Baloney. No one sells silicon, particularly something with as much IP
in it as a processor, for anything close to "production costs", to
*ANYONE*. Selling at close to cost isn't a good way to pay the note on
a multi-billion dollar fab, a hundred million in engineers salaries,
while leaving enough to keep the stock holders happy.
Of course you have to include the investment amortisation into the
"production cost" at some expected production volume. However, the
early "list price" charged from early adaptors can be quite arbitrary.
Maybe, sorta. Good enough for a first swag, anyway. Large caches give
opportunities to change these simple rules somewhat. But that has
nothing to do with the issue at hand.
This again depends if there are only core specific caches or just a
common cache on the chip.
Oh, so now they're going to increase the chip size yet again to add a
fourth core (and all the interconnect) knowing that they're going to
throw it away? Yikes! The cores we're talking about aren't a sq. mm.
Those 4 core chips are made for customers that really need them and
are willing to pay the premium price for them.
These customers might not be willing to pay anything for chips with
one or two defective chips. Instead of scrapping these chips, there
might be others that are paying something for chips with 2 or 3
functional cores. It makes perfectly sense for the chip maker to sell
these chips instead of scrapping them.
However, if the demand for three core chips exceeds the amount of
partially faulty four core chips, only then it makes sense to produce
chips with a mask for only three cores. This threshold depends how the
yield for the 4 core chips is developing and how fast the demand for 3
core chips is growing.
Sure, partial-good strategies have been used for memories for a *long*
time, but memory <> processors.
With a sufficiently large number of cores on a chip, we are
approaching a similar tradeoff.
Paul
