weakness in consumer demand during the holiday season, there can be
excess inventory of finished electronics goods in stores in the new year. Excess systems inventory results
in a drop in orders to systems manufacturers, which, in turn, results in a drop in demand for chips and a
drop in chip revenue.
Exhibit 23 though Exhibit 26 show a sample of inventory graphs related to various parts of the semiconductor
supply chain that we monitor.
In the downturn of 2001, inventory spiked at various points in the electronics supply chain and, in
particular, excess inventory built up at chip distributors and electronic systems contract manufacturers.
This excess inventory took about two years to work down, creating a significant headwind working against
a recovery in chip demand through 2001 and 2002.
Perhaps in part as a response to the inventory issues that developed in 2001, over the past few years there
has been a trend toward more internal inventory at chip companies and less at distributors and contract
manufacturers (as measured in days of inventory). This reduces the risk that excess inventory might build
up at distributors and contract manufacturers, catching the chip manufacturers unaware.
We believe that the memory of problems of 2001-02 may have been a factor behind the very decisive
action taken by participants in the electronics supply chain to cut back production and reduce inventory in
H2 2008 and in early 2009, when the impact of the global economic issues became clear. As shown in our
inventory graphs, the supply chain did successfully avoid a repeat of the inventory problems of 2001. We
credit the skillful management of inventory with supporting chip pricing and paving the way for the
rebound in sequential semiconductor shipment growth, which began around Q2 2009.
30
In the second half of 2011, concerns over a variety of factors including, we think, global economic concerns
and also disruptions to disk drive supply associated with flooding in Thailand, resulted in an inventory
correction throughout the electronics supply chain that extended through 2012.
At the time of writing of this report (November 2014), we believe that exiting the third calendar quarter of
2014, inventory levels were slightly elevated, due in part to a normal seasonal lift in builds and also to the
somewhat softer than seasonal outlook many chip companies have indicated heading into the December
2014 quarter.
A more detailed discussion of inventory trends and the various companies that we monitor to track inventory is
beyond the scope of this primer, but is contained in our Inventory Review, which we typically issue quarterly.
70%
90
60%
80
50%
70
40%
60
30%
50
20%
40
10%
30
0%
20
(10%)
10
(20%)
(30%)
Yr/Yr % Change
60%
$12,000
50%
$10,500
40%
$9,000
30%
$7,500
20%
$6,000
10%
$4,500
0%
$3,000
(10%)
$1,500
(20%)
$0
(30%)
Mar-99
Jun-99
Sep-99
Dec-99
Mar-00
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
$13,500
% Change
Days Of Inventory
Yr/Yr % Change
100
Jun-99
Sep-99
Dec-99
Mar-00
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
Inventory
Yr/Yr % Change
Qtr/Qtr % Change
* Compiled data are from INTC, TXN, MXIM, MU, STM, and LLTC.
**Micron has a fiscal quarter that ends one month before the calendar quarter. We have mapped each Micron number into
the closest calendar quarter.
Source: Company reports and Wells Fargo Securities, LLC estimates
31
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
32
60%
90
50%
80
40%
70
30%
60
20%
50
10%
40
0%
30
(10%)
20
(20%)
10
(30%)
0
(40%)
Days of Inventory
Inventory
4000
3800
3600
3400
3200
3000
2800
2600
2400
2200
2000
1800
1600
1400
1200
1000
800
600
400
200
0
70%
50%
30%
10%
(10%)
(30%)
(50%)
Yr/Yr % Change
*Compiled data are from Xilinx, Altera, Broadcom, Qualcomm, and Nvidia.
Source: Company reports and Wells Fargo Securities, LLC estimates
Qtr/Qtr % Change
Yr/Yr % Change
100
% Change
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
Semiconductors
Yr/Yr % Change
150%
130%
110%
90%
60%
90
50%
80
40%
70
30%
60
20%
50
10%
40
0%
30
(10%)
20
(20%)
10
(30%)
(40%)
Yr/Yr % Change
100
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
Yr/Yr % Change
$6,500
80%
$6,000
70%
$5,500
60%
$5,000
50%
$4,500
40%
$4,000
30%
$3,500
20%
$3,000
10%
$2,500
0%
$2,000
(10%)
$1,500
(20%)
$1,000
(30%)
$500
(40%)
$0
(50%)
% Change
Mar-99
Jun-99
Sep-99
Dec-99
Mar-00
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
Days of Inventory
Inventory
Yr/Yr % Change
Qtr/Qtr % Change
*Compiled data are from Arrow and Avnet. However, in addition to semiconductors and components, some portion of Avnet
and Arrows total sales are derived from computer products, services, and computer components. Therefore, the data in
these graphs really reflect a blend of semiconductor and computer systems business.
Source: Company reports and Wells Fargo Securities, LLC estimates
33
Mar-99
Jun-99
Sep-99
Dec-99
Mar-00
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
34
60%
90
50%
80
40%
70
30%
60
20%
50
10%
40
0%
30
(10%)
20
(20%)
10
(30%)
0
(40%)
Days Of Inventory
Inventory
3-Year Quarter Average
$12,000
200%
$11,000
180%
$10,000
160%
$9,000
140%
$8,000
120%
$7,000
100%
$6,000
80%
$5,000
60%
$4,000
40%
$3,000
20%
$2,000
0%
$1,000
(20%)
$0
(40%)
Yr/Yr % Change
Yr/Yr % Change
100
% Change
Jun-99
Sep-99
Dec-99
Mar-00
Jun-00
Sep-00
Dec-00
Mar-01
Jun-01
Sep-01
Dec-01
Mar-02
Jun-02
Sep-02
Dec-02
Mar-03
Jun-03
Sep-03
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
Semiconductors
Yr/Yr % Change
Qtr/Qtr % Change
*Compiled data are from Jabil, Flextronics, Sanmina, Celestica, and Benchmark Electronics. Jabil has a fiscal quarter that
ends one month before each calendar quarter.
Source: Company reports and Wells Fargo Securities, LLC estimates
Semiconductor Segments
Exhibit 27 through Exhibit 29 show, in a number of different ways, segment breakouts for integrated circuits.
We discuss some of what we consider to be the more important sub-segments further on in this section.
As discussed, semiconductors can be grouped in three broad categories:
Discrete devices (simple, single device products, as opposed to a complete circuit); and
ICs account for more than 82% of semiconductor sales. There are three main groups of ICs:
Analog (chips that act as an interface between the real world and the logic chips); and
Most logic is digital logic and most memory is digital. In other words, logic chips and memory chips are
designs to deal with two types of information: a 1 or a 0. Analog, on the other hand, is a type of chip that
can handle a continuous signal.
35
Semiconductors
Exhibit 27. Semiconductor Market Breakdown
Semiconductors
Integrated Circuits
Logic
Micro-components
Microprocessor
Microcontroller
DSP
Standard Logic
Display Drivers
General Purpose
MOS Gate Array
Digital Bipolar
Memory
Volatile Memory
DRAM
SRAM
Non-Volatile Memory
Flash
NAND Flash
NOR Flash
Analog
Optoelectronics, Sensors,
and Actuators
Discrete Components
Transistors
Rectifiers
Diodes
Thyristors
Other Discretes
Source: Semiconductor Industry Association and Wells Fargo Securities, LLC estimates
36
$252 billion
37
Logic
57%
MPU
29%
Gen. MOS
Logic &c.
59%
Analog
16%
Memory
27%
MCU
10%
DSP
2%
$24 billion
Automotive
24%
$40 billion
$67 billion
Communication
49%
Other
Memory
3%
Industrial &
Other
10%
Application
Specific
60%
NonVolatile
44%
Computer &
Peripherals
9%
General
Purpose
Analog
40%
Volatile
53%
Consumer
8%
$16 billion
$30 billion
$36 billion
SRAM
2%
Amplifiers
16%
Interface
12%
DRAM
98%
Power
Management
58%
NAND
Flash
92%
NOR
Flash
8%
Signal
Conversion
14%
Semiconductors
Exhibit 29. Types Of Integrated Circuits (Percent Of Sales In 2013)
Analog
16%
MPU
16%
MCU
6%
DSP
1%
DRAM
14%
NAND Flash
11%
NOR Flash
1%
Other logic
34%
Other Memory
1%
Analog
Analog semiconductors can be broken out into two main sub-segments: standard and application-specific.
Standard linear chips. These are generic products that are often sold through distribution.
Application-specific analog. These are chips that are designed for a specific customer or device. This
category includes mixed-signal analog chips, which combine analog and digital capability.
Power management. These are the chips that control the distribution of power through an electronic
systems, making sure that the voltages are the right voltages for each chip and that there is enough current
for each chips needs.
Data signal conversion chips (see Exhibit 30). There are two main types of data conversion chips:
o
Analog-to-digital converters (A/D converters). These take a continuous analog signal and
convert it to a stream of digital numbers by making measurements of the height (voltage of the signal)
at regular time intervals.
Digital-to-analog converters (D/A converters) These do the opposite; they take a stream of
numbers and make these into a continuous signal.
Data converters sit at the edge of many electronic digital systems. For example, todays cellphones are all
based on digital standards. When one speaks into a cellphone, the voice is a continuous signal. This has to be
converted to a stream of digital numbers for the digital chips inside the cellphone to process the information
and an A/D converter does this. At the other end, the receiving cellphone has a stream of numbers that has to
be converted back into a sound that the listener can hear. This job is done by a D/A converter.
38
Interface chips. Interface chips are analog chips that provide the interface onto standardized
communications signal lines. They are the chips responsible for driving the electrical voltages or currents
down the lines. For example, in a computer, there would be some PCMCIA chips for driving signals down
a PCMCIA bus.
Amplifiers. These are analog chips that make electrical signals bigger, while maintaining the same shape
as the original electrical signal.
0101, 0111
5 = 0101
7 = 0111
D/A
Converter
0101, 0111
Texas Instruments, Maxim, Linear, and Analog Devices are companies that make standard linear chips.
Communications applications often use application-specific solutions. A number of semiconductor companies
that are thought of as communications chip companies, rather than analog make mixed-signal chips that
fall into this category. High-frequency or radio frequency chips, for applications such as wireless and
microwave applications, are analog chips, too; so, while companies that make this class of chips (e.g.,
Skyworks, RF Micro Devices, Triquint) are also producing analog chips, they are not usually described as
analog companies.
As indicated by their relatively low ASP (roughly $0.35-0.45), analog chips are generally fairly small.
Consistent precise performance is often important for analog chips, but in terms of the number of transistors
and other circuit elements, analog chips tend to be relatively simple.
Logic
Although the manufacturing technology for making logic chips is similar for most logic sub-segments, the
design techniques differ and so, semiconductor companies that make logic chips often specialize in one or a
small number of sub-segments.
Logic--processors. Processors are chips that can think and also have a fair amount of programming
flexibility.
39
Semiconductors
Digital signal processors (DSP). DSPs are processors that are optimized for making sense of
communications signals. One important application for DSPs is in wireless handsets, but DSPs are also
used in a variety of communication, consumer, industrial, and other end markets. DSPs are moderately
complicated chips and have an ASP of about $6.00. Texas Instruments, Analog Devices, and Freescale
make DSPs.
Microcontrollers. Microcontrollers are processors that are used for simple control applications in a
wide range of devices, from washing machines and microwave ovens, to cars and industrial machinery. A
microcontroller is used when the amount of thinking needed is less than what a full-fledged
microprocessor might provide. Microcontrollers are relatively simple chips and generally do not require
very advanced manufacturing technology. They have an ASP of about $0.88. Examples of companies
making microcontrollers include Microchip, Freescale, Texas Instruments, Infineon, and Philips. The
processing horsepower of a microcontroller is dictated by word length (number of bits). Eight-bit
microcontrollers currently represent the low end of the market, and we think the market will continue to
migrate toward 32-bit over the next several years. Many microcontroller companies have adopted the
ARM architecture for 32-bit offerings.
Other processors. Two examples of other processors are graphics processors and network processors.
Graphics processors are processors optimized for generating or processing images (i.e., graphics
rendering). One important application for these chips is in computers. Graphics processors tend to be
quite large, complicated chips, and command prices of tens of dollars. Nvidia and AMD (ATI) are
companies that make graphics processors. Network processors are processors optimized for handling data
traffic in computer and communications networks. Makers of network processors include Broadcom, LSI
(Agere), and PMC-Sierra.
Standard logic. Standard logic chips are logic chips that are less flexible than microprocessors. They do logic
operations (i.e., they think), but in general, each specific type of standard logic chip is designed to always do
the same specific logic operation, unlike processors, which run software programs that determine what logic
operations they perform. Even programmable logic devices (PLD) are a type of standard logic, which are
designed to always do the same thing once they have been configured with a specific program.
Programmable logic devices (PLD). Programmable logic devices are devices that are not committed
to any specific logic operation; however, they can be configured with programming. This is different from
the way processors use software programs. Processors have circuitry that is designed in a specific way and
then this circuitry executes instructions according to its software program. PLDs are chips that obtain the
actual configuration of their circuitry (what is connected to what) from a program. Although in principle
they can be reprogrammed, each chip is typically programmed once and then operates like a chip with
fixed logic that always does the same thing. Field programmable gate arrays (FPGA) and complex
programmable logic devices (CPLD) are types of PLDs. FPGAs are typically used for prototyping or
relatively low-volume applications. Because PLDs can be configured to perform a wide range of logic
operations, they are often used instead of making a custom design of a chip, an application-specific
integrated circuit (ASIC). PLDs have a wide range of complexity; some of the high-end FPGAs are among
the largest chips made. PLD prices range from a few dollars to several hundred, depending on the size and
level of complexity. Xilinx and Altera are the two largest PLD companies, with a combined market share of
90% of the FPGA and PLD market. Smaller PLD companies include Lattice, Microsemi, Atmel, Chengdu
Sino Microelectronics System, and Cypress Semiconductor.
Standard logic. This is a catch-all describing a fairly broad range of standard logic chips that are neither
processors, nor PLDs. Standard logic chips tend to be fairly simple chips. A wide range of companies make
standard logic chips, including Texas Instruments.
ASICs and applications-specific standard products (ASSP). ASICs are chips that are custom designed
for a specific end user (e.g., Cisco designs many of its own ASICs). ASSPs are chips that are designed for a
specific application (as opposed to standard logic, which can be used for multiple applications), but not just for
one customer. The term ASSP is a fuzzy one and many chips that are termed ASSP might fall into the analog
group as mixed-signal chips or be considered standard logic.
40
Memory
DRAM and NAND flash are the two significant semiconductor memory segments, though there are a number
of other (legacy) types of memory that are in secular decline.
There are two main types of semiconductor memory:
(1) Volatile memory. Volatile memory forgets what it was supposed to remember when the power is
switched off. However, it is relatively easy to write information into volatile memory and to retrieve the
information from it; so volatile memory is used as the main working memory in a system such as a PC.
The two main types of volatile memory are DRAM and static random access memory (SRAM), though
from an investment point of view, we believe DRAM is by far the more important of the two, with many
major companies being involved in DRAM, whereas SRAM is made by a fistful of fairly small companies.
(2) Nonvolatile memory. Nonvolatile memory, on the other hand, remembers what it is supposed to
remember when the power is switched off. Traditionally, nonvolatile memory has been used for
permanent code storage in systems. For example, in a PC there is a small flash chip that contains the PC
BIOS, that is, all the crucial information a PC needs to know when it is being switched on. Similarly, in a
cellphone there is a flash chip that contains the code information (e.g., what the cellphone needs to know
about the global system for mobile [GSM] standard or code division multiple access [CDMA] standard so
that it can interpret the signals it receives). More recently, nonvolatile memory has also found a place in
the storage of data, such as photographs taken with digital photography. In general, it is more difficult to
write information to nonvolatile memory than to volatile memory, which is one reason that volatile
memory is still used in many electronic systems. The main type of nonvolatile memory is flash memory, of
which there are two types: NOR and NAND flash. Both of these are important in terms of total volume of
sales, but NAND flash has a strong growth dynamic, whereas NOR flash does not. There are a number of
legacy types of nonvolatile memory, none of which we believe are particularly important from an
investment point of view since they are made by a small number of fairly small companies.
Volatile Memory
DRAM. In a PC, the information is initially loaded into the dynamic random access memory from the hard
drive and also some information is provided by the microprocessor. It is dynamic because even when the
power is switched on, it forgets what it is supposed to remember and so each memory cell has to be
refreshed. Each DRAM chip contains circuitry to do this refresh tens to hundreds of times per second. It is
random access because any information can be retrieved from any place in the memory at a given time (as
opposed to serial, in which memory locations have to be looked at one after another in a certain order; NAND
flash is serial to some extent). DRAM memories come in different sizes. A typical DRAM memory chip for a PC
might be 4 Gigabits (Gigabit = Gb= 1 billion bits). This means that the chip can remember 4 billion ones or
zeroes. Most electronic systems, including PCs, have memory specified in bytes, not bits. A byte equals eight
bits. A 4-gigabyte (GB) PC might have eight different 4Gb DRAM chips in it. DRAM chips come in all sizes,
from 64 million bits (Mb)( it is possible to get smaller memories, too) to 4 Gb. DRAM chips are commodities
(in principle, a 1 Gb chip works exactly the same, regardless of which company one buys it from) and so the
price tends to be very volatile. Also, as technology progresses, more bits can be jammed on a chip, and so, for
the same price it is possible to buy a chip with more bits. At the time of writing this report, a 4 Gb DRAM chip
cost around $4.00 (we have graphs with pricing information further on in this report.) Samsung, Qimonda
(Infineon), Hynix, Elpida, and Micron are some of the bigger DRAM companies.
SRAM. SRAM is static because while the power stays on, it remembers what it is supposed to remember (it
does not need a refresh). Static random access memory has traditionally been used for very high-performance
applications, when very fast access to the information in the memory has been needed. SRAM is still used in
communications systems, but its general use is less and less widespread. Cypress Semiconductor, Renesas, and
Integrated Silicon Solutions are some companies that still make SRAM.
41
Semiconductors
Non-volatile Memory
NAND and NOR flash. Flash is the name of a type of memory cell (something that can store
a 1 or a 0). NAND and NOR refer to the logic functions (Not AND) and (Not OR), which describe the
circuitry that a flash chip uses to access the memory locations. Because of the difference in memory access
circuitry, NOR flash is more expensive than NAND flash and offers faster access to the memory. The largest
use of NOR flash today is in cellphones for code storage. NAND flash is used when large amounts of
information are to be stored and price is unimportant. One big use of NAND flash is in digital cameras for
storage of photographs. Apples iPad, iPod, and iPhone also generate large demand for NAND flash as a way to
store information to play back songs. In the future, NAND flash will likely be used to replace notebook hard
drives. Samsung, Toshiba, SanDisk (joint venture with Toshiba), Micron, Hynix, Intel (joint venture with
Micron), and Spansion are some companies that make NAND flash memory. Micron, Spansion, and Macronix
are some companies that make NOR flash memory.
Other types of nonvolatile memory. These include electrically erasable programmable read-only memory
(EEPROM) (pronounced E-squared PROM), erasable programmable read-only memory (EPROM), and readonly memory (ROM). None of these are very important today in terms of volume of sales, in our opinion.
Discrete Components
A discrete component is the most basic type of semiconductor device. A discrete component is a single
elementary electronic device (such as a transistor). An integrated circuit (IC) is a chip in which many
transistors and other devices (ranging from a few to more than a billion) are all connected together on a piece
of silicon (a chip). Examples of discrete components include transistors and diodes. Discrete components are
used for power management, voltage regulation, and to connect integrated circuits within a system board (i.e.,
printed circuit board). The discrete component market is approximately $18 billion in size (2013), or
approximately 6% of the overall semiconductor market. For the four-year period 2003-07 (between the two
downturns of the past decade), the discrete component market grew with a compound annual growth rate
(CAGR) of 6%, versus an overall semiconductor CAGR of 11%. This trend continued into the 2007-09 period,
with the discrete components market sliding 8%, versus a 6% decline for the semiconductor industry as a
whole. Coming off the 2007-09 downturn, the discrete component market grew at a CAGR of 6% from 2009 to
2013, versus an overall semiconductor CAGR of 8%. We think, in general, the slower growth rate relative to the
overall semiconductor market is in part the result of more chip designs including the functionality of discrete
components in integrated circuits.
The discrete component market is highly fragmented, with no participant controlling more than 10% market
share. Exhibit 31 highlights the top 25 participants in the discrete component market.
42
Semtech, 1%
Jilin Sino-Microelectronics,
1%
Others, 11%
Infineon Technologies, 8%
Sanken, 1%
Pan Jit, 1%
NXP, 7%
Avago Technologies, 1%
Toyota, 1%
Toshiba, 6%
Denso, 1%
Panasonic, 2%
ON Semiconductor, 6%
Shindengen Electric, 2%
Freescale Semiconductor,
2%
Mitsubishi, 6%
Diodes, 3%
Microsemi, 3%
STMicroelectronics, 6%
Fuji Electric, 4%
International Rectifier, 4%
Rohm, 5%
Fairchild Semiconductor, 5%
Vishay, 5%
Renesas Electronics, 5%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Devices, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
CAGR 20002003
(4%)
(5%)
(25%)
(21%)
4%
(17%)
CAGR 20092013
6%
6%
8%
9%
6%
12%
11%
(5%)
7%
6%
13%
8%
43
Semiconductors
Exhibit 32 shows historical growth by IC segment. We have broken out the past 23 years into five periods:
From 1990 to the beginning of 2000, ICs showed strong growth, with a CAGR of 16%, about 6 percentage
points higher than unit growth of about 10% (though with a cyclical overlay.)
(1) The past 13 years have been affected two large downturns, in 2001 and 2009. In our view, the 2000-03,
and 2007-09 growth numbers we have shown provide an interesting view into how the various segments
of semiconductors fare through downturns. However, there are differences between the two periods.
Communications-related chips were hit particularly hard in 2001 because of the communications
infrastructure bubble in the late 1990s. The 2009 downturn was more of a broad-based downturn.
(2) Our long-term projection for IC sales and semiconductor sales in dollar terms is 10-12% per year,
comparable to or slightly higher than underlying unit growth of 10% per year, which we think will
continue. The 1990-2000 period benefited for a richening mix of chips. Overall growth from 2000 to 2011
was affected by the two downturns and also by falling ASP. However, IC ASP is currently (2014) running
in the $1.20-1.30 range, similar to where it was in 1990. We believe that in the long run, ASP movements
will be roughly neutral to semiconductor sales growth. Over the next few years, we think that growth in
smartphones, tablets, and servers may help richen mix, driving up ASP.
(3) We think the period 2003-07 to be representative of a recent period of relatively normal growth.
1990-2000
As discussed in preceding sections, microprocessors grew strongly from 1990 to 2000 (30% per year
CAGR), driven by the growth of the PC market.
Digital signal processors grew 36% per year from a small base, driven by a variety of end markets. The use
of the DSP in wireless handsets was one driver of DSP growth during this period, we believe.
PLDs grew 29% per year, also from a small base. We think this was due to the emergence of PLDs as an
alternative to custom chips (ASICs) for prototyping and specialized chips with relatively small run
requirements. PLDs are used extensively in communications infrastructure and so, benefited from the
communications boom toward the end of the 1990s.
Analog and DRAM both had growth roughly equal to the overall IC growth of 16% from 1990 to 2000
(though pricing movements resulted in DRAM growth jumping in the first half of the decade and then
dropping back in the second half).
We believe that PLDs were hit particularly hard, declining with a negative 21% CAGR, because of the
overbuilding of communications infrastructure toward the end of the 1990s.
DRAM declined at a substantial 17% CAGR, demonstrating how commodity-like segments are particularly
sensitive. The large decline for DRAM through this period was in part the result of pricing compression
being above normal trend and in part, a reduction in unit (bit) growth.
ICs as a whole grew 12% per year from 2003 to 2007, below the 16% per year growth from 1990 to 2000.
As discussed, the data suggest that the underlying unit growth for the industry has changed much in the
past 25 years, running at a CAGR of 10%. The higher revenue growth number for 1990-2000 was in part
driven by rising ASPs, resulting from a mix shift related to PCs being a big driver of overall semiconductor
growth. In the 2003-07 period, wireless handsets were an important high-growth semiconductor end
market, which, in our view, did not provide quite as much of an opportunity for overall richening of mix
that PCs did in the decade of the 1990s.
DRAM did well, with a CAGR of 17% per year from 2003 to 2007, but then declined more than
semiconductors overall in the 2009 downturn (a decline of 15% per year CAGR in 2007-09), the effect
being better-than-trend pricing from 2003 to 2007, followed by a pricing collapse in 2007-09. For periods
of rapidly alternating upturns and downturns such as we have had from 2000 through 2009, we would
expect DRAM to outperform in the upturns and underperform in the downturns because of pricing
swings. For longer periods, such as 1990-2000, the long-term growth of DRAM roughly tracks the overall
semiconductor market.
44
Nestled within the overall flash CAGR numbers shown in Exhibit 32 are strong growth numbers for NAND
flash and declining sales for NOR flash. We do not have NAND data for 2003, but from 2004 through
2007, NAND flash sales grew at a CAGR of 31%.
Microprocessors underperformed with a 6% CAGR from 2003 to 2007. However, the primary end market
for microprocessors, PCs, was strong during this period, and the microprocessor CAGR was pulled down
by ASP contraction. We believe that this was driven in part by competition between Intel and AMD, and in
part by a mix shift toward lower-end desktops and notebooks, partially offset by a positive mix shift of
notebooks having higher growth than desktops (notebook microprocessors generally have higher ASPs
than desktops).
There are a number of theoretical reasons as to why PLDs might have better secular growth opportunities
than ICs in general. However, from 2000 to 2007, PLD growth was, in fact, below aggregate IC growth,
declining with a CAGR of negative 21% per year from 2000 to 2003 and then growing only 7% per year
from 2003 to 2007. We think that this may have been due in part to the need to grow into the considerable
overcapacity in communications infrastructure that was created toward the end of the 1990s.
As noted in the prior section, pricing declines drove down DRAM revenue in 2009, with the DRAM
segment declining with a CAGR of 15% from 2007 to 2009.
NAND flash actually achieved a slightly positive CAGR, of 1% from 2007 through 2009. NAND was less
than half of total flash sales in 2004 and grew to being more than three-quarters of the total flash market
by 2009.
Microprocessor pricing was stable through the depths of the downturn in 2009, and microprocessor
outperformed the overall semiconductor market from 2007 to 2009.
The PLD segment significantly underperformed the broader chip group in the downturn of 2000-03
because there was a large overbuild in one of the big PLD end markets (communications infrastructure).
However, there was no similar PLD-specific issue in the 2007-09 downturn, and PLD declines in this
period (a CAGR of negative 4%) were roughly the same as overall IC declines (a CAGR of negative 7%).
Analog pricing tends to be quite stable since much of the analog customer base is fragmented (few
large customers to demand price concessions) and analog prices tend to be relatively low (average
price of an analog chip close to $0.50, compared to about $1.30 ASP for ICs overall). This should
make the analog segment more resilient in downturns. Indeed, it appears as if the profitability of
analog companies does hold up very well in downturns. For example, Linear Technology, which does
all its own wafer manufacturing, reported a gross margin of 77% in the June 2008 quarter (prior to
the downturn, which gained momentum in December 2008), and lost just 3 percentage points of
gross margin to report a 74% gross margin in the March 2009 quarter, the depths of the downturn.
On the other hand, analog companies tend to sell a high percent of their product through distribution,
and many analog companies recognize at least their international distribution revenue on sell-in to
distribution. In a downturn, such companies are affected both by a reduction in end-market demand
and also a reduction on inventory in distribution.
In the 2000-03 drop, analog sales fell, with a CAGR of only negative 4%, compared to a bigger CAGR
drop rate of 8% of overall ICs. The corresponding recovery from 2003 to 2007 was, accordingly, more
muted, with a CAGR of 8%, compared to ICs growing at a CAGR of 12%. In the downturn of 2008 and
2009, analog declined with a CAGR of negative 6% from 2007 to 2009, close to ICs overall declining with
a CAGR of negative 7%.
2009-2013 (recovery followed by three soft years). The year 2010 was one of strong recovery, followed
by an inventory correction in the second half of 2011, a soft year for semiconductors in 2012, and a year of
moderate growth in 2013. From 2009 to 2013:
Flash memory continued to show the strong secular growth of NAND, with total flash CAGR of 11%, driven
by a 17% CAGR for NAND flash sales. However, NOR flash has been in secular decline, affected by wireless
handsets moving away from using NOR flash.
45
Semiconductors
The DRAM CAGR of 12% was above total IC growth, driven by strong pricing dynamics in 2013. DRAM
pricing has continued to be strong in 2014, leading to DRAM growth in that could well be close to 30% in
2014, resulting in a 2009-2014 CAGR of close to 10%. In our view, DRAM pricing is currently (November
2014) running at far above its equilibrium point, and we think that DRAM growth in recent years has been
unsustainably high. There are several dynamics at play in the DRAM industry at the moment, including
slowing bit growth demand, consolidation of DRAM providers, which could potentially help stabilize
pricing, and a slowing rate of technology transitions. We believe that overall DRAM revenue growth might
slow over the next few years, though it is possible that the profitability of the segment could remain solid.
PLD chips achieved a 9% CAGR, above overall IC growth of 7%. Altera and Xilinx have both suggested in
the past that the PLD segment could be at a point at which secular PLD growth could begin to run above
aggregate IC growth, with PLDs eating into the broader custom chip application-specific integrated circuit
(ASIC) market. The modest outperformance from 2009 to 2013 could be evidence of this, but we see it
more as a reversal from underperformance in the years 2000-07. It appears that in 2014 PLD growth will
match overall IC growth, with, we estimate, revenue for Altera+Xilinx, up 9%, compared to our
expectation that IC growth might be of the order of 8-10% in 2014.
Despite investor concerns about the cannibalization of computers by mobile devices, microprocessors
achieve a 6% CAGR from during this period, close to overall IC growth of 7%.
In the following sections we provide more detail on some of what we consider to be the more interesting
semiconductor segments.
Sub-Segments of Interest
Microprocessors -- Some Microprocessor Concepts And Terms
What is a microprocessor (CPU)? A microprocessor is a logic chip that provides the thinking
capability in a computer or other device. Another term often used for a microprocessor is a central
processing unit (CPU). Microprocessors accounted for 16% of total IC sales in 2013. We discuss
microprocessors in detail in our quarterly Processor Review.
x86 microprocessors
Intel is the worlds biggest maker of microprocessors. Most of the processors Intel makes are
based on the x86 instruction set and are called x86 microprocessors.
AMD and VIA also make microprocessor based on the x86 instruction set. Intel owns the
intellectual property and licenses the right to make x86 microprocessors to AMD and VIA. We
believe that under the terms of the agreement Intel has with AMD, the license automatically
renews periodically (Intel cannot decide to stop the license agreement). However, our
understanding is that there is a change-of-control clause in the license, resulting in the
termination of the license in the event of a change of control at AMD (another company cannot
acquire AMD and obtain the right to make x86 processors through such an acquisition).
ARM processors
ARM does not make microprocessors, but it does design microprocessor cores, which it
licenses to various chip makers.
Qualcomm, Apple, Texas Instruments, Nvidia, and MediaTek are some companies that
license ARM cores, which they include in their own microprocessor designs. ARM-based
microprocessors are used extensively in smartphones and in tablets.
PowerPC
46
IBM makes PowerPC processors. Apple used to use PowerPC in all its desktop and notebook
computers, but switched to x86 processors from Intel in 2006. PowerPC is primarily used by
IBM, itself, for its mainframe computers and its Unix servers.
Itanium
Intel makes Itanium processors for Unix servers used by companies like HP. HP also has server
lines based on x86 microprocessors.
SPARC
Oracle (Sun) makes SPARC processors, which it uses for its Unix server line. Oracle also has
server lines based on x86 microprocessors.
Other processors
There are a number of (relatively small) companies like MIPS and Tensilica that design
processor cores, which they license to other chip companies.
Chips. Almost all desktop and notebook computers contain a single microprocessor to do their thinking.
Some servers contain just one processor chip and are called 1-way servers. Most servers have two
processor chips and are called 2-way servers. Some servers have four chips or more.
Cores. A single processor chip contains several circuits on the same piece of silicon. The circuit on a
processor that does the thinking is the processor core, but there are also other parts of the chip, like
memory and other circuitry. Many microprocessors have multiple cores, many thinking circuits, each of
which can think by itself. Most desktop and notebook computers used dual-core processors (2 cores on a
processor). Higher end desktops have quad-core processors (4 cores on a processor). Sever
microprocessors typically have more cores. Currently Intel offers server chips with up to 10 cores. Thus, a
typical server might have several chips (e.g., two chips in a 2-way server) and each chip might have several
cores. So a 2-way server that is populated with 6-core microprocessors actually has 12 little brains (2x6).
Clock frequency. Microprocessors are clocked chips, which means that all calculations or other
operations in the circuit wait for a beat of a clock before they happen. This makes sure that signals are
transferred at the same time, avoiding errors that may occur as a result of one thing happening earlier
than something else. The higher the clock frequency, the more things that can happen per second and the
more the chip can achieve. Typical clock frequencies for Intel processors are 1.0-3.5 GHz (Gigahertz),
1.0-3.5 billion clock cycles per second. Therefore, billions of things happen within a microprocessor each
second. Higher frequencies in microprocessors thus allow processors to achieve more in a given amount of
time, resulting in an improvement in microprocessor performance.
Cache memory. Microprocessors have a certain amount of memory included on the microprocessor
chip, which is where information is temporarily stored while the microprocessor is working on the
information. Microprocessor performance is typically improved by having more cache memory available.
There are often as many as three layers of cache memory available: level 1, level 2, and level 3 (L1, L2, and
L3 cache). For a processor with more than one core, the extent to which the cache is shared affects how
much the cache helps performance. Cache memory that is shared among all the cores has more impact on
performance than cache memory that is restricted in its use to a subset of cores. L1 cache is typically small
and attached to a single core, while L2 and L3 cache is often shared. Todays microprocessors often have
several megabytes of cache memory. The cache memory is typically SRAM memory since SRAM circuitry
can be made with a standard logic process used for microprocessors, whereas DRAM circuitry needs
special processing.
Thermal design power (TDP). The peak power consumption of the microprocessor is often specified
in terms of thermal design power, or TDP. TDP is, in fact, a specification of how much heat is generated by
the microprocessor when it is running as hard as it can. It is the heat that a system needs to be designed to
remove from the chip. For server processors AMD uses an alternative specification, average CPU power
(ACP), with ACP generally being lower than TDP for any given chip since it is not the peak power
consumed (which is what TDP specifies), but the average power.
Die size. This represents the size of a chip. The smaller the chip, the lower the cost of the chip since more
chips can be printed on a silicon wafer.
47
Semiconductors
Chipsets. A microprocessor works with an additional chip or pair of chips, the chipset, which handle
most of the communications between the microprocessor and the rest of the computer. We discuss what
chipsets do in more detail further on in this report.
Integrated graphics versus discrete graphics. Computers use microprocessors for general purpose
thinking, but as well as this, many applications need some additional circuitry to do computations related
to displaying graphics on the computer screen. For many computers (one-half to three-quarters of
desktops and notebooks), the graphics computations are dealt with by integrated graphics, which is
circuitry that is included either with the chipset, or included with the microprocessor. For computer users
that are prepared to pay extra, a separate graphics card is included with the computer that contains a
stand-alone graphics chip and some memory. This is discrete graphics.
Tick-Tock (Intel). Intel has developed new microprocessors on what it calls a Tick-Tock schedule,
introducing a new microprocessor product family each year on a two-year cycle as follows:
o
60%
40%
20%
0%
(20%)
(40%)
Total ICs
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Microprocessors
48
Dec-05
Dec-04
Dec-03
Dec-02
Dec-01
Dec-00
Dec-99
Dec-98
Dec-97
Dec-96
Dec-95
Dec-94
Dec-93
Dec-92
(60%)
Dec-91
80%
Exhibit 33 shows the growth of microprocessors compared with overall integrated circuits. As shown in
Exhibit 32, microprocessors performed better than semiconductors as a whole from 1990 to 2003, and again
from 2007 to 2009. Despite investor concerns about the cannibalization of computers by mobile devices,
microprocessors achieved a 6% CAGR from 2009 to 2013, comparable to overall IC growth of 7%.
There are a number of different microprocessor architectures, but x86 processors account for the bulk of total
microprocessor sales. Intel and AMD are the main makers of x86 microprocessors (see Exhibit 34).
Exhibit 34. Total x86 Microprocessor Market Share Over Time
100%
80%
70%
60%
50%
40%
30%
20%
10%
0%
Q1 02
Q2 02
Q3 02
Q4 02
Q1 03
Q2 03
Q3 03
Q4 03
Q1 04
Q2 04
Q3 04
Q4 04
Q1 05
Q2 05
Q3 05
Q4 05
Q1 06
Q2 06
Q3 06
Q4 06
Q1 07
Q2 07
Q3 07
Q4 07
Q1 08
Q2 08
Q3 08
Q4 08
Q1 09
Q2 09
Q3 09
Q4 09
Q1 10
Q2 10
Q3 10
Q4 10
Q1 11
Q2 11
Q3 11
Q4 11
Q1 12
Q2 12
Q3 12
Q4 12
Q1 13
Q2 13
Q3 13
Q4 13
Q1 14
Q2 14
Q3 14
90%
VIA
Exhibit 34 shows that in H2 2005, Intel began losing microprocessor market share to AMD. AMDs share
gains were driven by some good design decisions on AMDs part that gave AMD a performance lead in
desktop and server processors. AMD also began to build out its notebook processor line, which further
fueled its market-share gains. AMDs market share peaked at about 25% in the December 2006 quarter
from about 16% in the June 2005 quarter. However, in 2006, Intel managed to re-establish a clear
performance leadership position in essentially every high-volume microprocessor segment with a series of
new microprocessor families, which resulted in AMDs overall processor market share falling steadily
since the end of 2006 through to today (2014).
Microprocessors are among the highest-priced chips. Exhibit 35 shows the average microprocessor selling
price as reported by Intel and AMD. Intels overall microprocessor ASP is about $140, with its highestpriced chips (top-end server microprocessors) commanding prices that stretch to several thousand dollars.
49
Semiconductors
Exhibit 35. Average Microprocessor Selling Price--Intel Versus AMD
$200
$180
$160
$140
$120
$100
$80
$60
$40
$20
Q1 01
Q2 01
Q3 01
Q4 01
Q1 02
Q2 02
Q3 02
Q4 02
Q1 03
Q2 03
Q3 03
Q4 03
Q1 04
Q2 04
Q3 04
Q4 04
Q1 05
Q2 05
Q3 05
Q4 05
Q1 06
Q2 06
Q3 06
Q4 06
Q1 07
Q2 07
Q3 07
Q4 07
Q1 08
Q2 08
Q3 08
Q4 08
Q1 09
Q2 09
Q3 09
Q4 09
Q1 10
Q2 10
Q3 10
Q4 10
Q1 11
Q2 11
Q3 11
Q4 11
Q1 12
Q2 12
Q3 12
Q4 12
Q1 13
Q2 13
Q3 13
Q4 13
Q1 14
Q2 14
Q3 14
$0
Intels overall microprocessor ASP (around $140) is currently more than double AMDs (around $60), due
to a better mix of higher performance chips.
Microprocessor ASP was fairly stable from 2001 through 2005, though it fell between 2005 and 2008 for
both Intel and AMD. Intels microprocessor ASP has for the most part been on an upward trend since
early 2009, while AMDs microprocessor ASP has drifted downward in this period, in part the result of
AMDs position eroding in the high ASP server processor market.
Exhibit 36 is a pricing table that shows an example of list prices posted on Intels and AMDs websites. List
prices are not necessarily the actual prices at which the microprocessor companies sell chips to large PC
makers, but we think that list pricing is a helpful indicator of general pricing trends and the pricing positioning
of Intel and AMD. We generally assume that the large customers get discounts to list prices of approximately
30%. However, we believe that it is difficult to get reliable information on the amount that chips are
discounted. We think that changes in discounting sometimes result in ASP movements for Intel and AMD that
are not reflected in list prices.
50
Price
$999
$583
$389
Price
X2 (Athlon 28nm)
450 (3.5GHz 2C 1MB L2 65W)
$303
$303
$303
$303
$213
$192
$182
$213
$213
$192
$192
$192
$182
$182
$149
$138
$138
$138
$117
$117
$117
$138
$117
AMD CPU
X4 (Athlon 28nm)
860K (3.7GHz 4C 4MB L2 95W)
840 (3.1GHz 4C 4MB L2 65W)
$86
$75
$75
$75
$75
$64
$64
$64
$64
$999
$583
$323
$358
$303
$303
$142
$132
$122
$91
$77
X4 (Athlon 32nm)
760K (3.8GHz 4C 4MB L2 100W)
750 (3.4GHz 4C 4MB L2 65W)
750K (3.4GHz 4C 4MB L2 100W)
740 (3.2GHz 4C 4MB L2 65W)
X2 (Athlon 32nm)
370K (4.0GHz 2C 1MB L2 65W)
350 (3.5GHz 2C 1MB L2 65W)
340 (3.2GHz 2C 1MB L2 65W)
FX Series (Vishera 32nm)
9590 (4.7GHz 8MB L2 8MB L3 220W)
9370 (4.4GHz 8MB L2 8MB L3 220W)
8370E (3.3GHz 8C 2MB L2 8MB L3 95W)
8370 (4.0GHz 8C 2MB L2 8MB L3 125W)
8350 (4.2GHz 8C 1MB L2 8MB L3 125W)
8320E (3.2GHz 8C 2MB L2 8MB L3 95W)
8320 (4.0GHz 8C 1MB L2 8MB L3 125W)
6350 (4.2GHz 6C 6MB L2 6MB L3 125W)
6300 (4.1GHz 6C 6MB L2 8MB L3 95W)
4350 (4.3GHz 4C 4MB L2 4MB L3 125W)
4300 (4.0GHz 4C 4MB L2 4MB L3 95W)
$220
$204
$194
$194
$173
$142
$142
$122
$101
$97
$97
$183
$153
$132
$112
$122
$112
$101
$122
51
Semiconductors
4770K (8M 4C 3.50GHz 84W)
4771 (8M 4C 3.50GHz 84W)
4770 (8M 4C 3.40GHz 84W)
4765T (8M 4C 2.00GHz 35W)
$339
$314
$303
$303
$276
$255
$213
$213
$242
$213
$192
$192
$192
$182
$182
$182
$182
$149
$138
$138
$122
$117
$86
$75
$75
$64
$64
$94
$94
52
$294
$332
$294
$294
$205
$225
$205
$205
$205
$205
$201
$184
$184
$184
$182
$182
$122
$142
$142
$97
$91
$97
$62
$49
$40
$40
$40
Virgo (Trinity32nm)
A10-5800K (3.8GHz QC 4MB 100W)
A10-5700 (3.4GHz QC 4MB 65W)
A8-5600K (3.6GHz QC 4MB 100W)
A8-5500 (3.2GHz QC 4MB 65W)
A6-5400K (3.6GHz DC 1MB 65W)
A4-5300 (3.4GHz DC 1MB 65W)
$122
$122
$91
$91
$57
$47
X6 (Phenom II - 45nm)
1100T (3.3GHz 6C 3M L2 6M L3 125W)
1090T BE (3.2GHz 6C 3M L2 6M L3 125W)
1075T (3.0GHz 6C 3M L2 6M L3 125W)
1065T (2.9GHz 6C 3M L2 6M L3 95W)
1055T (2.8GHz 6C 3M L2 6M L3 125W)
1045T (2.7GHz 6C 3M L2 6M L3 125W)
X4 (Phenom II - 45nm)
980 BE (3.7GHz QC 2M L2 6M L3 125W)
975 BE (3.6GHz QC 2M L2 6M L3 125W)
970 BE (3.5GHz QC 2M L2 6M L3 125W
965 BE(3.4GHz QC 2M L2 6M L3 125W)
955 BE(3.2GHz QC 2M L2 6M L3 125W)
910e (2.6GHz QC 8M)
905e (2.5GHz QC 8M)
850 (3.3GHz 2M L2 QC 95W)
840 (3.2GHz QC 8M)
X4 (Athlon II - 45nm)
651 (3.0GHz QC 4M 100W)
645 (3.1GHz QC 2M 95W)
641 (2.8GHz QC 4M 100W)
$91
$87
$101
$91
$77
$77
$77
$59
$40
$36
$194
$173
$163
$153
$163
$153
$142
$81
$81
$143
$100
$103
$92
$102
$81
$182
$182
$177
$98
$81
$79
$133
$122
$98
53
Semiconductors
54
x86 and other types of microprocessors and processor cores. There are a number of different
types of microprocessors, but the x86 microprocessors made by Intel and AMD account for the bulk of the
total microprocessor market.
o
x86 processors. x86 microprocessors are the microprocessors used in essentially every desktop and
laptop computer sold in the world today. They are also used in x86 servers and in some other
relatively minor applications. x86 microprocessors are microprocessors that implement the x86
instruction set. An instruction set is the set of commands that a microprocessor can carry out, the
language in which the microprocessor thinks. Intel owns the rights to the x86 technology and AMD
(and Via) manufacture microprocessors under license from Intel. Although Intel owns the rights to
the x86 license set and AMD can manufacture microprocessor only under license from Intel, there are
apparently terms in the license agreement that Intel has with AMD that require Intel to periodically
renew the license to AMD under reasonable terms. The name x86 comes from the series of Intel
processors that powered the early PCs, the 8086, followed by the 80286, 80386, 80486, etc.
ARM processors. ARM Holdings is a semiconductor company that licenses processor cores, rather
than making its own chips. Other semiconductor companies include the circuit designs from ARM in
their own chips and pay ARM a royalty. ARM built its business around the concept that a processor
design was needed that had low-power consumption, to serve markets in mobile electronics, where
battery life is an important consideration. ARM processors have a particularly important presence in
the wireless handset market, with chips like Samsungs Exynos chip, Qualcomms Snapdragon chip,
and Nvidias Tegra chips all using ARM processors.
Some companies, including Nvidia, Samsung, and Broadcom, use ARMs circuit design in
processor cores that they include together with other circuitry on their chips.
Other companies, for example Apple and Qualcomm, license the use of the structure of the
microprocessor from ARM (i.e., the instruction set), but design their own circuits to implement
this microprocessor structure.
PowerPC. IBM manufactures PowerPC to use in its own server products. At some point in the past,
Apple used PowerPC for all its computers, but several years ago, Apple switched to Intels x86
products.
SPARC. Sun Microsystems (Oracle) uses SPARC chips in its servers, as does Fujitsu. However, with
the decline of the Unix server market and the acquisition of Sun by Oracle, we expect SPARC chips to
(continue to) fade over time.
4-core
2-core
1-core
Atom
Xeon MP (4-way)
47%
53%
FY2007
3%
70%
27%
FY2008
First Atom
FY2009
FY2010
First 6-core
<1%
13%
81%
2%
4%
FY2011
FY2012
FY2013
<1%
19%
76%
<1%
32%
62%
43%
52%
5%
6%
4%
6%
83%
8%
79%
12%
<1%
First 4-core
First Atom
14%
80%
3%
3%
<1%
73%
16%
11%
First 4-core First 6-core
1%
65%
11%
23%
7%
65%
7%
21%
First 8-core
6%
75%
3%
16%
8%
81%
1%
10%
11%
84%
96%
100%
100%
100%
100%
16%
4%
First 6-core
100%
100%
100%
100%
77%
23%
6-core or
greater
4-core
2%
37%
2-core
38%
56%
1-core
60%
7%
Xeon DP (2-way)
First 4-core
6-core or
greater
4-core
2%
37%
2-core
38%
56%
1-core
60%
7%
Source: Mercury Research and Wells Fargo Securities, LLC
84%
16%
96%
4%
Servers and server processors--4-way and 2-way servers. An x86 server (a server that uses x86
processors) can be designed to use one, two, four, or more processors. A server that uses two processors is
called a 2-way server, and one that has four processors is a 4-way server. In addition to having multiple
processors in a server, each processor can have multiple cores. For example, a 4-way server might be filled
with 8-core processors, which would make a total of 4x8 = 32 processor cores in that particular server.
55
Semiconductors
Applications Processors
Application processors are processors used as the brains to run applications in tablets and smartphones.
Exhibit 38 shows details for leading application processors.
Most applications processors (except those made by Intel) have processor cores that are based on ARM
architecture.
o
Qualcomm and Apple implement the ARM processor core in their own circuit design. Nvidia and
Samsung use circuit designs which they licensed from ARM.
Qualcomm, Mediatek, Intel, and Nvidia are merchant chip companies that design processors which are
used in tablets and smartphones made by other companies.
Apple designs its own applications processors and has these manufactured at a foundry (currently TSMC
and Samsung). Apples applications processors are used solely in Apples own products, primarily iPad
and iPhone.
Samsungs processors are manufactured by Samsungs semiconductor division, which operates separately
from Samsungs smartphone and tablet businesses. In principle, we believe that these chips are available
for sale to any tablet or smartphone manufacturer, but we believe that in practice, Samsungs tablet and
smartphone processor are used primarily in Samsung devices. However, Samsungs systems divisions also
offer smartphones and tablets that use processors from other companies.
56
Die Size
(mm2)
Family
Follow-on
from SoFIA
Intro
2016E
Broxton
(Willow
Trail)
SoFIA
2015E
Broadwell Y
2H2014
Cherry Trail
Moorefield
2H2014E or early
2015E
2H2014E
Merrifield
Jan 2014
Bay Trail
2H2013
~100
Ivy Bridge Y,
Haswell Y
1H2013 2H2013
94,
177
Details
14nm integrated LTE/Processor
manufactured internally at Intel.
14nm
Atom
Goldmont,
converged cores for tablets and
smartphones. 64-bit.
For low end tablets and for
smartphones. Integrated global
3G HSPA+, connectivity with
Intel Atom. Manufactured at an
external foundry. Integrated LTE
expected in 1H15. Quad core
with integrated LTE for Rockchip
1H15, we expect there will also
be customized SoFIA chips for
Spreadtrum in 2015.
14nm, high end notebook-class
Core series processor cores, TDP
low enough for fanless tablets.
64-bit.
14nm
Atom
Airmont,
next
generation graphics. 64-bit.
Quad core 22nm Silvermont. 64bit.
4Q14E
Qualcomm
Clover Trail,
Clover Trail+
2H2012
1H2013
96
Medfield
1H 2012
63
Oak Trail
2011
65
Snapdragon
810/808
2H2014E
(sample),
2H2014/1H2015E
(production)
1H 2013
118
Snapdragon
800/600/400
Snapdragon
200
1H 2013
MSM8960
2012
MSM8660
Nvidia
Samsung
88
105
Tegra K1
1H2014 (32-bit)
2H2014 (64-bit)
Tegra 4i
1H2013
Tegra 4
1H 2013
~80
Tegra 3
2012
82
Tegra 2
2010
50
Tegra
2008
Exynos
1H15E
Exynos
ModAP
Currently
in
planning,
and
only available as
custom ASIC for
specific
customers
1H 2013
Exynos 5
32nm,
dual
core
in-order
Saltwell
Atom
x86
core,
Imagination PowerVR graphics,
64-bit.
32nm, single core Saltwell Atom
x86 core, 64-bit.
45nm, single core, Bonnell Atom
x86 core, 64-bit.
20nm, 64-bit, Hexa-core CortexA57 (2) with Cortex-A53 (4),
Adreno 430/418, integrated CAT
6 LTE-A
28nm HPm TSMC, quad/dual
Krait (proprietary ARM) core,
Adreno 3xx graphics, integrated
baseband.
8-core 64-bit 600
series chip announced in Feb
2014, expected to sample in
3Q2014
28nm LP TSMC, quad core ARM
Cortex-A5 processor, Adreno
203
graphics,
integrated
baseband
28nm TSMC, dual/quad core
Krait
processor,
integrated
baseband.
Qualcomm also
offers APQ series products which
are
standalone
processors
without integrated basebands
45nm TSMC, dual core Scorpion
processor, integrated baseband.
32-bit quad-core, 4-Plus-1 ARM
Cortex A15 CPU and Nvidiadesigned 64-bit dual Super Core
CPU (Denver). Kepler GPU.
28nm HPM, quad core Cortex A9
processor, GeForce graphics,
integrated
Icera
software
baseband
28nm HPL, quad core Cortex
A15 processor (with companion
core), GeForce graphics
40nm TSMC, quad core Cortex
A9 processor (with fifth control
core), GeForce graphics
40nm TSMC, dual core Cortex A9
processor, ARM v11, Windows
CE and Windows Mobile support,
GeForce graphics
65nm TSMC, ARM11 Core
14nm FinFET (with, we believe,
20nm metallization)
Integrated
LTE
Modem-AP
solution. Capable to support 4G
LTE Release 9 and Cat 4, at FDD
and TDD, in addition to legacy
2G and 3G mobile interfaces
28nm HKMG quad core Cortex
A15
processor/Cortex
A7
big.little (Octa variant with
Imagination graphics) or 32nm
Samsung, dual core Cortex A15
57
Semiconductors
processor
graphics.
Exynos 4
120
ARM
Mali
58
2011
with
Tablet Processor Market. Exhibit 39 shows Strategy Analytics estimates of worldwide tablet processor
market share in 2013 and estimates for H1 2014.
Strategy Analytics estimates that in Q2 2014, the tablet applications processor market grew 27% year over year,
to $945 million, with units rising 23% year over year, to 65 million units.
In the June 2014 quarter, Apple maintained its lead with 22% unit share in the tablet applications
processor market, with its A-series processors being the captive chips used in iPad. However, Apples
share was down from 25% in Q1 2014 and 30% for the full year 2013.
o We expect Apples share of the overall tablet processor share will continue to drop as Android
and Windows tablets ramp up in volume.
Intels share rose sharply, to 15% in Q2 2014 from 9% in Q1 2014 and 5% in 2013. The Strategy Analytics
numbers suggest that in the June 2014 quarter, Intel became the largest merchant chip supplier of tablet
processors, behind Apple, which makes tablet chips only for its own use.
Qualcomms unit share was flat sequentially at 12% in Q2 2014; however, this was an increase from 8% in
2013. Amazon switched from using TI to using Qualcomms processors in its Amazon Kindle Fire tablet,
and Qualcomm also appears to be establishing a foothold in some of Samsungs tablets.
MediaTeks share was also flat at 12% in Q2 2014, but was up from 8% in 2013, in part, we think, because
of ramping sales in tablets made and sold in China and elsewhere in Asia, both with Tier 1 tablet makers
such as Lenovo, as well as smaller tablet makers.
Rockchips share of 5% in the total global tablet market, with its business concentrated in Chinese tablets,
implies a strong market position in China.
Samsungs share declined slightly, to 4% in Q2 2014 from 5% in Q1 2014 and 9% in 2013. Samsung uses its
Exynos processors in many of its own tablets, though it also uses processors from other chip vendors in
some tablets.
Nvidias share was flat, at around 4% in Q2 2014, with no change from 2013 and down from 12% in 2012.
o Nvidia has said that it no longer considers the mainstream smartphone processor market to be
a good target for its Tegra processor. We wonder if Nvidia will be able to maintain an
appropriate level of tablet and smartphone business to support the investment needed to keep
its processors competitive in the tablet space.
Vendor
2Q14
Units
(M)
1Q14
Units
(M)
2Q13
Units
(M)
2013
Units
(M)
2Q14
Share
1Q14
Share
2Q13
Share
2013
Share
30%
Apple
14.4
13.5
14.4
72.5
22%
25%
27%
Intel
9.8
4.9
2.6
11.3
15%
9%
5%
5%
MediaTek
8.1
6.4
4.3
19.8
12%
12%
8%
8%
Qualcomm
8.0
6.7
3.1
18.3
12%
12%
6%
8%
Allwinner
5.8
5.5
5.9
25.6
9%
10%
11%
11%
Rockchip
3.5
3.4
3.3
13.9
5%
6%
6%
6%
Marvell
3.5
2.9
2.8
13.5
5%
5%
5%
6%
Samsung
2.8
2.5
5.0
20.9
4%
5%
9%
9%
Nvidia
2.4
2.1
1.9
8.8
4%
4%
4%
4%
Amlogic
1.3
1.2
1.6
6.2
2%
2%
3%
3%
Others
5.4
4.9
8.2
10
Total
64.9
54.1
53.0
Source: Strategy Analytics, Wells Fargo Securities, LLC estimates.
30.2
8%
9%
15%
13%
241.1
100%
100%
100%
100%
59
Semiconductors
Smartphone Processor Market. As shown in Exhibit 40, Strategy Analytics estimates show that in Q2
2014, the smartphone applications processor market grew about 20% sequentially and 22% year over year, to
$5.2 billion, while unit shipments increased approximately 17% sequentially and 25% year over year, to 364
million units.
For Q2 2014, Qualcomm expanded its share of the smartphone apps processor market with a 41% unit
share, up from 37% in Q1 2014 and 34% in 2013, followed by MediaTek, with 24% share, and Apple, with
12% share.
Strategy Analytics estimates that Samsungs smartphone applications processor shipments continued to
decrease, falling 20% sequentially and 49% year over year in Q2 2014, implying a shift by Samsung away
from use of its own Exynos processors in its smartphones.
Nvidias smartphone processor shares remains low, at about 1% in Q2 2014, Q1 2014, and for full year
2013.
o
The Strategy Analytics numbers suggest that Intel continues to have a very small position in the
smartphone market.
o
Nvidia has said that it no longer considers the mainstream smartphone market is a particularly good
market for Tegra, though it implied it still has an interest in the super phone market. We would not
be surprised if Nvidia eventually exited the phone processor market altogether.
Intels lack of an integrated baseband-processor product appears to be limiting its ability to penetrate
the smartphone processor market. We do not expect Intel to gain any significant share in the
smartphone processor market until perhaps 2015 at the earliest, with the launch of its SoFIA product
in late 2014.
The numbers we have compiled are for stand-alone and integrated applications processors combined.
o
The numbers shown for Apple, Samsung, and Intel are primarily stand-alone processors. We believe
that Apple iPhone and Samsung Galaxy account for the majority of stand-alone processor use.
The numbers shown for MediaTek, Spreadtrum, and Broadcom are primarily for integrated
processors (and hence, overlap with the baseband unit data we discuss in the next section).
Strategy Analytics estimates that about 10-15% of Qualcomms applications processor shipments are
stand-alone processors and 85-90%, integrated processors.
Apples smartphone processor shipments are all discrete processors (Apples A8, A7, etc.) used in Apple
iPhones, as are Samsungs (Samsungs Exynos processors). However, the units shipped by most of the other
major vendors are, we believe, for the most part integrated modem/processor chips or chipsets. Therefore, the
unit numbers included in Exhibit 40 are to some extent duplicated in the modem data we present in the next
section.
Exhibit 40. Total Smartphone Processor Unit Market Share
2Q14
1Q14
2Q13
Units
Units (M)
Units (M)
(M)
150.9
115.7
103.6
1
Qualcomm
85.8
76.5
53.7
2
MediaTek
45.4
35.9
34.5
3
Apple
35.8
30.1
45.8
4
Spreadtrum
15.6
13.4
10.0
5
Marvell
9.6
17.2
15.5
6
Broadcom
9.6
11.9
18.6
7
Samsung
3.1
3.0
2.1
8
HiSilicon
2.9
3.0
0.5
9
NVIDIA
1.6
1.0
0.7
10
Intel
4.1
4.0
5.6
Others
364.2
311.6
290.5
Total
Source: Strategy Analytics, Wells Fargo Securities, LLC estimates.
2Q14
Rank
60
Vendor
2013
Units (M)
2Q14
Share
1Q14
Share
2Q13
Share
2013
Share
397.6
219.9
162.9
144.2
38.3
77.9
73.0
7.5
6.6
2.2
26.9
1156.9
41%
24%
12%
10%
4%
3%
3%
1%
1%
0.4%
1%
100%
37%
25%
12%
10%
4%
6%
4%
1%
1%
0.3%
1%
100%
36%
18%
12%
16%
3%
5%
6%
1%
0.2%
0.3%
2%
100%
34%
19%
14%
12%
3%
7%
6%
1%
1%
0.2%
2%
100%
Basebands. Baseband processors (modems) are the primary communication chip for both voice and data in a
smartphones. Some tablets also have baseband modules for long-range communications, though many tablets
are offered with just short-range communications chips (WiFi).
Exhibit 41, Exhibit 42, and Exhibit 43 show baseband unit market share for 2013 and H1 2014 for the overall
cellphone market, the 3G (3rd generation) market and the 4G market, respectively. Originally wireless
handsets were used for voice communications. At the 3G level, the communications standard became capable
of supporting reasonably good data transmission, as well, and so smartphones, which have a fair amount of
computing capability in addition to supporting voice communications, tend to have at least 3G
communications capability. More recently, various regions of the world have been building out 4G networks,
which work on the long term evolution (LTE) standard.
According to Strategy Analytics estimates, total baseband shipments increased 5% sequentially and 6% year
over year in the June quarter; 3G baseband shipments increased 6% sequentially and 8% year over year; and
LTE baseband shipments increased 28% sequentially and 98% year over year. Strategy Analytics estimates that
Qualcomm captured 94% unit share in the LTE baseband market in H1 2013.
Qualcomm continues to dominate the baseband market, with its overall blended share rising to 39% in Q2
2014 (from 34% for Q1 2014 and 32% for full-year 2013), and its 3G share rising to 46% (from 42% in Q1
2014 and 45% for full-year 2013); however, its LTE share was flat sequentially at 85% and down from 93%
share for full-year 2013.
Strategy Analytics estimates that Marvells LTE shipments jumped 34% sequentially, to 5.4 million units
in Q2 2014, an impressive jump from 0.7 million units for the full year 2013.
o
Strategy Analytics is projecting Marvells LTE shipments might decline to 4.4 million units in Q3
2014.
Strategy Analytics estimates that Intel shipped 3.3 million LTE units in the June quarter (up 27%
sequentially).
Strategy Analytics estimates MediaTek shipped 1.3 million LTE units in the June quarter, though it
estimates MediaTeks 3G market share jumped to 27% in the June 2014 quarter, up from 18% for the full
year 2013 (flat sequentially).
In Q2 2014, Broadcom announced that it had decided to proceed with closing down its baseband business.
61
Semiconductors
Exhibit 41. Total Baseband Unit Market Share
2Q14
Rank
Vendor
2Q14
Units
(M)
1Q14
Units
(M)
2Q13
Units
(M)
2013
Units
(M)
2Q14
Share
1Q14
Share
2Q13
Share
2013
Share
Qualcomm
224.1
187.0
172.7
747.8
39%
34%
32%
32%
MediaTek
158.8
158.1
121.7
569.0
27%
29%
22%
24%
Spreadtrum
91.2
82.5
87.7
346.0
16%
15%
16%
15%
Intel
41.6
50.9
74.6
335.2
7%
9%
14%
14%
Marvell
19.6
17.0
11.0
50.5
3%
3%
2%
2%
RDA
15.9
20.2
38.5
138.8
3%
4%
7%
6%
Broadcom
9.6
17.3
16.4
79.9
2%
3%
3%
3%
VIA
5.1
5.0
5.0
20.9
1%
1%
1%
1%
HiSilicon
3.9
3.0
0.1
3.6
1%
1%
0%
0%
10
Samsung
2.9
2.5
2.5
10.7
0%
0%
0%
0%
Others
7.1
7.6
16.5
70.4
1%
1%
3%
3%
579.9
551.0
546.7
2372.8
100%
100%
100%
100%
Total
Vendor
Qualcomm
MediaTek
Spreadtrum
Marvell
Intel
Broadcom
VIA
HiSilicon
ST-Ericsson
RDA
Others
Total
2Q14
Units
(M)
1Q14
Units
(M)
2Q13
Units
(M)
2013
Units
(M)
2Q14
Share
1Q14
Share
2Q13
Share
2013
Share
119.5
70.9
28.0
14.1
9.8
9.1
5.1
1.3
0.6
0.4
3.5
262.4
104.8
66.6
23.6
12.8
13.7
16.3
5.0
1.1
0.6
3.4
247.9
113.4
40.1
28.4
10.6
22.4
15.8
5.0
0.1
4.3
2.6
242.8
445.8
175.3
99.4
48.1
88.7
78.6
20.9
2.2
19.4
12.9
991.2
46%
27%
11%
5%
4%
3%
2%
0%
0%
0%
1%
100%
42%
27%
10%
5%
6%
7%
2%
0%
0%
0%
1%
100%
47%
17%
12%
4%
9%
7%
2%
0%
2%
0%
1%
100%
45%
18%
10%
5%
9%
8%
2%
0%
2%
0%
1%
100%
2Q14
Units
(M)
1Q14
Units
(M)
2Q13
Units
(M)
2013
Units
(M)
2Q14
Share
1Q14
Share
2Q13
Share
2013
Share
Qualcomm
103.4
81.0
57.6
296.4
85%
85%
94%
93%
Marvell
5.4
4.0
0.7
4%
4%
0%
0%
2Q14
Rank
Intel
3.3
2.6
0.1
3.2
3%
3%
0%
1%
Samsung
2.9
2.5
2.5
10.7
2%
3%
4%
3%
HiSilicon
2.6
1.9
0.0
1.4
2%
2%
0%
0%
MediaTek
1.3
1%
0%
0%
0%
Altair
0.6
0.6
0.0
1.0
1%
1%
0%
0%
GCT
0.5
0.4
0.6
2.4
0%
0%
1%
1%
Broadcom
0.4
0.9
0.0
0%
1%
0%
0%
Sequans
0.4
0.3
0.0
0.2
0%
0%
0%
0%
Others
0.6
0.5
0.4
10
Total
121.4
94.7
61.3
Source: Strategy Analytics, Wells Fargo Securities, LLC
62
1.0
0%
1%
1%
0%
317.1
100%
100%
100%
100%
Although the microprocessor makers provide graphics circuitry with their microprocessor products (integrated
graphics), there is also a fairly large market for separate chips that just do graphics: discrete graphics processor
units (GPU). Nvidia and AMD are the primary makers of GPUs (see Exhibit 44). AMDs discrete graphics
product line (and the companys integrated graphics expertise) comes from its acquisition of ATI in 2006.
AMD still uses the ATI brand for its discrete graphics products.
Exhibit 44. Discrete Graphics Chip--Market-Share Trends
DiscreteGPUMarketShare
80%
70%
60%
50%
40%
30%
20%
10%
Nvidia
Q314
Q214
Q114
Q413
Q313
Q213
Q113
Q412
Q312
Q212
Q112
Q411
Q311
Q211
Q111
Q410
Q310
Q210
Q110
Q409
Q309
Q209
Q109
Q408
Q308
Q208
Q108
0%
AMD/ATI
The bulk of GPUs are used in desktop and notebook computers. Exhibit 45 shows an attached rate of GPUs to
microprocessors, our estimate of the proportion of microprocessors that are used together with a GPU in a
system.
Our graphs suggest that historically, some 30-40% of desktops, workstations, and servers, and about 2530% of notebooks feature discrete GPUs. Most microprocessors are sold with integrated graphics.
Therefore, the proportion of systems that contain both integrated graphics and discrete graphics is fair.
We think that one risk to the attach rate of GPUs is the recent action by Intel and AMD to bring graphics
circuitry onto the microprocessor chip. Prior to 2011, integrated graphics circuitry was part of the chipset.
Integrated graphics circuitry that is included on the microprocessor chip benefits from having access to
the memory included on the microprocessor chip and direct communications with the microprocessor
without the delays associated with driving signals from chip to chip. Intel has made the claim that its
integrated graphics performance is now equal to some GPUs. AMD already has significant graphics
expertise from its GPU product lines, and our understanding is that AMD is using similar circuitry in its
integrated graphics to what it has in its GPUs. If Intel and AMD truly do close the performance gap
between integrated graphics and GPUs, this could well result in a falling attach rate for GPUs in the future.
While driving graphics performance for PCs is the single biggest application for GPU chips, other
applications include graphics for workstations, i.e., improving the performance applications like
computer-aided design (CAD). Since graphics processing is very computationally intensive, GPUs have a
great deal of mathematical computation capability. One application for GPUs that Nvidia is promoting is
coprocessors in high-performance computing.
63
Semiconductors
Exhibit 45. Discrete Graphics Chip--Attach Rate
50%
40%
30%
20%
10%
Desktop/Workstation/Server
Source: Mercury Research and Wells Fargo Securities, LLC
64
Q214
Q314
Q413
NotebookExcl.Tablets
Q114
Q213
Q313
Q113
Q312
Q412
Q112
Q212
Q311
Q411
Q211
Q410
Q111
Q210
Q310
Q110
Q309
Q409
Q109
Q209
Q308
Q408
Q108
0%
Q208
DiscreteGPUShipmentsAsAPercentageOf
MicroprocessorShipments
60%
(1) Volatile memory. This memory forgets what it is supposed to be remembering when the power is
switched off. The main type of volatile memory is DRAM.
(2) Non-volatile memory. This memory recalls what it is supposed to be remembering even when the
power is switched off. The main type of nonvolatile memory is flash memory, which itself is divided into
NAND flash and NOR flash.
Exhibit 46. Memory Segment Breakout (2005 Through 2013)
NOR Flash
16%
Other
Memory
7%
Other Memory
9%
NAND Flash
22%
DRAM
53%
NAND Flash
33%
DRAM
50%
NOR Flash
10%
NAND
Flash
41%
DRAM
52%
NOR
Flash
3%
Source for all charts: Semiconductor Industry Association and Wells Fargo Securities, LLC
Exhibit 46 shows that DRAM is the largest component of the memory market, accounting for 52% of the total
memory market. Over the past few years, NAND Flash sales have ramped more quickly than the other memory
segments, climbing to 41% of the overall memory market in 2013 from 22% in 2005. NAND flash has probably
cannibalized NOR flash to some extent, while also serving new markets.
DRAM
There are two main types of volatile memory:
(1) DRAM. DRAM (or dynamic RAM) forgets what it is supposed to be remembering even when the power is
switched on, and so it needs to constantly remind itself of what it has been told (refresh) several hundred
times per second.
(2) SRAM. SRAM (or static RAM) remembers what it has been told as long as the power stays on.
DRAM is less expensive than SRAM and can be made in higher densities. Since the DRAM market is more than
10x the size of the SRAM market, we address only DRAM in this report.
DRAM is random access memory (RAM). This means that the information can be retrieved from any memory
cell without looking through all the other memory cells. DRAM and SRAM are random access. NOR flash is
random access, but NAND flash is serial access. For a NAND flash memory, all the cells in a whole line (a row)
must be looked at one after the other.
DRAM is used as the main type of temporary silicon memory in computers that the microprocessors use when
they are actively working. DRAM chips are usually described in terms of bits (the total number of digits that
65
Semiconductors
can be stored, with each digit being either 1, or o), whereas computer memory is described in terms of bytes.
There are 8 bits in a byte. A typical PC (e.g. a consumer laptop) might have 8 GB of DRAM. 8GB = 8 Gigabytes
= 8 billion bytes. A 4 Gb chip costs about $3.00-$4.00 today (late 2014). Therefore, a typical PC might contain
about $64 worth of DRAM chips in it, i.e., 16 chips of with 4 Gb of memory on each chip, or 32 chips with 2 Gb
of memory each.
DRAM is used in other electronic devices, too, including wireless handsets, networking systems, and
communications infrastructure systems. However, we estimate that more than half of the worlds DRAM
consumption is for computers.
Exhibit 47 shows the main DRAM makers. Samsung is the worlds largest DRAM vendor (and is also a
significant producer of flash memory), producing 36% of the worlds DRAM in 2013 (down from its 41% share
in 2012). In our opinion, Samsung is also a leader in DRAM manufacturing technology. In 2013, Hynix held
second place, with a 27% share, while Micron held third place with 22% share, followed by Elpida, in fourth
place, with a 7% share. However, Elpida filed for bankruptcy in February 2012, after which Micron acquired
the company, completing the transaction in 2013. Micron, with its Elpida acquisition, is now of comparable
size to Hynix in DRAM.
Exhibit 47. DRAM Market Share (2013)
DRAM Market 2013 (Revenue)
Total Market $34.9 billion
Micron Technology
22%
Elpida Memory
7%
Nanya Technology
4%
Winbond Electronics
1%
SK Hynix
27%
Samsung Electronics
36%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Devices, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
The semiconductor memory business is highly capital-intensive, but still commodity-like, and for this reason,
we think memory has historically had bigger cyclical swings than the rest of semiconductors (see Exhibit 48
and Exhibit 54). When the business environment is weak, DRAM pricing tends to fall, and when business is
strong, DRAM pricing rises. We think that these big swings in business, coupled with the capital intensity of
the DRAM business, make it difficult for most DRAM companies to achieve consistent value accumulation over
the long run. In recent years there has been a fair amount of consolidation in the memory industry, most
recently with the acquisition of Elpida by Micron. We think it is possible that as the number of major DRAM
makers shrinks and the growth of the market slows, pricing dynamics might stabilize.
66
Yr/Yr Growth
100%
50%
0%
Total ICs
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
Dec-01
Dec-00
Dec-99
Dec-98
Dec-97
Dec-96
Dec-95
Dec-94
Dec-93
Dec-92
(100%)
Dec-91
(50%)
DRAM
1,000,000,000
100,000,000
10,000,000
1,000,000
DRAM Mbits
77%/yr (1990-2002)
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
Dec-01
Dec-00
Dec-99
Dec-98
Dec-97
Dec-96
Dec-95
Dec-94
Dec-93
Dec-92
Dec-91
Dec-90
100,000
DRAM bit growth averaged around 77% per year from 1990 to 2001 and has been around 53% per year from
2001 through H1 2014. Our graph shows lower bit growth than 53% per year from H2 2012 to the present (H2
2014), it is unclear to us at the moment whether this further slowing is temporary or a secular step down.
These bit growth numbers are far higher than the unit growth rates of the electronic systems such as
computers that DRAM sells into. DRAM growth is driven not just by growth of system units shipped, but also
by the ever increasing amount of DRAM in the system. However, the number of DRAM chips in the system
does not typically increase with time. Instead, it is the number of bits on each DRAM chip that increases.
The commodity-like nature of DRAM might be expected to increase the price elasticity of the demand for the
product. However, Exhibit 49 shows that, historically, DRAM bit growth has been very consistent, with far less
variation than shown in the DRAM revenue graph of Exhibit 48.
67
Semiconductors
Exhibit 50 shows how DRAM price per bit has dropped over time. We have plotted price per bit, calculated
from total DRAM sales divided by total bit shipments. This aggregates DRAM prices over many different
densities of chips (i.e., today the average price per bit is calculated from sales of several different sizes of
memory: 2 Gb, 1 Gb, 512 MB, 256 MB, 128 MB, etc.). Lower density chips tend to have a higher price per bit
than the higher-density chips (there is less demand and hence, less pricing elasticity for older, legacy chips,
and for these chips, the somewhat fixed packaging and test costs are a higher proportion of total chip cost).
And so, the data of this graph often reflects a price per bit that is higher than the price per bit of the highdensity volume runner at any given time.
It can be seen that actual price per bit can deviate significantly from the longer-term trend, but invariably,
seems to return to the trend line. Our trend line shows that over the past two and a half years, DRAM price per
bit has been moving sideways. As we discuss below, it is very possible that price per bit might not be declining
as rapidly as it used to. However, we think that some investors have the unrealistic view that DRAM bit costs,
and hence, the price per bit trend, may have stopped going down altogether. We expect that in time, DRAM
prices will likely correct downward toward the longer term trend line.
We think the reason is that the trend line, in fact, approximates the cost per bit of making DRAM. Over time,
transitions to smaller line widths allow memory makers to reduce the size of each memory cell, packing more
memory (more bits) into the same silicon area. Since manufacturing cost is proportional to silicon area used,
this reduces price per bit. We think that the pace of technology transitions in DRAM might be slowing. Micron
has been noting for some time now that there are growing challenges associated with memory technology
transitions. Compared to logic circuits (e.g., chips made by companies like Intel, and foundries like TSMC),
memory chips present particular challenges. Not only does the memory chip manufacturer need to successfully
image the smaller geometries, it needs to ensure that as the memory cell shrinks it still holds enough electrical
charge to properly retain the information that its supposed to be remembering.
Exhibit 50. DRAM Price Per Megabit
$10.0000
$1.0000
DRAM Price/Mb
$0.1000
$0.0100
$0.0010
DRAM price/Mb
68
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
Dec-01
Dec-00
Dec-99
Dec-98
Dec-97
Dec-96
Dec-95
Dec-94
Dec-93
Dec-92
Dec-91
Dec-90
$0.0001
Exhibit 49 and Exhibit 50 allow us to calculate an underlying growth rate for overall DRAM segment revenue.
From 1990 to 2001, DRAM bit growth rate was about 77% per year and the trend for price per bit declines was
35% per year. A 77% per year bit growth rate, matched by a 35% per year price per bit drop, results in 15%
revenue growth per year:
1.77*(1-0.35) =1.15
From 2001 to today, the bit growth rate appears to have slowed to 53% per year, with the price per bit decline
moderating to 29% per year. This gives a revenue growth trend of 9% per year:
1.53*(1-0.29) = 1.09
If indeed the pace of memory transitions is slowing, we think that the trend toward lower DRAM price per bit
might also be moderating. We think it is possible that the trend line of DRAM price per bit might show less
than the 29% per year decline we have seen over the past ten years, but also that the trend line of DRAM bit
growth might be less than the 53% per year of the past ten years.
Exhibit 51 and Exhibit 52 show DRAM price data. DRAM can be bought on contract in which prices are
negotiated in advance or on the spot market, which offers a constantly changing price. Up to quite recently,
DRAM contract prices were renegotiated twice per month. However, since about mid-2010, our understanding
is that Micron has moved to setting contract prices once a month with most customers and that more recently
(2014), DRAM contract negotiations may have move to a quarterly basis for some customers.
DRAM is sold on the spot market and also by contract directly from the DRAM manufacturers to systems
makers. We believe that the bulk of DRAM is sold on contract. We estimate that Micron typically sells
70-90% of its DRAM in any quarter on contract. The spot market is often thought, correctly, we think, to be a
leading indicator of contract pricing. The reason is that spot prices, which change daily, quickly reflect the
current supply and demand situation.
When demand is rising and supply tightens, spot prices tend to rise quickly, while DRAM makers are
locked into lower prices for their contracts with large PC makers. Since the contract prices represent a
commitment to pay a certain price, the PC makers are reluctant to let contract prices rise at each
negotiation until evidence accumulates that there is indeed more than a temporary shortage of memory.
Contract prices then tend to begin rising toward spot prices, with a time lag.
Similarly, when demand is falling and there is plenty of supply, there is little demand for memory on the
spot markets and spot price falls. However, the large memory customers still need memory and memory
makers resist lowering prices for as long as possible. Eventually contract prices do fall though, relatively
slowly, toward the lower spot prices.
PCs represent the biggest end market for DRAM, and generally, PCs require high-density chips (chips with the
maximum amount of memory). Therefore, at any given time, one of the higher density memories is the volume
runner for the DRAM industry. As time passes, the DRAM industry transitions to higher and higher densities.
In the pricing figures we have plotted are the prices of the volume runners: 256 Mb chips from 2002 to 2004,
512 Mb chips from 2004 to 2008, 1 Gb chips from 2008 to the start of 2011, 2Gb from the start of 2011 to early
2013, and 4Gb from early 2013 to now (late 2014).
DRAM circuit technologies also change with time. Over the past few years, double data rate (DDR) DRAM, and
then DDR2 DRAM began to become mainstream. DDR is a circuit design that allows the memory to do
operations twice in a clock cycle rather than just once, doubling the number of calculations that the memory
can achieve in any given time. DDR2 is a type of DDR memory that operates at higher frequencies. Over 2010
and 2011, the DRAM industry transitioned to using DDR3 memory, which transfers data at twice the rate of
DDR2, while also featuring reduced power consumption.
Currently (November 2014), the DRAM contract price is running at close to $4.00 for a 4 Gb DDR3 chip. As
can be seen from the graphs, the price for the volume runner DRAM chip at any given point in time can
range from below $1.00 to as much as $5.00-6.00. Given that PCs typically contain 16 or 32 of the volume
runner DRAM chips (8 bits make a byte), the DRAM content of a PCs can fall to as low as $20 and rise to as
high as $200. We think about $50 of DRAM content per PC is typical.
69
Apr-15
Jan-15
Oct-14
Jul-14
Apr-14
Jan-14
Oct-13
Jul-13
Apr-13
Jan-13
Oct-12
Jul-12
Apr-12
Jan-12
Oct-11
Jul-11
Apr-11
Jan-11
Oct-10
Jul-10
Apr-10
Jan-10
Oct-09
Jul-09
Apr-09
Jan-09
Oct-08
Jul-08
Apr-08
Jan-08
Oct-07
Jul-07
Apr-07
Jan-07
Oct-06
Jul-06
Apr-06
Jan-06
Oct-05
Jul-05
Apr-05
Jan-05
Oct-04
Jul-04
Apr-04
Jan-04
Oct-03
Jul-03
Apr-03
Jan-03
Oct-02
Jul-02
Apr-02
70
$4.00
$3.00
$2.00
Contract Price
Jun-15
Mar-15
Dec-14
Sep-14
Jun-14
Mar-14
Dec-13
Sep-13
Jun-13
Mar-13
Dec-12
Sep-12
Jun-12
Mar-12
Dec-11
Sep-11
Jun-11
Mar-11
Dec-10
Sep-10
Jun-10
Mar-10
Dec-09
Sep-09
Jun-09
Mar-09
Dec-08
Sep-08
Jun-08
Mar-08
Dec-07
Sep-07
Jun-07
Mar-07
Dec-06
Sep-06
Jun-06
Mar-06
Dec-05
Sep-05
Jun-05
Mar-05
Dec-04
Sep-04
Jun-04
Mar-04
Dec-03
Sep-03
Jun-03
Mar-03
Dec-02
Sep-02
Jun-02
Mar-02
4Gb DDR3
2Gb DDR3
1Gb DDR3
1Gb DDR2
512Mb DDR2
512Mb DDR
256Mb DDR
4Gb DDR3
2Gb DDR3
1Gb DDR3
1Gb DDR2
512Mb DDR2
512Mb DDR
256Mb DDR
256Mb SDRAM
$4.00
$3.00
$2.00
Spot Price
$6.00
$5.00
$1.00
$-
$7.00
$6.00
$5.00
$1.00
$-
Flash memory. In the past there have been many types of non-volatile memory. Flash memory has emerged
as the most important type of non-volatile memory. Other, older types of non-volatile memory include readonly memory (ROM), programmable read-only memory (PROM), erasable read only memory (EPROM),
electrically erasable read-only memory (EEPROM), etc.
There are two main types of Flash memory:
(1) NAND flash. NAND flash is often used in consumer applications for the storage of large amounts of
digital data. Smartphones, digital cameras, iPods and USB computer drives are important end markets for
NAND flash. Tablets typically used Flash-based solid state drives (SSD) for system data storage and the
use of Flash-based solid state drives (SSD) instead of hard-disk drives in notebook computers is likely to
be an important emerging application for NAND flash over the next few years. (See our discussion in the
Selected Technology Topics section, which follows.)
(2) NOR flash. NOR flash has slightly better performance characteristics than NAND flash, but also a higher
cost. NOR flash is used primarily in applications that require relatively small amounts of non-volatile
memory.
From a circuit point of view, NOR flash differs from NAND flash because it is random access, which means that
information in any cell can be looked at, at any point in time. Therefore, there must be a connection (a bit line
contact) to every memory cell, increasing the size of the cell. NAND flash is serial access; a whole row of cells
must be read, one after another. This reduces the size of the NAND flash cell (no need to have a bit line contact
per cell) reducing the cost. However, it also reduces the speed of the NAND flash, as well as increasing the
error potential; if something goes wrong with one cell, it can become impossible to get information in and out
of a whole row.
NAND Flash.
Exhibit 53 shows the top NAND market players for 2013. Samsung holds approximately 32% revenue share,
followed by Toshiba, with a 26% share, and SanDisk, with a 16% share. From a technology point of view, the
equipment needed to make NAND flash is similar to the equipment needed for DRAM manufacturing, and so,
NAND capacity is fungible with DRAM. As a result, the largest NAND makers tend to also be large DRAM
producers.
Samsung and Toshiba have been significant manufacturers of NAND for some time.
SanDisk does not have direct ownership of memory manufacturing facilities, but participates in a number
of manufacturing joint ventures with Toshiba.
Micron was involved in manufacturing NAND flash for many years at a relatively low level. In early 2006,
Micron and Intel formed a joint venture, Intel-Micron Flash Technology (IMFT), which invested
substantial capital in NAND flash capacity. More recently, Micron began expanding capacity in the joint
venture with Intel by equipping a facility in Singapore. Since Intel decided not to contribute capital for the
manufacturing equipment in this facility, Micron has a right to the bulk of the output from the Singapore
NAND fab. In 2012, Intel sold some of its joint venture stake back to Micron, further increasing Microns
ownership of NAND facilities and reducing Intels.
71
Semiconductors
Exhibit 53. NAND Market Share (2013)
SanDisk, 16%
SK Hynix, 10%
Intel, 4%
Others, 1%
Toshiba, 26%
Samsung
Electronics, 32%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Devices, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
NAND flash is often viewed by investors as an emerging area that promises significant growth. Many memory
companies have been investing substantial capital in expanding NAND flash capacity in recent years.
Exhibit 55 shows NAND megabits shipped, and Exhibit 54 year-over-year growth rates for total IC and NAND
sales. Despite the large growth potential for NAND, large year-over-year NAND pricing declines have
pressured revenue growth in the NAND segment.
72
Yr/Yr Growth
50%
25%
0%
(25%)
(50%)
(75%)
Total ICs
Aug-14
Aug-13
Aug-12
Aug-11
Aug-10
Aug-09
Aug-08
Aug-07
Aug-06
Aug-05
(125%)
Aug-04
(100%)
NAND
10,000,000,000
1,000,000,000
100,000,000
10,000,000
NAND Mbits
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
1,000,000
73
Semiconductors
NAND bit growth averaged around 218% per year from 2004 to 2008 and has been around 57% per year from
2008 through 2014. The use of NAND in a variety of different applications has driven high bit growth for
NAND over the past several years. Wireless handsets, portable music players, and digital still cameras have
been big sources of NAND bit growth in the past. Today tablets, solid state drives, and smartphones are driving
high NAND bit growth. Using analysis similar to the DRAM bit growth analysis we provide in the preceding
text, we conclude from SIA data that NAND bit growth has been on average:
The cost dynamics of NAND are, we think, very similar to that of DRAM, and so it might be expected that the
trend of NAND cost per bit and price per bit declines might match the 29% per year trend for DRAM. However,
as we discuss in the NAND section, there is an additional technology factor associated with NAND, which is the
ability of a single NAND cell to hold more than one piece of information at a time. NAND flash can either be
single-level cell (SLC), or multi-level cell (MLC). Semiconductor memory is designed to store a 1 or a 0 in a tiny
piece of circuitry called a memory cell. For DRAM, each cell can store just a single 1 or 0. NAND flash can be
designed as a single-level cell (just one 1 or 0) or a multilevel cell (two 1s or 0s; i.e., 00, 010, 01 or 11) per
memory cell. The cell sizes of an SLC chip are the same as for an MLC chip. Therefore, an MLC chip can store
twice as much information as a similar-sized SLC chip, for the same manufacturing cost, or, looking at it in
another way, the cost of making a fixed amount of memory (e.g., 8 Gb) in an MLC chip is about half of what it
would cost to make the same amount of memory in an SLC chip. However, since the additional storage in an
MLC cell is achieved by using finer voltage divisions to decide what the stored information is, MLC chips have
a lower reliability (sometimes they remember things incorrectly) than SLC chips and so SLC NAND is still used
in some applications where high reliability is more important than cost.
Exhibit 56 shows how NAND price per bit has dropped over time. We have plotted price per bit, calculated
from total NAND sales divided by total bit shipments. This aggregates NAND prices over many different
densities of chips (i.e., today the average price per bit is calculated from sales of several different sizes of
memory: 32 Gb, 16 Gb, 8 Gb, 4 Gb, 2 Gb, etc.).
The transition from SLC to MLC technology contributed to price per bit declines of about 60% per year from
2004 through 2008. The possibility of future MLC technologies, i.e., the transition to triple-level cells (TLC)
from two-level cells could help offset struggles with moving to more advanced technologies, driving the trend
of NAND cost per bit and price per bit declines. In the past 3-4 years, the SIA data suggest that the price per
bit declines have been on the order of 20-25% per year.
74
NAND Price/Gb
$10.0000
$1.0000
$0.1000
price/Gb
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
$0.0100
Exhibit 57 shows pricing for NAND. At the time of writing of this report (H2 2014), the price of a 64Gb chip
was running at close to $3 (i.e., $3x8=$24 for 64GB, or $0.38 per Gigabyte).
Exhibit 57. NAND Pricing
$20.00
$18.00
$16.00
$14.00
Price
$12.00
$10.00
$8.00
$6.00
$4.00
$2.00
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Sep-14
Dec-14
Mar-15
Jun-15
$-
75
Semiconductors
SSDs use less power, in part because there are no mechanical parts that use power. An HDD has one or
more disks that rotate at high speeds.
SSDs are more robust and tolerant to movement. In an HDD there is a disk that rotates at high speeds
with a magnetic head that has to position itself a tiny distance from the disk to read the magnetic signals.
SSDs can access data more quickly. Among other things, this can improve the start-up time of a notebook.
This has become a more important factor in the past year as tablets, such as the iPad, have begun to
feature instant-on capabilities. As a result of these products, consumers have come to demand similar
features in their notebooks. Intel has made efforts to address this demand with its Ultrabook and 2-in-1
form factor concepts of thin sub $1,000 computers with SSDs, instant on capabilities and all-day battery
life.
These advantages are particularly important for notebooks. However, for solid state drives to achieve
widespread use in notebooks and in computers in general, they must have a cost that is close to that of a hard
disk drive. The following are some simple calculations and considerations related to the cost of HDDs and
SSDs
An HDD must have a magnetic platter (the hard disc) and a sophisticated read and write head etc., so
there is a minimum cost, no matter how little memory is on it. Standardized hard disk drives are not
generally sold with just 1 GB of memory on them, but if they were their cost or price would probably be
comparable with that of HDDs offering several tens of GB. It is only when a hard disk drive reaches high
capacities (100s of GB or more) that the cost of the hard disk drive goes up (see Exhibit 58).
Unlike a hard disk drive, flash memory has a low starting cost, maybe a few dollars for the controller. The
cost of any given amount of NAND flash scales fairly linearly with the amount of flash (though there is a
somewhat fixed packaging and test cost for the chips.) This is illustrated in Exhibit 58. The cost-tocapacity graph for an SSD is roughly a straight line with a fairly small initial cost. Today (November 2014)
the cost of 8GB of MLC NAND flash (a 64Gb chip) is about $3.00, so the NAND memory content of a
64GB drive costs about $24 (8 bits to a byte, so it takes eight 64Gb chips to make 64GB of memory).
A typical traditional notebook might have an HDD with capacity of the order of 500 GB. If we reckon that
the transition to SSDs might begin to gather momentum in mainstream notebooks when the price of a 256
GB SSD is roughly the same as that of an HDD, it might perhaps be necessary for the 256 GB of NAND
memory content in the drive to have a cost that is close to $30.
These calculations show that the price of NAND flash today is about 3x what it might need to be for SSDs
to begin matching the cost of HDDs in mainstream notebooks. However, NAND flash is widely used in
tablets.
76
Number of Gigabytes
NOR flash. Exhibit 59 shows 2013 market share for NOR flash. Micron and Spansion are two major players
in the NOR flash market, with a combined share of about 50%.
Exhibit 59. NOR Market Share (2013)
Others, 2%
Micron Technology,
27%
Integrated Silicon
Solution, 1%
Eon Silicon Solution,
2%
Microchip
Technology, 3%
GigaDevice
Semiconductor, 5%
Spansion, 24%
Samsung
Electronics, 5%
Winbond
Electronics, 14%
Macronix
International, 17%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Devices, Worldwide, 2013, Bryan Lewis et alia, March 31, 2014
77
Semiconductors
NOR Flash. NOR appears to have limited growth prospects and has shrunk to a total sales level that is about
one-tenth that of NAND flash. We think that the erosion in NOR demand over the past several years is
associated with NAND flash emerging as the technology of choice for addressing the bulk of the non-volatile
memory needs. NOR Flash typically has a higher cost, albeit also higher reliability than NAND flash. NAND is
used in the bulk of high-density, high-volume applications, NOR has been increasingly relegated to the task of
storing of critical system information. One common example of this in the past was in a smartphone in which
high-density NAND flash was used for the large amount of memory used for data recorded by the user (e.g.,
the storing of photographs taken on a phone), while lower density NOR flash was used for code storage of the
information needed for the communications protocols used by the phone. We believe that in recent years, even
the code storage function has been usurped by NAND flash in most phones.
However, we think that NOR flash has a long-term future in various applications that require low to medium
quantities of memory, which must function with high reliability over a long time period. This is true, for
example, of many industrial and automotive applications, as shown in Exhibit 60.
The maximum-density segment of NOR has moved up only one notch from 2004 through 2013, from >64
Mb to >128 Mb. Admittedly though, today NOR flash chips are available in densities of up to several
gigabit.
In 2004, the highest density segment (>64Mb) accounted for 1% of units (and by implication, quite a bit
more than 20% of total bits and of revenue). By 2013, the highest density segment (>128Mb) was 11% of
units, according to the Semiconductor Industry Association.
While the highest density segment has fallen as a percentage of units, the lowest density segment has been
slowly rising. The 2Mb and lower segment moved up to 15% by 2013 from 10% of units in 2004.
2004
> 64Mb
15%
2Mb
10%
> 128Mb
11%
4Mb
12%
2Mb
15%
128Mb
11%
64Mb
17%
4Mb
12%
8Mb
13%
32Mb
14%
16Mb
19%
64Mb
16%
8Mb
10%
32Mb
13%
16Mb
12%
*Note: Since these pie charts represent unit percentage, the actual revenue and total bit percentages are far higher for the
higher density segments than suggested in the figures.
Source for both charts: Semiconductor Industry Association, Wells Fargo Securities, LLC
78
Exhibit 61 shows NOR bit shipments; Exhibit 62 shows blended price/Gb; and Exhibit 63 shows chip ASP. We
believe that typically higher density products have higher chip ASP but lower price/Gb.
Worldwide NOR bit demand has been declining at a rate of about 4% per year since mid-2006.
NOR price/bit has also been falling, at a rate of about 16% per year.
The relatively slow rate of increase in NOR chip density, coupled with the steady decline in price/bit, has
driven the blended NOR flash ASP down to about $0.50 today from about $3.50 in early 2004.
100,000,000
10,000,000
NAND Mbits
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
1,000,000
* We have used an approximation that the chips in the >128Mb category of the SIA data have densities of 512Mb. We have
probably underestimated slightly the total Mbits shipped in recent years.
Source: SIA, Wells Fargo Securities, LLC
NOR Price/Gb
$100.0000
$10.0000
price/Gb
Dec-14
Dec-13
Dec-12
Dec-11
Dec-10
Dec-09
Dec-08
Dec-07
Dec-06
Dec-05
Dec-04
Dec-03
Dec-02
$1.0000
*We have used an approximation that the chips in the >128Mb category of the SIA data have densities of 512Mb. We have
probably overestimated slightly the price/Mb in recent years.
Source: SIA, Wells Fargo Securities, LLC
79
Dec-03
Mar-04
Jun-04
Sep-04
Dec-04
Mar-05
Jun-05
Sep-05
Dec-05
Mar-06
Jun-06
Sep-06
Dec-06
Mar-07
Jun-07
Sep-07
Dec-07
Mar-08
Jun-08
Sep-08
Dec-08
Mar-09
Jun-09
Sep-09
Dec-09
Mar-10
Jun-10
Sep-10
Dec-10
Mar-11
Jun-11
Sep-11
Dec-11
Mar-12
Jun-12
Sep-12
Dec-12
Mar-13
Jun-13
Sep-13
Dec-13
Mar-14
Jun-14
Semiconductors
80
$4.00
$3.50
$3.00
$2.50
$2.00
$1.50
$1.00
$0.50
$0.00
Analog
Analog chips are generally broken out into two main segments:
(1) Standard analog chips, which are generic products mostly sold through distribution, and
(2) Application-specific analog products, which often combine both analog and digital capability (mixedsignal analog chips).
The standard analog segment has some very good business characteristics:
Low capital requirements for manufacturing. The use of cutting-edge technology generally does
not improve the performance of standard analog chips in the way that it improves the performance of logic
chips. Making things smaller (which is what advanced technology facilitates) helps logic chips because it
helps them to work faster and do more things at the same time (packing more transistors in the same
space). The performance of most standard analog chips, though, is not so much related to speed or doing
many things, but in accurately measuring or reproducing the desired electrical shape. Since analog chips
do not need to use the latest technology, in general, the cost of the manufacturing equipment to make
analog chips is much lower than the cost of manufacturing equipment to make advanced logic chips. Many
of the larger analog companies do make their own chips, but they still have relatively low capital expenses,
versus logic or memory companies, which make their own chips. For example, in each of the past 14 fiscal
years (FY2001-FY2014), Linear Technology, an analog company, reported capital expenditure that ranged
from 1% to 13% of revenue, with this percentage being 6% or lower in 12 of the 14 years. In contrast,
Micron, a memory company, in the past 13 fiscal years (FY2001-FY2013) invested between 7% and 63% of
its revenue in capital expenditure, with this percentage being above 20% in 10 of the 13 years.
Long product lifetimes. In part because the performance of analog products is not dependent on
rapidly changing technology, most standard analog chips have long lifetimes, stretching from years to
more than a decade. In principle, this should result in a good return on circuit design costs, i.e., research
and development (R&D).
Broad base of applications and customers. Many standard analog products have generic functions;
they are not specific to a given standard or application. For this reason, most analog companies have a
broad base of served end markets and customers.
Relatively little advantage from scaling in size; it appears that both large and small analog companies can
coexist and flourish; and
Relatively little concentration of market or customer risk. Standard analog companies rarely have many, if
any, greater than 10% customers, and usually sell their products into multiple end markets.
There is, however, one notable risk for analog companies. Because of their breadth of applications and
customers, often a high percent of their revenue is sold through distribution. Some analog companies
recognize revenue on sell-in to distribution, rather than sell-through. This can periodically result in excess
inventory building up in distribution, which can affect revenue when the distributors take action to reduce
inventory.
There is some ambiguity as to what is characterized as standard analog versus application-specific analog,
leading to discrepancies in market size and market-share numbers quoted by various industry groups and
market research companies. In 2013, the top standard analog market-share leader by revenue was Texas
Instruments, with, according to Gartner, a 24% share, followed by Analog Devices, with a 10% share. Maxim
had an 8% share, while Linear came in fourth place, with a 6% market share (see Exhibit 64).
81
Semiconductors
Exhibit 64. Standard Analog (Excluding ASIC Analog) Market Share (2013)
Texas Instruments
24%
Others
35%
Analog Devices
10%
Richtek
Technology
2%
Maxim Integrated
Products
8%
Sanken
2%
Intersil
3%
Linear Technology
6%
Cirrus Logic
3%
STMicroelectronics
3%
ON Semiconductor
4%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Devices, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
Application-specific analog businesses often have quite different characteristics from the standard analog
businesses. Some application-specific analog chips are often mixed-signal chips (combining analog and digital
logic functions), and for such chips there can be a need to use advanced manufacturing technology. Some
application-specific analog companies do not do their own manufacturing, but instead use chip foundries.
Also, there is often a higher customer and end-market concentration for any given application-specific product
line, which can lead to pricing pressures. Lifetimes of applications-specific products are often considerably
shorter than those of standard analog products. Investors often think of applications-specific analog
companies not so much as analog, but more in terms of the end markets they serve, such as
communications chip companies or consumer chip companies. Exhibit 65 shows the end-market
distribution for the application-specific analog segment; there are five main markets, the largest of which is
Communication Applications, with about 49% market share based on revenue.
82
Consumer
8%
Industrial & Other
10%
Computer &
Peripherals
9%
Automotive
24%
Communication
49%
Source: Semiconductor Industry Association and Wells Fargo Securities, LLC estimates
83
Semiconductors
Over the past ten years analog sales growth has generally trended in line with overall IC sales growth (see
Exhibit 66).
Exhibit 66. IC And Analog Sales (Year-Over-Year Growth)
100%
80%
Yr/Yr Growth
60%
40%
20%
0%
(20%)
(40%)
Dec-92
Jun-93
Dec-93
Jun-94
Dec-94
Jun-95
Dec-95
Jun-96
Dec-96
Jun-97
Dec-97
Jun-98
Dec-98
Jun-99
Dec-99
Jun-00
Dec-00
Jun-01
Dec-01
Jun-02
Dec-02
Jun-03
Dec-03
Jun-04
Dec-04
Jun-05
Dec-05
Jun-06
Dec-06
Jun-07
Dec-07
Jun-08
Dec-08
Jun-09
Dec-09
Jun-10
Dec-10
Jun-11
Dec-11
Jun-12
Dec-12
Jun-13
Dec-13
Jun-14
Dec-14
(60%)
84
The PLD uses the information in the memory on board the chip to make connections between its logic
elements so that collectively they perform a specific electronic function.
Todays PLDs also have some specialized additional circuits, such as transceiver circuits, which the
connected logic elements can use to complete the functionality of the programmed PLD.
Many types of PLDs use volatile memory (static RAM, or SRAM) to store the program information on the
chip. Volatile memory loses its information when the power is switched off; therefore, these types of PLD
are often used together with a separate flash memory chip, which stores the program of the PLD when the
power is off (see Exhibit 67).
Exhibit 67. PLD Diagram
Logic
Logic
Logic
Logic
Logic
Element
Element
Element
Element
Element
Logic
Logic
Logi c
Logic
Logic
Element
Element
Element
Element
Element
Logic
Logic
Logic
Logic
Logic
Element
Element
Element
Element
El ement
Logic
Logi c
Logic
Logic
Logic
Element
Element
Element
Element
El ement
Logic
Logic
Logic
Logic
Logic
Element
Element
Element
El ement
Element
Memory
85
Semiconductors
Designers of electronic systems have to decide between whether to use a PLD instead of a custom-designed
chip, i.e., an application-specific integrated circuit (ASIC) or a standard chip designed for a specific function,
i.e., an application-specific standard product (ASSP).
An ASIC is a fully custom-designed chip. Therefore, there are up-front costs associated with paying for the
design of the ASIC and the initial manufacturing of the ASIC (the costs of buying the masks that contain
the circuit patterns). We believe that the up-front costs of designing an ASIC could run anywhere from
millions to tens of millions of dollars, while the time to design the chip could well be on the order of many
months. On the other hand, an ASIC does not require additional memory to store programming the way a
PLD does. Also, when a PLD is programmed for a specific function there is almost always a fair number of
logic elements and other pieces of circuitry that are not used for the particular application. Therefore, a
custom-designed ASIC would be expected to have a significantly smaller chip (die) size than a PLD used
for the same application. The cost of making each ASIC chip is accordingly lower than the cost of making
each PLD chip to do the same thing. Exhibit 68 shows this theoretical relationship. For small volume,
PLDs are more cost effective since there is no up-front cost associated with a PLD. However, since the
incremental cost of each chip is lower for an ASIC, for high-volume applications it can work out to be
cheaper to incur the expense of designing the ASIC rather than using a PLD. There are two other
considerations that play into the decision of whether to use a PLD instead of an ASIC:
o
The time taken to design an ASIC and the cost and time delays associated with having to fix a design if
the ASIC does not work correctly are disadvantages to using a PLD. Some systems makers use a PLD
for the initial launch of a system and then as a cost-saving measure switch over to using an ASIC once
the ASIC is ready.
Since an ASIC is specifically designed for a particular application or system, it is likely to have better
performance than a PLD used for the same application.
Some applications are based on widespread standards (e.g., communications standards such as Ethernet
chips). Instead of needing to make a custom-designed chip, a designer of an electronic system has a choice
of a chip that has already been designed for that application, an ASSP. Since the ASSPs are off-the-shelf
products, they do not have the same disadvantage of ASICs in having up-front costs to the systems
designs, or design risks and delays. For this reason, systems makers do not often use PLDs in applications
for which there are ASSPs.
There is another choice, a structured ASIC, which is essentially a semi-custom chip. It can be more
ASIC-like with a chip designed choosing from a library of standard circuits to put together to make a
unique chip design, or more PLD-like. In the past Altera offered its Hardcopy option as a cost-reduction
path for customers that have already designed solutions using Alteras PLDs. For this, Altera used a
generic chip with an array of logic elements and specialized circuitry, but no program-memory, and
designs for the customer some custom metal connections between the logic elements and other circuits.
The up-front cost and delay and the cost per chip of a structured ASIC are in between the corresponding
costs for a PLD and for an ASIC. Altera no longer offers HardCopy-structured ASIC products for new
design starts, but does continue to support HardCopy for existing designs.
86
PLD
Total Cost
ASIC
Number of Chips
Exhibit 32, presented in a preceding section of this report, contains data on PLD growth relative to other
semiconductor segments:
Between 1990 and 2000, PLD revenue grew at a compound annual rate of 29%, well in excess of 16%
compound annual growth for semiconductor integrated circuits overall.
However, toward the end of the decade of the 1990s, one big growth driver for PLDs was the
communications end market, which underwent a particularly big correction in the downturn of 2001. PLD
sales declined with a CAGR of negative 21% from 2000 to 2003, while the overall IC segment declined
with a CAGR of negative 8%.
From 2003 to 2007, as the electronics end markets recovered and the semiconductor industry grew, PLDs
grew with a CAGR of 7%, below the overall IC CAGR of 12%.
The PLD segment significantly underperformed the broader chip group in the downturn of 2000-03
because there was a large overbuild in one of the big PLD end markets (communications infrastructure).
However, there was no similar PLD-specific issue in the 2007-09 downturn and PLD declines in this
period (a CAGR of negative 4%) were slightly better than overall IC declines (a CAGR of negative 7%).
PLD chips achieved a 9% CAGR, outstripping overall IC growth of 7% from 2009 to 2013. Altera and
Xilinx have both suggested that the PLD segment could be at a point at which secular PLD growth could
begin to run above aggregate IC growth, with PLDs eating into the broader custom chip application
specific integrated circuit (ASIC) market. The growth from 2009 to 2013 could be evidence of this.
However, it appears unlikely that in 2014, PLD growth will be any better than overall IC growth and might
even trail total IC growth. Our current (November 2014) estimates for Altera and Xilinx imply an increase
in PLD revenue of about 9% year over year for 2014, compared to our current projection that the IC
market could increase 8-10% in 2014.
Altera and Xilinx have both been vocal in recent years about the potential of the PLD segment to outgrow the
overall semiconductor market. There are several arguments that could be made as to why this might happen.
These include the following:
Capturing more value by including specialized circuitry such as transceivers into the PLDs. For example,
Xilinx is attempting to penetrate the embedded processing space with its Zynq product line (a family of
PLDs with ARM processor cores).
Clearly from 2000 to 2007, these various factors were not enough to offset other forces that may have had a
suppressive effect on PLD growth. In particular, we think that over-investment in global communication
infrastructure toward the end of the 1990s took several years to work through, which hurt PLD growth because
of the high exposure of PLDs to the communications infrastructure segment. PLDs did outperform the broader
chip industry in 2010 and 2011, however, it appears that revenue for PLDs declined about 9% in 2012
87
Semiconductors
compared to overall IC sales falling about 4% and only increased 2% in 2013, while total IC sales increased by
around 6%, leading to a 2009-2013 CAGR that was slightly better than overall IC growth. It seems unlikely that
this outperformance will continue in 2014. While it is very possible that the data of recent years indicate that
PLD growth is accelerating above semiconductor growth, in our view, the evidence is not very strong. We are
choosing to assume for the moment that the growth potential of PLDs is comparable to the growth potential of
semiconductors overall, which we project to be 10-12% per year.
Exhibit 69 shows the top PLD market-share leaders in 2013 were Xilinx and Altera, with 51% and 37% revenue
share, respectively. Other PLD companies include Lattice, 7%, and Microsemi, 4%.
Exhibit 69. PLD Market Share (2013)
Altera
37.1%
Lattice Semiconductor
7.4%
Microsemi
4.4%
Others
0.2%
Xilinx
50.9%
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc., Market Share: Semiconductor Applications, Worldwide, 2013, Gerald Van Hoy et alia, March 31, 2014
Exhibit 70 shows PLD sales by end market for Xilinx and Altera for full-year 2013. Communications
infrastructure (telecom infrastructure, wireless base stations, and networking equipment) represents the
largest end market for PLDs, accounting for about 45% of Xilinxs revenue and about 41% of Alteras.
We discuss the PLD market, and Altera and Xilinx, in detail in our PLD Primer issued on October 24, 2014.
88
Industrial
Automation,
Military & Auto
22%
Communications
& Data Center
45%
Broadcast,
Consumer &
Automotive
16%
Industrial,
Aerospace &
Defense
36%
89
Semiconductors
Exhibit 71 shows Altera and Xilinxs historical quarterly revenue.
Exhibit 71. Altera And Xilinx Quarterly Revenue ($ Millions)
$700
$600
$500
$400
$300
$200
$100
Mar-90
Sep-90
Mar-91
Sep-91
Mar-92
Sep-92
Mar-93
Sep-93
Mar-94
Sep-94
Mar-95
Sep-95
Mar-96
Sep-96
Mar-97
Sep-97
Mar-98
Sep-98
Mar-99
Sep-99
Mar-00
Sep-00
Mar-01
Sep-01
Mar-02
Sep-02
Mar-03
Sep-03
Mar-04
Sep-04
Mar-05
Sep-05
Mar-06
Sep-06
Mar-07
Sep-07
Mar-08
Sep-08
Mar-09
Sep-09
Mar-10
Sep-10
Mar-11
Sep-11
Mar-12
Sep-12
Mar-13
Sep-13
Mar-14
Sep-14
$-
XLNX Revenues
90
ALTR Revenues
91
Semiconductors
Manufacturing Transitions Wafer Size and Line Width (Moores Law)
Wafer size is usually quoted as the diameter of the wafer (i.e., a 300 mm wafer has a diameter that is 300 mm).
Prior to 2004, MOS wafers were 200 mm wafers or smaller (see Exhibit 73). It was not until 2004 that 300
mm wafers hit the market in any real volume. The 200 mm wafers have accounted for more than half of total
semiconductor capacity from 2003 through 2007, with the crossover of 300 mm accounting for more capacity
than 200 mm occurring in 2008.
The data of Exhibit 73 are presented in millions of square inches, rather than in absolute numbers of wafers. A
300 mm wafer makes up 2.25x the area of a 200mm wafer (the square of the ratio of the diameters), so more
than 2x the number of chips can be made on a 300mm wafer. When comparing capacity of different wafer
sizes, the numbers are often expressed in equivalents of a specific size. One 300mm wafer counts as 2.25
200 mm wafer equivalents. We believe capacity and output are generally measured in wafer starts per week
(the number of wafers that processing is started on in any given week) or wafer outs per week (the number of
wafers that processing is completed on in any given week.) Sometimes memory companies describe their
capacity in terms of wafer starts per month.
The microprocessor makers (Intel and AMD) have fully transitioned to 300 mm and Intel, TSMC, and
Samsung have made agreements with ASML Holding to develop 450 mm technology. Memory makers benefit
from the lower cost of manufacturing chips on larger wafers and so, many of the major memory manufacturers
are in various stages of transition to 300 mm production. Analog companies tend to use older technology and
many analog companies are still making chips on 200 mm or even smaller wafers.
Exhibit 73. Worldwide Semiconductor Capacity by Wafer Size (Millions of square inches/Quarter)
2500.0
2000.0
1500.0
1000.0
500.0
<200 mm
200 mm
2Q14
4Q13
2Q13
4Q12
2Q12
4Q11
2Q11
4Q10
2Q10
4Q09
2Q09
4Q08
2Q08
4Q07
2Q07
4Q06
2Q06
4Q05
2Q05
4Q04
2Q04
4Q03
2Q03
4Q02
2Q02
4Q01
2Q01
4Q00
2Q00
4Q99
2Q99
4Q98
2Q98
0.0
4Q97
3000.0
300 mm
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc.: Forecast: Semiconductor Wafer Fab Capacity, Worldwide, 2Q14 Update, David Christensen et alia
Line width is the minimum size of a circuit shape that can be made (i.e., how thin a line can be made). Being
able to make smaller shapes allows a designer to make chips smaller or to pack more circuitry into the same
area. In the past, line width was measured in microns (um = millions of a meter. Line width is currently
measured in nanometers; nm = billionths of a meter).
Moores Law is a term often used to describe the ongoing ability of the semiconductor industry to make
things smaller (transition to smaller line widths), resulting in chips with lower cost and/or more capability. It
is named after one of the co-founders of Intel, Gordon Moore, and Intel identifies its origin as a 1965 paper by
Gordon Moore published in Electronics Magazine. Gordon Moores original contention was that the optimum
number of devices on an integrated circuit for lowest cost per device would double every year.
Intel currently (2014) manufactures its newest Broadwell products on 14nm technology and expects to begin
its transition to 10nm technology in 2016. Other leading-edge semiconductor companies design logic circuitry
in 20nm, 28nm, 32nm, 45nm, 65nm, and 90nm technology, with the expectation that leading-edge
manufacturing is to transition to 16nm or 14nm technology in 2015.
92
The technology requirements for memory circuits are a little different from logic circuits and hence, memory
chip companies sometimes have processes that are nominally more advanced that logic manufacturers and
sometimes less advanced. Samsungs terms its most advanced NAND memory process today (H2 2014) 10nm
class and its most advanced DRAM memory process 20nm class.
In recent years a number of chip companies, especially memory companies, have commented on the increasing
difficulty of moving toward more advanced technologies. Exhibit 74 shows a comparison of line-width
transitions for Intel and TSMC. While there may be increasing risk that Moores law might slow at some point
in the future, our table does not show any obvious slowing in Moores law as yet. In general, Intel has been
maintaining a cadence of moving to a smaller line width every two years.
Exhibit 74. Line-width Transitions for Intel and TSMC (measurements in nanometers)
2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015E 2016E 2017E
Intel1
90
65
453
32
224
14
10
2
TSMC
90
65
40
283
20
16/204
10
1For Intel, we have mapped line-width transitions into the year the corresponds with Intels announced launch dates.
2For TSMC we have mapped line-width transitions into the year that corresponds with the first reported quarter of sales for each line width.
3Introduction of metal gate transistors
4Introduction of 3D/FinFET transistors (expected for TSMC).
TSMCs 16nm process has 16nm transistors but uses the same metallization
(connections between transistors) rules as TSMCs 20nm process.
Sources: Intel, TSMC, and Wells Fargo Securities, LLC
Exhibit 75 shows worldwide semiconductor production by line width. At any given time there is a very broad
spread of different technologies in use. Exhibit 76 shows the distribution of semiconductor production by line
width at TSMC, the worlds largest chip foundry.
93
Semiconductors
100%
90%
14 nm
80%
22 nm
70%
32 nm
60%
45 nm
50%
65 nm
90 nm
40%
0.13 Micron
30%
0.18 Micron
0.25 Micron
20%
0.35 Micron
10%
0.5 Micron
4Q97
2Q98
4Q98
2Q99
4Q99
2Q00
4Q00
2Q01
4Q01
2Q02
4Q02
2Q03
4Q03
2Q04
4Q04
2Q05
4Q05
2Q06
4Q06
2Q07
4Q07
2Q08
4Q08
2Q09
4Q09
2Q10
4Q10
2Q11
4Q11
2Q12
4Q12
2Q13
4Q13
2Q14
0%
3000
14 nm
2500
22 nm
32 nm
2000
45 nm
65 nm
1500
90 nm
0.13 Micron
1000
0.18 Micron
0.25 Micron
500
0.35 Micron
0.5 Micron
4Q97
2Q98
4Q98
2Q99
4Q99
2Q00
4Q00
2Q01
4Q01
2Q02
4Q02
2Q03
4Q03
2Q04
4Q04
2Q05
4Q05
2Q06
4Q06
2Q07
4Q07
2Q08
4Q08
2Q09
4Q09
2Q10
4Q10
2Q11
4Q11
2Q12
4Q12
2Q13
4Q13
2Q14
Note: Chart created by Wells Fargo Securities, LLC (based on Gartner data)
Source: Gartner, Inc.: Forecast: Semiconductor Wafer Fab Capacity, Worldwide, 2Q14 Update, David Christensen et alia
94
0.25/0.35um
20nm
4%
9%
0.50um+
1%
0.11/0.13um
3%
28nm
34%
90nm
6%
65nm
13%
40nm
17%
Source: TSMC and Wells Fargo Securities, LLC
95
Semiconductors
A dual-core Intel Broadwell microprocessor has a die size of 82 mm2 (roughly the size of a thumbnail).
The area of a 300mm wafer (12 inch wafer) is 3.14x150x150 mm2 = 70,695 mm2.
However, as shown in Exhibit 72, which is a picture of a Broadwell dual-core wafer, since the chips are
rectangular and the wafer is round, not all the area around the edge is used and some of the chips on the
edge are no good because they are not full rectangles. There are, in fact, only 789 full rectangles printed on
the wafer.
Not all these chips will be good chips; some might not work properly. The number of good chips is the
number of chips printed on the wafer, multiplied by the yield.
We can extend these calculations to estimate wafer value and wafer manufacturing cost.
Perhaps the yield is about 90%; this would give about 710 good chips on a wafer. Using an ASP of about
$100 for a dual-core Broadwell chip gives us the value of the die on a typical Broadwell wafer might be
close to $71,000.
If we assume a gross margin of 75% for these chips, Intels total cost for making the chips on the wafer
might be close to $18,000. On top of the costs of processing the silicon wafer, there are packaging and test
costs for each of the good chips. Backing out this test and assembly cost, we arrive at a manufacturing
cost of the wafer that might be of the order of $14,000-$16,000.
Exhibit 20 of this report shows that the average selling price of a Taiwanese foundry wafer is close to
$1,000 for an 8-inch equivalent wafer. Since the Intel wafers are 12-inch wafers (300mm wafers), they
have double the area, so the equivalent foundry ASP of a 300mm wafer is about $2,000-2,500, still far
below the $71,000 we calculate for Intel. However, the foundry wafer ASPs are blended ASPs for both
newer and older technology of varying degrees of complexity (see Exhibit 75). The Intel Broadwell wafers
are made on the very most advanced technology (14nm) with a high degree of complexity (e.g., we believe
that Intel uses 10-11) layers of metal for Broadwell. The minimum number of layers of metal a
semiconductor wafer might have is two layers.
96
Transistor
Transistor
(Top-down view)
(Front view)
Gate
Drain
Source
Gate Oxide
u
So
e
rc
ai
Dr
Line width
Source: Wells Fargo Securities, LLC
Exhibit 77 shows a schematic of a transistor, which is one of the basic building blocks of an integrated circuit.
Integrated circuits are for the most part chips of silicon on which lots of transistors are fabricated, connected
by metal lines also made on the silicon chip. What we have shown in the diagram is MOS transistor. This is
made up of a gate, which is separated from the silicon below by very thin gate oxide, which separates a source
and a drain. MOS stands for Metal-Oxide-Semiconductor, which is a reference to the gate (metal; see
discussion on metal gates, which follows) sitting on top of an oxide, which sits on top of the semiconductor
(silicon). (It turns out though, despite the term MOS, for the past 30 years, most semiconductor technologies
have not used metal to make the transistor gate.)
A transistor is an on/off switch. If a certain voltage is put on the gate, current can flow from the source to the
drain; the source and drain are electrically connected. If there is a different voltage on the gate, then current
cannot flow and the source and drain are electrically disconnected. The gate of the transistor is typically the
narrowest line that is used in all the patterns that make up an integrated circuit, and so the line width of any
given semiconductor technology generally refers to the width of the lines used to make the transistor gates.
There are two types of MOS transistor: NMOS and PMOS. An NMOS transistor can be connected to a PMOS
transistor to make a complementary MOS (CMOS) gate. Today, the bulk of logic circuit design is done using
CMOS.
Metal Gate Mini-Primer
Intel was the first semiconductor company in the world to transition to using metal gate transistor technology,
which it introduced on its 45nm node in 2007. In this section we provide a brief explanation of what metal gate
transistors are and why the transition to metal gate transistors is so pressing.
What a metal gate transistor is. Exhibit 78 illustrates the difference between metal gate transistor and a
traditional polysilicon gate transistor. Transistor A shows a typical 65nm transistor structure.
The silicon dioxide (the gate oxide) acts as an insulator to stop electrical current running from the gate to
the transistor.
As each generation of technology has been made smaller, the gate oxide had to get thinner in order to
allow the transistor body to electrically see the voltage on the gate. At the 65nm nanometer level, we
believe that the gate oxide is about 30 angstroms thick. This is about six atoms thick!
97
Semiconductors
At the 45nm level, the silicon dioxide would be so thin that it would be a poor blocker of electrical current
and would allow a leakage current to flow through the gate. If, as an alternative, the gate dielectric were
thinned less than is normal for the scaling of the transistor, the transistor would have insufficient drive
current (the amount of current the transistor supplies when it is switched on), which would reduce the
speed of the circuits.
Also, as everything shrinks, the source (S) and the drain (D) get so close that a leakage current can flow
from source to drain when the transistor is supposed to be turned off.
The silicon dioxide has been replaced by a high-k material. The k value can be thought of in terms of
electrical density. A high-k material can be thicker physically, while looking like it is thin electrically to
electric fields. Therefore, it can be made thicker than the silicon dioxide gate, but still let the body of the
transistor see the gate voltage clearly.
Since the gate insulator is now thicker, this stops the leakage current from flowing through the gate. This
gives process engineers a lot more flexibility in deciding exactly how to adjust the thickness of the gate
insulator to get enough drive current out of the transistor to make the circuits fast.
Although Intel and other semiconductor manufacturers refer to this device as a metal gate transistor, we
believe that the gate material is not, in fact, a pure metal, but a metal alloy (we believe that it is a metal
silicide or metal silicate, which is a mixture of metal and silicon, and possibly, some other elements).
There are two types of transistors used in a CMOS process, an N transistor (NMOS) and a P transistor
(PMOS). The atomic properties (i.e., the work function) of the gate material (polysilicon or metal) affect
how tightly the transistor can be turned off and how hard it can be turned on. If two different materials
are chosen, one can be optimized for the N transistor, and the other can be optimized for the P transistor.
The result is that each transistor can be turned off more tightly (less source or drain leakage) and turned
on harder (more drive current) than if a polysilicon gate had been used.
Polysilicon
Gate
Transistor
Metal Gate
Transistor
Low Resistance
Layer
Low Resistance
Layer
Metal gate
S
Leakage currents
D
(A)
98
Polysilicon gate
D
(B)
After having transitioned to metal gate transistors about two years ahead of any other semiconductor
company, Intel made another big change in its transistor design, to 3-D transistors. Exhibit 79 shows the
difference between a traditional (Planar) transistor and a 3-D transistor. In a planar transistor the gate sits on
top of the place where the current flows (the channel). In a 3-D transistor the gate wraps around the channel.
Since the gate is controlling the channel from three sides (hence the alternative name, tri gate), it can turn on
the current more strongly (increasing the performance) or shut off the current more completely (reducing the
power consumption associated with leakage).
A 3-D transition is sometimes called a FinFET because the gate looks like a fin. FET stands for field effect
transistor.
Exhibit 79. Planar Transistors And 3-D Transistors
gate Oxide
te
ga
so
ur
gate
ce
ai
dr
drain
gate
source
cu
t
en
rr
gate
gate oxide
D
current
Silicon Substrate
Silicon Substrate
Source: Wells Fargo Securities, LLC
Intel moved to using 3D transistors on its 22nm technology node, launching the first products using 3D
transistors launched in April 2012. By the beginning of September 2014, Intel had launched 22nm products
using 3D transistors in all its main server, desktop, notebook and tablet product lines and introduced 14nm 3D
transistor products for fanless devices. To our knowledge, at the time of writing this report (November 2014)
there is no other chip manufacturer in volume production of chips with 3D transistors.
99
Semiconductors
Mask
Lens
Wafer
Translation Stage
Source: Wells Fargo Securities, LLC
Exhibit 80 is a very rough schematic of how circuit patterns get transferred to a wafer using a lithography tool.
These machines are commonly called steppers because they can image the patterns needed on only a small
portion of a wafer at any time, and so they step across the wafer exposing the patterns again and again until the
whole wafer is covered.
100
Today (2014), Intel has the most advanced technology among chip companies, manufacturing its
chips on 14nm line widths. The most advanced production-worthy steppers available today use
193nm laser light to create the images. The wavelength (color) of light used provides a limitation on
how small a pattern can be imaged (see Exhibit 81). An analogy to this is that a sharp pencil can draw
a smaller shape than a thick crayon. In principle, the smallest shape that can be imaged by light of a
given wavelength (the resolution of the light) is half the wavelength of the light used, and so, with
193nm steppers, it becomes very difficult to accurately create shapes of line width less than 100nm.
Exhibit 81. The Wavelength of Light Affects the Smallest Possible Pattern That Can Be Easily Imaged
Over the years, the semiconductor industry has implemented a number of imaginative technical tricks to
push the resolution of the steppers below the theoretical limits associated with the wavelength of the light
source. These include the following:
Immersion lithography--running the light through a liquid rather than air. This increases the numerical
aperture of the optical system, improving the ability of the system to image features that are smaller than
the wavelength of the light used.
Doubling patterning (to be followed by multiple patterning in the future). Double patterning involves
using the stepper twice to create two different patterns, creating very small shapes associated with the
overlap of the two separate patterns. Multiple patterning refers to using more than two separate
exposures to create the final pattern.
The move to double patterning roughly doubles the number of steppers needed for each of the smallest
patterns (each critical layer) needed in the wafer manufacturing process. Since the steppers are arguably the
most expensive individual pieces of equipment in a fab (the most advanced steppers are priced at tens of
millions of dollars each) the need for double and, in the future, multiple patterning, will likely drive up the
capital cost of equipping a fab.
Given that one of the key difficulties in the creation of the circuit patterns is that the steppers use a wavelength
of light that is too long to easily image the patterns, an obvious solution is move to a shorter wavelength of
light. This is what the semiconductor industry is trying to do, by moving to using Extreme Ultra Violet
(EUV) light. ASML is working with 13.5nm light sources for its EUV machines. Violet is the shortest
wavelength that a human eye can normally see, about 400nm. The 193nm used in todays steppers is called
ultra violet, a shorter wavelength than violet. Hence, the even shorter wavelength of 13.5nm is called extreme
ultra violet. There are many difficulties, though, in working with such a short wavelength of light. One of the
bigger ones is that it is difficult to make a laser that can produce enough brightness of 13.5nm light to expose
the pattern in a short time, resulting in low throughput of the stepper. As a result, it takes a long time to
process each wafer. There are other difficulties associated with working with EUV light, as well. For example,
13.5nm light gets blocked by the glass typically used to make lenses so that curved mirrors are needed to focus
the light (reflective optics), rather than passing the light through normal lenses (transmission optics).
101
Semiconductors
Appendix A. Semiconductor Companies (Figures In Millions Of Dollars, Except Per Share Data)
Mkt Cap
(millions)
Chip
Description
Revs ($
millions)
11/13/2014 2013
$161,392
$52,708
C hips For PC s
$30,636
Memory C hips
$114,349
$20,083
Makes Wafers For Fabless
C ompanies
$54,167
$12,205
Broad-Based
$11,277
Memory C hips
$116,864
$25,470
C hips For Wireless Handsets
Company
INTEL
SAMSUNG*
TSMC (Taiwan Semiconductor)
Ticker
INTC
Division of Samsung
TSM
TEXAS INSTRUMENTS
TOSHIBA*
QUALC OMM
TXN
Division of Toshiba
QC OM
STM
Not listed in US
Not listed in US
MU
BRC M
$6,072
$35,327
$24,361
$8,088
$7,979
$12,625
$11,281
$8,219
Broad-Based
Broad-Based
Memory C hips
Memory C hips
C hips For C ommunications
AMD
$2,101
$5,299
C hips For PC s
SNDK
Division of Matsushita
$20,483
$6,170
$2,018
Memory C hips
Broad-Based
FREESC ALE
FSL
$5,997
$4,186
NXPI
$16,937
$4,815
UMC
$5,378
$4,165
SONY*
Division of Sony
INFINEON
IFX
$10,655
$3,488
$5,290
MRVL
$6,857
$3,404
NVIDIA
MEDIATEK*
NVDA
Not listed in US
$10,646
$4,130
$4,586
ROHM*
ANALOG DEVIC ES
FUJITSU*
MAXIM INTEGRATED PRD
ON SEMIC ONDUC TOR
XILINX
SUNPOWER
SILIC ONWARE
AVAGO
SHARP*
ALTERA
Not listed in US
ADI
Division of Matsushita
MXIM
ONNN
XLNX
SPWR
SPIL
AVGO
Division of Sharp
ALTR
$10,650
$2,591
$2,640
$1,542
$2,419
$2,783
$2,297
$2,508
$2,333
$2,653
$1,619
$1,733
Not listed in US
ATML
$3,157
$1,497
$1,386
FC S
SMI
$1,881
$3,607
$1,405
$2,071
VISHAY*
MITSUBISHI*
LINEAR TEC HNOLOGY
ROBERT BOSC H*
Division of Vishay
Division of Mitsubishi
LLTC
Division of Bosch
MC HP
Not listed in US
Not listed in US
$8,718
$1,868
$1,519
$229
C ODE
TTMI
SQNS
SWKS
Not listed in US
$1,322
$578
$87
$11,559
$1,035
$1,368
$14
$1,844
$1,397
RF MIC RO DEVIC ES
OSRAM*
C REE
WINBOND ELEC TRONIC S*
RFMD
Dividion of OSRAM
C REE
Not listed in US
$3,852
$1,173
$1,328
$1,530
$961
CY
OVTI
$1,635
$1,554
$723
$1,459
INTERSIL HLDG
MAC RONIX INTERNATIONAL*
ISIL
Not listed in US
$1,695
$575
$670
102
$15,791
$8,332
$3,613
$11,467
$3,828
$4,400
$21,997
$10,282
$4,166
$1,357
$1,551
$1,317
$1,785
Broad-Based
Broad Based (C onsumer
Applications)
Makes Wafers For Fabless
C ompanies
C hips For C onsumer Electronics
Broad-Based, Esp. C ommunications
C hips For C ommunications &
Storage
Graphics C hips For PC s Etc.
C hips For C ommunications
Broad-Based
Broad-Based
Broad-Based
Broad-Based
Broad-Based
Broad-Based
Solar C hips
Packaging and Testing
C ommunications
Broad-Based
Broad-Based (Programmable Logic)
Chip Category
Microprocessors
DRAM, NAND
Wafer Foundry
Analog, DSP
NAND
Application Specific Analog,
Specialized Logic
Analog, Microcontrollers, ASSPs
Microcontrollers, Memory, Etc.
DRAM, NAND
DRAM, NAND
Application Specific Analog,
Specialized Logic
Microprocessors, Graphics Processors
NAND
Discretes, App. Specific Analog,
Microcontrollers Etc.
Application Specific Analog,
Microcontrollers Etc.
Application Specific Analog,
Specialized Logic
Wafer Foundry
Specialized Logic, Application Specific
Analog
Application Specific Analog,
Microcontrollers Etc.
Application Specific Analog,
Specialized Logic
Graphics Processors
Application Specific Analog,
Specialized Logic
IC s And Discretes
Analog, DSP
Microcontrollers, ASIC s
Analog
Analog, Discrete
PLDs
Solar
Optical C omponents
Optoelectronics, IC s
PLDs
Memory C hips
Broad Based (Microcontrollers,
Memory)
Broad-Based (Analog, Discretes)
Makes Wafers For Fabless
C ompanies
General Purpose
Sensors And General Purpose
Broad-Based (Analog)
Sensors, Application Specific
Devices
Broad-Based (Microcontrollers)
Optical Semiconductors (LEDs)
Memory
Dynamic RAM
Microcontrollers, EPROM
Memory C hips
Broad-Based
Wireless
Wireless
General Purpose Logic, Application
Specific Devices
Wireless Handsets
General Purpose
Light Emitting Diodes
Broad-Based
Analog, Discrete
Wafer Foundry
Discretes
Transistors, Optical Semiconductors
Analog
Sensors, Automotive ASSPs
Microcontrollers
Optical C omponents
Dynamic RAM, NAND-Based Flash
Not listed in US
PMC S
$1,523
ARMHY
Not listed in US
IDTI
$2,641
$1,119
$114
$476
$19,397
$950
$508
C ommunications
C ommunciations
DIODES
DIOD
$1,255
$827
IMOS
$649
$649
MSC C
TSEM
$2,606
$524
$1,021
$505
SUNE
$4,805
$1,985
SILIC ON LABORATORIES
BOOKHAM
OC LARO INC .
SEMTEC H
II-VI
STANDARD MIC ROSYSTMS*
SLAB
BKHM
BKHM
SMTC
IIVI
(Merged with Microchip)
$1,940
$179
$179
$1,736
$820
$580
$477
$477
$595
$623
$207
Not listed in US
SYXI
C RUS
$366
$1,235
$361
$313
$772
RAMBUS
TESSERA TEC HNOLOGIES INC .
RMBS
TSRA
$1,316
$1,664
$272
$169
POWER INTEGRATIONS
LATTIC E SEMIC ONDUC TR
MIC REL
DIALOG SEMIC ONDUC TOR*
RAYDIUM SEMIC ONDUC TOR*
SIGMA DESIGNS
POWI
LSC C
MC RL
Not listed in US
Not listed in US
SIGM
$1,566
$784
$709
$144
$347
$333
$237
$903
$367
$199
ISSI
AMC C
DSPG
MPWR
ANAD
C AVM
$417
$496
$225
$1,761
$68
$2,851
$310
$221
$151
$238
$134
$304
EMKR
Not listed in US
IKAN
SIMG
$162
$50
$431
$163
$103
$80
$276
INVENSENSE
INVN
$1,284
$249
Not listed in US
VTSS
Not listed in US
$237
$156
$105
$203
INPHI C ORPORATION
MONTAGE TEC HNOLOGY
IPHI
MONT
$535
$589
$103
$111
ENTR
PSEM
MLNX
EXAR
SIMO
$233
$251
$1,965
$420
$744
$259
$127
$391
$128
$225
OIIM
Not listed in US
$52
$74
$140
KOPIN C ORP.
SILIC ON INTEGRATED
SYSTEMS*
ADVANC ED ANALOGIC
TEC HNOLOGIES INC .*
PIXELWORKS
KOPN
Not listed in US
$242
$23
$17
PXLW
$118
$48
8X8
ON TRAC K INNOVATIONS
EGHT
OTIV
$712
$81
$122
$20
TRANSWITC H
C EVA
TXC C
C EVA
$0
$323
$18
$49
Not listed in US
QUIK
MOSY
ALSC
ADNC
SEMI
$178
$83
$26
$92
$691
$25
$26
$4
$10
$160
$921
Broad-Based
Broad-Based
Memory (Intellectual Property)
Memory
Audio
Semiconductor Materials
Acquired by Skyworks
$80
KNOWLES C ORP.
KN
$1,617
$1,215
INVENSENSE
AMBARELLA
INVN
AMBA
$1,284
$1,467
$249
$158
INPHI
SENSATA TEC HNOLOGIES
NEOPHOTONIC S
MAGNAC HIP SEMIC ONDUC TOR
M/A-C OM TEC HNOLOGY*
IPHI
ST
NPTN
MX
MTSI
$535
$7,767
$98
$401
$1,097
$103
$1,981
$282
$857
$159
Power Management
Analog, Discrete
Specialty Wafer Foundry
Makes Wafers For Fabless
C ompanies
Makes Silicon Wafers and sells solar Polysilicon and silicon wafers
energy solutions
Wireless
Application Specific Analog
Optical C ommunications
Optoelectronics
C ommunciations
Optoelectronics
Broad-Based
Analog And Mixed-Signal
Industrial Laser C omponents
Optoelectronics
Broad-Based
Application Specific Analog,
Specialized Logic
Broad-Based
Applications Specific Analog
Broad-Based
Power Mosfets And Mixed Signal IC s
Analog And Mixed-Signal IC s
C hips for Audio and Industrial
Applications
Memory (IP)
Intellectual Property - Memory
C hip Packaging Technology (IP)
Intellectual Property - C hip Packaging
Broad-Based
Broad-Based
Broad-Based
Broad-Based
C hips for displays (LC D Drivers)
C hips for consumer (IPTV And
Media Players)
Memory
C ommunications
Broad-Based
C hips for Power Management
Wireless C ommunciations
C ommunciations (Network
processors)
Fiber Optics, Solar Power
C hips for C onsumer Applications
C hips For C ommunications
C ommunications for consumer
(HDMI)
Motion sensing (audio) for
consumer
C omputing Applications
C ommunications
C omputing and C ommuncations
C omputing and C ommuncations
Digital set top boxes and server
DRAM
Set top box (C onsumer)
C ommunciations
C hips For C ommunications
C hips For C ommunications
NAND Flash, Digital C onsumer,
Handsets
Power Management
C omputer C ommunications
(chipsets etc.)
Semiconductor Materials
C omputer and consumer
applications
Power Management
Analog IC s
PLDs
Analog, Mixed-Signal IC s
Applications Specific C hips
Applications Specific Analog
Specialized Logic
SRAM and DRAM
Application Specific Analog
IC s
Analog
Applications Specific Analog
Specialized Logic
Optoelectronics
Application Specific
Applications Specific Analog
Application Specific Analog
Application Specific Analog
Processors and Logic
Application Specific Analog
Intellectual Property for Applications
Specific C hips
Analog/Mixed Signal Ics
Analog/Mixed Signal Ics
Analog/Mixed Signal Ics
Application Specific Analog
Application Specific Analog
Application Specific Analog
Specialized Logic
Analog
Specialized Logic
Wafer Manufacturer
Logic
Analog IC s
VoIP Services
C hips for Smart C ards
Analog
PLDs
Embedded DRAM
DRAM, SRAM (controllers and IP)
Wafer Manufacturer
Analog And Mixed-Signal IC s
Analog And Mixed-Signal IC s
Analog And Mixed-Signal IC s
Analog And Mixed-Signal IC s
Sensors
103
Semiconductors
MINDSPEED TEC H
MSTAR SEMIC ONDUC TOR*
ELPIDA
C AMBRIDGE SILIC ON RADIO*
INTERNATIONAL REC TIFIER
C ORP.
PEREGRINE SEMIC ONDUC TOR
APTINA
Acquired by M/A-C OM
Acquired by MediaTek
Acquired by Micron
In processor of being acquired by
Qualcomm
IRF, Infineon acquisition expected to
close late 2014
acquired by Murata
Acquired by ON Semiconductor
LSI LOGIC
Acquired by Seagate
Acquired
Acquired
Acquired
NSM
Acquired
Acquired
Acquired by ADI
Acquired by C hinese government
Acquired by Micrsemi
Merged with MediaTek
Acquired by Broadcom
Acquired by Qualcomm
NSM
Acquired by DTS
Acquired by C ypress
AC TEL
MIC ROTUNE
SILIC ON STORAGE
TEC HNOLOGY
C ALIFORNIA MIC RO DEVIC ES
C O C OM
ZILOG
C ENTILLIUM C OMMUNI.
C HARTERED SEMIC ONDUC TOR
$1,040
Acquired by ON Semiconductor
Acquired by IXYS C orporation
Acquired by TranSwitch C orporation
Acquired by GlobalFoundries
Privately Owned by Golden Gate
C apital
Acquired by Exar C orporation
Merged with Sanjan
Division of NEC
Filed for Bankruptcy in Multiple
Jurisdictions
Acquired by C SR PLC in 2009
Acquired by Intersil
Acquired by Novafora, C eased
Operations August 2009
Tvia, Inc. Reorganized Out of
Bankruptcy Aug. 3, 2011
Acquired by Synopsys
Acquired by Microsemi
Acquired by Advantest
Power Management
Application Specific Devices
Image Sensor IC s For Handsets
And C ameras
C hips For C omms And Storage
RF Wireless
Power Management
C ommunciations
Broad-Based
Power Management
C hips for Audio (C onsumer)
by Maxim
by C irrus
C ONEXANT SYSTEMS
$2,855
by RFMD
by Microchip
by Avago
LOGIC DEVIC ES
SIMPLETEC H
STAKTEK
MIPS TEC HNOLOGIES
TRIDENT MIC ROSYS
SIGE SEMIC ONDUC TOR
HI / FN INC C OM
MATHSTAR
NEC
QIMONDA AG SPONSORED ADR
$2,467
$1,025
Wireless
ASSPs
$1,437
IC s
IC s (FPOAs)
Very Broad
DRAM
Digital TVs
Specialized Logic
Memory (C ontrollers)
Defense, C ommercial
Test & Equipment
*For companies listed in the United States that are primarily chip companies, we have listed total revenue. For companies
not listed in the United States, or divisions of larger companies, we have included only semiconductor revenue.
Source: FactSet, Company reports and Wells Fargo Securities, LLC
104
Appendix B: Glossary
3G/4GThe third and fourth generations of mobile communications standards used for mobile web access, IP
telephony, gaming services, HD mobile TV, and cloud computing.
A/D converter--see analog-to-digital converter
analog--something that processes information in the form of continuous voltage and current. A circuit can be
analog or digital. A digital circuit thinks in terms of just a 1 or a 0, whereas an analog circuit has to deal
with everything in between (0.1, 0.15, 0.2 . etc.). Analog circuits have different requirements and sensitivities
from digital circuits and so require different design and manufacturing techniques. Typically analog chip
manufacturing does not require leading-edge equipment and so is less capital intensive than digital chip
manufacturing. Both logic chips and memory are typically digital.
analog-to-digital converter--data conversion chip that takes a continuous analog signal and converts it to a
stream of digital numbers by making measurements of the height (voltage of the signal) at regular time
intervals.
Applications processor--Applications processors are processors used as the brains to run applications in
tablets and smartphones.
ASIC--(Application Specific Integrated Circuit)--a custom-designed chip, as opposed to an off-the-shelf
generic chip.
ASSP--(Application Specific Standard Products)--a name for something that is not an ASIC a standard
product designed for a specific purpose.
Baseband processora device in a network interface that manages all of the radio functions, including wi-fi
and Bluetooth.
bit--see megabit
byte--a byte is 8 bits. See megabit.
capacitor--a device in electronic circuits that can store electric charge (electrons).
capacity utilization--actual output of a semiconductor fab or group of fabs as a percent of what the fab is
theoretically capable of. As a rule of thumb, we consider 90-100% capacity utilization as a healthy level of
utilization. In principle, 100% is the maximum capacity utilization a fab or a company can achieve. TSMC has
in the past quoted capacity utilization of as much as 110%.
chip--a piece of semiconductor containing a circuit. Also called a die.
clock, clock frequency--many chips are synchronous in operation, which means that a clock signal needs
to be generated and all calculations or other operations in the circuit wait for the next beat of the clock before
they happen. This makes sure that signals are transferred at the same time, avoiding errors that may occur as a
result of one thing happening earlier than something else. The higher the clock frequency, the more things that
can happen per second and the more the chip can achieve. Typical clock frequencies are 100s of megahertz
(hundreds of million times per second) or several gigahertz (several billion times per second).
computing cloud--a network of large data centers and high-end servers which facilitate remote access to
centrally located software programs via a suitable internet connection. The idea of a computing cloud is to
spread an extremely large amount of computing power to a variety of users who individually lack the resources
to acquire such a high volume of data.
CMOS--Complementary MOS. A type of integrated circuit. Most digital semiconductor circuits are CMOS
these days. A CMOS circuit uses two types of transistor, NMOS and PMOS. The predecessor to CMOS logic
was NMOS logic. A circuit might also be bipolar rather than CMOS. And bipolar circuits are occasionally used
for analog applications, but CMOS has also achieved widespread use even in analog. See definition of MOS.
contract price, contract market--semiconductor memory can be bought on contract, in which prices are
negotiated in advance, twice per month, or on the spot market, which offers a constantly changing price. The
bulk of DRAM is sold on contract, but the spot market is often thought to be a leading indicator of contract
pricing.
105
Semiconductors
106
gigabyte--A quantity of memory. 1000 megabytes (actually 1024 megabytes). See megabit.
gigahertz (GHz)--a billion times per second. See clock.
hard disk drive (HDD)--the standard bulk memory device in a personal computer. Most computer contain
DRAM for storing the information that the microprocessor needs immediately, when the computer is switched
on, and bulk memory to store all the files permanently, even when the computer is switched off. In a hard disk
drive the information is stored on one or more magnetic disks (hard disks).
HDD--see hard disk drive
IC--See integrated circuit.
integrated circuit (IC)--A silicon chip on which many (ranging from tens of transistors to hundreds of
millions of transistors) are connected together, all on the same chip. See discretes for comparison.
installed capacity--The manufacturing capability of the manufacturing equipment in a fab. This is
sometimes less that floor capacity (see the definition of floor capacity).
interface chips--Analog chips that provide the interface onto standardized communications signal lines; they
are the chips responsible for driving the electrical voltages/currents down the lines. For example, in a
computer, there would be some PCMCIA chips for driving signals down a PCMCIA bus.
linear--Another word for analog; see analog
linewidth--A semiconductor circuit is made by creating a series of patterns in a number of semiconductor,
insulator and metal films that are laid one on top of another. The narrowest width that can be used in a pattern
is referred to as the line width. As technology develops, the line width drops, and so each feature can be made
smaller, resulting in either smaller chips, or more typically, chips of the same size that contain more transistors
and other devices on them.
lithography toolmachines used in semiconductor manufacturing to transfer circuit patterns onto
semiconductor wafers. These machines are commonly called steppers because they can image the patterns
needed on only a small portion of a wafer at any time, and so they step across the wafer exposing the patterns
again and again till the whole wafer is covered.
logic chip--a chip that thinks (can do computations or make decisions), as opposed to a memory chip (a chip
that remembers) or an analog chip (a chip that deals with sending or receiving signals). Examples of logic chips
include microprocessors, microcontrollers and programmable logic devices.
LTE (Long Term Evolution)--The wireless standard for high-speed data for mobile phones and data
terminals.
maintenance capex (maintenance capital expenditure)the level of capital spending, by a chip
manufacturer, that is just enough to keep the manufacturing capacity constant. There is a certain minimum
level of spending needed, in part to replace old or obsolete manufacturing equipment. If capital spending drops
below maintenance capex, then total manufacturing capacity will fall.
megabit--a bit refers to a single 1 or 0. Memory size is measured in the number of bits (the number of
separate 1s or 0s) that a memory can store. A megabit is approximately a million bits. Because of the way
memory locations are defined (as a string of 1s and 0s), the amount of memory provided in a chip is not exactly
a million bits for a megabit, but actually some number that is 2n. In fact, a megabit is 1,048,576 bits. The fact
that the amount of memory is always defined as a power of two also explains the various choices of standard
memory size (e.g., 1 Gb, 2Gb, 4Gb etc.)
megabyte--A quantity of memory. 8 megabits make up a megabyte. See megabit.
megahertz (MHz)--a million times per second. See clock.
memory--a chip that remembers information. The two most common types of memory are DRAM (volatile
memory) and flash (non-volatile memory). Memory chips tend to be commodity-like in their sales
characteristics; prices change on a daily basis on the spot market and contract prices are generally renegotiated
twice per month. The commodity-like nature of memory makes memory manufacturing very capital intensive,
since leading-edge technology is needed to get the lowest costs (the smaller a chip can made, the less expensive
107
Semiconductors
it is.) Most of the major memory companies run their own fabs; there are relatively few fabless memory
companies.
memory cell--a piece of circuitry that makes up one unit of memory in a memory chip.
microcontroller--a simple thinking chip--like a microprocessor but smaller and less smart. Electronic
devices like washing machines have microcontrollers to do the simple thinking required to run themselves.
microprocessor--a chip that is used as the thinking part of a computer. It is sometimes call a CPU or the
central processor unit.
MLC--see multi-level cell
Modem a device that modulates an analog carrier signal to encode digital information and also demodulates
carrier signals to decode transmitted information, producing a signal that can be easily transmitted and
decoded to reproduce the original digital data.
Moores Law. Moores Law is a term often used to describe the ongoing ability of the semiconductor
industry to make things smaller (transition to smaller linewidths), resulting in chips with lower cost and/or
more capability. It is named after one of the co-founders of Intel Gordon Moore, and Intel identifies its origin
as a 1965 paper by Gordon Moore published in Electronics Magazine. Gordon Moores original contention was
that the optimum number of devices on an integrated circuit for lowest cost per device would double every
year.
MOS (metal oxide semiconductor)--This is a reference to a type of transistor (see definition of transistor
below), the one that is most commonly used in integrated circuits today. The term MOS is misleading since
almost all MOS transistors today are really silicon-oxide-silicon structures. Intel is transitioning to metalgate transistors, but we believe that even with these new transistors the metal gate is not a pure metal. There
are two types of MOS transistors: NMOS and PMOS. Together NMOS and PMOS transistors can be connected
to create a CMOS (complementary MOS) circuit. See definition of CMOS above.
multilevel cell (MLC)--a NAND memory cell that can store more than one bit--a 1 and/or 0 at the same
time. Today NAND MLC is invariably a two-level cell--two 1s and/or zeros stored in one cell: 00, 10, 01, or 11.
nanometer--a billionth of a meter, denoted by the symbol nm. Semiconductor manufacturing technology is
described by the smallest width of a line that can be created in that technology. Currently, the most advanced
technology is 45nm.
NAND flash--as opposed to NOR flash--NAND is a type of flash memory (non-volatile) that is used in nonvolatile memory applications in which low cost and/or a large amount of memory is needed, typically
consumer applications such as storage for digital cameras, music storage for iPods, and removable storage for
computers. From a circuit point of view, NOR flash differs from NAND flash because it is random access,
information in any cell can be looked at, at any point in time. Therefore, there must be a connection (a bit line
contact) to every memory cell, increasing the size of the cell. NAND flash is serial access, a whole row of cells
must be read, one after another. This reduces the size of the NAND flash cell (no need to have a bit line contact
per cell) reducing the cost. However, it also reduces the speed of the NAND flash, as well as increasing the
error potential; if something goes wrong with one cell it can become impossible to get information in and out
of a whole row.
Non-volatile memory--memory that retains its data when the power source is removed.
NMOS--a type of MOS transistor (see definition of MOS above) in which the electrical current comes from the
flow of electrons (negative charged particles, hence the N in NMOS).
NOR flash--as opposed to NAND flash. NOR flash is used mainly for code storage in cell phones. From a
circuit point of view, NOR flash differs from NAND flash because it is random access; information in any cell
can be looked at, at any point in time. Therefore, there must be a connection (a bit line contact) to every
memory cell, increasing the size of the cell. NAND flash is serial access; a whole row of cells must be read, one
after another. This reduces the size of the NAND flash cell (no need to have a bit line contact per cell) reducing
the cost. However, it also reduces the speed of the NAND flash as well as increases the error potential; if
something goes wrong with one cell it can become impossible to get information in and out of a whole row.
optoelectronics--a type of semiconductor chip/device, not an integrated circuit, that deals with converting
light to electronic energy and vice versa.
108
PLD (programmable logic device)--an integrated circuit that can be programmed to perform complex logic
functions.
PMOS--a type of MOS transistor (see definition of MOS above) in which the electrical current comes from the
flow of positive charges (hence the P in PMOS).
qualification(1) qualification of a machine in a semiconductor fab refers to the process of installing the
machine and running tests on the machine to confirm that it can reliably process semiconductor wafers.
Qualification of a machine may take several weeks. (2) qualification of a new semiconductor fab or
manufacturing line refers to running tests on the entire factory or manufacturing line to confirm that the
factory can reliably processor semiconductor wafers. This can take several months. (3) qualification of a new
process (often in a new factory) refers to running test on chips made using the new process to confirm that
these chips are good chips and chips made on this process will reliably last a certain minimum time. It can take
several quarters to years to develop a new process and several months to qualify the process.
RAM (random access memory)--a semiconductor memory in which the information can be retrieved in
from any memory cell without looking through all the other memory cells. DRAM and SRAM are random
access. NOR flash is random access, but NAND flash is serial access. For a NAND memory, all the cells in a
whole line (a row) must be looked at one after the other.
Random Access Memory (RAM)see RAM
semiconductor--a material that is neither a good conductor nor a good insulator, but can exhibit both
properties. The most commonly used semiconductor material is silicon.
sensor--a chip/device that can sense heat, light, or pressure and generate a corresponding signal that can be
measured or interpreted.
single-level cell (SLC)--as opposed to a multilevel cell, a single-level cell is a NAND memory cell that can
store just one memory bit (a 1 or a 0) at a time.
SLC--see single-level cell.
Solid-state drive (SSD)--a form of computer memory, replacement for a hard disk drive, made of NAND
flash memory.
spot price, spot market--semiconductor memory can be bought on contract, in which prices are negotiated
in advance, twice per month, or on the spot market which offers a constantly changing price. The bulk of
DRAM is sold on contract, but the spot market is often thought to be a leading indicator of contract pricing.
SDRAM--synchronous DRAM, not to be confused with SRAM. A type of DRAM. SDRAM was used in the
1990s, but in the past two years DDR has become the standard type of DRAM.
SRAM (static random access memory)--type of volatile memory that is generally faster and more reliable
than DRAM due to the fact that it does not have to be refreshed like DRAM. It is generally more expensive and
is commonly used as cache memory in a computer.
standard linear--standard analog chips for doing generic tasks (as opposed to applications-specific
functions), general sold through catalogs.
steppersee lithography tool.
synchronous--many chips are synchronous in operation, which means that a clock signal needs to be
generated and all calculations or other operations in the circuit wait for the next beat of the clock before they
happen. This makes sure that signals are transferred at the same time, avoiding errors that may occur as a
result of one thing happening earlier than something else. The higher the clock frequency, the more things that
can happen per second, and the more the chip can achieve. Typical clock frequencies are 100s of megahertz
(hundreds of million times per second) or several gigahertz (several billion times per second).
transistor--a basic building block of integrated circuits generally working as a switch (turning on or off the
flow of electrons) or as an amplifier (making a small voltage bigger). The term transistor comes from the
concept of trans-resistance, being able to influence a current in one place by a current or a voltage in another
place. The first transistors were not semiconductor transistors at all, but vacuum tubes.
109
Semiconductors
Tri-GateTri-Gate is Intels name for its 3D transistors. In a 3-D transistor the gate wraps around the
channel. Since the gate is controlling the channel from three sides (hence the alternative name, tri gate) it can
turn-on the current more strongly (increasing the performance) or shut off the current more completely
(reducing the power consumption associated with leakage). A 3-D transition is sometimes called a FinFET
because the gate looks like a fin. FET stands for field effect transistor.
volatile memory--memory that loses its data when the power is turned off.
wafer--a round semiconductor disk, about half a millimeter thick, most commonly with a diameter of 8 or 12
inches, on which semiconductor circuitry is made. After a wafer has passed through the circuit fabrication
process, it is cut up into squares, the semiconductor chips, each containing a circuit.
wafer capacity--the rate at which semiconductor wafers can be processed in a semiconductor fab, typically
described in wafer starts per week or wafer starts per month (how many wafers begin being processed in a
week or month).
wafer equivalent--Since the transition to different wafer sizes often takes place over the span of years, at any
given point in time the semiconductor industry as a whole is doing its manufacturing on more than one size of
wafer. A lot of manufacturing is done on 200mm (8 inch) and 300mm (12 inch) wafers, but there still remains
some manufacturing on 6-inch and even some 5-,4- and 3-inch wafer processing. To normalize all these
different wafer sizes, capacity numbers (and other things associated with wafers, such as, for example, prices
charged per wafer by foundries) are often couched in wafer equivalent numbers. A 300mm wafer has 2.25x
the area of a 200mm wafer, and hence, 2.25 more chips can be fabricated on such a wafer. Therefore, a
300mm wafer counts as 2.25 200mm wafers when adding up wafer capacity.
wafer starts, wafer output--semiconductor chip production can be measured in terms of wafer starts per
week or wafer starts per month (how many wafers begin being processed in a week or month). An alternative
measurement is wafer outs per week or wafer outs per month (how many wafers complete their process in a
week or month.) Since it typically takes about 13 weeks to process a wafer, the time difference between wafer
starts and wafer outs is about a quarter.
Gartner disclaimer: All statements in this report attributable to Gartner represent Wells Fargo Securities
interpretation of data, research opinion or viewpoints published as part of a syndicated subscription service by
Gartner, Inc., and have not been reviewed by Gartner. Each Gartner publication speaks as of its original
publication date (and not as of the date of this report. The opinions expressed in Gartner publications are not
representations of fact, and are subject to change without notice.
110
Required Disclosures
Additional Information Available Upon Request
I certify that:
1) All views expressed in this research report accurately reflect my personal views about any and all of the subject securities or
issuers discussed; and
2) No part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed
by me in this research report.
Wells Fargo Securities, LLC does not compensate its research analysts based on specific investment banking transactions.
Wells Fargo Securities, LLCs research analysts receive compensation that is based upon and impacted by the overall profitability
and revenue of the firm, which includes, but is not limited to investment banking revenue.
STOCK RATING
1=Outperform: The stock appears attractively valued, and we believe the stock's total return will exceed that of the market over the
next 12 months. BUY
2=Market Perform: The stock appears appropriately valued, and we believe the stock's total return will be in line with the market
over the next 12 months. HOLD
3=Underperform: The stock appears overvalued, and we believe the stock's total return will be below the market over the next 12
months. SELL
SECTOR RATING
O=Overweight: Industry expected to outperform the relevant broad market benchmark over the next 12 months.
M=Market Weight: Industry expected to perform in-line with the relevant broad market benchmark over the next 12 months.
U=Underweight: Industry expected to underperform the relevant broad market benchmark over the next 12 months.
VOLATILITY RATING
V = A stock is defined as volatile if the stock price has fluctuated by +/-20% or greater in at least 8 of the past 24 months or if the
analyst expects significant volatility. All IPO stocks are automatically rated volatile within the first 24 months of trading.
As of: November 14, 2014
47% of companies covered by Wells Fargo Securities, LLC
Equity Research are rated Outperform.
111
Semiconductors
112
Diane Schumaker-Krieg
Sam J. Pearlstein
Co-Head of Equity Research (212) 214-5054
sam.pearlstein@wellsfargo.com
(212) 214-5051
(212) 214-5029
(212) 214-8064
(212) 214-8050
(314) 875-2042
(212) 214-8018
(415) 947-5470
(415) 396-4328
(212) 214-5015
(617) 603-4260
(415) 396-3873
(314) 875-2041
(314) 875-2556
(314) 875-2557
(617) 603-4314
(617) 603-4315
(212) 214-5072
(212) 214-8065
(212) 214-8059
(415) 396-3938
(415) 222-1159
(415) 396-3194
(415) 396-3977
(212) 214-8024
(212) 214-5024
ENERGY
Exploration & Production
David R. Tameron
Gord0n Douthat, CFA
Brad Carpenter, CFA
Jamil Bhatti, CFA
Richard Vidal
Master Limited Partnerships
Michael J. Blum
Sharon Lui, CPA
Praneeth Satish
Eric Shiu
Ned Baramov, CFA
David Freeland
Sam Dubinsky
Amir Chaudhri
Utilities
Neil Kalton, CFA
Sarah Akers, CFA
Jonathan Reeder
Glen F. Pruitt
Peter Flynn
Oilfield Services and Drilling
Matthew D. Conlan, CFA
Oilfield Services and Equipment
Judson E. Bailey, CFA
Daniel Cruise
Coleman W. Sullivan, CFA
David Cheng
Lisa Hausner
Global Head of Publishing
(443) 263-6522
lisa.hausner@wellsfargo.com
ENERGY CONT.
CONSUMER
Beverage/Convenience Stores/Tobacco
Bonnie Herzog
Jessica Gerberi, CFA
Adam Scott
Cosmetics, Household & Personal Care
Chris Ferrara, CFA, CPA
Joe Lachky, CFA
Zachary Fadem, CPA
Education
Trace A. Urdan
Jeffrey Lee
Food
John Baumgartner, CFA
Homebuilding/Building Products
Adam Rudiger, CFA
Joey Matthews, CPA
Leisure
Timothy Conder, CPA
Karen Wang
Marc J. Torrente
Restaurants & Foodservice
Jeff Farmer, CFA
Imran Ali
Retail
Paul Lejuez, CPA, CFA
Tracy Kogan
Justin C. Matthews
Matt Nemer
Omair Asif
Maren Kasper
Kate Wendt
Evren Kopelman, CFA
Connie Wang
Todd M. Wickwire
Co-Head of Equity Research (410) 625-6393
todd.wickwire@wellsfargo.com
(303) 863-6891
(303) 863-6920
(303) 863-6894
(303) 863-6880
(303) 863-6816
(212) 214-5037
(212) 214-5035
(212) 214-8056
(212) 214-5038
(212) 214-8021
(212) 214-5050
(212) 214-5043
(212) 214-5045
(314) 875-2051
(314) 875-2040
(314) 875-2052
(314) 875-2047
(314) 875-2049
(212) 214-5044
(713) 577-2514
(713) 577-2515
(713) 577-2510
(713) 577-2516
(713) 577-2542
(713) 577-2543
FINANCIAL SERVICES
BDCs
Jonathan Bock, CFA
Ronald Jewsikow
Gregory Nelson
Brokers/Exchanges/Asset Managers
Christopher Harris, CFA
Andrew Bond
Robert Ryan, CFA
Insurance
John Hall
Elyse Greenspan, CFA
Kenneth Hung, CFA, ASA
Rashmi H. Patel, CFA
Specialty Finance
Joel J. Houck, CFA
Vivek Agrawal
Charles Nabhan
Max Maier
U.S. Banks
Matt H. Burnell
Jason Harbes, CFA
Jared Shaw
Timur Braziler
(443) 263-6410
(443) 263-6449
(704) 410-2197
(443) 263-6513
(443) 263-6526
(212) 214-5025
(212) 214-8032
(212) 214-8031
(212) 214-8023
(212) 214-8034
(443) 263-6521
(443) 263-6563
(443) 263-6578
(443) 263-6573
(212) 214-5030
(212) 214-8068
(212) 214-8028
(212) 214-5048
HEALTH CARE
Biotechnology
Brian C. Abrahams, M.D.
Matthew J. Andrews
Shin Kang, PhD
Ronald Hsu, M.D.
Healthcare Facilities
Gary Lieberman, CFA
Ryan Halsted
Healthcare IT & Distribution
Jamie Stockton, CFA
Stephen Lynch
Nathan Weissman
Life Science Tools & Services
Tim Evans
Luke E. Sergott
Managed Care
Peter H. Costa
Polly Sung, CFA
Brian Fitzgerald, CFA
Medical Technology
Larry Biegelsen
Lei Huang
Craig W. Bijou
David Y. Brill. M.D.
Pharmaceuticals
Michael Faerm
(212) 214-8060
(617) 603-4218
(212) 214-5036
(212) 214-5064
(212) 214-8013
(212) 214-8022
(901) 271-5551
(901) 271-5552
(901) 271-5553
(212) 214-8010
(212) 214-8027
(617) 603-4222
(617) 603-4324
(617) 603-4277
(212) 214-8015
(212) 214-8039
(212) 214-8038
(212) 214-8042
(212) 214-8026
INDUSTRIAL
Aerospace & Defense
Sam J. Pearlstein
Gary S. Liebowitz, CFA
Automotive/Electrical and Industrial Products
Rich Kwas, CFA
David H. Lim
Deepa Raghavan, CFA
Chemicals
Frank J. Mitsch
Sabina Chatterjee
Containers & Packaging
Chris Manuel
Gabe S. Hajde
Nikita V. Bely
Diversified Industrials
Allison Poliniak-Cusic, CFA
Michael L. McGinn
Machinery
Andrew Casey
Justin Ward
Sara Magers, CFA
Metals & Mining
Sam Dubinsky
Amir Chaudhri
Shipping, Equipment Leasing, & Marine MLPs
Michael Webber, CFA
Donald D. McLee
Transportation
Casey S. Deak
(212) 214-5054
(212) 214-5055
(410) 625-6370
(443) 263-6565
(443) 263-6517
(212) 214-5022
(212) 214-8049
(216) 643-2966
(216) 643-2967
(216) 643-2968
(212) 214-5062
(212) 214-5052
(617) 603-4265
(617) 603-4268
(617) 603-4270
(212) 214-5043
(212) 214-5045
(212) 214-8019
(212) 214-8029
(443) 263-6579
(415) 396-4478
(415) 396-6056
(212) 214-5010
(212) 214-5011
(212) 214-8033
(212) 2148044
Satellite Communications
Andrew Spinola
(212) 214-5012
Applied Technologies
Andrew Spinola
Communication Technology
Jess Lubert, CFA
Michael Kerlan
Gray Powell, CFA
Priya Parasuraman
E-commerce
Matt Nemer
Trisha Dill, CFA
Information & Business Services
William A. Warmington, Jr.
Bill DiJohnson
Internet
Peter Stabler
Ignatius Njoku
Steve Cho
IT & BPO Services
Ed Caso, CFA
Richard Eskelsen, CFA
Tyler Scott
IT Hardware
Maynard Um
Munjal Shah
Santosh Sankar
Semiconductors
David Wong, CFA, PhD
Amit Chanda
Parker Paulin
Software/Internet, Technology
Jason Maynard
Karen Russillo
Vilma Chuy
Transaction Processing
Timothy W. Willi
Robert Hammel
Alan Donatiello, CFA
ECONOMICS
(212) 214-5046
(212) 214-8030
(212) 214-8066
(562) 637-1371
(443) 263-6442
(443) 462-7354
(617) 603-4262
(617) 603-4233
(617) 603-4263
(443) 263-6568
(212) 214-5013
(212) 214-8052
(212) 214-8048
(617) 603-4269
(415) 396-3938
(312) 920-3594
(617) 603-4283
(617) 603-4271
(415) 396-4478
(415) 396-4064
(415) 396-6056
(443) 263-6524
(410) 625-6381
(443) 263-6540
(212) 214-8008
(212) 214-8061
(212) 214-8007
(212) 214-5007
(314) 875-2045
(212) 214-5066
(415) 947-5472
(415) 396-3505
(415) 396-3345
(314) 875-2044
(314) 875-2053
(314) 875-2054
STRATEGY
Equity Strategy
Gina Martin Adams, CFA, CMT
Peter Chung
Strategic Indexing
Daniel A. Forth
Gaming
Cameron McKnight
Rich Cummings
Tiffany Lee
Healthcare/Manufactured Housing/Self Storage
Todd Stender
Philip DeFelice, CFA
Jason S. Belcher
Lodging/Multifamily/Retail
Jeffrey J. Donnelly, CFA
Dori Kesten
Robert LaQuaglia, CFA, CMT
Tamara Fique
Office/Industrial/Infrastructure
Brendan Maiorana, CFA
Young Ku, CFA
Blaine Heck, CFA
(212) 214-5012
(212) 214-8043
(212) 214-8063
(704) 410-3233
Economists
John E. Silvia, PhD
Mark Vitner
Jay H. Bryson, PhD
Eugenio J. Alemn, PhD
Sam Bullard
Anika Khan
(704) 410-3275
(704) 410-3277
(704) 410-3274
(704) 410-3273
(704) 410-3280
(704) 410-3271
(410) 625-6378
(443) 263-6516
(443) 263-6564
(443) 263-6529
www.wellsfargoresearch.com
Copyright 2014 Wells Fargo Securities, LLC