NVIDIA's New Mini PC - With the GB10 Grace Blackwell Superchip

Video ID: GY8YudAnxu0

YouTube URL: https://www.youtube.com/watch?v=GY8YudAnxu0

Added At: 13-06-25 21:17:16

Processed: No

Sentiment: Neutral

Categories: Tech, Education

Tags: AI, Machine Learning, Data Science, Nvidia, Mini PC, GB10, 128GB RAM

Summary

Nvidia announces a new Mini PC called Project Digits, featuring the Nvidia GB10 processor and 128GB of RAM. It's aimed at data science, machine learning, and AI workloads. The device will be available in May for $3,000.

Transcript

so CES is in full swing in Las Vegas and
Nvidia has made its keynote
announcements now at the very end of the
presentation it announced the uh project
digit Mini PC that's right it's
announced a Mini PC it's got a new
processor in it called the
gb10 from Nvidia itself now if you want
to find out more please let me
[Music]
explain okay so let's dive into the
Nvidia Project digits digit will explain
what all that means in a moment the key
thing is here it features a new chip the
Nvidia gb10 that's the grace Blackwell
super chip and of course Nvidia have got
you know the GB 100 the gb2 200 for that
stuff it uses for its AI uh super
computer this is what we're looking at
we're looking at a mini uh PC it can be
used as a workstation or of course it
can be used you know connected remotely
remote desktop secure shell just
connecting to various web services
connecting to various services that it
offers for machine learning uh and so on
and the design that uh Nvidia are
currently showing here is it's kind of
like a shrunk down version of the the
dgx kind of supercomputers or modules
that they've got with all of their gpus
and so on inside of it so digits
WhatsApp we found out during the keynote
that it stands for deep learning GPU
intelligence training system digits that
was the internal name for basically all
of the different systems that Nvidia
have been working on for many many years
now uh that have resulted in well what
the djx system so it will shorten from
digits to dgx to make it the same as you
know RTX uh uh agx and so on uh but
really it's deep learning GPU system
that's really what it means so I don't
think that this Mini PC will actually be
called project digit in the end that's
its current name they will probably come
up with something like you know the dgx
mini or something like that I don't know
uh but the key thing is it features this
Nvidia Grace Blackwell super chip the
gb10 so it offers a Peta flop of AI
Computing performance at uh 4 bits so
that's really really quite Mega powerful
for especially for that form factor now
it's got an Nvidia blackw GPU so that's
the latest GPU uh architecture coming
out from Nvidia with the latest Cuda
calls fifth generation uh tensor cores
and then it's got the EnV link chipto
chip interconnect to connect to the GPU
to the CPU uh on the actual chip
something that Nvidia uses in many of
its other uh grace-based uh Super Chips
the Grace Hopper Super Chips and so on
now the CPU part is Arm based and we'll
talk more about that that in a moment
and it comes with 128 GB of unified
coherent memory what that means is like
other systems so type systems uh this
kind of the the memories on the chip so
it's a fixed amount of memory and
they've gone with 128 gabt because
you're going to need that if you're
doing AI type uh workloads and then up
to 4 terabytes of nvme storage there'll
be different versions available with
different amounts of storage again I
think that storage is on the chip itself
so it's kind of an so uh design now with
that amount of memory and of course if
you have the storage you can run up to a
200 billion parameter large language
model if you're using 4bit quantization
now 200 billion parameter today when
we're talking about you know running you
know llama 3.2 or something like that on
you know uh your desktop or on an Nvidia
Jetson Orin or something like that we're
talking about seven billion parameters
or three billion parameters so being
able to run a 200 billion parameter is a
whole different scale of things and
that's that's the combination of that
unified memory 128 GB of it and then
also the ability to speak very quickly
to the GPU and the tensor cores to get
that best performance out of that model
so the CPU is a 20 core Arm based CPU so
you get 10 cortex
x925 CPU calls I've got videos about the
x925 here on this channel and then you
get another 10 a725 CPU Calles and again
I got videos about that here on this
channel so it's uh heterogeneous in the
sense that you've got two different
types of core but of course they are uh
compatible architecturally but the a725
uh is slightly more power efficient
whereas the x925 gives you that maximum
uh performance and so they've given you
this 10+ 10 setup so a nice 20 core CPU
setup along with a black GPU is
certainly something to be interested in
now earlier I kept mentioning it's kind
of an so design and that was important
because this was made in collaboration
with mediatech that was even said during
the keynote so mediatech have had their
input into this the way of putting this
all together the interconnects and
everything uh they've worked with Nvidia
which personally I find surprising
because Nvidia have got a big history of
working with Arm based CPU so there was
obviously something they wanted from
mediatech and maybe we'll get more
details about that as time goes by
mediatech have also published a press
release saying they worked with Nvidia
but thin in the details but actually
what it was that they've provided that
will be interesting to find out and it
runs Nvidia dgx o Linux so it's a Linux
based system uh based on Ubuntu just
like other envidia AI products and it's
compatible that means you can run all of
the nvidia's AI stack all the stuff you
need to run all the cudic cor
programming the tensor programming all
of the things you could run on a bigger
djx system you can run on this because
it's running the same OS and of course
you've got an arm CPUs and also you've
got the same Nvidia gpus so it's
compatible across the the region because
this is a smaller one so if you put this
in comparison to some of the bigger ones
the new blackw gpus in their full spec
Out versions can offer maybe 40 times
more per performance so this is
something that's not in a huge briefcase
size rack size thing this is something
that's in a Mini PC but they really have
specked it out maximum specs for the
size of the little box that you're
getting so there was lots of talk in
2024 about mediate Tech's collaboration
with Nvidia we knew there was a
collaboration going on and the main
thrust was is because there's going to
be a Windows on arm PC from mediatech
with an Nvidia uh GPU that's not what's
been announced today this is a Linux
based machine that's coming purely from
Nvidia it's not being sold by mediatech
although mediatech have collaborated
with the design of thec as I said so is
Project digits this windows on arm PC
absolutely not this is definitely not
what it is could it be the foundation
for something like that maybe if Nvidia
and mediatech have been working together
maybe mediatech has got some kind of
deal where it can get access to these
chips because normally Nvidia chips are
in Nvidia products there are a couple of
cases where they're in other products
like for example the Nintendo switch so
maybe um Media Take and Nvidia have got
some kind of collaboration going on but
whatever it is it won't be this
configuration this uh is a very specific
configuration 128 GB of RAM if you're
selling a Windows on arm PC you're not
going to sell it with 128 GB of RAM
maybe 64 if you wanted to be crazy 32 16
that kind of thing 128 definitely not so
it's not going to be in this
configuration uh and there's other thing
and that's the price which we'll talk
about right now because these boxes will
be available in May of this year so
we're looking about another five months
away from here and it will be sold
directly from Nvidia and nvidia's
partners and they start at
$3,000 so you're going to really want to
have to want an AI
machine you can't afford a dgx rack size
thing you want something you can do uh
AI inference on your desktop uh and
you've opted not to go down the line of
an Nvidia or AMD GPU inside of a a beefy
PC you're going to have to Fork over
$3,000 which is a lot of money
particularly when you consider the price
of the new M4 Mac Mini which is quite an
amazing little device but of course this
is got the extra things all that memory
128 GB of memory and i' would imagine
the GPU is many many times uh faster but
if this was could be a Windows PC you're
not going to sell it at that price point
either no one's going to say yeah I'm
going to aband an x86 and go over to
your Mini PC at 3 Grand that's just not
going to happen so it will be
interesting to see what uh unfolds but
in terms of AI products this is
absolutely amazing 20 core CPU blackw
GPU 128 GB of RAM all in a Mini PC
running uh dgx OS Linux okay so there we
go project digits a Mini PC aimed
squarely at data science machine
learning and so on it's got that $33,000
price tag love to hear what you think
too expensive actually it's a lot
cheaper than a big djx system is it
worth it what about it being a stepping
stone over to Windows on arm in
collaboration with mediatech love to
hear your thoughts on that as well okay
that's it my name's Gary Sims this is
Gary explains I really hope you enjoyed
this video if you did please do give it
a thumbs up and if you like these kind
of videos then stick around subscribe to
the channel okay that's it I'll see you
in the next one
[Music]
he