Nvidia's Announces Game Changer AI Platform That Changes Everything For AMD & Intel, Worth Trillions

Millionaires Investment Secrets
20 Aug 202419:42

Summary

TLDRDieses Skript fasst die Entwicklung von Nvidia und die Bedeutung von Deep Learning in der Generativen KI zusammen. Es erzählt von der Geburt von Nvidia, der Erfindung des programmierbaren Shading GPUs, bis hin zu den Fortschritten in der Computergrafik und der Einführung von AI in der Industrie. Es präsentiert auch die Herausforderungen und Lösungen im Bereich der Energieeffizienz und zeigt die Vision einer zukünftigen KI, die nicht nur die Produktivität steigert, sondern auch neue Wissenschaft entdeckt und die Energieeffizienz verbessert.

Takeaways

  • 🚀 Die Beschleunigung des Deep Learnings um eine Million Mal hat die Schaffung großer Sprachmodelle und die Entstehung von generativer KI ermöglicht.
  • 📉 Der Kosten- und Energieverbrauch für die Entwicklung von generativer KI hat sich drastisch reduziert.
  • 🎨 NVIDIA hat sich durch die Erfindung des programmierbaren Shading und der GPUs für Computer-Grafiken einen Namen gemacht.
  • 🧠 Die erste Kontaktaufnahme mit KI erfolgte 2012 mit AlexNet, was ein Durchbruch in der Computer Vision darstellte.
  • 🔧 NVIDIA hat seine Forschung und Entwicklung auf Deep Learning umgestellt, um neue Software-Schreibweisen zu erforschen und zu nutzen.
  • 🤖 Die Einführung des DJX1 von NVIDIA im Jahr 2016 markierte einen Wendepunkt in der Entwicklung von KI für autonome Fahrzeuge und Robotik.
  • 🔄 Der Einsatz von DLSS (Deep Learning Super Sampling) ermöglichte eine enorme Geschwindigkeitssteigerung bei der Ray-Tracing-Grafik.
  • 🌐 Generative AI wird in naher Zukunft viele Branchen revolutionieren, einschließlich wissenschaftlicher Computing und autonomer Fahrzeuge.
  • 🛡️ Neuere Fortschritte in der KI, wie z.B. die Verstärkung durch menschliches Feedback und die 'Guard railing', haben dazu beigetragen, die Kontrollbarkeit und Genauigkeit von generativer KI zu erhöhen.
  • 🔍 Die Einführung von 'Retrieval-augmented generation' hat es möglich gemacht, die KI mit semantischen Datenbanken zu beliefern, um eine fundierterere und zielgerichtetere Antwortfähigkeit zu erreichen.
  • 🌍 Generative AI ist energieintensiv, doch es gibt Hoffnung, dass durch neue Technologien wie Blackwell die Energieeffizienz signifikant gesteigert werden kann.

Q & A

  • Wie wurde das Deep-Learning vor einigen Jahren beschleunigt?

    -Das Deep-Learning wurde durch eine Millionfache Beschleunigung vor einigen Jahren vorangetrieben, was es möglich machte, große Sprachmodelle zu erstellen.

  • Was war der Grund für die Entstehung von generativer KI?

    -Die Entstehung von generativer KI wurde durch die Millionfache Beschleunigung und Kostenreduktion ermöglicht, was die Entwicklung von allgemeiner generativer KI erlaubte.

  • Was ist der Unterschied zwischen traditioneller Computergrafik und der durch AI unterstützten Computergrafik?

    -Traditionelle Computergrafik basiert auf programmierbaren Schattierungen und Rendering-Techniken, während AI-unterstützte Computergrafik in der Lage ist, komplexe Aufgaben wie Ray Tracing in Echtzeit durchzuführen, was zu einer höheren Qualität bei interaktiven Visualisierungen führt.

  • Welche Rolle spielten die Tensor Core GPUs und das MVLink Switch Fabric in der Entwicklung von AI?

    -Tensor Core GPUs und das MVLink Switch Fabric sind grundlegende Bausteine für die Beschleunigung von AI-Anwendungen und haben es ermöglicht, Deep-Learning-Modelle effizienter und schneller zu trainieren und zu nutzen.

  • Wie hat sich die Einführung von DLSS (Deep Learning Super Sampling) auf die Computergrafik ausgewirkt?

    -DLSS hat es ermöglicht, hochauflösende, vollständig raytrace-basierte Simulationen mit einer hohen Framerate zu rendern, was durch die Verwendung von AI und generativer KI erreicht wurde, die die Qualität bei einer reduzierten Anzahl von berechneten Pixeln erhält.

  • Was ist der Unterschied zwischen Chat GPT und früheren AI-Modellen?

    -Chat GPT nutzt reinforcement learning with human feedback, um die AI an unseren Kernwerten auszurichten und die gewünschten Fähigkeiten zu erlernen, was zu einer größeren Kontrollbarkeit und Genauigkeit führt.

  • Wie wird die zukünftige Entwicklung von generativer AI durch die Energieeffizienz beeinflusst?

    -Generative AI hat das Potenzial, die Energieeffizienz in vielen Bereichen zu verbessern, indem sie die Notwendigkeit reduziert, Daten aus Datenzentren abzurufen und stattdessen lokal generiert werden kann.

  • Was sind die Vorteile von Domain-Specific Libraries (DSLs) in Bezug auf generative AI?

    -DSLs wie CDNN für generative AI und CDF für SQL-Verarbeitung ermöglichen spezialisierte Beschleunigung für bestimmte Anwendungen, was die Leistung und Effizienz in diesen Bereichen erhöht.

  • Wie plant NVIDIA, die Herausforderungen der Energieintensität von generativer AI zu bewältigen?

    -NVIDIA plant, durch die Entwicklung von Technologien wie Blackwell, die die Anwendungsleistung bei konstanter Energie und Kosten erhöht, sowie durch die Verlagerung von Datenzentren in Bereiche mit überschüssiger Energie, die Herausforderungen zu bewältigen.

  • Was ist das Konzept hinter Omniverse und wie wird es in der generativen AI eingesetzt?

    -Omniverse ist eine Plattform zur Zusammensetzung von multimodalen Daten, die es ermöglicht, Inhalte aus verschiedenen Quellen zu verbinden und zu steuern. In Kombination mit generativer AI kann dies zu einer präziseren und kontrollierteren Erstellung von Inhalten führen.

  • Wie wird die zukünftige Entwicklung von Hardware im Bereich der generativen AI aussehen?

    -Die Entwicklung von Hardware wird sich darauf konzentrieren, die Leistung von generativer AI weiter zu verbessern, indem man neue Prozessoren und Systeme entwirft, die speziell für die Anforderungen von AI-Anwendungen optimiert sind.

Outlines

00:00

🚀 Entwicklung von KI und Deep Learning

Dieser Absatz beschreibt die enorme Beschleunigung im Bereich des Deep Learnings, welche die Schaffung von großen Sprachmodellen und generativer KI ermöglicht hat. Es wird auf bedeutende Meilensteine in der Computerindustrie hingewiesen, darunter die IBM System 360, die Erfindung der modernen Computergrafik, Raytracing und die Programmiersteuerung von Shadern. Der Absatz erwähnt die Gründung von Nvidia durch Chris Curtis und den Redner selbst, die Erfindung des ersten programmierbaren Shader-GPUs und wie diese Technologien zur Entwicklung von Nvidia beigetragen haben. Zudem wird die erste Begegnung mit künstlicher Intelligenz durch AlexNet im Jahr 2012 und die darauf folgende Umorientierung des gesamten Unternehmens auf Deep Learning thematisiert.

05:00

🤖 KI in der Computergrafik und in der Realität

Dieser Absatz konzentriert sich auf die Anwendung von generativer KI in der Computergrafik und wie Nvidia mithilfe von DLSS (Deep Learning Super Sampling) die Leistung von KI zur Verbesserung von Rendering-Geschwindigkeiten genutzt hat. Es wird auch auf die Herausforderungen und die Optimismus bezüglich der Kontrollierbarkeit und Genauigkeit von generativer KI eingegangen. Der Redner erwähnt die drei Durchbruchstechnologien: Verstärkungslernen durch menschliche Feedback, 'Guard railing' zur Einengung der AI-Ausgaben und 'Retrieval-augmented generation' zur Verbesserung der Datensicherheit. Der Absatz endet mit der Einführung von Edify, einem von Nvidia entwickelten Modell, das 2D-Text in 2D-Bilder umwandelt.

10:01

🌐 Die Zukunft der generativen KI und ihre Kontrolle

In diesem Absatz wird auf die stetige Verbesserung der Kontrollierbarkeit und Genauigkeit von generativer KI durch spezifische Modelle und Verfahren eingegangen. Es wird die Verwendung von AI Foundry und Omniverse beschrieben, die es ermöglichen, generative KI mithilfe von Vorlagen und Anpassungen zu steuern. Der Fokus liegt auf der Schaffung von 3D-Modellen und der Kontrolle von Bildern, die aus generierten Texten resultieren. Der Redner diskutiert auch die Bedeutung der Softwareentwicklung für Nvidia und wie diese die KI-Technologie optimieren kann, um die Energieeffizienz zu erhöhen und neue Märkte zu erschließen.

15:01

⚡ Energieverbrauch und zukünftige Entwicklung der KI

Der vierte Absatz thematisiert den Energieverbrauch von generativer KI und wie dies in den kommenden Jahrzehnten eine Herausforderung darstellen könnte. Der Redner diskutiert die stetige Skalierung der KI-Modelle und die damit verbundene steigende Rechenleistung, die für deren Training erforderlich ist. Es wird auch auf die positiven Auswirkungen von generativer KI eingegangen, wie z.B. die Verringerung des Energieverbrauchs durch die Generierung anstatt des Abrufs von Daten. Der Redner betont, dass generative AI dazu beitragen kann, die Energieeffizienz in verschiedenen Sektoren zu verbessern und wie die Zukunft der Datenzentren möglicherweise von der Platzierung in Energieüberschuss-Regionen abhängen wird.

Mindmap

Keywords

💡Deep Learning

Deep Learning ist ein Teilbereich des maschinellen Lernens, der neuronale Netze verwendet, um komplexe Datenmuster zu erkennen und zu lernen. Im Video wird betont, wie die Beschleunigung des Deep Learnings das Erstellen großer Sprachmodelle ermöglicht hat. Ein Beispiel ist AlexNet, das als Durchbruch in der Computer Vision gilt und die Grundlage für neue Software-Schreibweisen bildet.

💡Generative AI

Generative AI bezieht sich auf künstliche Intelligenzen, die in der Lage sind, neue Inhalte zu erzeugen, anstatt nur auf bestehende zu reagieren. Im Video wird dies als eine Technologie beschrieben, die durch Verbesserungen in der Deep Learning-Beschleunigung möglich wurde und die Fähigkeit hat, Branchen wie die wissenschaftliche Computing oder das autonome Fahren zu transformieren.

💡Programmable Shading

Programmable Shading ist eine Technologie, die es erlaubt, die Oberflächen von Objekten in der 3D-Grafik zu beeinflussen. Im Video wird erwähnt, wie dies ursprünglich auf einem Supercomputer durchgeführt wurde und später zur Basis vieler heutiger animierten Filme wurde.

💡Nvidia

Nvidia ist ein führendes Unternehmen im Bereich der Grafikprozessoren und hat maßgeblich zur Entwicklung von Technologien beigetragen, die die Computergrafik und Deep Learning revolutioniert haben. Im Video wird auf die Rolle von Nvidia bei der Erfindung des ersten programmablen Shading GPUs und der Einführung von RTX als Echtzeit-Rendering-Plattform hingewiesen.

💡RTX

RTX ist eine von Nvidia entwickelte Technologie für Echtzeit-Rays-Tracing in der Computergrafik. Im Video wird RTX als ein wichtiger Schritt beschrieben, der es ermöglichte, die Prozesse zu beschleunigen, die für die Erstellung realistischer 3D-Grafiken notwendig sind.

💡DLSS

DLSS, Deep Learning Super Sampling, ist eine Technologie von Nvidia, die AI verwendet, um die Bildqualität in Echtzeit zu verbessern, indem sie ein Pixel rendert und die anderen mithilfe von KI abschätzt. Im Video wird DLSS als Beispiel für die Anwendung von Generative AI in der Grafik genannt.

💡Omniverse

Omniverse ist eine von Nvidia entwickelte Plattform für die Zusammenarbeit und die Erstellung von 3D-Designs und Simulationen. Im Video wird Omniverse als Werkzeug beschrieben, das zur Komposition von multimodalen Daten und zur Kontrolle von Generative AI beiträgt.

💡Energieverbrauch

Der Energieverbrauch von Generative AI und Datenzentren ist ein wichtiges Thema im Video. Es wird diskutiert, wie die fortschreitende Entwicklung von KI-Modellen den Energiebedarf erhöht und wie Generative AI letztendlich dazu beitragen könnte, den Energieverbrauch zu reduzieren, indem sie die Notwendigkeit von Datenabholung in Datenzentren verringert.

💡Tensor Core GPUs

Tensor Core GPUs sind spezielle Prozessoren, die für die Verarbeitung von Tensoroperationen optimiert sind, die im Zusammenhang mit Deep Learning und Generative AI häufig vorkommen. Im Video wird betont, wie diese GPUs eine wichtige Rolle bei der Beschleunigung von AI-Anwendungen spielen.

💡DSL

DSL steht für Domain-Specific Library und bezieht sich auf Bibliotheken, die speziell für bestimmte Anwendungsbereiche entwickelt wurden. Im Video wird erwähnt, wie DSLs wie CDNN für Generative AI oder CDF für SQL-Prozessierung die Beschleunigung von Anwendungen ermöglichen.

Highlights

Beschleunigung des Deep Learning um eine Million Mal, was es ermöglicht hat, große Sprachmodelle zu erstellen.

Millionfache Reduzierung von Kosten und Energie für die Entwicklung von generativer KI.

Einführung des ersten Computers für Deep Learning, der DJX1, für autonome Fahrzeuge, Robotik und generative AI.

Die Transformer-Technologie hat das moderne maschinelles Lernen revolutioniert.

Einführung von RTX, der weltweit ersten Echtzeit-Interaktiv-Rays-Tracing-Plattform.

DLSS (Deep Learning Super Sampling) reduziert die Energieverbrauch durch intelligente Pixel-Generierung.

Chat GPT als schnellstes wachsendes Service, der Industrien weltweit beeinflusst.

Optimismus bezüglich der Kontrollierbarkeit und Genauigkeit von generativer AI durch neue Technologien.

Einführung von Guard railing zur Fokussierung der AI-Antworten auf bestimmte Bereiche.

Retrieval-Augmented Generation zur Verbesserung der Datensicherheit und Autorität in AI-Systemen.

Entwicklung von Edify, einem 2D-Text-zu-2D-Bild-Foundation-Modell für generative AI.

AI Foundry zur Erstellung von Modellen mit Kundendaten für eine bessere Kontrolle und Anpassung.

Omniverse zur Zusammensetzung von multimodalen Daten für eine präzisere Kontrolle von generativer AI.

Erstellung digitaler Charaktere zur Interaktion mit AI und Verbesserung der Benutzererfahrung.

Die Bedeutung von Software-Entwicklung für die Zukunft von NVIDIA und die Erreichung von beschleunigten Anwendungen.

Einführung von DS-Libraries (Domain-Specific Libraries) für generative AI und andere Anwendungen.

Die Herausforderungen der Energieverbrauchseffizienz bei der Entwicklung von generativer AI.

Perspektiven für die Zukunft von Datenzentren in Bezug auf Energieverbrauch und Standortwahl.

Die Rolle von generativer AI in der Steigerung der Produktivität und der Förderung der Energieeffizienz.

Transcripts

play00:00

we've accelerated deep learning by a

play00:01

million times which is the reason why

play00:03

it's now possible for us to create these

play00:05

large language models a million times

play00:08

speed up a million times reduction in

play00:10

cost and energy is what made it possible

play00:12

for us to make General generative AI

play00:14

possible and so I made a cartoon for you

play00:18

I made a cartoon for you of our journey

play00:20

did you make it or did gener of AI I had

play00:22

it made I had it made that's what CEOs

play00:24

do we don't do anything we just have it

play00:26

be

play00:27

done and so this is this cartoon here is

play00:30

really terrific so these are some of the

play00:32

most important moments in the computer

play00:34

industry the IBM system 360 of course

play00:36

the invention of modern Computing the

play00:39

teapot 1975 the Utah teapot 1979 Ray

play00:44

tracing 1986 programmable shading of

play00:48

course most of the animated movies that

play00:50

we see today wouldn't be possible if not

play00:52

for programmable shading originally done

play00:54

on the CRA supercomputer and then in

play00:56

1993 Nvidia was founded Chris Curtis and

play00:59

I founded the company uh

play01:02

1995 Windows PC revolutionized the

play01:06

personal computer industry put a

play01:08

personal computer in every home and

play01:09

every desk 2001 we invented the first

play01:13

programmable shading GPU and that that

play01:16

really drove uh vast majority of

play01:19

nvidia's Journey up to that point But at

play01:21

the background of everything we were

play01:23

doing was accelerated Computing so that

play01:25

you can solve problems that normal

play01:27

computers can't and the application we

play01:30

chose first was computer graphics and it

play01:31

was probably one of the best decisions

play01:33

we ever made because computer Graphics

play01:36

was insanely computationally in

play01:38

intensive and remain so for the entire

play01:40

31 years that that uh Nvidia has been

play01:42

here it was also incredibly high volume

play01:45

because we applied computer Graphics to

play01:48

an application at the time that wasn't

play01:51

mainstream 3D Graphics video games the

play01:54

combination of very large volume very

play01:57

complicated Computing problem led to

play02:00

a very large R&D budget for us which

play02:03

drove the flywheel of our company one

play02:05

day in 2012 we made our first Contact

play02:09

you know Star Trek first Contact with

play02:11

artificial intelligence that first

play02:13

Contact was Alex net and was in 2012

play02:16

very big moment we made the observation

play02:18

that alexnet was an incredible

play02:21

breakthrough in computer vision but at

play02:23

the core of it that it was a new way of

play02:26

writing software instead of Engineers

play02:30

given input imagining what the output

play02:33

was going to be right algorithms we now

play02:36

have a computer that given an input an

play02:40

example outputs would figure out what

play02:43

the program is in the middle that

play02:45

observation and that we can use this

play02:47

technique to solve a whole bunch of

play02:49

problems that previously wasn't solvable

play02:52

was a great observation and we changed

play02:55

everything in our company to pursue it

play02:58

from the processor to the systems to the

play02:59

the software stack all the algorithms

play03:02

Nvidia basic research pivoted towards

play03:04

working on deep learning and so in 20

play03:07

2016 we introduced the first computer we

play03:10

built for deep learning and we called it

play03:12

djx1 and I delivered the first djx1

play03:15

outside of our company I built it for

play03:17

NVIDIA to build models for self-driving

play03:20

cars and Robotics and such and and

play03:21

generative AI for graphics somebody saw

play03:24

an example of djx1 Elon reached out to

play03:27

me and said hey I would love to have one

play03:29

of those for a startup company we

play03:31

starting and so I delivered the first

play03:33

one to a company at the time uh that

play03:34

knew nobody knew about called open Ai

play03:37

and so that was 2016 2017 was the

play03:39

Transformer that revolutionized modern

play03:42

machine learning modern deep learning in

play03:44

2018 right here at sigraph we announced

play03:48

RTX the world's first realtime

play03:51

interactive rate Tracer R tracing

play03:53

platform we call it RTX it was such a

play03:56

big deal that we changed the name of GTX

play03:59

which every body referred to our

play04:00

graphics cards as to RTX and you

play04:03

mentioned last year during your sigraph

play04:05

keynote that RTX R tracing extreme was

play04:09

one of the big important moments when

play04:11

computer Graphics met AI That's right

play04:14

but that had been happening for a while

play04:15

actually so what was so important about

play04:18

RTX in 2018 we made it possible to use a

play04:22

parallel processor to accelerate rate

play04:24

tracing um but even then we were rate

play04:28

tracing at about five frames every

play04:30

second depending on on how many Rays

play04:32

we're talking about tracing and we were

play04:34

doing it at 1080 resolution uh obviously

play04:37

video games need a lot more than that

play04:40

obviously realtime Graphics need more

play04:41

than that this crowd definitely knows

play04:43

what that means but for the folks who

play04:45

are watching online the rendering

play04:46

processes used to take a really long

play04:48

time when you were making something it

play04:49

used to take a Cay supercomputer to

play04:51

render just a few pixels and now we have

play04:54

our RTX to accelerate that rate tracing

play04:56

but it was interactive it was time but

play05:00

it wasn't fast enough to be a video game

play05:02

and so we realized that we needed a big

play05:06

boost probably something along the lines

play05:08

of 20x or so maybe 50x or so boost and

play05:13

the team uh invented dlss which

play05:16

basically renders one pixel while it

play05:19

uses AI to infer a whole bunch of other

play05:22

pixels and so we basically taught an AI

play05:25

that is conditioned on what it saw and

play05:28

then fills in the dots for everything

play05:30

else and now we're able to render fully

play05:33

rate Trace fully path Trace simulations

play05:35

at 4K resolution at 300 frames per

play05:39

second made possible by by Ai and so

play05:42

2018 came along 2022 as we all know chat

play05:46

GPT came out fastest growing service in

play05:49

history just about every industry is

play05:51

going to be affected by this whether

play05:52

it's scientific Computing trying to do a

play05:55

better job uh predicting the weather

play05:57

with uh a lot less energy and very

play06:00

importantly robotics self-driving cars

play06:02

are all going to be transformed by

play06:04

generative AI I've gotten the sense from

play06:06

talking to you recently that you are

play06:08

optimistic that this these generative AI

play06:11

tools will become more controllable more

play06:15

accurate we all know that there are

play06:17

issues with hallucinations low quality

play06:20

outputs that people are using these

play06:22

tools and they're maybe not getting

play06:25

exactly the output that they're hoping

play06:26

for right meanwhile they're using a lot

play06:28

of energy which which we're going to

play06:30

talk

play06:30

about why are you so optimistic about

play06:32

this what is what do you think is

play06:34

pointing Us in the direction of this

play06:36

generative AI actually becoming that

play06:37

much more useful and controllable the

play06:39

big breakthrough of Chad GPT was

play06:42

reinforcement learning human feedback

play06:45

which was the way of using humans to

play06:48

align the AI on our core values or align

play06:51

our AI on the skills that we would like

play06:53

it to perform other breakthroughs have

play06:56

arrived since then Guard railing which

play06:58

causes the AI to focus its energy or

play07:01

Focus its response in a particular

play07:03

domain so that it doesn't wander off and

play07:05

pontificate about all kinds of stuff

play07:07

that you ask it about it would only

play07:09

focus on the things that it's been

play07:10

trained to do align to perform and it

play07:13

has deep knowledge in the third

play07:15

breakthrough is called retrieval

play07:17

augmented generation which basically is

play07:20

data that has been embedded so that we

play07:23

understand the meaning of that data and

play07:25

so it's a more authoritative data set it

play07:28

goes beyond just the trained data set

play07:30

for example it might be all of the

play07:32

articles that you've ever written all of

play07:34

the papers that you've ever written and

play07:36

it could be essentially a a chatbot of

play07:39

you so everything that I've ever written

play07:41

or ever said could be vectorized and

play07:43

then created into a semantic database

play07:46

and then before an AI responds it would

play07:49

search the appropriate content from that

play07:52

Vector database and then augment it in

play07:55

its gener generative process and you

play07:57

think that is one of the most important

play08:00

factors these three combinations really

play08:02

made it possible for us to do that with

play08:04

text now the thing that's really cool is

play08:06

that we are now starting to figure out

play08:08

how to do that with visual right and so

play08:11

if you look at today's generative AI in

play08:13

this particular case this is a edify

play08:16

model that Nvidia created it's a 2d text

play08:19

to 2D Foundation model it's multimodal

play08:22

and we used we partnered with Getty to

play08:25

use their library of data to train an AI

play08:29

model and so this is a a text to uh 2D

play08:33

image and you also created this slide

play08:36

personally right I I had I personally

play08:38

had this slide created and

play08:40

so here's a prompt and this could be a

play08:43

prompt for somebody who owns a brand in

play08:45

this case Coca-Cola it could be a car it

play08:47

could be a luxury product it could be

play08:49

anything you use the prompt and generate

play08:51

the image however it's hard to control

play08:53

this prompt and it may hallucinate it

play08:56

may create it in such a way that it's

play08:58

not exactly what you want and to

play09:00

fine-tune this using words is really

play09:02

hard because it's very imprecise and so

play09:05

the ability for us to now control that

play09:07

image is difficult to do and so we've

play09:09

created a way that allows us to control

play09:12

and align that with more conditioning

play09:15

and so the way you do that is we create

play09:17

another model and it's edify 3D one of

play09:20

our foundation models we've created this

play09:22

AI Foundry where Partners can come and

play09:24

work with us and we create the model for

play09:26

them with their data we invent the model

play09:29

and they bring their data and we uh

play09:31

create a model that they can take with

play09:32

them is it their data only uses their

play09:35

data so this only uses all of the data

play09:37

that's available on Shutterstock that

play09:39

they have have the rights to to use the

play09:41

train and so we now use prompt generator

play09:44

3D we put that in a place where you

play09:46

could compose data and content from a

play09:49

lot of different modalities it could be

play09:50

3D it could be AI it could be uh

play09:52

animation it could be materials and so

play09:55

we use Omniverse to compose all of these

play09:57

multimodality data and now you can

play09:59

control it you could change the pose you

play10:01

could change the placement you could

play10:02

change whatever you like and then you

play10:04

take what comes out of Omniverse you now

play10:06

augment it with the prompt it's a little

play10:08

bit like retrieval augmented generation

play10:11

this is now 3D augmented generation the

play10:15

edified model is multimodal so it

play10:17

understand the image understands the

play10:18

prompt and it uses it in combination to

play10:21

create a new image so now this is a

play10:23

controlled image we can generate images

play10:26

exactly the way we like it just now I

play10:28

showed you Omniverse augmented

play10:31

generation for images this is a rag this

play10:34

is a uh retrieval augmented generative

play10:38

Ai and we've created this digital human

play10:41

front end basically the io of an AI that

play10:44

has the ability to speak make eye

play10:47

contact with you an animate in an

play10:49

empathetic way you could decide to

play10:51

connect your chat gbt or your AI to the

play10:55

digital human or you can connect your

play10:58

digital human to our retrieval augmented

play11:01

generation this breakthrough is really

play11:04

quite incredible and it makes it

play11:05

possible for us amazing Graphics

play11:08

researchers welcome to sigraph 2024 so

play11:11

it makes it possible to animate using an

play11:14

AI you you chat with the AI it generates

play11:17

text that text then is translated to

play11:20

sound text to speech that speech the

play11:23

sound then animates the face and then

play11:25

RTX path tracing does the rendering of

play11:28

the digital human what I hear you

play11:29

talking a lot about today these are

play11:32

software developments right they're

play11:33

relying on your gpus but ultimately this

play11:36

is software this is NVIDIA going further

play11:39

up the

play11:40

stack meanwhile there are some companies

play11:44

some folks in the generative AI space

play11:47

who are in software and cloud services

play11:48

but they're looking to go further down

play11:50

the stack right they might be developing

play11:51

their own chips or tpus that are

play11:53

competitive with what you're doing how

play11:56

crucial is this software strategy to

play11:58

Nvidia maintaining its lead and actually

play12:00

fulfilling some of these promises of

play12:02

growth that people are looking at for

play12:04

NVIDIA right now we've always been a

play12:06

software company and even first and the

play12:09

reason for that is because accelerated

play12:10

Computing is not general purpose

play12:12

Computing general purpose Computing can

play12:14

take any program Python and just run it

play12:16

and almost everybody's uh program can be

play12:19

compiled to run effectively

play12:21

unfortunately when you want to

play12:22

accelerate fluid dynamics you have to

play12:24

understand the the algorithms of fluid

play12:27

dynamics so that you could uh refactor

play12:30

it in such a way that it could be

play12:31

accelerated and you have to design an

play12:33

accelerator you have to design the coua

play12:36

GPU so that it understands the

play12:38

algorithms so that it could do a good

play12:40

job accelerating it and the benefit of

play12:42

course by redesigning the whole stack we

play12:45

can accelerate applications 20 40 50

play12:47

times 100 times over general purpose

play12:50

Computing in the case of deep learning

play12:52

over the course of last 10 to 12 years

play12:54

or so we've accelerated deep learning by

play12:56

a million times which is the reason why

play12:59

now possible for us to create these

play13:00

large language models a million times

play13:03

speed up a million times reduction in

play13:05

cost and energy is what made it possible

play13:07

for us to make General generative AI

play13:10

possible but that's designing a new

play13:12

processor a new system tensor core gpus

play13:16

the mvlink switch fabric is completely

play13:19

groundbreaking for AI and if you don't

play13:22

understand the algorithms the the

play13:23

applications above it it's really hard

play13:25

to figure out how to design that whole

play13:27

stack what is the most important

play13:29

part of Nvidia software ecosystem for

play13:32

nvidia's future it takes a new library

play13:35

we call it dsl's domain specific library

play13:38

in generative AI that DSL is called

play13:41

cdnn uh for SQL processing data frames

play13:45

is called CDF we got a whole bunch of

play13:47

coups every time we introduce a domain

play13:50

specific Library it exposes accelerated

play13:53

Computing to a new market but notice

play13:55

every single time we want to open up a

play13:57

new market like CF in order to do data

play13:59

processing data processing is probably

play14:01

what a third of the world's Computing

play14:03

every company does data processing and

play14:06

most companies data is in data frames in

play14:09

tabular format and so in order to create

play14:11

an acceleration library for tabular

play14:14

formats was insanely hard because what's

play14:16

inside those tables could be floating

play14:18

Point numbers 64-bit integers it could

play14:20

be letters and all kinds of stuff and so

play14:22

we have to figure out a way to go

play14:24

compute all that every single time we

play14:26

open up a new market it just requires us

play14:28

to reinvent everything of that Computing

play14:30

that's the reason why we're working on

play14:31

robotics that's the reason why we're

play14:33

working on autonomous vehicles to

play14:35

understand the algorithms that's

play14:37

necessary to open up that market and to

play14:41

understand the Computing layer

play14:42

underneath it so that we can deliver

play14:44

extraordinary results and so there's

play14:46

nothing easy about it generative AI

play14:49

takes up a lot of energy I'm just saying

play14:50

my job's super

play14:52

hard yeah go ahead let's talk about

play14:55

energy yeah generative AI incredibly

play14:58

energy

play14:59

intensive I am going to read from my

play15:01

note cards here According to some

play15:04

research chat gbt a single query takes

play15:07

up nearly 10 times the electricity to

play15:09

process a single Google search data

play15:11

centers consume 1 to 2% of overall

play15:14

worldwide energy but some say that it

play15:16

could be as much as 3 to 4% some say as

play15:18

much as 6% by the end of the decade data

play15:22

center workloads tripled between 2015

play15:24

and 2019 that was only 2019 AI

play15:27

generative AI is taking up a large

play15:29

portion of all of that is there going to

play15:31

be enough energy to fulfill the demand

play15:34

of what you want to build and do yes and

play15:37

um a couple of observations so first

play15:41

there there are three or four model

play15:43

makers that are pushing to Frontier a

play15:46

couple of years ago they're they're

play15:47

probably three times that many this year

play15:50

that are pushing the frontiers of of

play15:52

models and the size of the models are

play15:55

call it uh twice as large every year and

play15:58

in order to train a model that's twice

play15:59

as large you need more than twice as

play16:01

much data and so the computational load

play16:03

is growing call it a factor of four each

play16:06

year just for simple thinking now that's

play16:08

one of the reasons why Blackwell is so

play16:10

highly anticipated because we

play16:12

accelerated the application so much

play16:15

using the same amount of energy and so

play16:17

this is an example of accelerating

play16:19

applications at constant energy constant

play16:22

cost you're making it cheaper and

play16:23

cheaper now the important thing though

play16:26

is I've only highlighted 10 companies

play16:28

the world has tons of companies Nvidia

play16:30

is selling gpus to a whole lot of

play16:32

companies and a whole lot of different

play16:33

data centers and so question is what's

play16:35

happening at the core the first thing

play16:37

that's actually happening is the end of

play16:41

CPU scaling and the beginning of

play16:43

accelerated Computing text completion

play16:45

speech recognition recommender systems

play16:48

that are used in data centers all over

play16:50

the world everyone is moving from CPUs

play16:53

to accelerated Computing because uh they

play16:56

want to save energy accelerated

play16:57

Computing helps you save so much energy

play17:00

20 times 50 times and doing the same

play17:02

processing generative AI is probably

play17:05

consuming let's pick a very large number

play17:07

probably a 1% or so of the world's

play17:10

energy but remember even if the data

play17:13

centers uh consume 4% of the world the

play17:15

goal of generative AI is not training

play17:17

the goal of generative AI is inference

play17:20

and the inference ideally we create new

play17:23

models for predicting weather predicting

play17:26

new materials allow us to optimize our

play17:30

supply chain reduce the amount of energy

play17:32

consumed and wasted gasoline as we

play17:34

deliver products and so the goal is

play17:36

actually to reduce the energy consumed

play17:38

of the 96% the second thing the next

play17:41

thing I'll say about generative AI is

play17:42

remember in the the traditional way of

play17:45

doing Computing is called retrieval

play17:47

based Computing everything is

play17:48

pre-recorded all the stories are written

play17:50

pre-recorded all the images are

play17:52

pre-recorded all the videos are

play17:53

pre-recorded everything is stored off in

play17:55

a data center somewhere pre-recorded

play17:58

generative AI reduces the amount of

play18:01

energy necessary to go run to a data

play18:03

center over the network re retrieve

play18:06

something and bring it over the network

play18:08

don't forget 60% of the energy is

play18:10

consumed on the internet moving the

play18:13

electrons around moving the bits and

play18:14

bites around and so generative AI is

play18:17

going to reduce the amount of energy on

play18:19

the internet because instead of having

play18:21

to go retrieve the information we can

play18:23

generate it right there on the spot

play18:24

because we understand the context we

play18:26

probably have some uh content on the

play18:29

device and we can generate the response

play18:31

so that you don't have to go retrieve it

play18:33

AI doesn't care where it goes to school

play18:35

today's data centers are built near the

play18:38

power grid where Society is of course

play18:40

because that's where we need it in the

play18:42

future you're going to see data centers

play18:44

being built in different parts of the

play18:45

world where there's excess energy it's

play18:48

just that it costs a lot of money to

play18:50

bring that energy to society maybe it's

play18:53

in a desert maybe it's in places that

play18:56

has a lot of sustainable energy we can

play18:57

put data centers where there's less

play19:00

population and more energy there's a lot

play19:03

of energy in the world and what we need

play19:05

to do is move data centers out closer to

play19:07

where there's excess energy and not put

play19:10

everything near population AI doesn't

play19:12

care where it's trained part i' never

play19:14

heard that phrase before AI doesn't care

play19:15

where it goes to school and that's

play19:16

interesting yeah it's true I'm going to

play19:18

think on that generative AI is going to

play19:19

increase productivity it's going to en

play19:22

enable us to discover new science make

play19:24

things more energy efficient so that

play19:26

accelerated Compu lights just came on

play19:28

because what what we were talking about

play19:29

energy and all of a sudden it's like the

play19:31

Earth was

play19:32

like okay Jensen thank you so much I

play19:35

think we're probably going to get kicked

play19:36

off stage soon thank you everybody we'll

play19:38

be right back

Rate This

5.0 / 5 (0 votes)

Связанные теги
Künstliche IntelligenzComputergrafikEnergieeffizienzDeep LearningNVIDIAAI-TechnologieGenerative AIRTX-RenderingOmniverseInnovation
Вам нужно краткое изложение на английском?