2004-05-12 The Evolution of Intel's Copy EXACTLY! Technology Transfer Method [developer.intel.com]
2004-05-12 20:10 erik
On Thu, 6 May 2004, Lise Eisenberg wrote:
>> Subject: [s-t] Color therapy for integrated circuits?
>> >From today's New York Times, buried in an otherwise uninteresting article:
>> Even Intel doesn't know quite why some chip manufacturing processes work
>> better than others. In the late 1990's it instituted a program called
>> "Copy EXACTLY!," which required that new plants use equipment and
>> procedures replicated from existing plants, right down to the color of
>> paint on the wall.
For the Times to interpret this as Intel not knowing quite why their
processes work the way they do is to ignore why they have those processes
in the first place, and how they came to be. It's not guesswork, it's
As a former drone who extricated herself from the Intel collective over
three years ago, I retain certain imprinted memories of this process.
Intel routinely builds research fabs to develop the next generation of
chip processes. For each smaller-circuit process level, a whole new class
of smaller particles exist that were once benign on larger processes, but
are now damaging to smaller circuits; a pebble becomes a boulder. When
configurations, processes and training in that fab mature to reach an
acceptable yield, it can become a production fab. If more fabs with that
configuration are required, Intel "Copies Exactly" the configuration of
the successful production fab in its final form.
It's not that they don't understand their own processes, but experimenting
with processes, equipment, or even paint, in the critical production path
can perturb yields; when a fab suddenly loses yield capacity, it's called
"losing the recipe". I had friends whose sole job was to analyze failed
chips, then find what caused the flaw and how it got there. The arrival
of a new machine version is a great way to lose the recipe.
One time, my friend did her usual particle analysis on some fouled wafers
and discovered, much to her surprise, silver particles. She tracked it to
a newer version of a machine that did stuff at high vacuum. Eventually
she discovered that inside that machine were tiny silver screws, which let
a few silver particles loose and caused defects. Whoops. Now we know.
This is why they want to "Copy Exactly". Even something as innocent as
paint color could be a source of damaging particles in a fab. Her fab
used only special paper in the special printers that were supposed to cut
down on particles. Of course, no makeup, hair spray, perfume, cologne,
deodorant etc. were allowed in either. Every hour that the line is down
for crap like this can cost millions of dollars. Better to set up the
processes and machines and copy exactly then screw around with the
production line and suffer reduced yields and corresponding losses. All
new technology development costs are upfront, and are recouped at a rate
directly proportional to the fab's yield.
A new production fab costs billions of dollars. You decide whether saving
a buck or two on a machine or painting a room a prettier color is worth
risking the loss of yield for three days while you figure out why. Every
hour on a production line is worth hundreds of thousands, maybe millions
of dollars. Copy Exactly is Intel's best protection against Murphy's Law.