00:00if 2024's economy can be summed up by a
00:02single word it would probably be clean
00:04sweep by Nvidia because not only the
00:06entire US economy is being carried by
00:08this GPU designing company our entire
00:11net worth is probably on its back too
00:13and today this company that probably
00:15determine if he can retire in 5 years
00:17has its biggest annual event GTC with
00:20CEO Jensen Hong delivering the Keynotes
00:22of the major developments in Nvidia for
00:24this year while AI is still on the side
00:26is stacking more layers than ever Nvidia
00:29is keeping up with this Trend by
00:30announcing its newest AI super chip
00:32called Blackwell packed with 208 billion
00:35transistors and 192 GB of HPM this bad
00:38boy can run a staggering 10 pet flops on
00:41Flo point8 which is capable of running
00:43models up to 10 trillion parameters to
00:46put that number into perspective 4090
00:48has like 90 Tera flops on a good day so
00:51Blackwell GPU has like 100 times more
00:53compute than 490 or nearly 2.5 times
00:56more compute than a ag100 additionally
00:59with their new announced fifth gen EnV
01:01link and NV link switch 7.2 T which has
01:04like 1.8 TB pers second by directional
01:06bandwidth if we were to build such a
01:08chip we can have every single GPU talk
01:12to every other GPU at full speed at the
01:16same time these blackw bad boys can then
01:19be linked up with some great CPUs with a
01:212 to1 ratio to create gb2 200 then they
01:24stack 36 of these gb2 200s to create an
01:27absolute unit that is gb200 and V l72
01:30this Omega chungus of a machine can
01:33actually act as a singular GPU with 1.4
01:36EXA flops of inference speed and 30 TB
01:39of fast memory and when you combine
01:41eight of these you would get this new
01:43dgx superod which would be available at
01:45Major clouds with Blackwell their second
01:48generation Transformer engine enables
01:50two times more compute and bend withd on
01:52fp4 for Transformer models it improves
01:55the communication bottleneck that
01:56mixture of xert models have which can
01:59Ensure High speed communication across
02:01576 gpus for super largee models up to
02:0530 times speed up if thee model has 1.8
02:08trillion parameters however this is the
02:11only GPU Hardware announcement we got so
02:13the 50 Series might still be far away
02:15from us but they do have some other
02:17quote developements to accelerate the
02:18world of Robotics the new project group
02:21I am Group which stands for generalist
02:23robot 00 technology uses a multimodel
02:26Transformer based model that takes text
02:29video and demonstrations as inputs and
02:31outputs the actions for the robot Nvidia
02:34also unveiled a new robot computer
02:36called jeton Thor based on the Blackwell
02:38architecture which when combined with
02:40Groot Foundation models will accelerate
02:42the development of embodied AI robots
02:45however embodi robots need a massive
02:47amount of data and Gathering real life
02:49data is not as efficient either so they
02:51introduced ISAC lab which is designed
02:53specifically for reinforcement learning
02:55and is built on top of ISAC Sim that can
02:57run thousands of parallel simulations
02:59for robot learning in order to scale up
03:01training in addition Nvidia also
03:03announced ISAC manipulator and ISAC
03:06perceptor where ISAC manipulator offers
03:08state-of-the-art dexterity for robotic
03:10arms with 80 times speed up and ISC
03:13perceptor which uses 3D surround Vision
03:15capabilities from multi cameras to help
03:17improve performance for autonomous
03:19mobile robots but nvidia's Quest into
03:21improving factories efficiency does not
03:24stop here with their latest digital twin
03:26technology that creates a digital
03:27counterpart for a real world system
03:30nvidia's Omniverse and the newly
03:31announced Metropolis gives industrial
03:34operations a jaw-dropping amount of
03:36optimizations with workers simulations
03:39robot simulations and Warehouse digital
03:42twin a super large scale Advent planning
03:44system can easily be made and with
03:46hundreds of cameras sensors and
03:48real-time feedbacks it can improve the
03:50efficiency of managing fleets of robots
03:52in a factory and even improve human
03:54robot collaboration in places such as
03:56supply chain distribution centers and
03:58you can use it on Apple Vision Pro on
04:00the digital side Nvidia is also
04:02launching their own generative AI
04:04services at ai. nvidia.com where you can
04:07experience open source models like llm
04:09Tex to image audio synthesis or even
04:11medical models however Nvidia itself is
04:14a platform not a model Building Company
04:17because these are actually microservices
04:19deployed with their latest API called
04:21Nvidia Nim where it Bridges millions of
04:23products together and serves it to you
04:25with 4.5k per GPU per year or $1 per GPU
04:29per hour power aidia Nim connects
04:31pre-existing model Builders and
04:32accelerated Computing systems for you so
04:34you can run any AI related products as
04:37easy as printing hello world and if you
04:39really think about it Nvidia is
04:41literally everywhere they are probably
04:43the only company that works with every
04:44AI company right now which is nuts but
04:47yeah that's it for today's quick summary
04:48of GTC 2024 there are other interesting
04:51announcements like Earth digital twin
04:53for weather prediction AKA Earth 2
04:56wouldn't you like to know weather boy
04:57which is designed for extreme prec ition
05:00global weather prediction and there are
05:02some cool Omniverse rendering on the
05:04keynote they are not animations by the
05:06way so definitely check out the keynote
05:08and you can even attend the conference
05:09virtually for other interesting panels
05:11and have the chance to win an RTX 480
05:13super Nvidia actually sent me this a
05:15giveaway to you guys and to attend this
05:17giveaway you just need to attend at
05:19least one virtual GTC session between
05:21March 18th to March 21st and show a
05:24proof of you attending it which is
05:25pretty straightforward for the proof I
05:27just need you to take a selfie while
05:29watching any the live virtual GTC
05:31sessions and if you don't want to show
05:32your face you can just show a thumbs up
05:34with the live GTC session on your
05:36monitor and also very important you have
05:38to use my link down in the description
05:40to sign up for the virtual GTC session
05:42or else they wouldn't know I hooked you
05:44up with some good unvid workshops or
05:46panels because this time the original
05:48authors of the paper attention is all
05:50you need which gave rise to the
05:52Transformer architecture that every LM
05:54uses now is hosting a panel session on
05:57March 20th this is definitely going to
05:59be one of the highlights and highly
06:00recommend you to attend it virtually so
06:03add the event to your calendar now and
06:05take a selfie attending it virtually to
06:07win of 4080 super a big shout out to
06:09Andrew lelus Dean Chris Leo Alex Shay
06:14Alex mares migul liim fifal and many
06:18others that support me through patreon
06:19or YouTube follow my Twitter if you
06:21haven't and I'll see you all in the next
