IEEE Spectrum October, 2007 - 23

be able to withstand the high temperatures used in the annealing structure using the gate-last process flow. It was a gutsy call.
step-not a problem for polycrystalline silicon but potentially a Our team knew it was committing all of Intel's next generation
big one for some metals.
of microprocessors to the biggest change in transistor technolTo make a long story short, the search for gate electrode ogy in 40 years.
materials with both the right work function and tolerance to
The next key milestone was to demonstrate working test
high-temperature processing was very difficult and full of dead chips using the final scaled dimensions combined with the new
ends. Especially for the PMOS transistor.
transistor features. The traditional chip to test a new technology
Another transistor process sequence, dubbed "gate last," on is static random access memory, or SRAM, which is the type
circumvents the thermal annealing requirement by deposit- of memory collocated on the same chip with the microprocessor.
ing the gate electrode materials after the source and drain are Typically, microprocessor makers have designs for SRAM that
formed. However, many of our peers saw the gate-last process, are a year or more ahead of their processor designs. SRAM is a
which we ultimately adopted, as too much of a departure and very regular array of memory cells, each of which consists of six
too challenging.
densely packed and interconnected transistors. Because of their
Meanwhile, a third approach remarkable in its simplicity density and regularity, SRAM chips provide good data on how
emerged. Called fully silicided gates, it lets you follow the many defects a manufacturing process produces.
normal gate-first process but then lets you turn the polyOur first fully functional test SRAM chips with the new transilicon gate into a metal-silicide
sistors came off the line in January
gate, essentially replacing every
2006. They were of a 153-megabit
other silicon atom with metal (usudesign consisting of more than
ally nickel). Then, by doping the
1
bi l l ion t ra n si stors. Each si xPolysilicon gate
nickel silicide, you can alter its work
transistor memory cell in the chip
function for use in either an NMOS
occupied little more than one-third
Polysilicon
gate
device or a PMOS one. By late 2006,
of a square micrometer. This test chip
High-k dielectric
though, nearly everyone, includhad all the features needed to build
ing us, had given up on the fully
a 45-nm microprocessor, including
High-k dielectric
silicided gates approach. No one
the high-k plus metal gate transiscould move the silicide's work functors and nine layers of copper interChannel
tion quite close enough to where it
connects. Considering how new and
BUMPY RIDE: The particular density of electrons in a traditional
needed to be.
radically different the transistor and
polysilicon gate allowed inherent vibrations in the high‑-k dielectric
Nevertheless, the search goes on
manufacturing process were, it was
Channel
to move into the transistor channel and
disrupt the flow of current.
at other major chip makers to find
a surprise even to some of the engithe materials with the right work
neers in the development group that
Metal gate
function that could survive high
it all worked together so well. Even so,
temperatures and enable the industry
the development team still had a lot
standard gate-first process flow.
ahead of it to bring the performance,
gate
High-kMetal
dielectric
reliability, and yield of the process up
having builT well-funcTioning TransisTors
to the level needed for manufacturing
using old technology, in the second
microprocessors.
High-k dielectric
half of 2003 it was time to move
The new gate stack worked wonChannel
from research to development of
ders in battling leakage through the
high-k dielectric plus metal gate
gate, reducing it by more than a facChannel
tra nsistors, as we ca l led them.
tor of 10. But the gate oxide is not
Engineers began working to deter- The higher density of electrons in a metal screened out the
the only source of transistor leakmine whether these early transis- vibrations, allowing current to flow more smoothly.
age chip makers have to worry about.
tors could be scaled to the upcoming
The other significant leak is called
45-nm dimensions and still meet the rigorous performance, source-to-drain or subthreshold leakage. It's a trickle of current
reliability, and manufacturability requirements of an advanced seen even when the transistor is intended to be in the "off" state.
microprocessor technology.
Making transistors smaller has also meant steadily lowering the
It was no cakewalk. The research group engineers had pro- amount of voltage needed to turn them on, the threshold voltvided a critical lead in identifying promising high-k and metal age. Unfortunately, steadily lowering the threshold voltage lets
gate materials, but the NMOS and PMOS transistors had not more current slip through. For many years, each new generation
yet been combined on one wafer as they would be in a micro- of transistor would increase drive current (and improve perprocessor, using a manufacturing process that could make both. formance) by about 30 percent but would pay the price of about
What's more, there were hard questions still to be answered a threefold increase in subthreshold leakage. Leakage currents
about how many good chips we could expect for every bad one have reached levels high enough to be a noticeable portion of
(yield) and how reliable those chips would be.
total microprocessor power consumption.
During the months that followed, the team cracked one
The industry is now in an era where power efficiency and
problem after another-making changes to materials, chemical low leakage are more important than raw speed increases. But
recipes, and manufacturing processes. It wasn't until late 2004 a transistor can be designed to operate to favor either priorthat the team felt it had enough convincing data that the new ity by adjusting the channel length or adjusting the threshold
transistors could be made to work on our 45-nm technology. voltage. A shorter channel leaks more but allows for a higher
At that point, there was no turning back. Intel was now com- drive current. A higher threshold voltage pinches off the leak
mitted to making a high-k dielectric plus metal gate transistor but also throttles the drive current. Adjusting the threshold

- - - - - - - -- -- -- - - - - - -- - - - -

--------------------------------------------------

34 IEEE Spectrum | October 2007 | NA

www.spectrum.ieee.org

http://www.spectrum.ieee.org

Table of Contents for the Digital Edition of IEEE Spectrum October, 2007