00:00:03
uh hi i'm here with uh Israel from
00:00:05
Nvidia welcome um what is your role on
00:00:08
the DJX Spark well I'm I'm doing the
00:00:10
tech marketing for it so I'm the person
00:00:13
that bridges the technical teams and the
00:00:16
marketing teams so I know enough about
00:00:20
the hardware but also in a way that it
00:00:22
can be uh digested I guess by Yeah by
00:00:26
the non- tech people yeah so maybe let's
00:00:29
talk about that right this is a $3,000
00:00:31
device uh starting starting starting at
00:00:33
3,000 wait what does it go up to 4,000
00:00:36
okay so like just extra upgrades of
00:00:37
course that might be a good first
00:00:38
question for you right what what's the
00:00:40
difference between the 3000 and the
00:00:42
4,000 model sure i don't I don't know
00:00:44
actually just extra memory storage
00:00:46
four four terabytes on the four uh on
00:00:49
the $4,000 one and one TB on the $3,000
00:00:53
and what uh part of the market is this
00:00:55
designed for like do you have a name for
00:00:57
like I guess proumer yeah it's we call
00:01:00
it a a mini AI supercomputer right it is
00:01:05
it is really in many ways a miniaturized
00:01:08
version of these things you see behind
00:01:11
me here yeah the real data center
00:01:12
equipment so the idea is you're taking a
00:01:15
a nibble of the that data center to your
00:01:18
home so you can develop uh this is
00:01:21
really a developer box to focus on on
00:01:25
individual developers so they can do
00:01:28
work on a platform that is really
00:01:32
100% derived from from these right the
00:01:36
architecture the CPU instruction set
00:01:38
software stack yeah the drivers the
00:01:41
network
00:01:42
component networking accelerations
00:01:46
um even the cross memory right that that
00:01:49
we have here the shared memory uh layout
00:01:52
that that's the machine that you would
00:01:54
find on a Grace Blackwell data center
00:01:57
machine
00:01:58
and you can once you you learn how to
00:02:02
develop here you're and you write code
00:02:06
here this code is ready to go there into
00:02:09
production to large scale deployment
00:02:12
without having to worry about is this
00:02:14
going to work you know is is my stack
00:02:16
ready for for that no it's the same
00:02:18
stack so is this for how much is is
00:02:21
hobbyist usage and how much is that kind
00:02:24
of workflow where they're actually
00:02:25
prototyping for the big data centers uh
00:02:27
honestly the the amount of things you
00:02:29
can do here I wouldn't call it a
00:02:31
hobbyist anymore because you you can do
00:02:33
uh some really serious AI uh work
00:02:37
finetuning which is very demanding right
00:02:39
and this is one of our key use cases
00:02:42
when we developed this product so uh
00:02:46
this is yeah I would I would say this is
00:02:48
one step further from a hobus and and
00:02:51
today there's so much you can do in in
00:02:54
terms of bit real business right um over
00:02:57
say a 32 billion LLM right just by
00:03:00
adjusting that to a certain demand using
00:03:03
fine-tuning that's already a product in
00:03:05
many ways now the only thing that you
00:03:07
cannot do here compared to these things
00:03:09
is yeah this is not designed to serve I
00:03:12
bet yeah 500,000 client at the same time
00:03:15
right so um you do this at home you test
00:03:18
and you and you run or even on a small
00:03:21
company and once you your stack is good
00:03:24
this can be pushed out to a a larger reg
00:03:27
either on prem or or a cloud resource
00:03:30
that will run and like I said the the
00:03:34
beauty is you don't have to tweak your
00:03:37
code it's it's just going to move over
00:03:40
to the next step right whether that is
00:03:42
the workstation that we have there or a
00:03:45
real data center uh grade equipment yeah
00:03:48
I was actually kind of wondering so I
00:03:50
know about the Jetsson Nano which you
00:03:52
released a few months ago um and then we
00:03:54
have the 5090s over there um what is the
00:03:57
sort of per the the the product suite
00:04:00
like how would you sort of grade it from
00:04:03
like the small end to the high end yeah
00:04:05
this is a little hard to compare with
00:04:06
Jetson because it's all new right this
00:04:10
is a Grace Black wall so all this all
00:04:14
the silicon here is uh newer than than
00:04:16
what you would find um but the the idea
00:04:21
is that this would be your starting
00:04:23
point for
00:04:25
Go ahead oh okay all right he's he's
00:04:26
doing it right okay cool sorry keep
00:04:29
going yeah the idea is Yeah you this is
00:04:32
really your starting point as a as a
00:04:35
developer for for the Grace Blackwell
00:04:37
architecture right
00:04:39
um we we will continue to support
00:04:43
different ranges of products for
00:04:46
different types of development right so
00:04:48
we have the IGX is more oriented towards
00:04:51
enterprise work and
00:04:53
um it's it's more of a professional tool
00:04:56
right than than this and this is really
00:04:58
more oriented towards a home user one
00:05:01
feature that you find here that
00:05:04
it's here we'll I'm I cannot wait for
00:05:07
this to reach homes and people start
00:05:10
playing with but this is also a a 5000
00:05:14
series Blackwell in many ways right it
00:05:16
has DLSS it has RTX and it has a very
00:05:19
capable display output subsystem and
00:05:22
right over here right so there are
00:05:24
things you can do here that we're not
00:05:26
even touching just yet um in terms of of
00:05:29
marketing right because for now we're
00:05:30
focusing more like ML and AI right but
00:05:35
um this is still a very capable home
00:05:38
computer well yeah gaming comes with
00:05:41
with you know we're this ships with
00:05:43
Linux right so that's your you start
00:05:46
from there and it's an ARM machine so
00:05:49
code right it has to run and it it it's
00:05:53
not today you know the the Linux game
00:05:55
and on ARM is a still like a this moving
00:05:59
piece but uh we've seen effort from
00:06:01
Valve right there are companies that are
00:06:03
working on on on that but I mean this
00:06:06
was not designed for it but sure still
00:06:08
you know capable telling you what's in
00:06:10
here silicon wise and I'm pretty sure
00:06:13
people will find interesting uses for
00:06:14
that as well yeah so um I mean this is
00:06:17
the first time like usually we see these
00:06:18
boxes this is the first time you sort of
00:06:19
open up the that I've seen sort of this
00:06:22
open up anything interesting that we
00:06:23
should look at this is literally the
00:06:25
first time we're showing this board yeah
00:06:27
so the board the board layout is
00:06:28
actually very simple okay so on the top
00:06:30
here we have the GB10 SOC
00:06:33
uh that we built in partnership with
00:06:34
MediaTek mhm around it you have the
00:06:38
LPDDR5X memory modules for your shared
00:06:41
128 gigabit gigabyte memory space that
00:06:45
these two chips can reach out to without
00:06:48
having to create copies of the memory
00:06:50
content right so that alone should give
00:06:53
you an enormous advantage on certain
00:06:55
workloads that you have a lot of transit
00:06:57
between your your RAM and your
00:07:00
VRAM compared to say even a powerful
00:07:03
workstation but that you're still using
00:07:06
a regular PCI Express card right so it's
00:07:08
DDR5
00:07:10
uh yeah it's LP LP because uh DDR this
00:07:13
is more like a laptop memory in the
00:07:15
electrical sense than uh so it's low
00:07:17
power DDR
00:07:18
but the the performance is great it's
00:07:20
actually a little faster than a socketed
00:07:22
um dim comparable with like using the
00:07:25
same type of uh chip right right right
00:07:28
and we have the
00:07:29
C2C uh interconnect that's the tech this
00:07:32
is Nvidia right it's our design that
00:07:37
connects the GPU to the CPU and allows
00:07:40
for we're estimating there's somewhere
00:07:42
in five times faster than PC express
00:07:45
communications here between these two
00:07:46
chips so they share the same memory
00:07:48
controller and everything right yeah uh
00:07:51
the there there's the the memory
00:07:53
controller is basically provides access
00:07:55
to both GPU and CPU
00:07:58
yeah this this is um it's actually very
00:08:00
interesting topic but uh it's very deep
00:08:03
as well so I I I can only get this this
00:08:07
far because I'm a marketing guy so what
00:08:10
is the networking on it is that so yeah
00:08:12
oh that's another crazy part okay so
00:08:14
yeah we were just talking about this
00:08:15
right so far we're talking about this
00:08:18
so let's move over this part of the PCB
00:08:20
here so we have a this is
00:08:22
enterprisegrade connect X7 dual port 200
00:08:26
Gbits per second Ethernet 200 Gbit per
00:08:29
port per port okay
00:08:31
which is wild right i don't from top of
00:08:35
my memory I can't think of a box this
00:08:36
small with a network this fast now of
00:08:40
course we don't expect people to have
00:08:41
200 gigabit ports on their own that'll
00:08:44
be actually nice but uh yeah that's not
00:08:46
happening but this is designed for the
00:08:48
scale out option so using um this is
00:08:53
basically a cluster to network them yeah
00:08:55
yeah okay so
00:08:57
using distributed workload balancing
00:08:59
with TRTLM
00:09:01
and all our software stack for that
00:09:03
which is the same enterprise software
00:09:06
you can develop something at home that
00:09:08
would run on say the NVL72
00:09:12
so is it shipping with like Kuntu is
00:09:14
that yeah so uh we call it DJXOS but
00:09:17
this is Ubuntu 244 LTS okay the only
00:09:20
thing is yeah we we add a few extras to
00:09:23
it so we add um I don't know from top I
00:09:27
have everything that we add to it but it
00:09:29
has some performance optimizations it
00:09:31
has our repositories and our software
00:09:33
preloaded so uh driver driver for the
00:09:36
connect 7 all of these pieces are there
00:09:39
for you those modules are so did you
00:09:41
guys announce when it's going to be
00:09:42
available or price i Yeah so it is
00:09:45
available right now for a weight list
00:09:47
right so people that want to get on the
00:09:49
wait list they they can get in uh our
00:09:52
partners are also enabling their weight
00:09:54
lists one thing that I want all of you
00:09:56
uh media to be aware is the what what
00:09:59
our partners Asus HP Dell and Lenovo
00:10:03
these are the four partners we have
00:10:04
enabled right now we're selling the same
00:10:07
thing okay the the only difference is
00:10:09
they they're going to have different
00:10:11
case designs maybe different cooling
00:10:13
solutions
00:10:14
But uh this board is the same the
00:10:16
feature set is the same the only
00:10:18
variable that we're we have between
00:10:20
models today is uh we on the bottom of
00:10:23
the the board that you can't see cuz
00:10:25
it's glued in the base here um there's
00:10:27
an M.2 for NVME right it's PCI Express
00:10:31
um there's a one TB disc for the entry
00:10:35
model and a 4 TB version for the top
00:10:38
model that's only the only difference
00:10:40
how much memory are upgradeable
00:10:43
uh I don't I don't know if we're going
00:10:45
to state it that way okay but it's it's
00:10:47
it's M.2 yeah how much memory does it
00:10:50
come with the base model 1 TB and the
00:10:54
one TB of D RAM no how much DRAM sorry
00:10:57
oh okay no yeah it's 128 for all trims
00:11:00
all trims 128
00:11:02
that's what I'm saying here this is this
00:11:03
is something that I really want people
00:11:04
to be clear about so what's the largest
00:11:07
model that will run on that well we're
00:11:10
expecting uh for the single unit
00:11:13
somewhere in the 200 uh billion in FP4
00:11:16
which is yeah try doing that on a laptop
00:11:19
it's going to be really hard and for the
00:11:22
stackup it's it's 400 right with with
00:11:24
two and all like the Python software for
00:11:26
like the data scientist that works like
00:11:29
the QDF stuff oh I'm I'm going to have a
00:11:31
guy in the booth later today afternoon
00:11:34
uh when we're done with the press that
00:11:36
he is working on the Python optimization
00:11:38
you can okay he can give you a much
00:11:40
better answer than I why is that a
00:11:42
question for you why is that a question
00:11:44
yeah why is just the I mean typically
00:11:47
x86 is the platform for a lot of yeah
00:11:50
stuff so just making sure that like if
00:11:52
I'm writing Python code and I need to
00:11:56
move it to this it just works right
00:11:58
awesome how many can you run in parallel
00:12:00
uh well so far we're supporting two
00:12:02
right but um again this is Ethernet and
00:12:05
and the scale out is using is software
00:12:08
based okay we're offering you the the
00:12:11
same software that we use for very large
00:12:14
clustered distributed workloads on on
00:12:17
systems like the NVL72 behind us here to
00:12:21
use here so that that's the interesting
00:12:24
part you have a mini lab at your home
00:12:25
that's using basically the best scale
00:12:29
out technology available today if you
00:12:32
learn here you can just go there and do
00:12:34
the same with the same code with the
00:12:36
same
