Intel Has Already Won the War
The semiconductor war is over and Intelo won it.
I can only describe as amusing the description of ARM Holdings
as a competitor to Intel. ARM is a designer of microprocessor cores that can be
bought and used by semiconductor manufacturers and fabless semiconductor
companies. Since ARM has no fabrication capability, they can never compete
directly with Intel, never. Apparently many of the writers and commentators
don’t understand that Qualcomm, Nvidia, TI, and others are dependent not only
on ARMH for their microprocessor design, but they are dependent on wafer
foundries, such as TSMC, Global Foundries, UMC, and even Samsung to provide the
bleeding edge manufacturing technology to make their products a reality. All of
these foundry sources, repeat, ALL of them are at least two generations of
process technology behind Intel. In the semiconductor business, the last thing
you want is to be four years or more behind the leader in manufacturing
technology. Being two generations behind means that an application processor of
equivalent function from Intel would be ¼ the physical size of a processor
from, for example, Qualcomm who receives their wafers form TSMC. The Intel
device would also be much higher performance and much lower power, both
desirable characteristics for battery operated mobile products.
To understand how bad this really is, a little primer on the
unusual semiconductor business model is in order:
Semiconductors, otherwise known as integrated circuits, are
manufactured in what is called a “batch process”. The ICs are manufactured on
silicon wafers that are, in the case of Intel, 300mm (12”) in diameter. The cost
of manufacture of the wafers is a constant. It costs the same amount to make a
wafer with 100 ICs on it or to make a wafer with 1000 ICs on it. The more ICs
that can be made at the same time on a wafer, the cheaper each IC will be. So,
the ongoing miniaturization of ICs is more than just a “fun” technological
thing to do, it reduces the cost of ICs and, of course, the price of the
equipment that the ICs go into.
An added benefit of the smaller ICs is that the performance
goes up and the power requirements decline. Modern microprocessors can have
over a billion transistors on them. It is absolutely essential that these chips
be made on the most advanced processes available. The history of
microelectronics is littered with the corpses of semiconductor companies that
fell behind in manufacturing technology.
Today, the cost of leading edge technology semiconductor
plant is several billion dollars. The cost is so high that most companies can
no longer afford to own a manufacturing plant. This has led to what is
confusingly referred to as a “fabless semiconductor company”. These companies
design their own products, but essentially time-share a manufacturing plant
owned by a semiconductor “foundry” to produce those products. Some names of the
foundry companies are TSMC, UMC, Global Foundries and even Samsung is acting as
a foundry for certain customers.
Some names of the fabless semiconductor companies are
Qualcomm, Nvidia, AMD, Broadcom, Marvel, even TI uses foundries to make their
leading edge mobile products.
The products of the fabless semiconductor companies can only
be as good as the technology available at the semiconductor foundry companies
used to manufacture them.
A word about “process nodes”:
Process nodes are the actual size of the transistors on the
ICs. Process nodes are continually reduced in size in order to gain performance
and cost advantages. Today the process nodes of interest are 32/28nanometer and
22nanometer. For example a 32nm transistor is 32nm by 32nm in size, or 1024
square nanometers in area. A 22nm transistor is 22nm by 22nm or 484 square
nanometers in area. As can be seen, shrinking the edge dimension of the
transistor by 1/3 reduces the area of the transistor by more than half. That
means that more than twice as many of the same ICs can by produced a 22nm wafer
than can be produced on a 32nm wafer. That further means that the cost of the
22nm IC is less than half the cost of a 32nm IC. Of course, the 22nm IC is
faster and lower power than the 32nm IC.
Gate-first vs. gate-last issue:
Less well understood by investors or discussed by the
commentators is the issue of gate first vs. gate last.
Several years ago a process consortium led by IBM attempted
to standardize on the basic process used by foundries to manufacture ICs. Many
of the members of the consortium decided on a “gate first” process. The only
two major manufacturers who decided to move to a “gate last” process are Intel
and TSMC. As it turned out the gate first approach would make it difficult or
impossible to advance the process node below 32nm. Changing from gate first to
gate last is extremely complex.
In order for companies like Samsung and Global Foundries to reduce their node
to 22nm, they will likely have to convert their existing 32nm gate first
process to 32nm gate last process and run it long enough to confirm the
integrity of the process and then shrink it to 22nm and below. This process
could take as long as four year to accomplish. All this time Intel and TSMC can
simply move to the more advanced nodes in much less time than the other companies.
Intel is already considered to be two generations ahead of the closest
competitor. TSMC, although using gate last processing, has run into serious
problems running their 32/28 nm fab. Since the other foundries run gate first
processes the TSMC customers’ designs would not be compatible with processes
run at Samsung and Global Foundries.
Some discussion in support of the Intel win:
The new Apple iPad is using an update of the Apple designed
A5 Application Processor called the A5X. The A5X is still manufactured by
Samsung on their 45nm node..(see, now you know what a node is), the update of
the chip increased the size to 165 sq mm. In the world of ICs that is a huge
and expensive chip. It also increased power consumption to the point that Apple
had to increase the size of the battery in the iPad. One would have expected
such an important part to be produced on the Samsung 32nm process, if Samsung
actually had a 32nm process node in high volume production. Apparently they
The financial media insists on referring to the “declining”
PC market. Gartner predicts a 4.4% growth in the worldwide PC market this year.
In top of that unit growth, Intel will receive generally higher prices for
their PC chips since the new Ivy Bridge will have a far better graphics
processing unit (GPU) integrated onto the CPU. The more powerful GPU will cost
AMD and Nvidia a great deal of the discrete GPU business. The value thus added
could increase CPU revenue 25% over the next couple of years. “Post PC” is
another fanciful term coined by Steve Jobs, rest his soul, but it just ain’t
The CEO of Nvidia recently suggested, in public, that Intel
to its fabs to the fabless semiconductor companies. What makes this a new high
water mark in desperation is the past relationship between Nvidia and
Intel. In early 2011 Intel and Nvidia
settled a lawsuit
in which Intel paid Nvidia $1.5 billion. There is no love lost between these
AMD, in the strangest transaction recently, “sold”
their interest in Global Foundries (GF used to be the AMD fab until they had to
sell it in order to stay in business). In this context “sold’ is an interesting
concept since AMD sent the amount of $425 million to the other owners of GF in
order to be released from the commitment to buy their chips from GF! Apparently
AMD doesn’t think the GF gate first technology can keep up with Intel. The plan
is to have TSMC to build AMD processors! That gives new meaning to the
expression, “jumping from the frying pan into the fire”.
As evidence of the fact that TSMC is in no position to
support large volume of 28 nm products, the TSMC CEO, Morris Chang, has urged
his customers to seek other suppliers so they are not totally dependent on
TSMC. In the same article Mr. Chang cautioned that falling behind Intel in
manufacturing technology would have consequences that would be “horrible
to imagine” [for TSMC and their customers].
TSMC predicts that 28nm will be 10% of their revenue this
year. What they don't say is that the 28nm wafers they are able to ship will cost
about five times the cost of an average 12” wafer from TSMC. So 28nm will
amount to about 2% of the wafer volume of TSMC....and that will be with bad
yields. 2% of TSMC’s wafer output is about 80,000 wafers. Total 2012 demand for
28nm and below wafers will be about 1 million wafers. To build the TSMC
capacity to one million wafers per year would cost $6-8 billion and take 2-3
years to get up and running. Good luck on getting those new Qualcomm and Nvidia
28nm application processors and AMD CPUs
To sum things up, the gate first decision was a disaster,
TSMC screwed up by underinvesting in 28nm capacity and is having yield
problems with the capacity that they do have, even AMD doesn’t trust GF, and
Samsung can’t build enough 32nm A5X chips to satisfy Apple.
Meanwhile, Intel is in the process of ramping up $40 billion
worth of 22nm capacity to serve all of the end customer markets that the
fabless semiconductor model will be leaving in the lurch. Intel finds itself in
the incredibly fortunate position of basically having no competition on the
most advanced process node for at least four years.
There is no way for the foundries to leap frog process nodes
in order to catch up with Intel. Even TSMC has to think seriously about
spending the money for more 28nm capacity when anything they build will be
obsolete before ground is broken on the new plant.
The bottom line is that the ongoing 28/32nm discussion is
irrelevant. Intel is moving to 22nm
Trigate this year. That is effectively two generations ahead. Next year will
see the first shipments of Intel 14nm products. No foundry, including Samsung,
will ever catch up with Intel on process technology, therefore the foundries
and their customers have already lost the war. How silly to discuss their
progress on 32/28nm when Intel is shipping 22nm Trigate products and is a year
away from 14nm! It's over, and Intel has won.
What to do?
If you own Qualcomm, Nvidia, TI, TSMC, Broadcom, Marvel,
ARMH or any other public fabless semiconductor company that needs leading edge
foundry technology, sell all holdings immediately. Put 70% of those funds on
Intel shares, 10% on Intel Jan 35 calls for some extra “juice”. Sit on 20% dry
powder for Micron. That story is for another day.
As Warren Buffet says, “Diversification is for chickens”.