MOORE IS NO
LONGER LESS
SMALLER
TRANSISTORS
COST BIG MONEY
If the demise of Moore’s Law has been called a little
prematurely, one thing’s for sure: keeping it going is
becoming increasingly costly. For instance, Intel says it
plans to spend $20 billion on a pair of 7nm fabrication units,
colloquially known as fabs. Meanwhile, just one of TSMC’s
latest 5nm manufacturing plants, known as fab 18, is said to
have cost £17 billion. That’s big money.
Those huge costs are the major reason why the industry
is consolidating down to just a few major players. The
price of entering the game, let alone of maintaining some
competitiveness, is incredibly prohibitive. Another intriguing
indicator of just how resource-intensive chip manufacturing
has become can be found in the example of ASML (Advanced
Semiconductor Materials Lithography). A Dutch company
founded in 1984, it specializes in the photolithography
machines used by all the big players in chip production. If
you want to knock out semiconductors using the latest EUV
processes, you need an ASML machine.
But here’s the incredible bit. ASML’s role is so important,
its market capitalization or overall value is now greater than
Intel’s. So, the broader economics of chip manufacturing is
changing fast—and not for the better. Taking TSMC as an
example (it’s harder to make these judgments based on Intel
because, until recently, it has only produced chips for itself
and not acted as a foundry for third parties), consider its
transition from 65nm to 28nm. The cost of a wafer went up
from around $2,000 for 65nm to $3,000 for 28nm. However,
because so many more chips can be produced on a 28nm
wafer than an equivalent 65nm wafer, the price of individual
chips dropped by two-thirds.
However, if you take TSMC’s more recent progression
from 10nm to 7nm and then 5nm, wafer prices have
ballooned from $6,000 to $17,000. The consequence is that
chip prices have only shrunk by about 15 percent. The moral
of the story is that if wafer prices keep going up, it won’t
matter whether it’s technically possible to keep Moore’s Law
alive. Nobody will be able to afford the chips, anyway.its latest MacBook Pro computers, packs a slightly crazy 57 billion
transistors, for instance. But for as long as we can remember,
Intel has been the bellwether for the industry. It’s also the outfit
that has suffered the most well-publicized chip-production woes
of late.
Back in 2015, Intel confirmed that the cadence of Moore’s Law,
as far as it was concerned, had slowed to two and a half years
with the transition from 22nm silicon chip production to 14nm.
But it was the jump from 14nm to its latest 10nm node that has
proved the most problematic. To understand just how wrong it
has gone for Intel, you have to look a little further back in time. In
2012, Intel was expecting to have 10nm processors in production
by 2015, with 7nm coming online in 2017 and an even smaller node
available in 2019.
But here we are in 2022 with Intel’s first 10nm desktop
processor having gone on sale just a few months ago. Intel’s 10nm
is at least five years late. You can read all about what went wrong
with Intel 10nm in the boxout on page 32, but suffice to say that
7nm is also proving problematic and is subject to delays. Indeed,
the situation has gotten bad enough that Intel has felt the need to
rebrand its production nodes to compensate for a perceived lag
versus the competition.
To cut a long story short, Intel is now calling its 10nm node
‘Intel 7’, while the upcoming 7nm node is rebranded ‘Intel 4’. If that
seems like a desperate attempt to use alternative facts to cover
up a technological shortcoming, perhaps there’s some truth to
that. However, node nomenclature such as ‘10nm’ and ‘7nm’ has
long become disconnected from the size of any actual features
inside a chip. The gate pitch of most existing production nodes
marketed as 7nm is roughly 50nm, for example. So, node names
are more a rough guide to comparative density than indicative of
actual feature sizes.
On that note, Intel’s 10nm tech (the one now known as Intel 7)
is said to be good for around 106 million transistors per square
millimeter of semiconductor. That’s actually slightly better than
the 95 million and 97 million achieved by Samsung and TSMC’s
7nm nodes, respectively. So, in comparative terms, ‘Intel 7’
certainly makes sense. That said, both Samsung and TSMC
are already producing chips on 5nm processes. TSMC’s 5nm
node packs in an incredible 173 million transistors per squareMAR 2022MAXIMU MPC 31
© ASML
Above: ASML’s lithography
machines are now worth
more than Intel’s CPUs.Left: Intel now knocks out 10nm
wafers with desktop CPU dies.