The world runs on electronics and semiconductor
chips. Many of these chips are made in massive, multi-billion facilities run by foundries
like TSMC, Samsung, and GlobalFoundries. Right now the world is going through a
protracted shortage of these semiconductors. Especially in the case of TSMC, the
most advanced foundry out there, there is only a limited amount of capacity
available for all of its hundreds of customers. So why can't TSMC just build another factory? Obviously more factories
need to be brought online. Just how hard can it be? In this video I want to talk about the logistical
and financial challenges that foundries tackle when building and running a new fab.
What are the thorny construction issues and financial tradeoffs involved? I have done these engineering and construction
videos before for the Hong Kong Airport and the Taipei MRT. I will link them in the description
below. Super fun but they are always complicated and so much is left on the cutting room
floor. Please forgive me for any errors. Let us begin with the basics. A fab is shorthand for a semiconductor
fabrication facility. It houses offices, locker rooms, storage facilities, test
and packaging rooms, and the cleanroom. These are massive facilities - aircraft carriers
on land. Take for example, TSMC's Fab 14, a Giga-Fab located in my hometown of Tainan.
Completed in 2008 over the span of two years, the multi-story building has over
114,000 square meters in floor space. And like an aircraft carrier, Fab 14 houses
its own cafes, cafeterias, convenience stores, bakeries and bookstores. Unlike an aircraft
carrier, it has car parking. Everything needed to manufacture wafers at scale. At full capacity, Fab
14 churns out 30,000 pizza-sized wafers a month. This is an amazing, spare-no-expense facility.
How does a company create such a thing? Imagine that you want to build and rent out rooms
in a 1,000 room resort hotel to take advantage of a hot tourist region. Constructing the whole thing
all at once would take years. You are also not sure how long the tourist boom is going to last
and if it will still be around when you are done. So ... what if you build and complete just part
of the resort first? You tell your construction crews to start and finish, say, 300 rooms.
That gets done in a year and you can start making money on those rooms right away. Once
the money is in hand you build the rest. This is, in short, how TSMC and other foundries
build a fab. Fab construction in its entirety takes about 3 years from start to end. This
is a long time no matter how you spin it. Semiconductor chips and the process nodes that
made them become outdated extremely quickly. The profitable economic life of a
process flow is about 3-4 years. Afterwards, the unit cost of
a chip drastically declines. Considering their staggering cost and the nature
of the business, you want to be able to get paid back as soon as you can on your multi-billion
dollar investment. No company can have its $20 billion fab sitting idle at the site. TSMC needs
to get it up and running as soon as possible. So the company breaks down the entire construction
into separate and parallel packages or "phases". Construction begins long before
the whole design is complete. The process gets to production much faster, but
as a result, gets fantastically more complex. You start by figuring out where to put your new
fab. You are considering a variety of factors when trying to decide on a site. What
are the tax incentives? How much land is available to build on? Are there enough qualified,
well-educated workers around that you can hire? Available infrastructure like constant water
and power is especially critical. When AMD (now GlobalFoundries) announced that it would build a
fab in New York's Luther Forest Technology Center, local officials said that it would take
two years to get the necessary water, sewage and road connections completed.
It ended up taking seven years. For that particular project, water
was the biggest obstacle. The local, state and federal governments as
well as various private entities pitched in over $65 million to build a
28 mile pipeline and a purification plant on the Hudson river capable of purifying 14
million gallons of water a day. And that is in New York, which for practical purposes has an
unlimited amount of water for industrial use. TSMC's fabs use a stunning amount of water. In 2013, those fabs by themselves
generated 19 million tons of wastewater. It is undoubtedly higher today. Tainan the entire
city that same year generated 29 million tons. Taiwan gets a lot of rain but it suffers
from its own water supply issues. As a result, Tainan has requested that TSMC's
fabs try to recover 85% of the water it uses. To meet those goals, the company built
a large rain collection system to capture rainwater and store it in 700 ton
tanks. Wastewater generated from industry is treated for 25 different
chemicals and can be reused. If not reusable, then it is put into the
cooling tower to help maintain temperatures. Geographic and geological concerns also play a
role in placing the site. Taiwan happens to be a very earthquake-prone location. I have another
video about the earthquake risk to TSMC's fabs but to summarize, the company builds dampers
into the facility to reduce tremors by up to 40%. These of course all cost money and
need to be done right and on time. With all of these considerations, the
cooperation (and tax incentives) of local government is critical. California's Silicon
Valley no longer has fabs for a good reason. They all had to leave for environmental
and resource constraint reasons. I did a video previously about TSMC's move to
Taiwan's south. There is also another regarding its decision to enter the China market. Both are
worth checking out and are in the TSMC playlist. At the heart of a semiconductor fab is the
cleanroom. The cleanroom is where the chip fabrication happens. Because a single
piece of dust can render a chip useless, pure air is constantly pumped in and
anyone entering needs to wear bunny suits. The size of these rooms - TSMC's Fab 14
cleanroom is 31,000 square meters large or over five and a half soccer fields
- combined with increasing demands on air purity make clean room construction
pricier than ever. Their cost skyrocketing as expensive HVAC machines have to be brought in
to maintain purity throughout such larger spaces. Fab cleanrooms have three types of
configurations: A ballroom style, tunnel style and mini-environment style. The
ballroom allows for the maximum flexibility. The tunnel, less so, but is also cheaper. The mini-environment, which is what TSMC
uses, comprises of compartments segmented by the level of cleanliness. So you have one
compartment with class 1, another with class 10, and class 1000. This helps keep costs down
as HVAC machines only need to maintain strict levels of cleanliness in a fraction
of the overall room. It also allows TSMC to start production in one environment while crews
simultaneously fill out the other environments. Speaking of filling out ... The equipment inside the fab
typically makes up to 75% of the cost of the entire facility. So
selecting and acquiring the right equipment is critical in making sure that the whole
fab is delivered on time and within budget. Tradeoffs need to be made everywhere throughout
the process. Here is one small example. You are bringing in your equipment and
need to place it inside your cleanroom. Where do you locate it? At the time of
cleanroom design, this might not be clear. So you decide to expand the range of your piping, HVAC and utilities throughout
all your micro-environments. This offers you maximum future flexibility to
deal with any unforeseen issues, but at the same time bloats the budget. But if you don’t do this
right and mess up the placements then there’s the potential of an even more expensive work stoppage.
So it is all about making the right trade-off. A cutting-edge ASML EUV machine can
cost over $150 million by itself. The company only makes 50 a year. Making
sure that it is properly brought over and installed in the cleanroom without
defects must be a nerve-wracking process. Accidents can happen. In one
anonymous insurance report, I read about a $5 million piece of equipment that
got dropped as it was unloaded from an elevator. The device looked fine on the outside but
inside it had suffered trauma and needed to be returned to Japan for recalibration at
a cost of $200,000 and incalculable lost time. Generally, almost all semiconductor
foundries run their fabs on top of precise data management and tracking. They
want to know how well their processes are yielding and whether or not there can be
room for improvement. You want to get to a good place as soon as possible so that the
fab can start making money for the company. And there’s always room for improvement. The
equipment is brand new. The process steps are brand new. There is a lot of fat to cut within the
thousand or so steps that go into making a chip. So teams on the ground should
have both the technical competence and the leeway to make changes whenever they
recognize the opportunity. The engineers work together with the operators and the technicians
to recognize, diagnose and solve such problems. Internal foundry teams also work very
closely with the vendors that make the equipment they are using. The vendors know
the equipment best, after all. They made it. They also work with the customer to
make sure that the product they get is the one that they want. This means that
for the most advanced nodes you want to make sure that you are working with customers
you really trust. Newer, more unfamiliar customers might start with more mature nodes
first before progressing to the hot stuff. As TSMC embarks on building its fab in Arizona -
its first advanced construction project situated on the other side of an ocean - it is hard to
appreciate just how rough and unforgiving of a task the whole thing is. I suppose it is
like what they say about ducks: You see a duck floating on the water, but have no idea how
frantically it is paddling beneath the surface. To construct, equip and optimize a fab so
that it can start making on-time deliveries of a fantastically advanced product
is hard. You are building a boat, launching it into the sea, and setting sail before
you have finished installing half the engines. And now they got to do it 11,400
kilometers away from home.
