Rod Pemberton wrote:
> I don't translate to English through an intermediate language. I
> don't mentally process concepts in a non-English language.
I don't say so. You just want to misunderstand people and therefore
choose the most inapropriate meaning of an ambiguous term in the next
round.
> You have a few spelling problems and a few
> comprehension issues. I've mentioned one of your misspellings more
> than a few times.
And you even needed to correct your own spelling... anal-retentive
stuff. Nobody else bothers here about the occasional mis-spelling. I
occasionally misspell in my own language, too, we all do, we respond to
that by doing best effort to try to understand what was written. This
is not a problem, apart for anal-retentive spelling Nazis, as we call
them in Germany. Why do you do it? It contributes *nothing* to the
discussion.
> They understood the literal meanings of words, but had
> problems grasping the concept of the words together, such as a phrase,
> sentence, or
> paragraph. You do the same, but at a higher, more functional level.
> I.e.,
> it's less noticeable, but still present.
I don't translate English into something else, particularly not into
Chinese. When I speak Chinese (my fourth language), I try not to
translate, either, it's a rather futile thing to do (Chinese words are
about as ambiguous as English words, but the different meanings they
cover is completely orthogonal - one English word translates to 5
Chinese words, and one of these Chinese words translate to 5 different
English words, which clearly aren't synonyms). I try to think in
Chinese. It's one of the things I would encourage every language
student: Try to think in the target language, don't try to translate.
However, I think you don't follow Postel's advice for communication: If
you speak to someone else, try to understand, not try to misunderstand.
If there's an interpretation that fits your understanding of things,
it's probably that you and I are in agreement. There's no need to find
the interpretation where we disagree.
Maybe you are not aware of the ambiguousities, because you don't know
any other language. Particularly Chinese is an eye-opener for that,
because you suddenly realize how context-sensitive the words in your own
language are, and how much you need to know the context to understand
Chinese, because the words there simply don't map to our words. Words
have a meaning only in context.
> Actually, talking to you sometimes is a bit like talking to two
> people: one who is fairly adept at English, and one who learned
> English as a second language. Do you have someone helping you?
No. I have definitely learned English as a second language, and don't
have someone helping me. It's just sometimes you don't understand me;
and I don't think in these cases it is my fault.
>> It's another misunderstanding of you: You absolutely have to
>> fill your fab. For memory makers, it's cut-throat; they have
>> been dying like flies in the past years.
>
> What's wrong with making a smaller fab? It's one way to fill your
> fab.
It's less efficient, i.e. the price per die will be higher. They don't
build bigger fabs for the fun of it, but because they want to be cheaper
than the competition. Furthermore, there aren't many companies who make
equipment; for leading edge processes, you essentially have to buy from
one company in the Netherlands (ASML); they have a sort-of (natural)
monopoly on the most important component - the stepper. That however
means that the equipment is designed for those who are willing to pay
the price - and those are the big manufacturers who want big fabs.
>> The semiconductor
>> industry did something completely different: They separated making
>> and designing.
>
> So did the automotive industry, circa the early 1990's. Many of them
> moved design from Detroit to California, Italy, Taiwan, Korea, China,
> etc.
Stop saying "automotive industry" when you want to say "Detroit". I
applied Postel's principle, and I clearly figure out that you use the
words "automotive industry" as synonym for "Detroit", and you ignore the
rest of the world, which also has their automotive industries.
Only Detroit destroyed itself by outsourcing everything to some other
place. Furthermore, when you say something like "moved design to
Italy", it sounds as if Italians haven't made cars before. FIAT is 9
years older than GM (the old GM, the one founded in 1908 which went
bankrupt 100 years later). You probably wouldn't say GM "moved design
to Germany", because, yes, they do (Opel), but that would make you sound
silly. Outsource it to the country where the car was invented? Maybe a
good idea (but Opels are rather lousy cars for German standards, though
they are clearly better than GM cars in America).
>> This is very natural, as the semiconductor industry is a "printer"
>> industry. The making is litography, i.e. printing.
>
> You mean lithography, not "litography". Specifically, it's
> photolithography.
See, you are doing it again. You perfectly understood me, but you think
it's worth to show off that your English is better than mine. What's
the point? It's supposed to pronounce "lit^{h}ography", but you
probably pronounce it with a fricative "th" (as in modern greek and
English, but this is supposed to be pronounced as ancient greek, where
theta was a plosive).
>> There are some businesses in the printing industry where you need
>> the press in-house (newspapers, e.g.), but usually, the design
>> (the writing) is separated from the replication (the printing).
>>
>
> There you go again with the "lump and dump" - conflating lithography
> with photolithography.
You apparently have an abstraction/analogies problem, you see the
differences but miss the similarities. The essence of lithography and
photolithography is that both make cheap replications possible -
replications of drawings, of books, of integrated circuits, and that the
replication process is independent of the creation process.
> Photolithography is too highly specialized to compare it to the
> printing of
> book and newspapers. Basically, you're conflating a simple process of
> placing ink on paper with a complex process such as chemical etching
> and vapor deposition of elements, like silicon, gallium, or arsenic,
> and also re-classifying the work numerous Electrical Engineers as that
> of mere wordsmiths.
Uh-oh. A writer or a painter who creates an *original*, which later is
replicated by printing, is an artist, his profession is not to do the
replication, but to create the original. As is the EE's task, who
creates a circuit, which is then "printed" in the fab. Both professions
require skills, quite different skills, but good writers are even rarer
than good EEs.
The fab however is not doing EE. They are doing lithography and
chemistry, their expertise is to control the process and keep the yield
high - they essentially operate really, really high-tech printers. It's
a well more advanced way of doing it than printing a newspaper, but the
essence is doing many, many cheap replications.
I've designed chips from top to botton, including (one time) device
engineering, which means creating a new type of transistor that wasn't
available in the design kit before.
However, I never was directly involved in *any* sort of actual
processing of the silicon in the fab, even during the years at Zetex,
which had their own fab (used e.g. for the low-scale integration analog
compagnion of the digital audio amplifier I developed there). Fab
people are not EEs. You confuse a printer with a writer. We visited
fabs to see how they do it and of course we talked to the fab people, so
we need some understanding of what they do. As in writing, between the
EE and the fab, there is a layouter - named the same way as the layouter
who does the typesetting of the words the writer has written down.
>> You can make smaller fabs, but these are less efficient - older
>> processes, smaller wavers, less output. The people who have small
>> fabs use old equipment and have their niche where this is profitable
>> - discretes, small analog components.
>>
>
> Small doesn't mean less output. It depends on your rate of
> production. Fast machines and three shifts a day can produce as much
> as a single shift plant with large, slow machines.
Rest assured that all high-end fabs around the world have the fastest
machines available, and are operated 24/7.
> Most large factories use increased
> scale as a replacement for speed, i.e., large, slow machines instead
> of fewer smaller, faster machines.
Have you ever been inside a fab? The machines are not particularly
large, they are room-height, the steppers are 2-3x5-10m or such, the
rest of the equipment is slightly smaller. In the free space, they have
robots to carry wafer trays from one machine to the next (the robots
don't look like R2D2, they are just a cross between cranes and small
railroads).
> Do you know what a related-rate problem is?
Yes. But this is completely irrelevant here.
> These machines are all custom.
Actually not. They are mass produced. For a not that big amount of
"mass", but they are definitely not built to order. Just look up
asml.com, they have nice photos, where you also can look inside the
machines. The steppers are the most essential machines for a fab, there
are other machines for etching and CVD, which however look similar from
the outside (they are all sealed to keep the dust out). The stepper is
doing the actual copying: mask to wafer. They can expose >100 wafers
per hour, so they aren't really slow. The rest of the equipment, the
CVD and etching, have comparable throughput rates, but you should
understand that each individual wafer stays for hours in any such
machine. The size of each of these things is similar to a stepper, see
for example this one, which set a record of 120wph:
http://www.novellus.com/products/product_lines/altus/
> There is nothing requiring them to be
> constructed for large capacities. A bread baking machine in bread
> factory or a sheet glass machine in a glass factory are sometimes over
> thousands of
> feet long and hundreds of feet wide. But, nothing requires that they
> be
> made that large. The can be made for whatever size is required. The
> process and machine size are not linked. It's just a matter of
> planning, design, and money.
Well, the semiconductor industry buys standard off-the-shelf products.
Lego bricks to build your fab, so to speak. The bricks get bigger and
bigger, though. When I visited the .35µm fab of AMS 10 years ago, the
stuff was a lot smaller than it is now for 32nm and below.
>> > If they don't need the volume that a fab can produce for their own
>> > production, they're better off selling it.
>>
>> Yes, so you finally got it. That's why they went fabless.
>
> That just indicates they overbuilt.
They have to buy and install new equipment every two years, so many
chances to correct that mistake. I don't think they overbuilt.
Let's look at the data: TSMC's 28nm fab has now 24k wafer starts per
month (and that's actually rather small, this is still in learning and
growing mode; GlobalFoundries combined 32/28nm fab has 50k wafer starts
per month and wants to grow to 80k, because now they have the customers,
not just AMD). ASML's twinscan XL, which is the appropriate stepper for
that fab, has 150wph, or 108k per month. It takes around 40 masks to
process a complete wafer (20 of them are metals and vias), so TSMC needs
only 10 of these steppers, if they are 100% utilized.
There isn't much room for less. Several process steps like the tungsten
contacts (done with the Altus linked above) are only done once per wafer
(contacts and local interconnects), and you want to utilize the entire
fab 100%. The single Altus with its 120wph is good for more than 80k
wafers per month. They won't sell that many of them; low two-digits, I
guess. I hope this illustrates the problem: The machines are quite
small and fast, but really, really expensive.
> There is still a financial advantage to
> owning your own fab. They can make them smaller or more efficient, if
> they want.
It's a fairly trivial calculation: Suppose your external fab has a cost
disadvantage of 30% per wafer - that's the profit they want to make.
Because they fab for everybody, they can fill their big, big fab, and
make it very efficient, by having 100% utilization for everything, and a
good yield. If they succeed to be 30% more efficient than your small
fab, the deal is done: Go fabless.
There was a time where you did need your own fab, that was during the
GHz race. Fabs like TSMC didn't offer the special processes you needed
for high-end processors; their processes were tuned for general purpose
ASICs, which didn't need that speed.
This has changed since. TSMC still has slower transistors, but Samsung,
GlobalFoundries, and IBM can provide you with the right processes to
build your supercomputer. And they are all open to the public, only
Intel's one half step ahead process isn't. And that lead is in danger:
ASML has troubles to make the EUV steppers production ready, and now all
the big fab companies and Intel did a joint investment into ASML, and
the result will be that everybody will get the new steppers at the same
time - Intel always paid a premium to get them first.
IMHO, apart from Intel and the memory makers, nobody today benefits from
having their own fab. The memory makers benefit, because they have
special needs (separate steps for the capacitors, which nobody else
makes, and still only two metal layers), and cut-throat competition.
Maybe with RRAM, that will go away - apparently it is possible to make
resistive RAMs with standard processes, i.e. without any special
equipment, and you also benefit from the many metal layers.