00:00:02
Today, Intel seeks to remake itself into a
full-service foundry with Intel Foundry Services.
00:00:09
But did you know that this is not
their first foundry? Some 15 years ago,
00:00:15
Intel had Intel Custom Foundry, or ICF.
00:00:20
Most people today hardly remember
ICF. Intel was so low-key about it.
00:00:26
But at its peak it had a billion
dollars in revenue contracts,
00:00:30
strong support by the ecosystem, over
a thousand multi-national employees ...
00:00:35
And ten plus customers - many won over from TSMC.
00:00:40
In today's video, the rise and
fall of Intel’s first foundry.
00:00:45
And the lessons to be learned for its second.
00:00:49
## The Best Factories
Intel in the 2000s believed that it had two Crown
00:00:52
Jewels. The first of which was its
leading edge manufacturing tech.
00:00:57
For two decades, Intel under its CEOs Andy Grove,
00:01:01
Craig Barrett and Paul Otellini
developed and scaled a brand new
00:01:06
technology to billions of chips. And did this
every two years to hit the Moore's Law cadence.
00:01:14
Such technologies spanned disciplines
from lithography like the excimer lasers
00:01:19
to new deposition methods like Atomic Layer
Deposition to new materials like Hafnium Oxide.
00:01:26
Several times, the company's
engineers made astute,
00:01:30
bold technical choices that yielded
where others did not. Like for example,
00:01:36
the choice of "Gate Last" over the "Gate First"
method, backed by the tech giant IBM and its
00:01:42
allies. Even TSMC pursued "Gate First" before
abandoning it halfway through development.
00:01:50
Such choices granted the company a strong
1-2 generation lead over its competitors.
00:01:56
And Intel's people certainly liked to
hammer that point in their presentations.
00:02:02
But Intel's top management also knew
that fabs were getting increasingly more
00:02:07
expensive. And at some point, Intel's PC market
revenues alone cannot justify the next node.
00:02:15
And in those years, the semiconductor industry was
still hoping to make the jump to 450 millimeter
00:02:21
wafers. Larger wafers reduce per-die cost
but with the drawback of massive investment.
00:02:28
The proper strategic response
then would be to roll up demand
00:02:32
from outside fabless customers and
use that to help fund the next node.
00:02:38
## Intel Architecture
00:02:38
What Intel saw as its other crown jewel was the
x86 chip architecture, or Intel Architecture.
00:02:46
Before Otellini took over the CEO spot from Craig
Barrett, Intel offered an Arm-based product called
00:02:53
XScale. Otellini sold XScale to Marvell in
2006, seeking to refocus the company on x86.
00:03:02
In an August 2007 interview
with the Financial Times,
00:03:06
he explained why he was throwing the
company's whole weight behind x86:
00:03:13
> And rather than rethink or rewrite the internet,
00:03:16
which today runs on X86 machines,
between us and our competition. I mean,
00:03:22
the entire internet is coded for that base. All
the browsers, all the Flash applications ...
00:03:28
You catch that? I admit it is
a bit garbled. But Otellini is
00:03:34
asserting that the whole internet was
written on x86 machines, referring to
00:03:39
Intel's then growing strength in Linux and
Xeon-powered internet servers. He continues.
00:03:45
> Rather than rewrite the internet to a new
architecture, it’s a lot easier to – from my
00:03:52
perspective, to do derivatives of our architecture
to meet new form factor and power requirements
00:04:00
The iPhone proved Otellini’s sentiments incorrect,
00:04:03
and Intel sensed the
immediate shift away from x86.
00:04:08
In response, they rushed out the Intel Atom
brand of low-power processors. These later
00:04:15
powered small, cheap netbooks, but
that wasn’t the original intention.
00:04:20
## Custom Intel Architecture Foundry
Intel's first foundry began as a simple initiative
00:04:24
by Paul Otellini to monetize
these two crown jewel assets.
00:04:29
A small team was formed to provide
"turnkey custom foundry services"
00:04:36
for internal and external customers to
produce custom or semi-custom chips.
00:04:42
The foundry's original name was "Custom
Intel Architecture Foundry" or CIAF.
00:04:49
The name hints at the goal: Monetize Intel’s
manufacturing factories, and also defend the
00:04:58
threat to the Intel Architecture franchise
from the trend of low-power systems-on-chip.
00:05:05
The effort essentially kicked off in 2008
with an Intel Technology & Manufacturing
00:05:11
VP discretely approaching existing customers,
00:05:15
and touting the benefits of customizing Intel
products using Intel manufacturing processes.
00:05:23
This VP shortly left this role and was replaced
by Sunit Rikhi, a vice president from Intel's
00:05:31
Technology and Manufacturing division.
Rikhi saw the foundry as an opportunity
00:05:37
to build a new business and would serve as
its general manager for the next seven years.
00:05:44
## Building a Foundry
00:05:44
Yes, Intel could be called the leader
in silicon manufacturing technology.
00:05:50
But that technology had been tightly optimized
by Intel’s Technology and Product groups to
00:05:56
make fast x86 CPUs and not much else than
that. It was less a question of ability,
00:06:04
than rather those processes being in no
shape to present to external customers.
00:06:11
Think of Intel’s factories as
like a pizza kitchen. For years,
00:06:15
that pizza kitchen - its equipment, ingredients,
00:06:19
processes and cooks - produced pepperoni pizzas
per a recipe made by another internal Intel team.
00:06:27
The two groups know each other well
and how they worked. And together they
00:06:32
made the best pepperoni pizzas in the industry.
00:06:36
But now we must reveal the kitchen’s capabilities
to external customers, who then might use it to
00:06:43
make cheese pizzas, Hawaiian pizzas, Chicago
deep-dish pizzas, or New York style pizzas.
00:06:51
Might even get the occasional
weirdo asking for a calzone.
00:06:56
A foundry like TSMC, UMC or GlobalFoundries
will have done the extensive work needed to
00:07:05
lay out its kitchen’s capabilities
for a customer to pick and choose
00:07:09
their pizza the way they want it
to be. Intel needed to match that.
00:07:16
Moreover, the Intel foundry team had to
also equip itself with complementary IPs
00:07:22
that its customers can use. Think
toppings like olives, sausages,
00:07:28
and onions. Building this whole
ecosystem required a lot of work.
00:07:34
In the case of TSMC, they had been
doing this for over three decades.
00:07:39
And those years of cumulative
knowledge have yielded a vast,
00:07:44
widely varied menu as well as a kitchen capable
of producing precisely what was ordered off
00:07:50
it. A different mindset, and approach,
from what Intel had always done before.
00:07:56
## Ditching Intel Architecture
00:07:56
Custom Intel Architecture Foundry was the
second prong of a two-pronged strategy by
00:08:02
Otellini to bring a low-power x86 Intel
Architecture, Atom, to mobile products.
00:08:09
The first prong had been an alliance
with TSMC. In this partnership,
00:08:14
fabless customers can license Intel Atom IP
and build systems-on-chips on top of them.
00:08:22
Less than a year later, however, this
collaboration was put on hold due to low customer
00:08:28
demand. The team at Custom Intel Architecture
Foundry experienced the same low demand issues.
00:08:36
This low demand stemmed from the Atom CPU's
immaturity as an ecosystem or platform for
00:08:43
building products. And even if there had
been customer interest, it would have been an
00:08:49
immense technical challenge to turn the CPU into
usable IPs akin to those already offered by Arm.
00:08:58
So some time in 2009, Otellini was convinced to
drop the Intel Architecture part of the name.
00:09:06
They would move forward with
just Intel Custom Foundry.
00:09:11
He reportedly did this reluctantly, and
held it in his heart until his retirement
00:09:16
that Intel Custom Foundry should be about
Intel Architecture designs rather than that
00:09:21
of others like Arm. Regardless, he gave ICF
support and should be credited for as such.
00:09:29
## Competing With Customers
00:09:30
There remained one major thing. Intel going into
00:09:33
the foundry business raises the question
of competing against your own customers.
00:09:40
This issue dates back to the days of the Japanese
semiconductor-makers, who gladly rented out spare
00:09:46
capacity on their lines in exchange for
a pound of flesh, your stock or your IP.
00:09:53
TSMC was founded on the basis of never doing this,
00:09:57
and it remains enshrined
in their corporate charter.
00:10:01
Intel would have been very much
competing with many of its foundry’s
00:10:05
largest potential customers. Companies like
AMD, Nvidia, and increasingly, Qualcomm.
00:10:13
Even if Intel was not in that business
right then, there lingered the worry that
00:10:19
that might change. The company’s future
aspirations spread quite far and wide.
00:10:26
So the Intel management chose to do a slow
and discreet rollout. They restricted the
00:10:33
young foundry to only engaging with
customers in areas that Intel does
00:10:38
not compete in and had no aspirations for.
Which was not a very large playing field.
00:10:46
In the end, the Intel Custom Foundry team
chose to compete in a specific type of chip
00:10:51
called a Field-Programmable Gate Array,
or FPGA. That was a large enough space
00:10:59
that Intel then had no intention of getting into.
00:11:04
Later. In the first two quarters after
Intel went public with the news that it was
00:11:09
getting into foundry, analysts asked twice
about Intel foundry’s impact on capacity.
00:11:16
Then-CFO Stacy Smith, who is now on the board,
00:11:20
replied both times that Intel
was not focused on that business,
00:11:25
was only talking to very specialized companies,
and was not building a broad-based foundry.
00:11:32
## Big Customers
00:11:33
That turned out to be alright,
00:11:34
because Intel Custom Foundry was nowhere
near ready to accept big customers.
00:11:41
Shortly after news first emerged of Intel
offering foundry services in late October 2010,
00:11:48
Apple emerged as a potential foundry client.
00:11:52
There was a brewing conflict between
themselves and their current SOC foundry,
00:11:57
Samsung. Later in 2011, Apple
sued Samsung for infringing
00:12:03
on Apple's intellectual property.
It was time to go somewhere else.
00:12:08
Morris Chang wrote in his autobiography - which
00:12:12
for now is only available in
Chinese - that in November 2010,
00:12:16
Terry Guo of Foxconn rung him to say that
he would bring an Apple executive to dinner.
00:12:24
That executive was Jeff Williams, today
Apple's COO. Williams showed interest in
00:12:31
engaging TSMC to produce iPhone SOCs and
the two sides started working on a deal.
00:12:39
But just as this work ramped up,
it was suddenly put on pause. Jeff
00:12:44
Williams called Morris to tell him that Paul
Otellini was negotiating with CEO Tim Cook
00:12:51
on a potential foundry deal. The TSMC-Apple
collaboration would be paused for two months.
00:12:58
Morris then wrote about visiting Apple
in April 2011 asking about things. Cook
00:13:05
simply tells him that "Intel is not good
at 'contract manufacturing'" (他們不擅做代工).
00:13:11
Let's drill into that a bit. In late 2010,
00:13:15
Intel Custom Foundry was only about
two years old. They were profoundly
00:13:20
unprepared to take on what is perhaps the most
demanding tech company client in the world.
00:13:27
I want to remind you. The most important
decision that any fabless company can make
00:13:33
is their choice of foundry partner. Get
that wrong, and nothing else matters. For
00:13:39
such a company-defining decision,
you don't want to take any risk.
00:13:45
Intel Custom Foundry, after two or so years
of work, was not yet 90% close to TSMC.
00:13:54
Even if they were, that last 10% would be a big
enough deal that Apple will still go with TSMC.
00:14:03
Apple's people took the meeting because their top
management asked for it. But in terms of actual
00:14:10
KPIs and milestones to get the deal, Intel was
nowhere near ready yet. It was dead on arrival.
00:14:18
There was another opportunity - maybe -
for an Intel-Apple deal three years later,
00:14:25
but I will get to that in a bit.
00:14:28
## Startup
00:14:28
So Intel Custom Foundry was not
ready for the major leagues.
00:14:33
But there were a few customers amenable to
what Intel was offering. Those guys were
00:14:38
the startups. They were tired
of being so beholden to TSMC.
00:14:43
They were also more tolerant of the
competitive risks of working with Intel,
00:14:48
thanks to intensely structured contracts that
held Intel to severe punishments if they did
00:14:55
not protect customer IP from Intel's Product
divisions or provide the contracted capacity.
00:15:03
The thing most important to these
startups was Intel's league-leading
00:15:07
silicon manufacturing. Intel Custom
Foundry's first Process Design Kits
00:15:12
were for the 32 nanometer node, but
they quickly switched to 22 nanometers.
00:15:18
Because customers wanted to bring out
a league-leading product and leapfrog
00:15:23
their competitors. Considering how long it
would take to bring a chip to the market,
00:15:28
the 32 nanometer process node
would already be lagging.
00:15:33
Intel Custom Foundry's first reported launch
customer was a small fabless FPGA startup called
00:15:40
Achronix. Achronix's FPGAs were for specific
tasks like network traffic or data encryption.
00:15:48
The Wall Street Journal reported in late
October 2010 that Intel agreed to give the
00:15:54
startup access to about 1% capacity of its
leading-edge 22-nanometer FinFET process.
00:16:02
In an interview, Achronix's chief executive
acknowledged that the Intel wafers cost more,
00:16:08
but said, "This is a historic development. Intel
00:16:12
is far ahead of anyone else
in new process generations".
00:16:17
## 22 nanometers
Intel's process lead on its competitors
00:16:19
seemed the greatest when it came to the
critical transition to 3D transistors.
00:16:25
In late 2011, Intel announced
its 22 nanometer process node
00:16:29
equipped with its "Tri-Gate" variant of FinFETs.
The FinFET is a type of transistor which covers
00:16:36
the channel on three sides, offering
superior power consumption and speed.
00:16:42
It took over a decade for Intel to master
the various bits of making the FinFET,
00:16:47
which included controlling the variations you
tend to get in the manufacture step. This work
00:16:54
remains one of the company's crowning
achievements in semiconductor history.
00:16:59
Intel originally scheduled to ship its
"Ivy Bridge" 22-nanometer CPUs in April
00:17:04
2012. But they encountered some problems,
and the ship date slipped to June 2012.
00:17:11
A few weeks delay, but no big deal at the
time. Intel was so far ahead of everyone
00:17:18
else - at least two generations
- that nobody thought more of it.
00:17:23
For Intel Custom Foundry, Achronix’s Speedster22i
FPGAs started to ship in volume in early 2013.
00:17:34
And those chips were good - ICF always
did make good silicon once it was all
00:17:38
said and done. These FPGAs were notable
for consuming half the power of other
00:17:45
high-end chips from other FPGA
makers like Altera and Xilinx.
00:17:52
The best working process node those other
two FPGA companies had access to in those
00:17:57
days was TSMC's 28-nanometer node
- which was then still somewhat
00:18:02
volume-constrained. So 22 nanometer
was maybe one or two generations ahead.
00:18:08
In February 2012, Intel added a second
foundry customer, Tabula. Like Achronix,
00:18:15
Tabula was another FPGA-maker startup, and they
had been using TSMC's 40-nanometer process node.
00:18:23
And then in April 2012, another start-up
customer called Netronome. They previously
00:18:29
used TSMC's 65-nanometer process node.
The news makes it clear. From the start,
00:18:36
ICF was stealing customers from TSMC.
00:18:41
## Design-Manufacturing Integration
Intel's external spokespeople regularly
00:18:43
touted their design-manufacturing
integration as the future of foundry.
00:18:49
In October 2012, Intel scientist
and senior fellow Mark Bohr said:
00:18:54
> The traditional foundry model
is running into problems. In order
00:18:58
to survive, the foundries will have to become
more like an integrated device manufacturer.
00:19:04
Even some of the chief spokespeople for
the foundries have said something similar.
00:19:09
> The foundry model worked well when traditional
scaling was being followed and everybody knew
00:19:15
where we were headed. In this era, where you
continually have to invent new materials and
00:19:20
new structures, it’s a lot tougher being a
separate foundry and maskless design house.
00:19:26
> Being an IDM, we have design
and process development under
00:19:30
one roof. That’s really a significant advantage.
00:19:35
A fully-focused Intel foundry can provide
more than just wafer fabrication. They can
00:19:40
also leverage the organization's
full line of design and advanced
00:19:44
packaging services too. They can do
the whole course from soup to nuts.
00:19:50
Some of the packaging stuff was special.
A bit later, Intel's teams invented a
00:19:55
new type of advanced packaging called
Embedded Multi-die Interconnect Bridge.
00:20:00
Per the name, this format embeds a small silicon
chip into the package to connect different dies
00:20:06
without the need for a more expensive silicon
interposer. None of the Intel product teams
00:20:12
could use it, but Foundry had customers who did. A
perfect example of Intel's integration strengths.
00:20:20
## Behind the Scenes
So theoretically, this all made sense.
00:20:22
But GlobalFoundries sang the same
tune of having IDM-like capabilities
00:20:27
when they first spun off from AMD.
Only to discover that in practice,
00:20:32
the two sides - locked in their individual
kingdoms - would struggle to come together.
00:20:38
Intel was no different. Many within Intel thought
that Intel Custom Foundry was a waste of money,
00:20:45
a side project distraction from
what was really generating revenue.
00:20:50
They were also worried what it
might mean for their own KPIs.
00:20:54
Intel’s Product people worried
that having a foundry might raise
00:20:58
outside concerns regarding Intel's own
internal evaluation of its products.
00:21:04
Imagine if Intel Custom Foundry taped out
another CPU and it turned out to be faster
00:21:09
than what Intel Product could do. Imagine
the egg on their faces. So they cried foul.
00:21:16
And then Intel's Technology group. They
sponsored and funded Intel Custom Foundry
00:21:21
in part because they thought it would showcase
the company's superior technology ... and maybe
00:21:26
put those pesky Intel Product guys back
in their place for complaining so much.
00:21:33
But when that did not immediately turn
out to be the case - when customers
00:21:38
had feedback on things they wanted changed -
the Technology Group's support fell out too.
00:21:45
This left the Intel Custom Foundry team
struggling to get their peers to change in
00:21:50
order to better serve the outside world. The
Technology people took a request from Intel
00:21:56
Product at least somewhat seriously because
it impacted revenue. But Foundry? No chance.
00:22:03
For example. Most larger fabless customers -
from their experiences working with TSMC and
00:22:09
other foundries - are used to getting some
level of insight into a node. But the Intel
00:22:15
Technology R&D team was extremely secretive
and often refused to divulge such details.
00:22:23
They also resisted making transistor changes
along the power-performance curve to accommodate
00:22:28
customers - seeing such things as distracting from
their main goal of advancing the leading edge.
00:22:36
To manage customer frustration, the
Intel Custom Foundry team regularly
00:22:41
brought in senior management like
the aforementioned Mark Bohr to
00:22:45
meet with such customers and use his
leverage to get R&D to cooperate.
00:22:51
## Altera
00:22:51
The FPGA market's two largest
players were Altera and Xilinx.
00:22:56
The two were fierce rivals. Altera had been a
long-time big customer of TSMC - one of their
00:23:03
top ten customers, actually. Meanwhile,
Xilinx used TSMC's rival foundry UMC.
00:23:10
But when UMC started to wither - a
rumored issue at its 65 nanometer
00:23:15
node delayed several Xilinx
products - Xilinx defected.
00:23:20
In February 2010, they announced that they had
added Samsung and TSMC as their new leading edge
00:23:26
partners. In an attempt to win the deal
over Samsung and their 40-nanometer node,
00:23:33
TSMC gave Xilinx access to its
then-leading edge 28-nanometer node.
00:23:39
This pissed off Altera, which now needed
a technical advantage they can leverage
00:23:44
against Xilinx in the high-end FPGA
space. This brought them to Intel.
00:23:51
In 2013, Altera agreed to make FPGAs
on Intel's 14-nanometer node - which
00:23:57
followed 22 nanometers. Some trailing edge
stuff would still be made at TSMC. CEO
00:24:04
John Daane cited Intel's process
lead as the reason for the move:
00:24:09
> "Intel’s 14-nm is a second
generation FinFET process,
00:24:14
while others are just starting
to implement their first"
00:24:18
At the start, TSMC did not take ICF all
that seriously. I heard second-hand of
00:24:24
a prominent former executive
there being quite dismissive.
00:24:28
So the Altera defection came as
a real shock for the Taiwanese.
00:24:33
Daniel Nenni recalled in a blog
post on his SemiWiki website:
00:24:37
> I was having coffee with a friend in TSMC Fab
12 when it was announced. If my memory serves
00:24:44
it was Dr. Morris Chang who made the announcement
and it honestly felt like parents were divorcing.
00:24:51
In the company's next earnings call on April
2013, Morris Chang said this about the loss:
00:24:58
> We have gained many customers in the
last few years, but I really hate to
00:25:02
lose even a part of an old one. We
want them all really. I regret it.
00:25:08
And because of this we have thoroughly
critiqued ourselves ... it's a lesson
00:25:13
to us and ... we'll try our very best not to
let similar kinds of things to happen again
00:25:20
Chang then addresses Intel Custom
Foundry's growing impact to TSMC itself.
00:25:26
> I still view Intel as a selective picker
among customers. As a foundry competitor
00:25:33
they will pick their targets and so on ... And
I don't view them as a general competitor ...
00:25:39
> But they are a very serious
competitor to our customers.
00:25:43
That really I would say applies even
greater pressure on us than Intel as
00:25:48
a direct foundry competitor. They are a
very serious competitor to our customers.
00:25:54
You see what he did there?
Morris is a very careful speaker.
00:25:58
He dismisses ICF as a general foundry competitor,
00:26:02
and then shifts the message again to his core
point: Intel competes with its own customers.
00:26:09
To me, it says everything: Intel Custom Foundry's
progress was indeed making an impact on TSMC.
00:26:18
## 14 Nanometers
00:26:18
Altera was without a doubt ICF’s biggest
win yet. And it was not the only one.
00:26:24
These early wins were promising
enough to Intel senior management
00:26:28
to slowly loosen the reins on the foundry's
sector restrictions. New business came in
00:26:34
from Panasonic and then LG, which contracted
ICF to make specific system-on-chips for them.
00:26:42
Switching to a new foundry always takes time. But
00:26:45
it took a few months for Altera to even
get started on their 14-nanometer design.
00:26:51
This was because another FPGA project -
the Arria 10, built on TSMC's 20-nanometer
00:26:58
process node - had sucked out design
resources. The Intel Custom Foundry
00:27:03
team had to dedicate many of their own design
resources to ready the design for the fab.
00:27:10
The design delay turned out to be fortunate
because Intel's 14-nanometer node was
00:27:15
a year late. The node's technology
definition had been too aggressive.
00:27:21
Not fatally so, but aggressive enough to force
00:27:25
additional experiment cycles and thus
extra time to mature into good yield.
00:27:31
At the end of 2012, Intel announced
that 14-nanometer should be ready by
00:27:37
2013. It did not reach high
volume production until mid
00:27:40
to late 2014 - entirely missing
the two-year Moore's Law cadence.
00:27:46
The year-long delay also gave foundries like
TSMC and GlobalFoundries - whose 14-nanometer
00:27:53
node was being done in collaboration with
Samsung - some time to close the gap.
00:27:59
But like I said, not a fatal error. Just
took some extra time to get out. Once mature,
00:28:06
the node was fine - a second-generation FinFET
process denser than TSMC’s first-gen FinFET
00:28:13
16-nanometer. The main issue was the
impact on the node that came after it.
00:28:19
## 10 Nanometers
00:28:19
In 2013, Intel and Apple met once more for
discussions on a possible foundry deal.
00:28:27
Apple was again in the market. Perhaps with the
00:28:30
intention to dual-source their next
iPhone chip, the A9. Rumors abounded
00:28:35
in the financial media that Intel was
targeting Apple as their big customer.
00:28:42
They were probably just kicking the tires.
Intel Custom Foundry had gotten a few
00:28:47
customers by now and was learning from
them, but it was still a bit too early.
00:28:53
The A9 second-source deal
eventually went to Samsung.
00:28:58
Critically in those meetings, Apple
indicated that it wasn't interested
00:29:02
in the Intel 14-nanometer node, which in 2013
was then going through its slow maturation.
00:29:09
But rather the node after that: 10-nanometer,
00:29:12
originally due in 2016. Had it ramped on
time, it would have done so 2 years ahead
00:29:18
of TSMC's N7 node - which it was quite
similar to. But famously, it didn't.
00:29:26
What went wrong with 10-nanometer?
Intel has never said publicly,
00:29:31
and the story is probably best reserved
for another day. This is what I understand.
00:29:37
Intel develops its process nodes
with two teams working in parallel.
00:29:42
One team works on what would be 14 nanometer
gen+1 while the other works on 10, gen+2.
00:29:48
So by the time senior management realized that
14-nanometer had a serious problem - perhaps
00:29:56
some time in 2014 - the definitions for
10-nanometer had already been solidified.
00:30:03
And 10-nanometer was a full 50%
area shrink - with a 2.7x increase
00:30:10
in transistor density - on top of what
had already been a very aggressive 14.
00:30:17
SemiAnalysis's Dylan Patel added in a
2022 report that Intel’s use of cobalt
00:30:23
in the vias and interconnects
played a major factor too. The
00:30:28
deposition tools for that were not ready
yet, but Intel blazed ahead despite warnings.
00:30:34
And meanwhile, EUV - which I remind you that
Intel funded and always believed in - was
00:30:40
not ready in 2014 or even 2015. It would not
be inserted into a process node until 2019,
00:30:46
when TSMC did it for their N7+ node. Even
if it was available earlier than that,
00:30:53
it wouldn’t have been trivial
to redesign the node to add EUV.
00:30:58
Long story short, it was too much.
A step of hubris that took them off
00:31:03
the cliff. But Intel's Product group had
already taken the assumptions Technology
00:31:08
group made for what 10-nanometer
was promised to offer and developed
00:31:13
products with them. Only to find out
that those promises cannot be kept.
00:31:19
It was like building a 50-story
skyscraper only to find out on the
00:31:23
20th floor that the foundations can't
support the whole building. You can't
00:31:28
just tweak this. You must tear it all
down and start again - a 2-3 year process.
00:31:34
Once the news about 10-nanometer's delays
started reaching outside parties in 2015,
00:31:40
the backlog of available business for the
foundry rapidly began to dry up. By the time
00:31:46
10-nanometer fabbed Intel chips finally started
to ship in 2019, the foundry no longer existed.
00:31:53
## The Altera Acquisition
00:31:53
And then in June 2015, Intel
announced that it would buy its
00:31:57
biggest foundry customer Altera for $16.7 billion.
00:32:03
CEO Brian Krzanich’s galaxy brain thinking
behind buying Altera was that they can
00:32:09
integrate FPGAs into their data center CPUs. Such
an arrangement could let people offload certain
00:32:16
workloads to those FPGAs - which meant that they
wouldn’t need to buy a GPU to do those workloads.
00:32:24
A relatively sound idea. And
indeed after AMD acquired Xilinx,
00:32:29
there have come out some interesting
SoCs with similar arrangements.
00:32:33
But for Intel, the work of merging these
two previously discrete chips - CPUs and
00:32:39
FPGAs - turned out to be too much.
The strategic notion soon dissolved
00:32:45
and the mooted product never made it to market.
00:32:49
From the perspective of Intel Custom Foundry,
00:32:52
the Altera acquisition was a disaster.
Assimilating Intel’s marquee external
00:32:58
foundry customer sends a bad message to
all of the foundry’s potential customers.
00:33:05
It implied to those potential customers -
guys like Qualcomm and such who were only
00:33:10
starting to open up to an Intel foundry
- that using ICF might mean Intel later
00:33:16
trying to buy them too. Not to mention all
the competition issues Morris talked about.
00:33:23
## End
00:33:23
Alone, the Altera buy was a bad
move. Combined with 10-nanometer's
00:33:28
ongoing failure, it was crippling.
00:33:31
In 2015, a number of major retirements,
including that of General Manager Sunit
00:33:36
Rikhi - who had run Intel Custom Foundry for
nearly eight years - delivered the final blow.
00:33:43
He was replaced by someone more amenable to
Technology's perspectives. In other words, making
00:33:49
the people in the Technology look good, rather
than advocating for the actual foundry customers.
00:33:56
So when the customers inevitably started to get
angry with their inability to get through Foundry,
00:34:02
senior management arranged for those
customers to work directly with the
00:34:06
Technology people. In other words, Intel
Custom Foundry got cut out of the loop.
00:34:12
With the writing on the wall now obvious,
the division - which once employed over
00:34:17
1,100 people around the world - quietly
began to dissolve. By 2018 it was no more.
00:34:25
## Conclusion
Circumstances have greatly changed between
00:34:27
Intel Custom Foundry in 2008 and
Intel Foundry Services in 2021.
00:34:34
It seems like Pat Gelsinger's approach
upon returning to Intel was to rebuild from
00:34:39
scratch. I think that was a mistake. Basically
throwing out the years of progress made by ICF.
00:34:45
Intel Custom Foundry was no joke. Over eight
years, they built up a book of 10+ customers
00:34:51
and a billion plus dollars of contracted
revenue. If TSMC considered them a real
00:34:56
competitive threat, then we should take
their lessons and progress seriously too.
00:35:01
So I will share a few takeaways of my own. First,
00:35:05
the foundry service needs to have organizational
support from across the whole organization. It
00:35:11
should not be left awkwardly stuck between
Technology and Product. At the very least,
00:35:17
it should have the autonomy
to modify its own technology.
00:35:22
Second, Intel's most significant technological
pull is a working leading-edge node. That is
00:35:29
what gets people in the door. Intel
Custom Foundry could lean on Intel's
00:35:34
multi-year process node lead on TSMC,
Samsung and the rest of the industry.
00:35:40
That lead no longer exists, and retaking
it will be extremely difficult. Get 18A
00:35:47
and 14A working right as a node for
foundry customers - meaning more
00:35:52
than just for making CPUs. That will not be easy.
00:35:57
Third, the only thing that closes the book
is bankruptcy. Prepare for the long run. Let
00:36:03
investors know that this is a very long
term story. TSMC was not built in four,
00:36:09
five or even ten years. Competing
with them will be a ten year endeavor.
00:36:15
Fourth, a little humility. Intel Custom
Foundry felt like an unusually humble part
00:36:21
of the company. Perhaps because it was a ragtag
side project that had to serve outside customers.
00:36:29
Perhaps because they hired over half of their
workforce out of the fabless-foundry ecosystem.
00:36:34
Perhaps because they closely studied Morris Chang
whenever he talked about the foundry business.
00:36:41
Bring that humility to the rest of the company.
