Intel is Gunning for NVIDIA

in blasted on stage at its Vision 2024

event with some news Papa's little baby

here man look at this big boy

yeah and you know I I sort of looked at

the prior decade as sort of

boring you know we made PCI gen a little

bit faster we made DDR the next

increment you we kicked up the thermal

envelope a little you know we

added a few more ques and we said okay

it's a good chip boring

definitely a different approach from

Nvidia let's see if they're excited

about anything gouty

[Applause]

[Music]

[Applause]

[Music]

3 you know huge advancements okay still

a different approach but that's

better I am the sound effect but what we

really care about is the marketing music

behind it you might remember nvidia's

brainwashing music

previously this the part

this and we're concerned about if intel

has an answer to that in the township of

quo where many folks lacked Vision Chief

geek patch and sales guy shell were

shopkeepers on a mission they thought

what quo needed was a spark of

inspiration so they whipped up a a batch

of woohoo chips they knew would cause a

sensation a business exploded with lions

out the door what the

Intel just got done with its Vision

2024 keynote event and you can guess

what they were talking about ai ai ai a

a ai ai ai ai ai ai ai ai ai ai but

Intel is absolutely taking this

seriously and actually so is Google both

companies announced their plans for AI

EX accelerators and both of them appear

to be aiming at Nvidia to try and knock

them down a few pegs if not maybe later

eventually try and take the crown Nvidia

has been asserting absolute and total

dominance over the AI market for some

time now and has been expanding more

into the software side of things which

is probably caused for concern for its

competitors because they start to create

this ecosystem where if you want AI

anything you're attached somehow to

Nvidia and in CEO Pat Galer took some

direct shots in the vision 2024 keynote

industry is quickly moving away from

proprietary Cuda models and gouty is the

only Benchmark alternative to the Nvidia

h100 for training llms and it's actually

not the first time that Galer made

comments about Cuda having maybe a

limited lifespan back in around 2008 or

so before he was CEO at Intel he had a

comment about Cuda being an interesting

footnote in the annals of of History uh

so and looking forward to seeing how

Google autot transcribes that word we'll

find out I'm not going to change it so

whatever it comes up with we'll leave it

there maybe it'll be like when nvidia's

Auto transcript from the keynote said to

connect anwers to the Omniverse digital

twin fun fact they changed it after our

video I was sad before that this video

is brought to you by lean Lee and the

o1d Evo RGB case the o1d Evo RGB is an

updated entry to the famed lineup

retaining heavy support for fan mounts

Drive Mount locations and flexibility on

component mounting such as two options

to the power supply the o1d Evo rgb's

dual chamber approach aims to maximize

cable storage on the backside to

streamline Cable Management coupling

this with a unique vertical GPU Mount to

showcase the most expensive part in most

systems learn more at the link in the

description below regardless of whether

you're using any of this Hardware today

or you're even in that part of the

industry you will probably be using it

in some form

at some point that's either going to be

directly as all of this stuff comes down

to Consumer it does tend to go that way

just less powerful we see npus and the

CPUs today which are AI accelerators

built into Hardware that are meant for

accelerating these so-called AI tasks uh

and even if you're not going to use that

stuff probably consumption via services

online is the next likely route so we

haven't seen Intel AMD and Nvidia be

this amped up and aggressively

competitive in quite some time it's

clear that they see a huge money

opportunity here and they're all going

for a land grab and what they see as a

new world actually there's some new

markets here too not just software and

selling those solutions to companies but

also there's some talk of manufacturing

targeting things that would be

traditionally seen as not a consumer of

this Ultra high-end type of uh

accelerator CPU GPU whatever it may be

so anyway Intel's event this year

debuted its gudy 3 accelerator and also

some new CPU so those are the primary to

announcements they also talked about

software they talked about ethernet as

an alternative to EnV link and pushing

some development in the ethernet space

and Intel directly targeted nvidia's

h100 as a head-to-head comparison for

performance on the hardware side so

that's kind of big news because the h100

has been backordered into Oblivion

there's a big opportunity there for any

other Hardware vendor to strike for

companies that either can't wait online

because they need it now or companies

that aren't big enough to get a spot

high up enough in the line

uh so it will now have Blackwell to

compete with as well with its gouty 3

parts and Intel seems to be aware we'll

talk about this later in this video uh

that they are not going to be the number

one immediately but they are still

trying to line up as competition Google

and AMD are ramping up their efforts

this year as well in the same space and

we expect to see more from AMD at compy

Tex this year now for G3 it was joined

by Zeon 6 announcements the Zeon 6

it lacks a prime number you might have

noticed that from the six and that to us

is concerning because Intel has threes

and fives and sevens if intel has

depleted its prime number

Supply it's

over I guess there's the nine but the

vision 2024 keynote with CEO Pat Galer

wasn't as information dense as nvidia's

GTC keynote but there was still a lot of

good detail for us to cover along with

our Impressions on Intel's General

positioning strategy in the exploding AI

compute market and because gaming and

crypto have gotten too boring for them

to talk about boring AI is the newest

craze we'll see to what extent it sticks

around in this capacity but that's what

everyone's talking about now and at

least some of these things can be

deployed for non deep learning or

machine learning use cases as well but

everybody is gunning for NVIDIA right

now currently they have a combination of

leading hardware and software Solutions

but also a massive head start on some

in-house designed Solutions like IO

nvlink for example was huge for NVIDIA

in this year's Blackwell keynote and the

company also went multi-chip with its

approach to Silicon Intel is countering

with a push for more open Solutions

including relying on ethernet instead of

NV link and it's taking an angle that's

actually somewhat familiar to amd's

approach to marketing gpus and

developing them which is that both

companies are promoting a level of

openness and focusing marketing on

non-proprietary Solutions and that is

clearly aimed at Nvidia so or at least

we think so anyway it seems like it it

may be a sort of enemy of my enemy type

of situation for these companies uh and

one of the biggest advancements for

NVIDIA again based on our exposure to

the side of the market has been the the

multiple uh silicon approach to

Blackwell and also the EnV link solution

we talked about that in our Nvidia

coverage giving some opinions and uh

some of the detail about the news from

that event but Jensen's modern

definition of a GPU has changed somebody

used to say you know you guys make gpus

and we do but this is what a GPU looks

like to me when somebody says GPU I see

this and that's going to be a

consideration for all of its competitors

because regardless of what you call it

it is actually uh not just the literal

GPU silicon anymore that matters Jensen

Juan was absolutely right in that sense

because these devices need to talk to

each other at speeds that don't

hamstring them significantly and so

that's where these IO Solutions and

interconnects and things like that come

in and that's something that Intel

talked about in its Vision keynote other

critical things that customers are

saying is boy how do I Stitch these

together well there's some proprietary

Solutions available we also see that the

ultra ethernet Consortium and the work

that we're driving is standing up to

fill this scale up and scale out

networking domain and through UE Intel

is

revolutionizing building on ethernet

networking for AI fabrics for the future

and we'll be introducing an array of AI

optimized ethernet Solutions this lineup

will include Cutting Edge AI Nick cards

that will be delivering standard Nick

Solutions it will also include AI

chiplets that will enable for our

customers and partners in Intel Foundry

we're not experts in this space of AI we

talk about gaming Hardware this stuff's

cool to us it's fun to cover because

it's interesting uh technologically it's

fun to to study and read about and some

of it does come to Consumer but as far

as we understand this Market again

coming from our perspective of

non-experts in this so-called AI era uh

as far as we understand it it seems to

me that the approach to ethernet or some

open standard here would be critical for

all these non- Nvidia companies to in

some form collaborate on because it may

need to be a strength and numbers effort

against the Behemoth that is NVIDIA now

and that if they all go off and develop

their own standards and proprietary

Solutions then uh they they'll lack that

focused effort in numbers but ultimately

what we took away from Intel's key note

was that if you are a company with big

compute or AI needs you go to Intel for

the same reason that you go to Waffle

House at 3:00 a.m. they're open and

they're cheap it's not an insult by the

way that if I weren't a fan of 3: a.m.

Waffle

House why would I

inex why would I inexplicably have this

sign that I probably shouldn't have the

answer is there's a security company

founded by X Waffle House staff that's

how I knew they were good at security

moving on CEO Pat Galino ped how

beneficial open platforms like pytorch

are and repeated stated the value

proposition and lower total cost of

ownership or TCO of its solution versus

the proprietary and expense of

competition the competition again of

course being Nvidia everyone already

knows which company is number one right

now in the compute Market it's not even

a question you pull up Nvidia stock

numbers and it becomes pretty clear what

is uncertain though is which company is

going to be number two in the longer

term because everybody is trying to get

into this even companies that are aren't

traditional silicon manufacturers like

AMD Intel and Nvidia we don't think

Intel or AMD would admit what's going on

here is a fight for second place but

that's kind of what's happening right

now you have to get there before

fighting for first but that doesn't mean

Intel isn't trying to win the top spot

according to Galer and Intel the new

gouty 3 accelerator chip can do llm

related generative AI tasks at what they

say in at least some scenarios is 40 to

50% faster than nvidia's h100

we don't have a way to validate that

that's not the kind of testing we do but

that's the claim and Intel is also

trying to reach manufacturing clients uh

which this we thought was pretty

interesting because this is a trend

we've seen in Nvidia as well where uh

Nvidia kind of kicked this off but

targeting places like factories

warehouses uh facilities that

traditionally you might not associate

with these super intensive computational

needs maybe Amazon kind of a side for

their logistical efforts and Company

like that but they are both targeting

this sector which opens up probably some

new uh client base that they might not

have had before and we think that's due

to a recognition that uh in this

instance you could start using things

like AI deep learning machine learning

whatever to either fuel things like

robotics so like Amazon does with uh

using robots in the workforce and some

level of AI and automation mixed

together to pick in place or pack the

right products and it can definitely

also be used for more advanced Logistics

like routing for all of the various

delivery vehicles uh and the various

modes of transport involved in between

point A and B this is a a quote Intel

was direct and its Outreach to that

sector here's what they said normally

right you know remember manufacturing

Supply chains those kind of things you

know the developers get the cool PCS and

then the mediocre ones you know go to

the white collar workers and then the

crap ones okay we'll give those to the

manufactur facturing

no call your head of it it is time for a

pcai refresh because use cases like this

show extraordinary value for Fleet

upgrades now there's more to all this

than just Hardware of course so they

also talked about software but we'll

start with the gouty 3 accelerator this

was Intel's biggest announcement for the

event as far as uh we're concerned

anyway and technically super micro

accidentally leaked it early on stage we

also showing a real G 3 system the

world's first G 3 system new I haven't

even announced it

yet we're a little bit ahead okay I

encourage everybody to go to our boost

just check it out genuinely a really

good save though the package has 16

total silicon tiles or chiplets the two

largest in the middle contain 64 tensor

cores eight Matrix math engines a media

engine and 96 MB of SRAM cache there's

also IO which is made up of 24 200 GB

ethernet connections along the left and

16 pcie Gen 5 Lanes on the right the top

and bottom have links for the hbm2e

memory which are the uh eight

mediumsized tiles that you see along the

edges there the total memory capacity

for gudy 3 is 128 GB with 3.7 terab per

second of bandwidth and it'll be

available as a pcie addin card with a

600 watt TDP so that's again the big

difference from Nvidia not using nvlink

of course for performance a single gouty

3 is listed at 1835 teraflops of fp8

performance scaling up to Universal

baseboard which as we understand it is

what would go inside a compute server

with eight accelerators bring that up to

14.6 pedop flops of fp8 it also have

over 1 terab of hbm2e with 29.6

terabytes per second of bandwidth and

9.6 tabes per second of bidirectional

networking just a quick note here as

well it appears that Galaxy might have

misspoken on stage he said that the

memory is hbm 3E all of the slides and

the white paper though indicate hbm2 e

and Nvidia made a a similar

transpositional mistake with hbm in its

presentation so maybe it's just an hbm

thing and Intel is excited about it man

look at this big boy

yeah I always bring Big Brother along

with me nothing says giant mega

Corporation hosts AI keynote quite like

the phrase Big Brother most of the

performance metrics that Intel provided

were listed as comparisons to nvidia's

h100 or just sort of Standalone Intel

claims gudy 3 is 40% faster in time to

train and 50% more power efficient than

nvidia's h100 in large language models

that's in a specific workload though as

this chart clearly shows that it's

actually a give and take between gudy 3

and the h100 marked as the Nvidia 100

here to make sure at least one person

gets confused and thinks that means

means black well one more thing on gudy

after its stage announcement Intel also

posted its gudy 3 AI accelerator white

paper with significantly more detail a

lot of this goes beyond the scope of

what we typically cover but there's a

couple interesting diagrams Worth

showing this block diagram shows the

multi- diey approach that Intel is

taking in this diagram you can see two

dieses separated horizontally across the

center and flanking hbm 2E modules on

the sides the orientation is different

from the one shown in some of the slides

previously the whole white pap paper

gives a relatively deep detailing of the

capabilities including a ton of

first-party numbers against the h100 and

also additional diagrams showing PCB

layout that one's kind of cool uh they

also show workflow canonical pipelines

and detail the capabilities of the

hardware we thought this diagram of

their scaleout cluster architectures was

cool as well acting like a sort of map

between uh clusters here but we'll leave

the white paper here since some of the

details in the document go beyond our

scope and knowledge we did want to bring

it to everyone's attention though just

in case interested in this type of

Technology briefing it's out there

available publicly Intel also announced

it's Zeon 6 CPU so we'll get into those

now uh the first one will be Sierra

Forest or Papa's little baby why what

why do all the executives do this Papa's

here it's Intel's first volume part made

with the Intel 3 process node that we're

aware of and it'll be available this

quarter it's entirely composed of ecores

leading to a quoted 2.4x Improvement in

performance per watt versus the previous

generation a Intel did a comparison of

200 racks of servers going down to 72

with Zeon 6 based on the performance per

watt metrics there's also a PE core

design or performance core design that

I'll follow up later that is called

Granite Rapids and I had to double check

my notes to not get swapped on the names

because AMD has CPUs in the 9,000

desktop series coming out probably it's

not like 100% confirmed that's the name

of them but it's new CPUs coming out

later this year will be called Granite

Ridge Intel Granite Rapids the way to

think about it is Intel has lots of

lakes and water featured code names so

Rapids

Water they've got a nice theme going on

Intel didn't share performance or

efficiency numbers for those likely

because they're probably more similar to

the outgoing generation there was also a

small segment on Intel's upcoming mobile

client CPUs code named lunar Lake these

are Central to what Intel is calling

that AI PC it's a magical new Revolution

and personal computing and it's

something that every business should buy

immediately tell made that very clear

it's like if you don't buy these you're

behind just think of all the other

people buying AI PCS for their employees

at their businesses you can do so much

AI with the AI and the AI PC it's just

there's so much AI in those PCS we we

have to get it out we have to rescue it

Papa's little baby seriously Galer was

doing everything except begging all the

CEOs in the audience to buy new laptops

and Michael Dell the founder of Dell

even got one for free from this guy it's

it's been a lot of fun you know working

super closely together lunar lake is

quoted as having three times the AI

performance as the current core Ultra

lineup with 45 npu tops of performance

3x is a huge leap but it's not really

surprising to us since this is a

follow-up to a first generation product

and uh things are accelerating rapidly

right now Intel listed Microsoft

co-pilot the Adobe site zoom and teams

as the applications with new AI

functions and finally new CPUs with

third generation npus code named Panther

Lake are in production now according to

Intel now for the software stack so

Intel also talked a little bit about

this uh was light on the details it's an

area where Nvidia is making big moves

right now they spent a lot of their

presentation talking about it Intel's

jumping in on the software side and uh

the main takeaway is that Intel is also

providing these curated pip lines in a

similar way to nvidia's Nims which is

their inference microservices that we

covered last month and Intel

additionally talked about rag tell me

they have garbage data yes it's clear

that the business world is legitimately

excited about retrieval augmented

generation or rag where you have a

pre-trained llms response take a

separate data set or document into

consideration in practice this means an

llm shouldn't need to be trained with

proprietary or sensitive data uh in

order to reference or know it and

critically it can cite a source for the

data Intel showed this off by having an

employee working in one of its silicon

Fabs using an llm running locally on the

laptop to identify Parts information

about a contamination incident this was

clearly staged but llms with rag really

do seem to have potential for being

useful in a business setting so some

closing thoughts here again uh we like

to cover this stuff because it's

interesting it's different it is cool to

learn about technology in general and a

lot this is hardware even with the

software side stuff uh but because we're

not an AI focused Outlet sometimes it

gives us maybe a unique or a different

perspective on these types of events

where we're looking at it more strictly

from a what does this mean for a

consumer ultimately at the end of the

day whether that's uh way down the line

as a user of a service that uses these

products and software Solutions or if

it's literal use of a solution that

eventually gets sold to you like

something uh such as an npu in a laptop

uh so that's kind of our approach to

this type of coverage and um that's how

we like to think about it now Intel

knows it's going to need time to catch

up to Nvidia and it wasn't hiding that

fact in a statement actually uh it

sounded like geler was almost asking

investors to just hold on a little bit

longer we're getting there you know

everything I've talked about today it's

not going to happen overnight but it's

also not going to take long and that

statement is probably the most

indicative from this keynote of the

company's General mindset right now

which is that it's not positioning

itself as having bested Nvidia it's not

saying we've defeated them that's it

we're number one now uh what it is

instead saying is that it's working on

it and that acknowledgement is kind of

interesting because that's not typical

it's become more common which is

probably a good thing uh but Intel in

years past say about a decade ago

actually let's go back to the snake oil

thing it's like months ago but a decade

ago in particular I remember being at

event where they clearly did not have

anything noteworthy it was a refresh of

a refresh of a refresh this is like

can't make 10 nanometer work era of

Intel so we're talking 14 nanometer

quadruple plus and I remember being at

some of those events and uh Intel

talking big game about how much of a

market leader it is how revolutionary

these products are and it does seem like

there's a little bit more humility in

this announcement they were still still

bragging about a lot but it's a

different tone than the Intel that we

encountered uh in the consumer space

until AMD kind of kicked it in the ass a

little bit and and maybe taught it to

not sit too long on Old technology so

it's we'll see where it's going it does

seem like all of these companies are are

behaving a little bit differently lately

with this new AI craze there is a

certain level of competitive aggression

not necessarily in a negative sense but

in a they're they're trying to take each

other's market share in a new space

sense and um that indicates to us that

they see this as obviously a massive

Market Nvidia has proven that uh

companies like Google and Amazon are

jumping in they don't want to get left

behind and there's attempts at

Acquisitions like we saw previously with

Nvidia attempting to acquire arm uh that

indicate a potential future of a

consolidation like we see in most

Industries where if you can't beat your

competitors or if you're concerned about

them just buy them make them go away or

become part of your outfit so it's a

really interesting territory right now

that will affect consumers uh whether

again that's in our actual diypc space

or just in daily life as users of all

this stuff indirectly but anyway Intel's

announcements do appear to point toward

an attempt at fighting for second place

and that is what we think most of these

companies are going to be doing right

now whether or not they say that they

want to establish themselves as the

other one with Nvidia being the primary