INTEL’S 10NM
NIGHTMARE
WHAT WENT WRONG
WITH INTEL’S
LATEST NODE?
As far back as July 2015, Intel publicly confirmed that its
10nm node was in trouble. Fast forward nearly seven years
and we’ve only just got our hands on its first desktop 10nm
processors. Moore’s Law dictates a doubling of transistor
densities every two years, implying a new process node on
roughly the same cadence. So something has gone horribly
wrong. But what, exactly?
The first problem was ambition. Intel simply went for
too big a leap with 10nm. In 2009, for instance, Intel made
the jump from 45nm to 32nm. With that came an increase in
transistor density from 3.3 million per square millimeter
to 7.5 million, a little over double the density. When 22nm
arrived in 2011, it was a similar story. It upped the transistor
density ante to 15.3 million, or slightly less than double.
Fast forward to 2014, one year later than you’d expect
a new node from Intel, and 14nm delivered 37.5 million
transistors per square millimeter. That’s well over double
what 22nm delivered. 14nm was the first troubled node from
Intel, and it’s likely no coincidence that it also attempted to
more than double transistor density.
Then there’s 10nm. The first version of 10nm, which was
used in tiny numbers to produce a dual-core mobile chip that
almost nobody bought in late 2018, was good for a massive
100.8 million transistors per square mm. That’s a huge 2.7x
scaling compared to 14nm and a big ask indeed.
The other problem was that Intel opted not to go with
EUV or extreme ultraviolet lithography technology. As
we’ve explained elsewhere, the wavelength of light used in
a lithography process ultimately dictates feature size. And,
currently, EUV tech is the shortest wavelength and most
advanced available for commercial production.
The reasons why Intel didn’t go with EUV are quite
technical and are related to the fact that it tends to produce
chips using single-die masks. Any damage to that mask
threatens to kill the entire wafer. By contrast, the likes of
TSMC and Samsung tend to use multi-die masks that can
sustain some damage while maintaining reasonable yields.
Also, EUV-compatible tools to protect masks, known as
pellicles, only became available in 2019.
In short, Intel decided not to go with EUV, making that
ambitious 2.7x density scaling even more technically
challenging. Intriguingly, Intel’s next node, formally known
as 7nm and now branded Intel 4, is expected to be less
ambitious in ‘only’ increasing transistor density to 180
million per square millimeter, less than double that of its
existing 10nm node. It will also be the first Intel node to use
EUV technology.millimeter, which is what makes that crazy 57 billion transistor
count viable for the Apple M1 Max chip in a portable MacBook Pro
computer.
Indeed, if there’s a single outfit remaining in the chip industry
that best represents the notion that Moore’s Law is alive and
well, it’s surely Taiwan Semiconductor Manufacturing Company
(TSMC). It’s currently the largest chip foundry in the world and
the undisputed leader when it comes to squeezing ever-smaller
features into viable, commercial semiconductors that go into
products you can actually buy. While Intel has fallen from its
leadership position, TSMC has kept on trucking. In September
2020, Apple’s A14 chip debuted TSMC’s 5nm process, putting it at
least a full node ahead of Intel.
Moreover, TSMC is incredibly bullish about its ability to keep
Moore’s Law going. Philip Wong, vice president of corporate
research at TSMC doesn’t just think Moore’s Law is alive and well,
he reckons it could keep on delivering for another three decades.
“It’s not dead,” Wong says of Moore’s Law, “it’s not slowing down.
It’s not even sick.”
As for how TSMC has achieved its lead, a major factor has been
its aggressive adoption of so-called EUV or extreme ultraviolet
lithography technology. In simple terms, semiconductor
manufacturing comes down to etching details on silicon wafers
by shining light through patterned masks onto a substrate. The
substrate (or wafer) is coated with a light-sensitive photoresist
chemical that hardens when exposed to ultraviolet light, while
the remainder of the coating is removed by solvents, leaving a
finished circuit.Intel’s 10nm node arrived at least five years late.Tools such as masks are critical to the chip production process.Moore’s law
32 MAXIMU MPCMAR 2022
© INTEL, ASML