WEBVTT

00:00:00.160 --> 00:00:10.519
here it is my friends concrete proof that satire truly is dead the GPU beside

00:00:06.600 --> 00:00:12.880
me contains 36 NVIDIA Grace Blackwell

00:00:10.519 --> 00:00:18.680
Super Chips and is estimated to cost over $3

00:00:15.240 --> 00:00:21.279
million now obviously the big heat sink

00:00:18.680 --> 00:00:25.199
on the side is illustrative you won't be installing one of these in your gaming

00:00:22.880 --> 00:00:31.960
rig unless uh you happen to have a 100,000 watts of power on tap and a

00:00:28.400 --> 00:00:33.559
building scale liquid cooling system but

00:00:31.960 --> 00:00:38.360
many of the Technologies NVIDIA is introducing here will benefit Gamers the

00:00:36.640 --> 00:00:45.360
the biggest one isn't really obvious until you go under the hood in my hands

00:00:43.039 --> 00:00:52.199
is one gb200 super chip now some of this we've

00:00:48.600 --> 00:00:55.840
seen before like this 72 core NVIDIA

00:00:52.199 --> 00:00:59.239
Grace CPU but these puppies right here

00:00:55.840 --> 00:01:02.199
these are all new and very very exciting

00:00:59.239 --> 00:01:07.520
do you guys see this tiny tiny line here thinner than the width of a human hair

00:01:04.479 --> 00:01:10.840
that is the gap between the two black

00:01:07.520 --> 00:01:16.799
well dyes that make up a b200

00:01:10.840 --> 00:01:21.400
GPU wait a second GPU is that not two

00:01:16.799 --> 00:01:23.640
gpus yes but also no while SLI might be

00:01:21.400 --> 00:01:29.040
dead for consumers NVIDIA has been hard at work creating interconnects that run

00:01:25.640 --> 00:01:32.799
at absolutely dizzying speeds allowing

00:01:29.040 --> 00:01:37.399
multiple G GPU dies to now act as a

00:01:32.799 --> 00:01:41.280
single GPU that allow multiple gpus to

00:01:37.399 --> 00:01:45.439
act as a single super chip and that

00:01:41.280 --> 00:01:49.520
allow multiple Super Chips to act as a

00:01:45.439 --> 00:01:51.799
single oh no to act as a single cohesive

00:01:49.520 --> 00:01:57.360
processing unit and it is going to unlock gaming experiences and more that

00:01:54.520 --> 00:02:03.200
are going to blow your mind you speak English right yes I speak English

00:02:00.479 --> 00:02:03.200
how can I help

00:02:05.280 --> 00:02:11.680
you can you also speak segue to our

00:02:08.360 --> 00:02:13.840
sponsor no I am also skilled in survival

00:02:11.680 --> 00:02:18.480
and op yes you can yes you can it's fine Ridge ridge's got your last minute

00:02:16.040 --> 00:02:21.959
father's Day's gift covered with a big sale click on our Link in the

00:02:19.920 --> 00:02:26.400
description and get up to 40% off their Rings their wallets and more

00:02:33.319 --> 00:02:38.360
we'll get to the Demos in a bit but first let's take a closer look at the

00:02:36.920 --> 00:02:43.720
product that is turning Global Tech media into Jensen hang's swifties NVIDIA

00:02:41.720 --> 00:02:49.560
chose not to disclose the number of cacor tenores or even the size of the

00:02:46.280 --> 00:02:52.040
ony caches of their new b200 Blackwell

00:02:49.560 --> 00:02:58.360
GPU but they did give us some numbers to work with Apples to Apples they expected

00:02:54.560 --> 00:03:00.959
to hit around 10 petaflops at fp8 sparse

00:02:58.360 --> 00:03:07.560
which puts it roughly 2 and half times faster than last gen Hopper also each of

00:03:04.640 --> 00:03:13.319
these is expected to draw about 1,000 Watts hence the

00:03:09.400 --> 00:03:16.519
uh liquid cooling each of our gpus gets

00:03:13.319 --> 00:03:19.640
192 GB of hbm 3E high-speed memory

00:03:16.519 --> 00:03:22.440
running at a casual 8 terabytes a second

00:03:19.640 --> 00:03:27.360
and is equipped with 1.8 terabyte per second Envy link and these numbers get

00:03:24.400 --> 00:03:34.959
even more ridiculous when we look at the super chip as a whole each super chip

00:03:31.120 --> 00:03:40.319
has two b200 gpus and a gray CPU for a

00:03:34.959 --> 00:03:44.439
total of 72 ARM CPU cores 864 GB of RAM

00:03:40.319 --> 00:03:48.040
and draws a total of 2700 Watts oh and

00:03:44.439 --> 00:03:50.519
by the way each of the 18 of these

00:03:48.040 --> 00:03:58.959
Blackwell compute nodes that make up an nvl 72 rack contains two Super Chips

00:03:55.720 --> 00:04:01.120
good Lord in California the rack that I

00:03:58.959 --> 00:04:08.079
was standing to in the intro would cost a whopping $30 an hour to run or about

00:04:05.599 --> 00:04:12.360
A4 million doar a year assuming you're paying residential energy rates speaking

00:04:10.280 --> 00:04:16.479
of running they literally need to take these demos to another room so I'm going

00:04:14.400 --> 00:04:21.040
to have to tell you about the spline on our way out of here we got our hands

00:04:18.759 --> 00:04:28.720
however temporarily on what NVIDIA is calling the spline of their nvl 72 rack

00:04:25.199 --> 00:04:31.520
this here contains 5,000 wires totaling

00:04:28.720 --> 00:04:36.919
over 2 miles mil and is cleverly laid out to optimize the latency and power

00:04:34.120 --> 00:04:41.039
efficiency see the networking all goes in the middle right here and the

00:04:39.240 --> 00:04:45.680
Blackwell compute nodes like the one they just took from us go at the top and

00:04:43.160 --> 00:04:51.160
the Bottom now they could have used fiber optics except that uh that would

00:04:48.800 --> 00:04:56.639
have cost them a casual 20,000 watts of additional power

00:04:53.479 --> 00:04:59.360
consumption so uh clever layout for the

00:04:56.639 --> 00:05:05.320
win put it all together and you've got a w

00:05:00.360 --> 00:05:09.360
72 Blackwell gpus 2600 gray CPU cores

00:05:05.320 --> 00:05:11.280
132 terab of hbm 3E memory with over

00:05:09.360 --> 00:05:16.840
half a pyte per second of aggregate bandwidth that's good for 720 pedop

00:05:14.560 --> 00:05:22.960
flops of fp8 training delivering results upwards of 30 times faster than a

00:05:19.400 --> 00:05:25.680
previous generation hgx h100 and if you

00:05:22.960 --> 00:05:30.560
didn't notice with perfect linear scaling something that is only possible

00:05:28.240 --> 00:05:34.840
when integrating your system this tightly even the placement of the

00:05:32.440 --> 00:05:39.600
individual blades matter on our super micro rack that we're looking at here

00:05:36.759 --> 00:05:43.800
you can see that they've got 10 up top and eight at the bottom with the nine

00:05:41.800 --> 00:05:48.720
nvlink switch units sandwiched in between that's because timing the

00:05:45.880 --> 00:05:52.560
electrical signals matters a lot and is easier on a more symmetrical setup now

00:05:51.080 --> 00:05:56.919
unfortunately NVIDIA didn't have a switch for us to show you they'll have

00:05:54.360 --> 00:06:03.520
to be represented by this piece of plastic but each of the nine units can

00:05:59.280 --> 00:06:06.560
handle 14.4 terabytes per second of NV

00:06:03.520 --> 00:06:10.199
link it is so integrated that NVIDIA

00:06:06.560 --> 00:06:14.840
says they think of this entire rack as

00:06:10.199 --> 00:06:17.160
one massive power hungry single GPU and

00:06:14.840 --> 00:06:20.560
it's kind of hard to argue otherwise other than that most of the time it's

00:06:18.759 --> 00:06:26.400
not doing graphics and I thought that's what the G was for and the craziest part

00:06:24.199 --> 00:06:33.400
is we haven't even looked at the craziest systems yet that was all mgx a

00:06:31.199 --> 00:06:41.240
standard set of reference designs that's intended to be compatible with multiple

00:06:35.720 --> 00:06:47.919
Generations hence the mg hgx is a whole

00:06:41.240 --> 00:06:51.080
different beast in this or on it it is

00:06:47.919 --> 00:06:52.560
eight Blackwell b200 gpus with a

00:06:51.080 --> 00:06:59.720
combined 1.44 terabytes of GPU memory absolutely

00:06:57.360 --> 00:07:03.960
ridiculous but the difference between this and what I just showed you is

00:07:01.280 --> 00:07:10.080
completely gone is any trace of Grace CPUs this is purely a GPU board because

00:07:07.879 --> 00:07:15.599
this insanity is meant to be integrated into a partner system like say from

00:07:12.400 --> 00:07:18.759
Super Micro now NVIDIA does sell their

00:07:15.599 --> 00:07:20.160
own dgx unit with this board and the

00:07:18.759 --> 00:07:25.360
rest of the components that make up a complete system but that's mostly

00:07:22.360 --> 00:07:29.160
intended to be a reference system these

00:07:25.360 --> 00:07:32.280
8 gpus get combined with Envy link just

00:07:29.160 --> 00:07:35.680
like the the rack setup for a whopping

00:07:32.280 --> 00:07:38.240
72 peda flops of FPA training while

00:07:35.680 --> 00:07:45.560
drawing nearly really 10,000 Watts now

00:07:42.400 --> 00:07:47.639
naturally this much power is a little

00:07:45.560 --> 00:07:51.960
hard to cool which is why it's so massive but the good thing about it is

00:07:50.080 --> 00:07:58.039
it doesn't require messing around with water or racks with 120,000 watts of

00:07:56.080 --> 00:08:02.800
power if you're installing these into an existing data center since

00:08:00.080 --> 00:08:05.840
those practically don't exist so you've got to spread them out a little bit

00:08:04.319 --> 00:08:11.159
according to your power budget which means you need networking and that is

00:08:09.120 --> 00:08:15.000
where NVIDIA's new hardware networking products come in this Ethernet switch

00:08:13.479 --> 00:08:20.319
will do something in the neighborhood if I think it's like 50 terabits per second

00:08:17.919 --> 00:08:24.720
of switching which is all really cool but what are we doing with this exactly

00:08:22.720 --> 00:08:28.319
I don't know how about Healthcare the tools I'm looking at right here use

00:08:26.360 --> 00:08:33.479
machine learning to approximate viral protein folding generate potential drug

00:08:30.840 --> 00:08:38.599
molecules to disable them and then test rapidly accelerating drug development oh

00:08:36.039 --> 00:08:41.680
and this is cool finding exactly what it is you're supposed to be taking a

00:08:39.839 --> 00:08:45.480
picture of with the ultrasound wand can take a bit of time why not let the

00:08:43.519 --> 00:08:51.680
machine identify it for you that's a left ventricle right cool know what else

00:08:48.680 --> 00:08:54.800
is cool simulations like the one we're

00:08:51.680 --> 00:08:56.640
living in behind me is Earth 2 a climate

00:08:54.800 --> 00:09:00.480
and weather simulation program that can run at such a high resolution that you

00:08:58.640 --> 00:09:06.000
can determine what's going to happen on a 1 km x 1 km basis is which is pretty

00:09:04.440 --> 00:09:11.760
cool but what if you need to simulate the movement of hot and cold air on a

00:09:07.959 --> 00:09:14.800
molecular level well you can do that too

00:09:11.760 --> 00:09:18.040
that is nuts which is all cool but what

00:09:14.800 --> 00:09:21.320
if I can't afford a dgx or an mgx to

00:09:18.040 --> 00:09:24.839
train those data sets well you can still

00:09:21.320 --> 00:09:28.360
use or experience NVIDIA's new Nims or

00:09:24.839 --> 00:09:30.440
NVIDIA inference microservices Nims are

00:09:28.360 --> 00:09:34.320
pre-trained and pre-optimized containerized AI models that you can

00:09:32.399 --> 00:09:39.079
download and deploy for any number of use cases and if that all sounded like

00:09:36.680 --> 00:09:43.560
gobleg um let's go back to that bilingual demo for a second facial

00:09:41.480 --> 00:09:47.600
animations can be an extremely timec consuming component of game development

00:09:45.680 --> 00:09:54.000
and are one of the big reasons that localization can be such a challenge the

00:09:50.800 --> 00:09:56.800
Nim in use here allows automatic mapping

00:09:54.000 --> 00:10:00.959
of speech to mouth animations and facial expressions meanwhile this guy takes

00:09:58.800 --> 00:10:05.560
things to two steps further using Nims for automatic speech recognition and

00:10:03.399 --> 00:10:09.839
facial animations but also a third one that I think is perhaps the most

00:10:07.720 --> 00:10:13.839
interesting to me there's a major concern right now in the games industry

00:10:11.920 --> 00:10:20.000
that AI is going to take jobs away from writers but this guy uses a n for data

00:10:16.800 --> 00:10:22.800
retrieval that is part of inworld ai's

00:10:20.000 --> 00:10:27.680
platform and this is really cool instead of him just crapping out whatever

00:10:24.880 --> 00:10:33.240
response chat GPT might throw at you he's actually got an extensive backstory

00:10:31.040 --> 00:10:38.720
that does need to be written by a human writer in order to give you a

00:10:35.399 --> 00:10:40.720
personality and context specific

00:10:38.720 --> 00:10:45.950
information that will help you advance the story have you ever tried any of the

00:10:42.800 --> 00:10:47.839
fine merchandise from LTT

00:10:47.839 --> 00:10:52.480
store.com now that is really cool and

00:10:50.880 --> 00:10:56.040
we're just scratching the surface right now over time NVIDIA is going to be

00:10:54.360 --> 00:11:00.880
looking to The Gaming Community both developers and Gamers for inspiration

00:10:58.560 --> 00:11:04.880
for what to do with these and oh I've got one more really cool demo G assist

00:11:03.120 --> 00:11:11.680
here might just be a tech demo at the moment but it's a pretty darn compelling

00:11:07.480 --> 00:11:16.200
one how do I craft a stone

00:11:11.680 --> 00:11:19.120
axe okay that is cool but what dinosaur

00:11:16.200 --> 00:11:19.120
am I looking at right

00:11:20.320 --> 00:11:27.079
now okay that's kind of sick and what's

00:11:24.240 --> 00:11:33.360
cool is the image recognition and I believe the voice to text are both

00:11:29.480 --> 00:11:34.760
running locally on this RTX series GPU I

00:11:33.360 --> 00:11:41.800
don't know about you guys but I think this is so cool that I don't know what to say way to our sponsor back Blaze

00:11:39.440 --> 00:11:45.920
losing your data is never fun so having solid backups of everything is super

00:11:44.120 --> 00:11:49.720
important and back Blaze is an affordable easyto ouse cloud backup

00:11:47.880 --> 00:11:54.720
solution with plans that start at just $9 a month you can backup almost

00:11:52.519 --> 00:11:58.279
anything from your Mac or PC and access it anywhere in the world with their web

00:11:56.399 --> 00:12:02.320
and mobile apps and they've restored over 55

00:11:59.760 --> 00:12:06.040
billion files with multiple options for how you can retrieve your data including

00:12:04.079 --> 00:12:09.399
having them send a physical hard drive straight to your door and if you're

00:12:07.800 --> 00:12:14.560
worried about accidentally deleting files you can increase your retention

00:12:11.279 --> 00:12:16.440
history to one year for free plus for

00:12:14.560 --> 00:12:20.279
organizational and business purposes their Advanced admin controls are

00:12:17.880 --> 00:12:24.279
designed for security scalability and ransomware resilience back blazs has

00:12:22.360 --> 00:12:29.320
over three exabytes of data under their management and has the trust of over

00:12:26.399 --> 00:12:33.160
half a million customers including us that's right we not only work with them

00:12:31.040 --> 00:12:38.000
on a sponsored basis we actually back up our servers nightly to back Blaze so

00:12:35.760 --> 00:12:42.079
starting at $9 a month it is hard to find a better investment than your peace

00:12:39.639 --> 00:12:47.680
of mind so sign up today and get a free 15-day trial at back blaze.com

00:12:44.880 --> 00:12:51.959
if you guys enjoyed this video uh why not check out our video from last

00:12:49.279 --> 00:12:55.680
computex showing off the grace CPUs and their last gen Super Chips we got a

00:12:53.880 --> 00:12:59.040
little bit more into the weeds and it was very very cool