Intel's CEO Shares His Plan To Win The AI Chip War — With Pat Gelsinger

Channel: Alex Kantrowitz

Published at: 2023-12-14

YouTube video id: d07wy5AK72E

Source: https://www.youtube.com/watch?v=d07wy5AK72E

we're here with Intel CEO Pat gelsinger
so Pat just to um you know go back to
basics can you also tell us what the
difference is between a CPU and a
GPU well CPU is uh this uh general
purpose compute device you know if you
think about it it runs everything you
know it could run a web service you know
it could run you know an application uh
you know it runs Zoom right it runs uh
you know every you know your uh recipe
program it runs everything and all this
software so it's called a general
purpose CPU a GPU onstead is really
built for a very specific class of
workloads and generally those have been
called throughput workloads so it does
lots of floating Point processing and
Matrix operations you know and so it's
very dedicated for things like Graphics
Matrix and now right is worked out to be
uniquely good at things like Ai and so
it's a very specific set of apps that
have become very important and why is
the GPU so good for
AI well AI tends to have very specific
operations that it's doing and that all
it's doing is compute compute compute
compute right whereas a CPU is sort of
saying oh if then and jump over here run
this application so it's very specific
you know and uh largely emerged from the
whole graphic space where all it's doing
is vector graphics you know
rasterization of V very narrow set of
compute workloads so it's designed
basically you can sort of think about it
you know as your general purpose sedan
that's sort of the CPU and the GPU all
it does it gets on the F1 track and all
it does is go fast on very specific
workloads interesting and obviously it
worked really well for gaming and that's
kind of was nvidia's specialty um how
Nvidia just ended up running away with
the game here was that they built this
GPU for gaming and it ended up being
they kind of lucked into it being good
for
AI yeah and it very much is that way and
Jensen and I you know we've known each
other for 35 years you know this general
purpose workload and we always are
adding more capabilities to the CPU but
over here it was always just go really
fast for graphics and then you got
really lucky that the AI workload sort
of looked a lot like the graphics
workload so as I joke with Jensen I said
you know you just were really true to
that mission of throughput computing and
graphics and then you got lucky on AI
and he said no no no Pat I got really
lucky on AI and but now it's interesting
because you have Nvidia okay they're the
clear leader but every single day it
seems like another company is announcing
their own GPU I know that Intel's had
its own Pono chip uh in development but
also you have accelerators right which
is basically ways that um companies like
Amazon and Google will modify chips in
order to be able to run AI workloads in
fact Google just trained it apparently
its entire Gemini model on its own
accelerator not needing envidia at all
so um just take us into that race a
little bit and does it seem like I mean
nvidia's lead for a long time has seemed
steep but it seems like less so now yeah
and what we expect you know and when we
think about AI workloads you could think
about training and inferencing and you
can think about that like a weather
model how many people create the weather
model that's training versus how many
people use the weather modeling oh
that's lots of people you know local
forecasts you know scheduling route maps
all that kind of stuff use weather
models for the training application you
now have what Nvidia does accelerators
like what we're doing with gudy but also
then the TPU from Google you know the
tranium from Amazon what Microsoft just
announced with Maya what AMD announced
because the soft Ware there right you
know is very specific in this class so
if I can run that python code as it's
called you know the key language in this
case then H I'm going to go compete at
that and those machines are getting big
and fast so a lot of people are pursuing
that but in the inferencing then you
sort of say Hey how do I mainstream that
application and that's an area that
actually is just another workload and
we're going to do a lot of inferencing
on our standard CPUs or the Zeon product
line as well you know so we expect that
there's going to be a lot of competition
in the AI space and finally for Intel
you know we're also going to be a
Foundry we're going to be the
manufacturer for many of those chips as
well so we want to be the manufacturer
for NVIDIA for AMD for Google for Amazon
we want to be their manufacturing
partner even if we're not using our
design chips yeah I'm definitely going
to talk to you about the manufacturing
in a bit but let's let's just stick with
the design here so T I mean it does seem
like what you're saying though is that
this landscape is going to be a lot more
competitive than it has been previously
I mean you have a company like Nvidia
that added what 600 billion dollars to
its market cap in one year like there
are going to be others that are going to
be trying to get in does that sound
right to you absolutely you know and I
sort of put them into two classes uh
Alex there's going to be those that
build their own you know and that's what
you see Amazon Microsoft Google are
doing they're going to say hey I'm going
to own this and do this myself and then
there's going to be the general
providers in the marketplace and that's
going to be Intel AMD Nvidia I think
will be the three big ones in that space
so there's going to be do it ourselves
we're going to own the full stack of
hardware and software you know which is
the big cloud guys and then there are
those who say hey I'm going to sell my
chips to everybody and I expect those
would be the big three okay so wait
which which part are you going to
compete in then both or both yeah you
know we're going to you know because I
want to be a Foundry to what Amazon does
what Microsoft does what Google does and
I'm going to sell my chips and I'm going
to sell my chips to the Enterprise
customers who want to do this with their
data on premise as well as to the big
cloud guys as well and today you know
biggest customers for NVIDIA today are
probably Microsoft who's putting up
their big farms but they're saying hey
no I'm going to build my own ship I'm
going to build Maya so that I do just
like what Google is doing with their own
TPU as well I want to own that margin
and I'm going to do it on my
architecture as well so Intel I think
uniquely has two bytes of the Apple here
to pursue right it's interesting to hear
you talk about how the AI inferencing or
the running of these models is going to
be done on CPU chips I mean it's you
just kind of explain the architecture
and the use of a
CPU um and it seems like even still a
GPU would be better for AI functionality
but are you saying that actually the CPU
will will be fine it seems like it
different yeah and what's going to
happen is AI going to get added to every
application right so everybody's going
to start saying how do I bring AI into
my apps so imagine I'm running sap right
I'm going to do a lot of my normal sap
and all of that runs my CPUs today but
then I want to add some inferencing
capabilities into my sap
environment you know we believe and
we're adding these Matrix functions onto
our CPUs so we're extending the
workloads of our general purpose CPUs to
do a better job at Ai and so if the
workload is just running inferencing I
would have probably run better on a GPU
but if it's running a lot of things
we're going to make it just run great on
the CPU we're finding great interest
from customers to do that today you know
and for my standard CPUs today in the
data center you know we see about a
third of the purchases are being based
for AI workload so we're already seeing
that characteristic emerge quite
strongly today so explain this one to me
then because Intel has its own ponio
trip which is apparently you know a GPU
trying to do some of this other stuff um
how but how are you you know how are you
going to balance that with running AI on
your CPU chips yeah you know so some you
know if the workload is running on the
CPU it's just going to stay running on
the CPU right right but then we're also
going to for these environments where
all that you're doing is running AI then
we're going to offer our accelerators as
well and ponio and gouty we're bringing
those together into a single product
line you know going forward because
we're going to compete in that space as
well you know we're going to be building
you know components whether they're gpus
or CPUs to capture as much of the market
as possible okay so I know about Pono
it's a it's a GPU uh gouty CPU that runs
AI functions it's a you it's called an
accelerator really is designed for these
unique Matrix functions that are seen in
AI workloads so just give us an honest
assessment of where Intel stands today
compared to Nvidia like are you getting
close in terms of like the volume or
where where is where how do you compare
right now to them yeah no Nvidia is the
runaway market share leader today right
we give them credit for that um you know
we're now seeing our growth rate and
quarter to quarter we approximately
doubl the growth rate you know but we're
still small marker share you know today
but we're Rising quickly because
customers are looking for Alternatives
you know today because this is demand
and they also want better price and
different you know features so our
business here is growing very rapidly
you know but it's from a much smaller uh
base but we are now winning some of the
performance benchmarks so all of a
sudden customers are saying huh you know
they're showing up winning some of the
benchmarks you know I want an
alternative Nvidia is short on Supply
and I'm getting much better TCO from
Intel you know hey let's go start
testing this and we're getting a lot of
interest in our value proposition yeah
in some ways this uh Supply crunch you
know really can end up working in your
favor because people absolutely
something yeah and it's both a supply
crunch for the supply chain right and
some of our packaging and wafer
capabilities people are saying hey can
you help us people who might not have
considered Intel as a Foundry supplier
or all of a sudden saying hey can you
manufacturer right even people we
compete with on the product side saying
hey can I be your manufacturer but you
know if your chips are working today and
I can build my next uh AI Farm you know
using you a lot of interest there as
well so unquestionably the supply crunch
is working for us okay great so we've
talked now about design I think we've
done enough on that we can talk now
about manufacturing um people are
talking a lot about tsmc Taiwan
semiconductor and the fact that a like I
think during Co a lot of people realized
that it was a strategic liability to
have core manufacturing for the US be
done offshore now it's going to take a
while for us to get to that point but
there are there's legislation there's
funding the chips act that's going to
give companies like yours an opportunity
to start to build some serious Foundry
capabilities in the US one one question
to you to start Intel has tried a couple
times to build Foundry capabilities for
others I think twice before and it
hasn't worked out it's very different to
basically manufacture your own chips
than to manufacture other people's chips
it takes you know off the-shelf
technology process that stuff so what
gives you confidence that this time is
going to be different for
Intel well several several things that
we're doing differently this time and
the first you know I'll say on the first
attempts they were Hobbies right it was
sort of like ah let's go try it you know
we really weren't taking it seriously as
a company this time I have bet the
future of the company that we are
GNA pardon why why not like why was this
a hobby and then we can talk a little
bit more about why this is so important
now and you know fundamentally the Intel
business was going really well before
and this Foundry business model was
still pretty nent right so it was sort
of like ah that model's emerging tsmc is
doing pretty good you let's go try it in
a few places but it wasn't taken deeply
intentional as a core part of the
strategy and not very profitable from
what I understand is that this is kind
of like the least profitable part of the
whole process the manufactur well hey
tsmc has gotten pretty good profits now
they figured it out yeah yeah they have
figured out how to make good profits
here so this is now a very profitable
business um you know almost as
profitable as the chip business itself
in many respects you know secondly the
ecosystem has become much more mature
and Intel before was very proprietary so
if you wanted to use my Foundry you had
to be proprietary on me well now we have
standardized our processes like the rest
of the industry so it's much easier to
use us uh as a Foundry you know third
I'd say everybody postco realizes oh my
gosh you know we desperately need a
western Foundry at scale you know this
is super important and we're finding
that uh interest from customers because
they see their supply chains as very
fragile you know and they have become so
dependent on one company one Island one
port you there's a lot of Industry
interest as well as Government interest
to build up us as a worldclass Founder
so we're well in the way and you know
it's become a key piece of the strategy
that I've laid out how would you assess
the geopolitical risk to Taiwan I mean
you said one country one Island
obviously that's Taiwan we've seen
already you know R Russia invade Ukraine
that put a lot of people's antennas up
this might happen in Taiwan what's your
perspective on how serious we should
take this well you know this is one
where it's going to take years to
rebuild these Supply chains right it
took us three it took us three decades
to have our supply chains move to Asia
you know what we've said is hey we've
gone from 8020 to 2080 in Asia wow by
the end of the decade so you know seven
eight years I think we can get close to
5050 by the end of the decade and if we
accomplish that right over a seven or
eight year period I think the world is
going to sleep much better at night you
know because hey this is you know a
blockade of the Taiwan Straits and all
of sudden the island Browns out in 30
days you know this becomes very
precarious right and we can fix it and
you know not just the economic but the
National Security benefits of this are
huge why is it going to take so long it
takes five years to build the new Fab
really right so you know what I've
described you know we're a couple of
years underway on this but you know if
we accomplish this by the end of the
decade as we've laid out you know that
is spectacular uh you know for a Layman
why does it take five so long to build
one of these factories well you know
these factories you know first they are
just amazing right and I you know I just
love people to come and visit the
factories these are the largest
construction projects on Earth today
building the smallest things that have
ever been built on Earth it really is
amazing the Precision manufacturing the
chemicals and so on you know it takes us
about 5 years to have one of these
factories up uh up and running on a
Leading Edge process technology you know
the total project is about $30 billion
right to build one of these Factory
complexes you know so it's an enormous
capital investment right and uh you know
I end up with like 7,000 trades people
you know to work for almost four years
to build one of these locations it truly
is right A U manufacturing Marvel
building the most advanced uh science
that's ever been done on earth now Pat
I've spoken with people in the early
days who were there at Texas Instruments
and were part of this offshoring of
chips to Taiwan and basically what they
said was um it was so it was the least
the least profitable part of the whole
process and they just didn't care they
didn't think about it strategically um
I'm curious if you think that it was a
mistake to let so much
manufacturing uh leave the US and then
also like it you know it it seems like a
good hedge to have a plant here but just
in terms of um a profitable business
line and a good business like it's going
to be very expensive to run in the US
don't you think curious to have you wait
for those yeah two things one is yes it
was a mistake and I think the world
realized how big a mistake it was in the
middle of covid you know that we allowed
our supply chains to become so fragile
and as I say what aspect of your life is
it more digital everything's more
digital going forward right and
everything digital needs semiconductors
you know where oil reserves are has
defined geopolitics for 50 years where
technology and Fabs are for the future
is more important you know so with that
in mind yes it was a mistake and now you
know with the chips act you know we've
taken the most significant industrial
policy legislation since World War II to
correct that error now part of the chips
Act was to level the playing field right
it was to close that economic uh Gap
that we see with IIA today so it's
designed to right bring those back on
parody so that the Investments that
we're making you know are competitive
with those of Asia in the world and I
believe as the decade goes forward we
rebuild the ecosystems we can
systemically and structurally be closing
those gaps in addition to the aid that's
been needed to immediately fix some of
that huge economic gaps that we have
today okay so there's there's some
subsidy there that sort of makes the
economics work yes
and so um we are big technology podcast
so I have to ask you about apple right
they've started designing their own
chips and they're doing pretty good job
of it um so I'm curious from from your
perspective as a chip manufacturer what
do you think about the position that
Apple's in today and I mean clearly the
performance is quite good on the chips
sa design is that something that you I
mean obviously not every company is
going to do it but uh what how would you
assess their effort well you know they
used to use Intel chips and when Intel
stumbled Apple stepped in and did their
own chips so ultimately my objective is
build better chips that they want to use
our chips versus doing it themselves
right but it also shows that this idea
of The Foundry ecosystem has become very
mature you know that a company like them
could step in and build very good chips
and remember they build chips for their
applications so they highly optimize
them just for the mac and for the iPhone
as well they'll do everything like you
know the Intel chips do across many
different markets they optimize them
solely for their applications and
products and they've done a super good
job and I'd say over time hey I'd hope
to give them a better product that they
could use my chips again but I still
want to be a manufacturer for them even
if they choose to keep designing their
own chips uh going forward I want to
become a Foundry for them just like they
use tsmc today and now have you talked
to them about that or are we still seven
years away from that being a reality of
course I've talked to them I've talked
to everybody in the industry you know
qualcom Nvidia AMD Google uh Apple
broadcom Etc I want them all running on
our factories because that is better for
them to have our technology better for
them to have more resilient Supply
chains and I'm going to make it a good
business proposition for them as well
okay you have a big AI event uh coming
up this week can you tell us a little
bit about what people can expect there
yeah you know we call it AI everywhere
right and in this sense you know AI
isn't just going to be for these big
high-end cloud and training environments
but how do we make it available across
every PC across every Edge device as
well as our chips for the data center
and we'll be introducing you know two
new Chips one is our main CPU right Zeon
Gen 5 that we have further enhancements
for the AI workload and we'll be
introducing core Ultra you know which is
uh for the client to put AI capabilities
directly into your
DC okay sounds good anything you think I
missed or anything else we should know
yeah and I just say for this AI
everywhere thing you know Intel's
showing up here saying about we're the
volume provider you know and much like
the Cino event was you know 20 plus
years ago that made Wi-Fi and access
points and every coffee shop had to have
you know Wi-Fi service you know it just
changed the not just the PC but the
entire way that people use used
Computing we see this AI PC having that
same kind of shift where all of a sudden
maybe I don't type to my computer
anymore I just talk to it in the future
it knows when I'm there it translate
languages you know it has new insights
and capabilities it becomes my personal
bot you know we just see it ushering in
a new generation and Andy Grove one of
the founders of Intel described the PC
as the ultimate darwinian device and we
think we're about to go through a you
know major evolutionary step in the life
of the PC and that begins today one
quick followup what is an AI computer
you mentioned an AI computer yeah you
know think about your PC today that now
has builtin AI capabilities where all of
a sudden instead of having to go to the
cloud to get a model all of a sudden my
PC is able to record translate summarize
you know be Vision tracking in Flight
you know where you could be speaking in
uh right Korean and I could be hearing
you in English and vice versa in real
time I'd look away from the screen right
and it would summarize the conversation
when I'm outside of the meeting you know
before the call you know before my next
call with you it would say Hey you know
on uh this date in December you spoke to
Alex and remember this is his birthday
coming up and don't forget to remind him
to bring flowers home there you know all
of those kind of things would be part of
that AIP PC experience as well as we see
it shifting and changing the form
factors as well cool stuff Pat thank you
so much for joining hope to keep up this
conversation as we go forward look
forward to it as well thank you so much
all right everybody thanks for listening
we'll see you next time on big
technology podcast