 |
You may be the foremost materials expert in the world, but if you try to
leave your mark on chip development without regard to some boundary conditions
of a more economical nature, you will not
achieve much. And if you are the manager
(which you should be with the kind of education you get here), you better be
aware of the following points that are special to research, development and
manufacture of (memory) chips. |
|
 |
There is no other product with quite such brutal
requirements, even considering that all
technical product development must follow similar (but usually much more
relaxed) rules. |
|
1. A new generation with four-fold
capacity will appear on the market every three years. |
|
|
That is an expression of "Moore's law". It is, of course, not a "law"
but an extrapolation from observation and bound to break down in the not so
distant future (with possible disastrous consequences to the economy of the
developed countries). |
|
|
The original observation made in 1965 by
Gordon Moore,
co-founder of Intel, was that the number of
transistors per square inch on integrated circuits had doubled every year since
the integrated circuit was invented. Moore predicted that this trend would
continue for the foreseeable future. In subsequent years, the pace slowed down
a bit, but data density has doubled approximately every 18 months, and
this is the current definition of Moore's Law, which Moore himself has blessed.
Most experts, including Moore himself, expect Moore's Law to hold for at least
another two decades. |
|
|
Here is a graphic representation for
microprocessors: |
| |
|
|
|
|
|
|
|
|
|
|
Not bad - but of course Moore's law must break
down sometime (at the very latest when the feature size is akin to the size of
an atom (when would that be assuming that the feature size of the P7 is
0,2 µm?). This is
illustrated in a separate module. |
|
|
Still, as long as it is true, it means that you
either have your new chip generation ready for production at a well-known time
in the future, or you are going to loose
large amounts of
money. There are some immediate and unavoidable consequences: |
|
|
You must spent
large amounts of money to develop the chip
and to built the new factory 2 - 3 years before the chip is to appear on
the market, i.e. at a time were you do not know if chip development will be
finished on time. And large means several
billion $. |
|
|
The time allowed for developing
the new chip generation is a constant: You can't start early, because
everything you need (better lithography, new materials,) does not exist then.
But since chip complexity is ever increasing, you must do more work in the same
time. The unavoidable conclusion is more people and shift work, even in research and development.
|
|
|
It follows that you need ever
increasing amounts of money for research and development of a new chip
generation (there is a kind of Moore's law for the costs of a new generation,
too). Look at it in another
way in a separate module. |
|
2. The market for chips grows
exponentially |
|
|
That is an expression of the insatiable demand
for chips - as long as they provide more power for less money! That this
statement was true is shown below. Note that the scale
is logarithmic! |
|
|
|
|
|
|
|
|
|
|
Shown is the total amount of money
grossed in the semiconductor market from 1960 - 2000 (Source: Siemens /
Infineon). |
|
|
The essentially straight line indicates
exponential growth - including the foreseeable future. Extrapolations, however,
are still difficult. The two extrapolated (in 2000) blue lines
indicating a difference of 100.000.000.000 $ of market volume in
2005 are rather close together. The error margins in the forecast thus
correspond to the existence or non-existence of about 10 large
semiconductor companies (or roughly 150.000 jobs). |
|
|
We also see the worst downturn ever in
2001. Sales dropped from 204 billion $ in 2000 to
139 billion $ in 2001, causing major problems throughout the
industry. |
|
More specific, we can see the
exponential growth of the market by looking at sales of DRAMs. Shown is
the total number of memory bits sold. Note
that just a fourfold increase every three years would keep the number of
chips sold about constant because of the
fourfold increase of storage capacity in one chip about every three years. |
|
|
|
|
|
|
|
|
|
|
|
The unavoidable consequence is:
Your production capacity must grow exponentially, too, if you just want to keep
your share of the market. You must pour an exponentially growing amount of money into
investments for building factories and hiring people, while the returns on
these investments are delayed for at least 2 - 3 years. In other words:
the difference of what you must spent and what you earn increases, very
roughly, exponentially. This is not a healthy prospect for very long, and you
must make large amounts of money every now
and then (e.g. by being the first one on the market with a new chip or by
having a quasi monopoly). |
|
|
You must make (and sell at a
profit) an exponentially increasing number of
chips (since the prize per chip is roughly constant) to recover your
ever increasing costs. Since chip sizes increase and prices must stay halfway
constant, you must use larger wafers to obtain more chips per processing. This
puts a lot of pressure on developing larger Si crystals and equipment to
handle them. |
|
|
You must produce as many chips
as you can during the product life time (typically 5 years). Continuous
shift work in the factory (7 days a week, 24 hours a day) are
absolutely mandatory! |
|
3. Chip prices for memories decay
exponentially from the time of their market introduction (roughly $60)
by two orders of magnitude within about 5 years (i.e. at the end you get
$1). |
|
|
The prize development up to the 16 Mbit
DRAM can be seen in an
illustration via the link. Microprocessors may behave very differently (as
long as Intel has a quasi monopoly). The rapid decay in prizes is an expression
of fierce competition and mostly caused by: |
|
|
1. The "learning curve", i.e. the increase of the
percentage of good chips on a wafer (the yield) from
roughly 15% at the beginning of the production to roughly 90% at
the end of the product life time (because you keep working like crazy to
improve the yield). |
|
|
2. Using a "shrink strategy" . This means you use
the results of the development efforts for the next two generations to make
your present chip smaller. Smaller chips mean more chips per wafer and
therefore cheaper chips (the costs of making chips are mostly the costs of
processing wafers). |
|
|
An immediate
consequence is that if you fall behind the mainstream for 6 months or so
- you are dead! This can be easily seen from a simple graph: |
|
|
|
|
|
|
|
|
|
|
|
The descending black curve shows the expected
trend in prizes (it is the average exponential decay from the
illustration). The ascending
curve is the "learning curve" needed just to stay even - the cost of
producing one chip then comes down exactly as the expected prize. |
|
|
Now assume you fall behind 6 month - your
learning curve, i.e. your yield of functioning chips does not go up. You then
move on the modified blue learning curve. The prize you would have to ask for
your chip is the modified red prize curve - it is 30 % above the
expected world market prize in the beginning (where prizes are still moderately
high). |
|
|
Since nobody will pay more for your chips, you
are now 30 % behind the competition (much more than the usual profit
margins) - you are going to loose large amounts of
money! |
|
|
In other word: You must meet the learning curve goals! But that is
easier said then done. Look at real yield curves to appreciate the
point |
© H. Föll
