- il y a 2 jours
Watch Party - Nvidia GTC Paris Keynote
Catégorie
🤖
TechnologieTranscription
00:00:01Sous-titrage Société Radio-Canada
00:00:30Tokens transform images into scientific data
00:00:34Charting alien atmospheres
00:00:36And guiding the explorers of tomorrow
00:00:43They probe the Earth's depths
00:00:47To seek out hidden danger
00:00:51They turn potential into plenty
00:00:58And help us harvest our bounty
00:01:05Tokens see disease before it takes hold
00:01:12Cure with precision
00:01:19And learn what makes us tick
00:01:27Tokens connect the dots
00:01:30So we can protect our most noble creatures
00:01:38Tokens decode the laws of physics
00:01:42To move us faster
00:01:49And make our days more efficient
00:01:57Tokens don't just teach robots how to move
00:02:00But to bring joy
00:02:04And comfort
00:02:06Hi, Marota
00:02:07Hi, Anna
00:02:09Are you ready to see the doctor?
00:02:11What's that?
00:02:12It's my enchanted jewel
00:02:19Tokens help us move forward
00:02:24One small step for man
00:02:29Becomes one giant leap for mankind
00:02:34So we can boldly go
00:02:36Where no one has gone before
00:02:49And here
00:02:52Is where it all begins
00:03:10Welcome to the stage
00:03:11NVIDIA founder and CEO
00:03:13Jensen Wong
00:03:22Hello, Paris
00:03:34Bonjour
00:03:38NVIDIA's first GTC in Paris
00:03:42This is incredible
00:03:46Thank you for all the partners
00:03:47That are here with us
00:03:51We have so many people
00:03:52That we work with over the years
00:03:54In fact, we've been in Europe
00:03:55For a very long time
00:03:56Even though this is my first GTC in Paris
00:03:58I have a lot to tell you
00:04:02NVIDIA
00:04:04Once upon a time
00:04:05Wanted to create
00:04:07A new computing platform
00:04:10To do things
00:04:11That normal computers cannot
00:04:14We accelerated
00:04:15The CPU
00:04:16Created a new type of computing
00:04:18Called accelerated computing
00:04:20And one of our first applications
00:04:22Was molecular dynamics
00:04:23We've come a long way since
00:04:28So many different libraries
00:04:29And in fact
00:04:31What makes accelerated computing special
00:04:33Is
00:04:34It's not just
00:04:36A new processor
00:04:37That you compile software to
00:04:40You have to reformulate
00:04:42How you do computing
00:04:43You have to reformulate
00:04:45Your algorithm
00:04:46And it turns out
00:04:47To be incredibly hard
00:04:48For people
00:04:50To reformulate software
00:04:52And algorithms
00:04:53To be highly paralyzed
00:04:54And so we created libraries
00:04:57To help
00:04:58Each market
00:04:59Each domain of application
00:05:01Become accelerated
00:05:02Each one of these libraries
00:05:04Opens up
00:05:06New opportunities
00:05:07For the developers
00:05:08And it opens up
00:05:10New opportunities
00:05:11For growth
00:05:11For us
00:05:12And our ecosystem partners
00:05:14Computational lithography
00:05:16That probably
00:05:17The single most important
00:05:19Applications
00:05:19In semiconductor design today
00:05:21Runs in a factory
00:05:22At TSMC
00:05:24Samsung
00:05:25Large semiconductor fabs
00:05:28Before the chip is made
00:05:30It runs through
00:05:30An inverse physics algorithm
00:05:32Called
00:05:33Coolitho
00:05:34Computational lithography
00:05:36Direct sparse solvers
00:05:39Algebraic multi-grid solvers
00:05:42Coolopt
00:05:44We just open sourced
00:05:45Incredibly exciting
00:05:47Application
00:05:48Library
00:05:49This library
00:05:50Accelerates
00:05:52Decision making
00:05:53To optimize
00:05:54Problems
00:05:56With millions of
00:05:57Variables
00:05:58For millions of
00:05:59Constraints
00:06:00Like traveling salespeople
00:06:01Problems
00:06:02Warp
00:06:03A pythonic
00:06:05Framework
00:06:06For expressing
00:06:08Geometry and physics
00:06:09Solvers
00:06:10Really important
00:06:11QDF
00:06:13QML
00:06:14Structured databases
00:06:16Data frames
00:06:18Classical
00:06:19Machine learning
00:06:20Algorithms
00:06:21QDF
00:06:22Accelerates
00:06:23Spark
00:06:24Zero lines
00:06:25Of code
00:06:26Change
00:06:26QML
00:06:27Accelerates
00:06:28Sidekick
00:06:29Learn
00:06:29Zero lines
00:06:30Of code
00:06:31Change
00:06:31Dynamo
00:06:33And QDNN
00:06:34QDNN
00:06:35Is probably
00:06:35The single
00:06:36Most important
00:06:37Library
00:06:37NVIDIA
00:06:38Has ever created
00:06:39It accelerates
00:06:40The primitives
00:06:41Of deep
00:06:42Neural networks
00:06:43And Dynamo
00:06:45Is our
00:06:45Brand new
00:06:46Library
00:06:46That makes
00:06:47It possible
00:06:48To dispatch
00:06:49Orchestrate
00:06:50Distribute
00:06:51Extremely complex
00:06:52Inference workloads
00:06:54Across an entire
00:06:55AI factory
00:06:57Equivariance
00:06:58And QTensor
00:07:00Tensor contraction
00:07:01Algorithms
00:07:02Equivariance
00:07:04Is for
00:07:04Neural networks
00:07:05That obey
00:07:06The laws of
00:07:07Geometry
00:07:07Such as
00:07:08Proteins
00:07:09Molecules
00:07:10Ariel and
00:07:11Shiona
00:07:11Really important
00:07:13Framework
00:07:13To enable
00:07:14AI
00:07:15To run
00:07:166G
00:07:17Earth 2
00:07:19Earth 2
00:07:20Our simulation
00:07:20Environment
00:07:21For foundation
00:07:23Models
00:07:24Of weather
00:07:25And climate
00:07:26Models
00:07:26Kilometer square
00:07:28Incredibly high
00:07:29Resolution
00:07:30Moni
00:07:31Our framework
00:07:32For medical
00:07:33Imaging
00:07:34Incredibly popular
00:07:35Parabricks
00:07:36Our solver
00:07:37For genomics
00:07:39Analysis
00:07:40Incredibly successful
00:07:41Cuquantum
00:07:43Cuquantum
00:07:43I'll talk about
00:07:45In just a second
00:07:45For quantum
00:07:46Computing
00:07:47And cu
00:07:48Pinumeric
00:07:49Acceleration
00:07:50For numpy
00:07:51And scipy
00:07:52As you can see
00:07:54These are just a few
00:07:55Of the examples
00:07:56Of libraries
00:07:57There are 400
00:07:59There are 400
00:07:59Others
00:07:59Each one of them
00:08:01Accelerates a domain
00:08:02Of application
00:08:03Each one of them
00:08:04Opens up new
00:08:05Opportunities
00:08:06Well one of the
00:08:07Most exciting
00:08:10Most exciting
00:08:11Is
00:08:11CudaQ
00:08:13CudaX
00:08:14Is this suite
00:08:15Of libraries
00:08:15A library suite
00:08:17For accelerating
00:08:18Applications
00:08:19And algorithms
00:08:20On top of
00:08:21Cuda
00:08:21We now have
00:08:23CudaQ
00:08:24CudaQ
00:08:25Is for
00:08:26Quantum
00:08:27Computing
00:08:27For
00:08:28Classical
00:08:29Quantum
00:08:31Quantum
00:08:32Classical
00:08:32Computing
00:08:33Based on
00:08:34GPUs
00:08:34We've been
00:08:36Working on
00:08:36CudaQ
00:08:37Now for
00:08:37Several years
00:08:38And today
00:08:39I can tell you
00:08:40There's an
00:08:41Inflection point
00:08:42Happening
00:08:43In quantum
00:08:43Computing
00:08:44As you know
00:08:45The first
00:08:47Physical
00:08:48Qubit
00:08:48Was demonstrated
00:08:49Some nearly
00:08:5030 years
00:08:50Ago
00:08:51An error
00:08:52Correction
00:08:53Algorithm
00:08:53Was invented
00:08:54In 1995
00:08:56And in
00:08:572023
00:08:57Almost 30
00:08:59Years later
00:08:59The world's
00:09:01First
00:09:01Logical
00:09:03Qubit
00:09:03Was demonstrated
00:09:04By Google
00:09:05Now since
00:09:06Then a couple
00:09:07Of years
00:09:08Later
00:09:08The number
00:09:09Of logical
00:09:09Qubits
00:09:10Which is
00:09:12Represented by
00:09:13A whole lot
00:09:13Of physical
00:09:14Qubits
00:09:14With error
00:09:15Correction
00:09:15The number
00:09:16Of logical
00:09:17Qubits
00:09:17Are starting
00:09:18To grow
00:09:18Well
00:09:20Just like
00:09:20Moore's Law
00:09:21Would
00:09:21I could
00:09:23Totally expect
00:09:2410 times
00:09:25More logical
00:09:26Qubits
00:09:26Every 5
00:09:27Years
00:09:27100 times
00:09:28More logical
00:09:29Qubits
00:09:29Every 10
00:09:30Years
00:09:30Those logical
00:09:31Qubits
00:09:32Would become
00:09:33Better error
00:09:33Corrected
00:09:34More robust
00:09:35Higher performance
00:09:36More resilient
00:09:37And of course
00:09:38Will continue
00:09:39To be scalable
00:09:40Quantum computing
00:09:41Is reaching
00:09:42An inflection
00:09:43Point
00:09:43We've been
00:09:44Working with
00:09:45Quantum computing
00:09:45Companies
00:09:46All over the world
00:09:47In several
00:09:48Different ways
00:09:49But here
00:09:49In Europe
00:09:50There's a large
00:09:51Community
00:09:52I saw
00:09:53Pascal
00:09:54Last night
00:09:54I saw
00:09:55Barcelona
00:09:56Supercomputing
00:09:57Last night
00:09:57It is clear
00:09:58Now
00:10:00We're within
00:10:01Reach
00:10:01Of being
00:10:01Able
00:10:02To apply
00:10:02Quantum
00:10:03Computing
00:10:03Quantum
00:10:04Classical
00:10:05Computing
00:10:05In areas
00:10:06That can
00:10:07Solve
00:10:07Some
00:10:07Interesting
00:10:07Problems
00:10:08In the
00:10:08Coming
00:10:08Years
00:10:10This is
00:10:11A really
00:10:12Exciting
00:10:12Time
00:10:12And so
00:10:13We've been
00:10:14Working
00:10:14With all
00:10:15Of the
00:10:15Supercomputing
00:10:16Centers
00:10:16It's
00:10:16Very clear
00:10:17Now
00:10:17That over
00:10:18The next
00:10:19Several
00:10:19Years
00:10:19Or at
00:10:20Least
00:10:20The next
00:10:20Generation
00:10:21Of
00:10:21Supercomputers
00:10:22Every
00:10:22Single
00:10:22One
00:10:23Of them
00:10:23Will
00:10:23Have
00:10:24A
00:10:24QPU
00:10:24Assigned
00:10:25And
00:10:25QPU
00:10:26Connected
00:10:26To
00:10:26The QPU
00:10:28Will
00:10:28Do
00:10:28Quantum
00:10:28Computing
00:10:29Of
00:10:29Course
00:10:29And
00:10:31The
00:10:31GPU
00:10:31Will
00:10:31Be
00:10:31Used
00:10:32For
00:10:32Pre-processing
00:10:33For
00:10:33Control
00:10:33For
00:10:34Error
00:10:34Correction
00:10:35Which
00:10:35Will
00:10:35Be
00:10:35Intensely
00:10:36Computationally
00:10:37Intensive
00:10:38Post-processing
00:10:39And
00:10:39Such
00:10:39Between
00:10:40The
00:10:40Two
00:10:41Architectures
00:10:41Just
00:10:42As
00:10:42We
00:10:43Accelerated
00:10:44The
00:10:44CPU
00:10:44Now
00:10:45There's
00:10:46QPU
00:10:46Working
00:10:47With
00:10:47The
00:10:47GPU
00:10:48To
00:10:48Enable
00:10:49The
00:10:49Next
00:10:49Generation
00:10:50Of
00:10:50Computing
00:10:51Well
00:10:51We've
00:10:52Today
00:10:52We're
00:10:53Announcing
00:10:53That
00:10:54Our
00:10:55Entire
00:10:56Quantum
00:10:57Algorithm
00:10:58Stack
00:10:58Is
00:10:59Now
00:10:59Accelerated
00:11:00On
00:11:01Grace
00:11:01Blackwell
00:11:02200
00:11:02And
00:11:03The
00:11:03Speed
00:11:03Up
00:11:04Is
00:11:04Utterly
00:11:04Incredible
00:11:05We
00:11:06Work
00:11:06With
00:11:06The
00:11:07Quantum
00:11:07Computing
00:11:08Industry
00:11:08In
00:11:09Several
00:11:09Different
00:11:09Ways
00:11:09One
00:11:10Way
00:11:10Is
00:11:11Using
00:11:11QQuantum
00:11:12To
00:11:13Simulate
00:11:14The
00:11:14Qbits
00:11:14Or
00:11:15Simulate
00:11:16The
00:11:16Algorithms
00:11:17That
00:11:17Run
00:11:17On
00:11:18Top
00:11:18Of
00:11:18These
00:11:19Quantum
00:11:19Computers
00:11:20Essentially
00:11:20Using
00:11:21A
00:11:21Classical
00:11:22Computer
00:11:23To
00:11:23Simulate
00:11:24Or
00:11:25Emulate
00:11:25A
00:11:26Quantum
00:11:26Computer
00:11:26At
00:11:28The
00:11:28Other
00:11:29Extreme
00:11:29Extremely
00:11:30Important
00:11:31Is
00:11:31Cuda
00:11:31Q
00:11:32Basically
00:11:33Inventing
00:11:34A
00:11:34New
00:11:34Cuda
00:11:34That
00:11:35Extends
00:11:36Cuda
00:11:36Into
00:11:36Quantum
00:11:37Classical
00:11:38So
00:11:39That
00:11:39Applications
00:11:39That
00:11:40Are
00:11:40Developed
00:11:41On
00:11:41Cuda
00:11:41Q
00:11:41Can
00:11:42Run
00:11:43Before
00:11:44The
00:11:44Quantum
00:11:45Computer
00:11:45Arrives
00:11:46In
00:11:46An
00:11:46Emulated
00:11:47Way
00:11:47Or
00:11:48After
00:11:48The
00:11:48Quantum
00:11:49Computer
00:11:49Arrives
00:11:50In
00:11:50A
00:11:51Collaborative
00:11:51Way
00:11:52A
00:11:53Quantum
00:11:53Classical
00:11:54Accelerated
00:11:55Computing
00:11:56Approach
00:11:56And
00:11:57So
00:11:58Today
00:11:58We're
00:11:58Announcing
00:11:59Cuda
00:11:59Q
00:11:59Is
00:12:00Available
00:12:00For
00:12:01Grace
00:12:01Blackwell
00:12:02The
00:12:02Ecosystem
00:12:03Here
00:12:03Is
00:12:04Incredibly
00:12:04Rich
00:12:05And
00:12:05Of
00:12:06Course
00:12:06Europe
00:12:07Is
00:12:07Deep
00:12:07With
00:12:08Science
00:12:08And
00:12:08Deep
00:12:08With
00:12:09Super
00:12:09Computing
00:12:09Expertise
00:12:10And
00:12:10Deep
00:12:10With
00:12:11Heritage
00:12:11In
00:12:11This
00:12:11Area
00:12:12And
00:12:12It's
00:12:12Not
00:12:13Surprising
00:12:13To
00:12:14See
00:12:14Quantum
00:12:15Computing
00:12:16Advance
00:12:16Here
00:12:17In
00:12:17The
00:12:18Next
00:12:18Several
00:12:18Years
00:12:18We're
00:12:19Going to
00:12:19See
00:12:19A
00:12:19Really
00:12:19Fantastic
00:12:20Inflection
00:12:21Point
00:12:21So
00:12:22Anyways
00:12:23For
00:12:23All
00:12:23Of
00:12:23The
00:12:23Quantum
00:12:24Computer
00:12:24Industry
00:12:25That
00:12:25Have
00:12:25Been
00:12:26Working
00:12:26On
00:12:26This
00:12:26For
00:12:26Three
00:12:27Decades
00:12:27Now
00:12:28I
00:12:28Congratulate
00:12:29You
00:12:29For
00:12:29Just
00:12:30The
00:12:30Incredible
00:12:30Accomplishment
00:12:31And
00:12:32The
00:12:32Milestones
00:12:32Today
00:12:32Thank
00:12:33You
00:12:42Let's
00:12:43Talk
00:12:43About
00:12:43AI
00:12:44You
00:12:45Might
00:12:45Be
00:12:45Surprised
00:12:47That
00:12:47I
00:12:47Would
00:12:47Have
00:12:47I
00:12:48Be
00:12:48Talking
00:12:48To
00:12:48You
00:12:48About
00:12:49AI
00:12:49The
00:12:50Same
00:12:50The
00:12:51Same
00:12:51GPU
00:12:52That
00:12:53Ran
00:12:54And
00:12:54Enabled
00:12:55All
00:12:56Of
00:12:56These
00:12:56Applications
00:12:57That
00:12:57I
00:12:57Mentioned
00:12:57That
00:12:58Same
00:12:58GPU
00:13:00Enabled
00:13:00Artificial
00:13:01Intelligence
00:13:02To
00:13:03Come
00:13:03To
00:13:03The
00:13:03World
00:13:03Our
00:13:04First
00:13:05Contact
00:13:05Was
00:13:05In
00:13:062012
00:13:06Just
00:13:07Prior
00:13:07To
00:13:07That
00:13:07Working
00:13:09With
00:13:09Developers
00:13:10On
00:13:10A
00:13:10New
00:13:10Type
00:13:10Of
00:13:11Algorithm
00:13:11Called
00:13:11Deep
00:13:11Learning
00:13:12And
00:13:13Enabled
00:13:14The
00:13:14Alex
00:13:14Net
00:13:15Big
00:13:16Bang
00:13:16Of
00:13:17AI
00:13:182012
00:13:18In
00:13:19The
00:13:20Last
00:13:2015
00:13:21Years
00:13:21Or
00:13:22So
00:13:22AI
00:13:23Has
00:13:23Progressed
00:13:23Incredibly
00:13:24Fast
00:13:24The
00:13:25First
00:13:25Wave
00:13:26Of
00:13:26AI
00:13:26Was
00:13:26Perception
00:13:27For
00:13:28Computers
00:13:29To
00:13:29Recognize
00:13:30Information
00:13:31Understand
00:13:32It
00:13:32The
00:13:33Second
00:13:33Wave
00:13:33Which
00:13:34Most
00:13:34Of
00:13:34Us
00:13:34Were
00:13:35Talking
00:13:35About
00:13:35The
00:13:36Last
00:13:36Five
00:13:37Years
00:13:37Or
00:13:37So
00:13:37Was
00:13:38Generative
00:13:38AI
00:13:39It's
00:13:40Multimodal
00:13:40Meaning
00:13:41That
00:13:41An
00:13:42AI
00:13:42Was
00:13:43Able
00:13:43To
00:13:43Learn
00:13:43Both
00:13:44Images
00:13:44And
00:13:45Language
00:13:45Therefore
00:13:46You
00:13:47Could
00:13:47Prompt
00:13:47It
00:13:47With
00:13:48Language
00:13:48And
00:13:48It
00:13:48Could
00:13:49Generate
00:13:49Images
00:13:50The
00:13:51Ability
00:13:51For
00:13:51AI
00:13:52To
00:13:53Be
00:13:53Multimodal
00:13:53As
00:13:54Well
00:13:54As
00:13:55Able
00:13:55To
00:13:55Translate
00:13:56And
00:13:56Generate
00:13:57Content
00:13:57Enabled
00:13:58Generative
00:13:59AI
00:13:59Revolution
00:14:00Generative
00:14:01AI
00:14:01The
00:14:02Ability
00:14:02To
00:14:02Generate
00:14:02Content
00:14:03Is
00:14:03Fundamentally
00:14:04Vital
00:14:05For
00:14:06Us
00:14:06To
00:14:06Be
00:14:06Productive
00:14:07Well
00:14:08We
00:14:08Are
00:14:09Starting
00:14:10A
00:14:10New
00:14:10Wave
00:14:11Of
00:14:11AI
00:14:11In
00:14:12This
00:14:12Last
00:14:12Couple
00:14:13Of
00:14:13Years
00:14:13We
00:14:13Seen
00:14:14Enormous
00:14:15Progress
00:14:15In
00:14:16AI
00:14:16Ability
00:14:18Fundamentally
00:14:19Intelligence
00:14:20Is
00:14:27Planning
00:14:27Planning
00:14:28A
00:14:28Task
00:14:29How
00:14:29To
00:14:29Solve
00:14:29A
00:14:29Problem
00:14:30And
00:14:30Then
00:14:30Executing
00:14:31The
00:14:31Task
00:14:31Perception
00:14:33Reasoning
00:14:34Planning
00:14:34The
00:14:35Fundamental
00:14:36Cycles
00:14:36Of
00:14:36Intelligence
00:14:37It
00:14:38Allows
00:14:38Us
00:14:38To
00:14:39Apply
00:14:39Some
00:14:40Previously
00:14:40Learned
00:14:41Rules
00:14:41To
00:14:42Solve
00:14:42Problems
00:14:42We've
00:14:43Never
00:14:43Seen
00:14:43Before
00:14:44That's
00:14:45Why
00:14:47Intelligent
00:14:47People
00:14:48Are
00:14:48Considered
00:14:48Intelligent
00:14:49To
00:14:49Be
00:14:49A
00:14:50To
00:14:50Take
00:14:50A
00:14:50Complicated
00:14:51Problem
00:14:51Break
00:14:52It
00:14:52Down
00:14:52Step
00:14:52By
00:14:53Step
00:14:54Reason
00:14:55About
00:14:55How
00:14:55To
00:14:56Solve
00:14:56The
00:14:56Problem
00:14:56Maybe
00:14:57Do
00:14:58Research
00:14:58Maybe
00:14:59Go
00:14:59Learn
00:14:59Some
00:14:59New
00:15:00Information
00:15:00Get
00:15:01Some
00:15:01Help
00:15:01Use
00:15:02Tools
00:15:02And
00:15:03Solve
00:15:04Problems
00:15:04Step
00:15:04By
00:15:04Step
00:15:05Well
00:15:06The
00:15:06Words
00:15:06That
00:15:06I
00:15:06Just
00:15:07Described
00:15:07Are
00:15:08Fundamentally
00:15:09Possible
00:15:09Today
00:15:09With
00:15:10What
00:15:10Is
00:15:10Called
00:15:11Agent
00:15:11AI
00:15:11And
00:15:12I'll
00:15:12Show
00:15:12You
00:15:12More
00:15:12In
00:15:13Just
00:15:13A
00:15:13Second
00:15:13In
00:15:14The
00:15:14Physical
00:15:15Implementation
00:15:16Of
00:15:16That
00:15:16The
00:15:16Embodiment
00:15:17Of
00:15:17That
00:15:18Agent
00:15:18AI
00:15:19And
00:15:20The
00:15:20Motion
00:15:20Now
00:15:21The
00:15:21Generative
00:15:22Capability
00:15:23Is
00:15:23Generating
00:15:24Motion
00:15:24Instead
00:15:25Of
00:15:25Generating
00:15:25Videos
00:15:26And
00:15:26Generating
00:15:26Images
00:15:27Or
00:15:27Generating
00:15:27Text
00:15:28This
00:15:29AI
00:15:29Generates
00:15:30Local
00:15:31Motion
00:15:31The
00:15:31Ability
00:15:32To
00:15:32Walk
00:15:32Or
00:15:33Reach
00:15:33Out
00:15:33And
00:15:33Grab
00:15:33Something
00:15:34Use
00:15:34Tools
00:15:34The
00:15:36Ability
00:15:36For
00:15:36AI
00:15:37To
00:15:37Be
00:15:37Embodied
00:15:38In
00:15:38A
00:15:38Physical
00:15:38Form
00:15:39Is
00:15:39Basically
00:15:40Robotics
00:15:41These
00:15:42Capabilities
00:15:43The
00:15:43Fundamental
00:15:44Technology
00:15:44To
00:15:45Enable
00:15:45Agents
00:15:46Which
00:15:47Are
00:15:47Basically
00:15:47Information
00:15:48Robots
00:15:49And
00:15:50Embodied
00:15:51AI
00:15:51Physical
00:15:52Robots
00:15:53These
00:15:54Two
00:15:54Fundamental
00:15:55Capabilities
00:15:55Are
00:15:56Now
00:15:56Upon
00:15:57Us
00:15:57Really
00:15:58Really
00:15:58Exciting
00:15:58Times
00:15:59For
00:15:59AI
00:15:59But
00:16:00It
00:16:00All
00:16:01Started
00:16:02With
00:16:03GeForce
00:16:04And
00:16:05GeForce
00:16:06Brought
00:16:06Computer
00:16:06Graphics
00:16:07This
00:16:07Is
00:16:07The
00:16:07First
00:16:08Accelerated
00:16:09Computing
00:16:09Application
00:16:10We
00:16:10Had
00:16:10Ever
00:16:10Worked
00:16:11On
00:16:11And
00:16:12It
00:16:12Is
00:16:12Incredible
00:16:13How
00:16:13Far
00:16:14Computer
00:16:14Graphics
00:16:15Has
00:16:15Come
00:16:16GeForce
00:16:17Brought
00:16:18Cuda
00:16:18To
00:16:18The
00:16:18World
00:16:19Which
00:16:19Enabled
00:16:20Machine
00:16:21Learning
00:16:21Researchers
00:16:22And
00:16:22AI
00:16:23Researchers
00:16:23To
00:16:24Advance
00:16:24Deep
00:16:25Learning
00:16:26Then
00:16:27Deep
00:16:27Learning
00:16:28Revolutionized
00:16:29Computer
00:16:29Graphics
00:16:30And
00:16:30Made
00:16:31It
00:16:31Possible
00:16:31For
00:16:32Us
00:16:32To
00:16:32Bring
00:16:32Computer
00:16:33Graphics
00:16:33To
00:16:34A
00:16:34Whole
00:16:34New
00:16:34Level
00:16:35Everything
00:16:36I
00:16:36Going
00:16:36To
00:16:36Show
00:16:37You
00:16:37Today
00:16:37Everything
00:16:38I
00:16:38Going
00:16:39To
00:16:39Show
00:16:39You
00:16:40You
00:16:40A
00:16:40Preview
00:16:40Of
00:16:40What I'm
00:16:41Going
00:16:41To
00:16:41Show
00:16:41You
00:16:41Is
00:16:43Computer
00:16:45Simulation
00:16:46Not
00:16:46Animation
00:16:49Photon
00:16:49Simulation
00:16:50Physics
00:16:51Simulation
00:16:52Particle
00:16:52Simulations
00:16:53Everything
00:16:54Everything
00:16:54Everything
00:16:55Is
00:16:55Fundamentally
00:16:56Simulation
00:16:56Not
00:16:57Animation
00:16:58Not
00:16:58Art
00:16:58It
00:16:59Just
00:16:59Looks
00:17:00Incredibly
00:17:00Beautiful
00:17:01Because
00:17:01It
00:17:02Turns
00:17:02Out
00:17:02The
00:17:02World
00:17:03Is
00:17:03Beautiful
00:17:03And
00:17:04It
00:17:04Turns
00:17:04Out
00:17:04Math
00:17:05Is
00:17:05Beautiful
00:17:05So
00:17:06Let's
00:17:06Take
00:17:06A
00:17:07Look
00:17:07To
00:17:35To학
00:17:36Thingству
00:17:37состо
00:17:37Youth⋅
00:18:07...
00:18:37...
00:18:38...
00:18:41...
00:18:42...
00:18:43...
00:18:46...
00:18:47...
00:18:52...
00:18:58...
00:18:59...
00:19:02...
00:19:04...
00:19:14...
00:19:16...
00:19:16...
00:19:19...
00:19:20...
00:19:24...
00:19:25...
00:19:25...
00:19:25...
00:19:26...
00:19:26...
00:19:26...
00:19:27...
00:19:28...
00:19:28...
00:19:29...
00:19:29...
00:19:29...
00:19:29...
00:19:31...
00:19:31...
00:19:32...
00:19:32...
00:19:32...
00:19:33...
00:19:34...
00:19:34...
00:19:35...
00:19:35...
00:19:35...
00:19:37...
00:19:37...
00:19:37...
00:19:39...
00:19:39...
00:19:39...
00:20:05...
00:20:06...
00:20:07...
00:20:35...
00:20:37...
00:20:37...
00:20:37...
00:20:39...
00:20:40...
00:20:42...
00:20:43...
00:20:43...
00:20:45...
00:20:45...
00:20:45...
00:20:46...
00:20:46...
00:20:47...
00:20:48...
00:20:49...
00:20:49...
00:20:49...
00:20:50...
00:20:51...
00:20:51...
00:20:51...
00:20:53...
00:21:23...
00:21:24...
00:21:24...
00:21:24...
00:21:26...
00:21:26pour notre propre mind, avant de produire.
00:21:29Et donc, la machine thinking est vraiment, architecturally,
00:21:32ce que Grace Blackwell était créé.
00:21:34C'était créé un grand GPU.
00:21:37J'ai comparé la G-Force pour une bonne raison.
00:21:40G-Force est un GPU.
00:21:43Et donc est la G-B200.
00:21:45C'est un grand virtual GPU.
00:21:48Nous avons dû dégrader dans un groupe de components,
00:21:52créé un montant de nouvelles technologies,
00:21:55une technologie de la technologie,
00:21:57un très haut de l'énergie,
00:21:59de l'énergie de l'interconnex,
00:22:02pour connecter tous ces chips et systems
00:22:05ensemble dans une virtuale GPU.
00:22:08C'est la version de Hopper.
00:22:12C'est le fameux Hopper.
00:22:15Il y a des 8 GPUs,
00:22:17qui sont connectés ensemble sur NVLink.
00:22:19Ce qui n'est pas shown ici,
00:22:21c'est un CPU tray.
00:22:23Un CPU tray avec dual CPU et system memory,
00:22:26qui sit sur le dessus.
00:22:28Together,
00:22:29ce représentait un node
00:22:31d'un AI super-computer.
00:22:34C'est un peu de 500 millions d'euros.
00:22:36C'est le Hopper système.
00:22:38C'est le système
00:22:39qui nous a mis en face à l'intérieur de l'AI.
00:22:43Et c'est un peu d'allocation pour un très long temps
00:22:47parce que le marché s'arrêtait très vite.
00:22:50Mais c'est le fameux Hopper système.
00:22:52Alors,
00:22:53c'est l'entire système,
00:22:55including le CPU,
00:22:56c'est remplacé par cette
00:22:58Grace Blackwell node.
00:23:00C'est l'un compute tray.
00:23:03ce qui se passe,
00:23:04c'est remplacé par cette système.
00:23:06C'est remplacé tout de suite,
00:23:09et les CPUs sont assวยés,
00:23:11connectés à les GPUs.
00:23:13Donc, vous pouvez voir que 2 CPUs et 4 GPUs,
00:23:16c'est plus performant
00:23:17que cette système,
00:23:20Mais ce qui m'amuse, c'est que nous voulions connecter tous ces systèmes ensemble.
00:23:25Comment vous connectez tous ces systèmes ensemble, c'est vraiment difficile pour nous imaginer.
00:23:29Donc nous désagregons-nous.
00:23:32Ce que nous avons fait, c'est que l'ensemble du motherboard,
00:23:35nous désagregons-nous, c'est ce et ce.
00:23:39C'est le révolutionnaire NVLink système.
00:23:43Scaling out computing n'est pas très difficile.
00:23:46Just connect more CPUs with Ethernet, scaling out is not hard.
00:23:51Scaling up is incredibly hard.
00:23:54You can only build as large of a computer as you can build.
00:23:57The amount of technology and electronics that you can fit into one memory model
00:24:03is incredibly hard to do.
00:24:05And so what we decided to do was we create a new interconnect called NVLink.
00:24:10NVLink is a memory semantics interconnect.
00:24:13It's a compute fabric, not a network.
00:24:16It directly connects to the CPUs of all of these different NVLink systems, compute nodes.
00:24:23This is the switch.
00:24:25Nine of these stands on top.
00:24:31Nine of it sits on the bottom.
00:24:33In the middle are the NVLink switches.
00:24:36And what connects it together is this miracle.
00:24:42This is the NVLink spine.
00:24:45This is 100% copper.
00:24:49Copper coax.
00:24:50It directly connects all of the NVLink chips to all of the GPUs.
00:24:56Directly connected over this entire spine.
00:24:59so that every single 144 Blackwell dies in 72 different packages are talking to each other
00:25:10at the same time without blocking all across this NVLink spine.
00:25:16The bandwidth of this is about 130 terabytes per second.
00:25:32130 terabytes per second.
00:25:34If it's in bits,
00:25:38130 terabytes per second.
00:25:41It is more than the data rate of the peak traffic of the world's entire internet traffic on this backplane.
00:25:51Yeah.
00:25:55So this is how you shrink the internet into 60 pounds.
00:26:02NVLink.
00:26:03And so we did all that.
00:26:04We did all that because the way a computer is considered,
00:26:09the way you think about computers is going to be fundamentally different in the future,
00:26:12and I'll spend more time on this,
00:26:14but it was designed to give Blackwell a giant leap above Hopper.
00:26:20Remember, Moore's Law, Semiconductor Physics,
00:26:24is only giving you about two times more performance every three to five years.
00:26:29How could we achieve 30, 40 times more performance in just one generation?
00:26:35And we need a 30, 40 times more performance because the reasoning models are talking to themselves.
00:26:42Instead of one-shot ChatGPT,
00:26:45it's now a reasoning model,
00:26:47and it generates a ton more tokens when you're thinking to yourself.
00:26:51You're breaking the problem down step by step.
00:26:54You're reasoning.
00:26:55You're trying a whole bunch of different paths.
00:26:57Maybe it's chain of thoughts.
00:26:59Maybe it's tree of thoughts.
00:27:00Best of end.
00:27:01It's reflecting on its own answers.
00:27:03You've probably seen these research models reflecting on the answers,
00:27:08saying, is this a good answer?
00:27:10Can you do better than that?
00:27:11And they, oh yeah, I can do better than that.
00:27:13Goes back and thinks some more.
00:27:14And so those thinking models, reasoning models,
00:27:18achieved incredible performance,
00:27:21but it requires a lot more computational capability.
00:27:25And what net of result, MVLink 72, Blackwell's architecture,
00:27:32resulted in a giant leap in performance.
00:27:34The way to read this is the x-axis is how fast it's thinking.
00:27:40The y-axis is how much the factory can output supporting a whole bunch of users at one time.
00:27:47And so you want the throughput of the factory to be as high as possible,
00:27:51so you could support as many people as possible,
00:27:53so that the revenues of your factory is as high as possible.
00:27:57You want this axis to be as large as possible because the AI is smarter here than it is here.
00:28:06The faster it can think, the more it can think before it answers your answer.
00:28:11And so this has to do with the ASP of the average selling price of the tokens,
00:28:17and this has to do with the throughput of the factories.
00:28:20These two combined, in that corner, is the revenues of the factory.
00:28:26This factory, based on Blackwell, can generate a ton more revenues as a result of the architecture.
00:28:34It is such an incredible thing what we built.
00:28:37We made a movie for you just to give you a sense of the enormity of the engineering
00:28:42that went into building Grace Blackwell.
00:28:44Take a look.
00:28:49Blackwell is an engineering marvel.
00:28:53It begins as a blank silicon wafer.
00:29:00Hundreds of chip processing and ultraviolet lithography steps build up each of the 200 billion transistors,
00:29:08layer by layer, on a 12-inch wafer.
00:29:10The wafer is scribed into individual Blackwell die, tested and sorted, separating the good dies to move forward.
00:29:20The chip-on wafer on substrate process attaches 32 Blackwell dies and 128 HBM stacks on a custom silicon interposer
00:29:30wafer.
00:29:33Metal interconnect traces are etched directly into it, connecting Blackwell GPUs and HBM stacks into each system and package unit,
00:29:43locking everything into place.
00:29:45Then the assembly is baked, molded and cured, creating the Blackwell B200 Superchip.
00:29:52Each Blackwell is stress-tested in ovens at 125 degrees Celsius and pushed to its limits for several hours.
00:30:05Robots work around the clock to pick and place over 10,000 components onto the Grace Blackwell PCB.
00:30:15Meanwhile, custom liquid cooling copper blocks are prepared to keep the chips at optimal temperatures.
00:30:28At another facility, Connect X7 SuperNICs are built to enable scale-out communications,
00:30:35and Bluefield 3 DPUs to offload and accelerate networking, storage and security tasks.
00:30:42All these parts converge to be carefully integrated into GB200 compute trays.
00:30:56NVLink is the breakthrough high-speed link that NVIDIA invented to connect multiple GPUs and scale up into a massive
00:31:05virtual GPU.
00:31:05The NVLink switch tray is constructed with NVLink switch chips, providing 14.4 terabytes per second of all-to-all
00:31:15bandwidth.
00:31:16NVLink spines form a custom blind-mated backplane with 5,000 copper cables connecting all 72 Blackwells or 144 GPU
00:31:26dies into one giant GPU,
00:31:28delivering 130 terabytes per second of all-to-all bandwidth, more than the global Internet's peak traffic.
00:31:36From around the world, parts arrive to be assembled by skilled technicians into a rack-scale AI supercomputer.
00:31:53In total, 1.2 million components, 2 miles of copper cable, 130 trillion transistors, weighing nearly 2 tons.
00:32:06Blackwell is more than a technological wonder.
00:32:09It's a testament to the power of global collaboration and innovation,
00:32:13fueling the discoveries and solutions that will shape our future everywhere.
00:32:18We are driven to enable the geniuses of our time to do their life's work.
00:32:24And we can't wait to see the breakthroughs you deliver.
00:32:37Grace Blackwell Systems, all in production.
00:32:41It is really a miracle.
00:32:42It's a miracle from a technology perspective, but the supply chain that came together to build these GB200 systems,
00:32:49two tons each, we're producing them now, 1,000 systems a week.
00:32:56No one has ever produced mass-produced supercomputers at this scale before.
00:33:01Each one of these racks is essentially an entire supercomputer.
00:33:06Only in 2018, the largest Volta system, the Sierra supercomputer in 2018, is less performant than one of these racks.
00:33:16And that system was 10 MW.
00:33:19This is 100 kW.
00:33:22So the difference generationally between 2018 and now, we've really taken AI supercomputing to a whole new level,
00:33:30and we're now producing these machinery at enormous scales.
00:33:33And this is just the beginning.
00:33:36In fact, what you've seen is just one system, Grace Blackwell.
00:33:42The entire world is talking of this one system, clamoring for it to get deployed into the world's data centers
00:33:48for training and inferencing and generative AI.
00:33:51However, not everybody and not every data center can handle these liquid-cooled systems.
00:33:58Some data centers require enterprise stacks, the ability to run Linux Red Hat, or Nutanix, or VMware.
00:34:06Storage systems from Dell EMC, Hitachi, NetApp, VAST, Weka.
00:34:12So many different storage systems, so many different IT systems.
00:34:16And the management of those has to be done in the way that's consistent with traditional IT systems.
00:34:22We have so many new computers to ramp into production.
00:34:26And I'm so happy to tell you that every single one of these are now in production.
00:34:32You haven't seen them yet.
00:34:33They're all flying off the shelves, flying off the ramps, the manufacturing lines, starting here.
00:34:40DGX Spark enables you to have essentially the Grace Blackwell system on your desktop.
00:34:47In the case of Spark, desktop. In the case of DGX Station, desk side.
00:34:52This way you don't have to sit on a supercomputer while you're developing your software, while you're developing your AI.
00:34:59But you want the architecture to be exactly the same.
00:35:02These systems are identical from an architecture perspective.
00:35:07From a software developer's perspective, it looks exactly the same.
00:35:12The only difference is scale and speed.
00:35:14And then on this side are all the x86 systems.
00:35:17The world's IT organization still prefers x86 and appreciates x86.
00:35:23Wherever they can take advantage of the most advanced AI-native systems, they do.
00:35:27Where they can't and they want to integrate into the enterprise IT systems, we now offer them the ability to
00:35:33do so.
00:35:34One of the most important systems, and it's taken us the longest to build,
00:35:39because the software in the architecture is so complicated,
00:35:42is how to bring the AI-native architecture and infuse it into the traditional enterprise IT system.
00:35:50This is our brand new RTX Pro server.
00:35:54This is an incredible system.
00:35:56The motherboard is completely redesigned.
00:36:02Ladies and gentlemen, Janine Paul.
00:36:11This motherboard looks so simple.
00:36:13And yet, on top of this motherboard are eight SuperNIC switches
00:36:19that connect eight GPUs across a 200 gigabits per second state-of-the-art networking chip
00:36:26that then connects eight of these GPUs and these Blackwell RTX Pro 6000 GPUs.
00:36:33Brand new, just entered into production.
00:36:37Eight of these go into a server.
00:36:38Now, what makes it special?
00:36:40This server is the only server in the world that runs everything the world has ever written and everything NVIDIA
00:36:48has ever developed.
00:36:50It runs AI, Omniverse, RTX for video games.
00:36:58It runs Windows.
00:36:59It runs Linux.
00:37:00It runs Kubernetes.
00:37:01It runs Kubernetes and VMware.
00:37:04It runs basically everything.
00:37:06If you want to stream Windows desktop from a computer to your remote device, no problem.
00:37:12If you want to stream Omniverse, no problem.
00:37:15If you want to run your robotic stack, no problem.
00:37:17Just a QA of this particular machine is insane.
00:37:21The applications that it runs are basically universal.
00:37:25Everything the world's ever developed should run on here.
00:37:28Including, if you're a video gamer, including Crisis.
00:37:34And so...
00:37:39If you can run Crisis, you can run anything.
00:37:44Okay, this is the RTX Pro server, brand new enterprise system.
00:37:51So something is changing.
00:37:55We know that AI is incredibly important technology.
00:37:59We know for a fact now that AI is software that could revolutionize, transform every industry.
00:38:06It can do these amazing things.
00:38:08That we know.
00:38:10We also know that the way you process AI is fundamentally different than the way we used to process software
00:38:17written by hand.
00:38:19Machine learning software is developed differently and it runs differently.
00:38:24The architecture of the systems, the architecture of the software, completely different.
00:38:28The way the networking works, completely different.
00:38:31The way it acts as storage, completely different.
00:38:35So we know that the technology can do different things.
00:38:39Incredible things.
00:38:41It's intelligent.
00:38:42We also know that it's developed in a fundamentally different way.
00:38:46It needs new computers.
00:38:47The thing that's really interesting is what does this all mean to countries, to companies, to society.
00:38:55And this is an observation that we made almost a decade ago that now everyone is awakening to.
00:39:02That in fact these AI data centers are not data centers at all.
00:39:07They're not data centers in the classical sense of a data center.
00:39:11Storing your files that you retrieve.
00:39:13These data centers are not storing our files.
00:39:17It has one job and one job only.
00:39:20To produce intelligent tokens.
00:39:23The generation of AI.
00:39:26These factories of AI look like data centers in the sense that they have a lot of computers inside.
00:39:34But that's where everything breaks down.
00:39:37How it's designed.
00:39:39The scale at which it's manufactured or scaled, designed and built.
00:39:45And how it's used.
00:39:47And how it's orchestrated and provisioned, operated.
00:39:51How you think about it.
00:39:52Now, for example, nobody really thinks about their data center as a revenue generating facility.
00:40:02I said something that everybody goes, yeah, I think you're right.
00:40:07Nobody ever thinks about a data center as a revenue generating facility.
00:40:11But they think of their factories, their car factories, as revenue generating facilities.
00:40:17And they can't wait to build another factory.
00:40:19Because whenever you build a factory, revenue grows shortly after.
00:40:24You could build more things for more people.
00:40:26Those ideas are exactly the same ideas in these AI factories.
00:40:32They are revenue generating facilities.
00:40:35And they are designed to manufacture tokens.
00:40:38And these tokens could be reformulated into productive intelligence for so many industries that AI factories are now part of
00:40:50a country's infrastructure.
00:40:51Which is the reason why you see me running around the world talking to heads of states.
00:40:56Because they all want to have AI factories.
00:41:00They all want AI to be part of their infrastructure.
00:41:03They want AI to be a growth manufacturing industry for them.
00:41:08And this is genuinely profound.
00:41:11And I think we're talking about, as a result of all that, a new industrial revolution.
00:41:17Because every single industry is affected.
00:41:20And a new industry.
00:41:22Just as electricity became a new industry.
00:41:27At first when it was described as a technology and demonstrated as a technology.
00:41:32It was understood as a technology.
00:41:34But then we understood that it's also a large industry.
00:41:37Then there's the information industry.
00:41:39Which we now know as the internet.
00:41:42And both of them, because it affected so many industries, became part of infrastructure.
00:41:46We now have a new industry.
00:41:48An AI industry.
00:41:50And it's now part of the new infrastructure called intelligence infrastructure.
00:41:54Every country, every society, every company will depend on it.
00:41:59And you can see its scale.
00:42:00This is one that's being talked about a lot.
00:42:03This is Stargate.
00:42:04This doesn't look like a data center.
00:42:06It looks like a factory.
00:42:07This is one gigawatts.
00:42:09It will hold about 500,000 GPU dyes.
00:42:12And produce an enormous amount of intelligence that could be used by everybody.
00:42:19Well, Europe has now awakened to the importance of these AI factories.
00:42:25The importance of the AI infrastructure.
00:42:27And I'm so delighted to see so much activity here.
00:42:31This is European telcos building AI infrastructure with NVIDIA.
00:42:37This is the European cloud service providers building AI infrastructure with NVIDIA.
00:42:43And this is the European supercomputing centers building next generation AI supercomputers and infrastructure with NVIDIA.
00:42:50And this is just the beginning.
00:42:53This is in addition to what will come in the public clouds.
00:42:58This is in addition to the public clouds.
00:43:01So indigenous built AI infrastructure here in Europe by European companies for the European market.
00:43:09And then there's 20 more being planned.
00:43:1320 more AI factories and several that are gigawatt gigafactories.
00:43:19In total, in just two years, we will increase the amount of AI computing capacity in Europe by a factor
00:43:30of 10.
00:43:30And so the researchers, the startups, your AI shortage, your GPU shortage will be resolved for you soon.
00:43:39It's coming for you.
00:43:48Now we're partnering with each country to develop their ecosystem.
00:43:54And so we're building AI technology centers in seven different countries.
00:44:00And the goal of these AI technology centers is, one, to do collaborative research, to work with the startups, and
00:44:08also to build the ecosystem.
00:44:09Let me show you what an ecosystem looks like.
00:44:12In the UK, I was just there yesterday, the ecosystems are built on top of the NVIDIA stack.
00:44:18So for example, every single NVIDIA, as you know, NVIDIA is the only AI architecture that's available in every cloud.
00:44:26It's the only computing architecture, aside from x86, that's available everywhere.
00:44:32We're available.
00:44:33We partner with every cloud service provider.
00:44:35We accelerate applications from the most important software developers in the world.
00:44:40Siemens here in Europe, Cadence, Red Hat, ServiceNow.
00:44:45We've reinvented the computing stack.
00:44:48As you know, computing is not just a computer, but it's compute, networking, and storage.
00:44:52Each one of those layers, each one of those stacks has been reinvented.
00:44:56Great partnership with Cisco, who announced a brand new model yesterday at their conference based on NVIDIA.
00:45:02Dell, great partnerships.
00:45:04NetApp, Newtonix, a whole bunch of great partnerships.
00:45:08As I mentioned earlier, the way you develop software has been fundamentally changed.
00:45:14It's no longer just write C program, compile C program, deliver C program.
00:45:20It's now DevOps, MLOps, AI Ops.
00:45:24So that entire ecosystem is being reinvented, and we have ecosystem partners everywhere.
00:45:30And then, of course, solution integrators and providers who could then help every company integrate these capabilities.
00:45:37Well, here in the UK, we have special companies that we work with.
00:45:41Really terrific companies from researchers to developers to partners to help us up-skill the local economy
00:45:51and up-skill the local talent, enterprises that consume the technology, and of course, cloud service providers.
00:45:57We have great partners in the UK.
00:45:59We have great partners in Germany.
00:46:02Incredible, incredible partnerships in Germany.
00:46:05We have great partnerships in Italy.
00:46:07And we, of course, have amazing partnerships here in France.
00:46:21That's right. Go France.
00:46:37President Macron is going to be here later on.
00:46:40We're going to talk about some new announcements.
00:46:43So we have to show some enthusiasm for AI.
00:46:47Okay?
00:46:51There you go.
00:46:53Show him some enthusiasm.
00:46:56So great partnerships here in France.
00:46:59One particular one I want to highlight, our partnership with Schneider.
00:47:03We're building the, even building these AI factories.
00:47:07We build them digitally now.
00:47:09We design them digitally.
00:47:10We build them digitally.
00:47:12We operate them or optimize them digitally.
00:47:15And we will even eventually optimize them and operate them completely digitally in a digital twin.
00:47:20These AI factories are so expensive.
00:47:25$50 billion sometimes.
00:47:27$100 billion in the future.
00:47:29If the utilization of that factory is not at its fullest, the cost to the factory owner is going to
00:47:37be incredible.
00:47:38And so we need to digitalize and use AI wherever we can.
00:47:42Put everything into Omniverse so that we have direct and constant telemetry.
00:47:47We have a great partnership here that we're announcing today.
00:47:50A young company.
00:47:52A CEO I really like.
00:47:55And he's trying to build a European AI company.
00:48:00The name of the company is Mistral.
00:48:11Today we're announcing that we're going to build an AI cloud together here.
00:48:15To deliver their models as well as deliver AI applications for the ecosystem of other AI startups.
00:48:22So that they can use the Mistral models or any model that they like.
00:48:26And so Mistral and us, we're going to be partnering to build a very sizable AI cloud here.
00:48:31And we'll talk about more of it later on today with President Macron.
00:48:43AI technology is moving at light speed.
00:48:47And what I'm showing you here, proprietary models on the left, moving at light speed.
00:48:52However, the open models are also moving at light speed.
00:48:56Only a few months behind.
00:48:59Whether it's Mistral, Lama, Deep Seek R1, R2 coming, Q1.
00:49:07These models are all exceptional.
00:49:10Every single one of them exceptional.
00:49:12And so we've dedicated ourselves over the last several years to apply some of the world's best AI researchers.
00:49:18To make those AI models even better.
00:49:22And we call that NemoTron.
00:49:24Basically what we do is we take the models that are open sourced.
00:49:29And of course, they're all built on NVIDIA anyhow.
00:49:32And so we take those models open sourced.
00:49:34We then post train it.
00:49:36We might do neural architecture search.
00:49:40We might do neural architecture search.
00:49:43Provide it with even better data.
00:49:45Use reinforcement learning techniques.
00:49:48Enhance those models.
00:49:50Give it reasoning capabilities.
00:49:52Extend the context so that it could learn and read more before it interacts with you.
00:49:58Most of these models have relatively short context.
00:50:01And we want it to have enormous context capability.
00:50:04Because we want to use it in enterprise applications.
00:50:07Where the conversation we want to have with it is not available on the internet.
00:50:11It's available in our company.
00:50:12And so we have to load it up with an enormous amount of context.
00:50:15All of that capability is then packaged together into a downloadable NIM.
00:50:21You can come to NVIDIA's website and literally download an API.
00:50:25A state of the art AI model.
00:50:28Put it anywhere you like.
00:50:30And we improve it tremendously.
00:50:33This is an example of NemoTron improvement over LAMA.
00:50:37So LAMA, 8B, 70B, 405B.
00:50:42Improved by our post training capability.
00:50:45Extension of reasoning capability.
00:50:48All the data that we provide.
00:50:50Enhanced it tremendously.
00:50:51We're going to do this generation after generation after generation.
00:50:55And so for all of you who uses NemoTron.
00:50:59You will know that there's a whole slew of other models in the future.
00:51:02And they're open anyway.
00:51:04So if you would like to start from the open model.
00:51:06That's terrific.
00:51:07If you'd like to start with the NemoTron model.
00:51:08That's terrific.
00:51:10And the NemoTron models.
00:51:11The performance is excellent.
00:51:13In benchmarks after benchmarks after benchmarks.
00:51:17NemoTron performance has top of the leaderboard all over the place.
00:51:21And so now you know that you have access to a enhanced open model.
00:51:26That is still open.
00:51:28That is top of the leader chart.
00:51:30And you know that NVIDIA is dedicated to this.
00:51:32And so I will do this for as long as I shall live.
00:51:35Okay?
00:51:41This strategy is so good.
00:51:44This strategy is so good.
00:51:45That the regional model makers.
00:51:49The model builders across Europe.
00:51:50Have now recognized how wonderful this strategy is.
00:51:54And we're partnering together.
00:51:55To adapt.
00:51:56Enhance.
00:51:57Each one of those models.
00:51:58For regional languages.
00:52:01Your data belongs to you.
00:52:04Your data belongs to you.
00:52:06It is the history of your people.
00:52:08The knowledge of your people.
00:52:09The culture of your people.
00:52:10It belongs to you.
00:52:11And for many companies.
00:52:13In the case of NVIDIA.
00:52:15Our data is largely inside.
00:52:1833 years of data.
00:52:20I was looking up this morning.
00:52:22Siemens.
00:52:23180 years of data.
00:52:25Some of it.
00:52:26Written down on.
00:52:28Papyrus.
00:52:31Roland Bush is here.
00:52:33I thought I'd pick on Roland Bush.
00:52:35My good friend.
00:52:36And so.
00:52:37You'll have to digitize that.
00:52:39Before the AI can learn.
00:52:40And so.
00:52:41The data belongs to you.
00:52:43You should use that data.
00:52:45Use an open model.
00:52:46Like Lemotron.
00:52:47And all the tool suites.
00:52:48That we provide.
00:52:49So that you can enhance it.
00:52:50For your own use.
00:52:52We're also announcing.
00:52:54That we have a great partnership.
00:52:55With Perplexity.
00:52:56Perplexity is a reasoning search engine.
00:52:59Yep.
00:53:05The three models I use.
00:53:06I use.
00:53:07Chat GPT.
00:53:08Gemini Pro.
00:53:08And Perplexity.
00:53:09And these three models.
00:53:10I use interchangeably.
00:53:12And Perplexity is fantastic.
00:53:13We're announcing today.
00:53:15That Perplexity will take these.
00:53:17Regional models.
00:53:18And connect it.
00:53:19Right into Perplexity.
00:53:21So that you could.
00:53:22Now.
00:53:23Ask and get questions.
00:53:25In the language.
00:53:26In the culture.
00:53:27In the sensibility.
00:53:28Of your country.
00:53:30Okay.
00:53:30So Perplexity.
00:53:31Regional models.
00:53:38Agentic AI.
00:53:41Agentic AI.
00:53:42Agents.
00:53:43Is a very big deal.
00:53:44As you know.
00:53:45In the beginning.
00:53:46With pre-trained models.
00:53:48People said.
00:53:49But it hallucinates.
00:53:50It makes things up.
00:53:52You're absolutely right.
00:53:54It doesn't have access.
00:53:56To the latest news.
00:53:57News.
00:53:58And data.
00:53:59Information.
00:54:00Absolutely right.
00:54:02It gives up.
00:54:03Without reasoning.
00:54:04Through problems.
00:54:06It's as if.
00:54:07Every single answer.
00:54:07Has to be memorized.
00:54:09From the past.
00:54:10You're absolutely right.
00:54:12All of those things.
00:54:13You know.
00:54:14Why is it trying.
00:54:15To figure out.
00:54:15How to add.
00:54:16Or count.
00:54:17The count numbers.
00:54:18And add numbers.
00:54:19Why doesn't it use.
00:54:20A calculator.
00:54:21You're absolutely right.
00:54:23And so.
00:54:24All of those capabilities.
00:54:26Associated with intelligence.
00:54:27Everybody was able to criticize.
00:54:29But it was absolutely right.
00:54:31Because everybody.
00:54:32Largely understand.
00:54:33How intelligence works.
00:54:34But those.
00:54:36Technologies.
00:54:36Were being built.
00:54:38All around the world.
00:54:39And they were all coming together.
00:54:41From retrieval.
00:54:43Augmented generation.
00:54:44To web search.
00:54:46To multimodal.
00:54:48Understanding.
00:54:49So that you can read.
00:54:50PDFs.
00:54:51Go to a website.
00:54:52Look at the images.
00:54:53And the words.
00:54:54Listen to the videos.
00:54:56Watch the videos.
00:54:59And then.
00:54:59Take all of that.
00:55:01Understanding.
00:55:01Into your context.
00:55:02You could also.
00:55:04Now understand.
00:55:05Of course.
00:55:05A prompt.
00:55:06From almost anything.
00:55:07You could even say.
00:55:08I'm going to ask you a question.
00:55:09But start from this image.
00:55:11I could say.
00:55:12Start from this.
00:55:13Start from this text.
00:55:15Before you answer.
00:55:16Answer the question.
00:55:17Or do what I ask you to do.
00:55:19It then goes off.
00:55:20And reasons.
00:55:21And plans.
00:55:22And evaluates itself.
00:55:24All of those capabilities.
00:55:25Are now integrated.
00:55:26And you can see it.
00:55:27Coming out into the marketplace.
00:55:29All over the place.
00:55:30Agentic AI.
00:55:31Is real.
00:55:32Agentic AI.
00:55:33Is a giant step function.
00:55:35From.
00:55:36One shot AI.
00:55:37The one shot AI.
00:55:38Was necessary.
00:55:39To lay the foundation.
00:55:41So that we can teach the agents.
00:55:43How to be agents.
00:55:44You need some basic understanding.
00:55:46Of knowledge.
00:55:47And basic understanding.
00:55:48Of reasoning.
00:55:49To even be able.
00:55:50To be.
00:55:50Teachable.
00:55:51And so pre-training.
00:55:53Is about.
00:55:54Teachability.
00:55:55Of AI.
00:55:56Post-training.
00:55:57Reinforcement learning.
00:56:00Supervised learning.
00:56:01Human demonstration.
00:56:04Context provision.
00:56:06Generative AI.
00:56:08All of that.
00:56:08Is coming together.
00:56:10To formulate.
00:56:11What is now.
00:56:11Agentic AI.
00:56:12Let's take a look.
00:56:13At one example.
00:56:14Let me show you something.
00:56:15It's built.
00:56:16On perplexity.
00:56:17And it's super cool.
00:56:21AI agents.
00:56:23Are digital assistants.
00:56:25Based on a prompt.
00:56:27They reason through.
00:56:28And break down problems.
00:56:30Into multi-step plans.
00:56:31They use the proper tools.
00:56:33Work with other agents.
00:56:35And use context from memory.
00:56:37To properly execute the job.
00:56:40On Nvidia accelerated systems.
00:56:42It starts with a simple prompt.
00:56:45Let's ask perplexity.
00:56:46To help start a food truck in Paris.
00:56:49First.
00:56:50The perplexity agent.
00:56:52Reasons through the prompt.
00:56:54And forms a plan.
00:56:56Then.
00:56:57Calls other agents.
00:56:58To help tackle.
00:56:59Each step.
00:57:00Using many tools.
00:57:02The market researcher.
00:57:04Reads reviews and reports.
00:57:05To uncover trends.
00:57:07And analyze.
00:57:08The competitive market.
00:57:10Based on this research.
00:57:12A concept designer.
00:57:14Explores local ingredients.
00:57:16And proposes a menu.
00:57:18Complete with prep time estimates.
00:57:20And researches palettes.
00:57:22And generates a brand identity.
00:57:26Then.
00:57:27The financial planner.
00:57:29Uses Monte Carlo simulations.
00:57:31To forecast profitability.
00:57:32And growth trajectory.
00:57:35An operations planner.
00:57:37Builds a launch timeline.
00:57:38With every detail.
00:57:40From buying equipment.
00:57:41To acquiring the right permits.
00:57:44The marketing specialist.
00:57:46Builds a launch plan.
00:57:47With a social media campaign.
00:57:49And even codes.
00:57:51An interactive website.
00:57:52Including a map.
00:57:54Menu.
00:57:54And online ordering.
00:57:57Each agent's work.
00:57:59Comes together.
00:58:00In a final.
00:58:01Packaged proposal.
00:58:02And it all started.
00:58:03From a single prompt.
00:58:13One prompt.
00:58:15One prompt.
00:58:16Like that.
00:58:16On the original.
00:58:18Chatbot.
00:58:19Would have generated.
00:58:20A few hundred tokens.
00:58:22But now.
00:58:22With that one.
00:58:23Single prompt.
00:58:24Into an agent.
00:58:26To solve a problem.
00:58:27It must have generated.
00:58:28Ten thousand times.
00:58:30More tokens.
00:58:31This is the reason.
00:58:32Grace Blackwell.
00:58:33Is necessary.
00:58:33This is the reason.
00:58:34Why.
00:58:35We need performance.
00:58:36And that systems.
00:58:38To be so much.
00:58:38More performance.
00:58:39Generationally.
00:58:40And.
00:58:40This is how.
00:58:41Perplexity builds.
00:58:43Their agents.
00:58:44Every company.
00:58:45Will have to build.
00:58:46Their own agents.
00:58:47It's terrific.
00:58:48You're going to be.
00:58:49Hiring agents.
00:58:50From.
00:58:50Open AI.
00:58:51And.
00:58:52Gemini.
00:58:53And.
00:58:53Microsoft.
00:58:54Co-Pilot.
00:58:56And.
00:58:56Perplexity.
00:58:57And.
00:58:58Mistral.
00:58:58And.
00:58:59There'll be.
00:58:59Agents.
00:59:00That are built.
00:59:00For you.
00:59:01And they might help you.
00:59:02Plan a.
00:59:03Vacation.
00:59:04Or.
00:59:05You know.
00:59:05Go.
00:59:05Do some research.
00:59:07So on.
00:59:07So forth.
00:59:08It.
00:59:08However.
00:59:09If you want to build.
00:59:10A company.
00:59:11You're going to need.
00:59:12Specialized agents.
00:59:13On specialized tools.
00:59:14And.
00:59:15Using specialized tools.
00:59:16And.
00:59:17Specialized skills.
00:59:18And so the question is.
00:59:19How do you build.
00:59:20Those agents.
00:59:20And so we created.
00:59:22A platform for you.
00:59:23We created.
00:59:24A framework.
00:59:24And a set of tools.
00:59:26That you can use.
00:59:26And a whole bunch of partners.
00:59:27To help you do it.
00:59:28It starts with.
00:59:29On the very bottom.
00:59:31On the very bottom.
00:59:32The ability to have.
00:59:33Reasoning models.
00:59:33That I spoke about.
00:59:35NVIDIA's.
00:59:36Nemo.
00:59:38Language models.
00:59:39Are world class.
00:59:40We have.
00:59:41Nemo.
00:59:41Retriever.
00:59:42Which is a.
00:59:43Multimodal.
00:59:44Search engine.
00:59:45Semantic search engine.
00:59:46Incredible.
00:59:47And we built.
00:59:48A blueprint.
00:59:49A demonstration.
00:59:51That is operational.
00:59:53That is essentially.
00:59:53A general agent.
00:59:55We call it.
00:59:56IQ.
00:59:56AI.
00:59:57AIQ.
00:59:58And on top.
00:59:59We have.
01:00:00A suite of tools.
01:00:02That allows you.
01:00:03To onboard.
01:00:04An agent.
01:00:05A general agent.
01:00:06Curate data.
01:00:07To teach it.
01:00:09Evaluate it.
01:00:10Guard rail it.
01:00:12Supervise.
01:00:12Train it.
01:00:14Use reinforcement.
01:00:15Learning.
01:00:15All the way.
01:00:16To deployment.
01:00:17Keep it secure.
01:00:18Keep it safe.
01:00:20That suite.
01:00:21Of toolkits.
01:00:22Is integrated.
01:00:23Those libraries.
01:00:24Are integrated.
01:00:25Into the AI.
01:00:26Ops.
01:00:26Ecosystem.
01:00:27You can come.
01:00:28And download it.
01:00:28From our website.
01:00:29Yourself.
01:00:29As well.
01:00:30But it's largely.
01:00:31Integrated.
01:00:31Into AI.
01:00:32Ops.
01:00:32Ecosystem.
01:00:33From that.
01:00:34You could create.
01:00:35Your own.
01:00:36Special agents.
01:00:38Many companies.
01:00:39Are doing this.
01:00:39This is.
01:00:40Cisco.
01:00:41They announced it.
01:00:42Yesterday.
01:00:43We're building.
01:00:43AI.
01:00:44Platforms.
01:00:45Together.
01:00:45For security.
01:00:46Now.
01:00:47Look at this.
01:00:49AI agents.
01:00:50And not.
01:00:51One model.
01:00:52That does.
01:00:53All of these.
01:00:53Amazing things.
01:00:54It's a collection.
01:00:55A system.
01:00:56Of models.
01:00:56It's a system.
01:00:57Of AI.
01:00:58Large language models.
01:00:59Some of them.
01:01:00Are optimized.
01:01:01For certain.
01:01:02Type of things.
01:01:03Retrieval.
01:01:03As I mentioned.
01:01:04Performing skills.
01:01:05Using a computer.
01:01:07You don't want to.
01:01:08Bundle all of that stuff.
01:01:09Up into one.
01:01:10Giant.
01:01:11You know.
01:01:11Massive.
01:01:12AI.
01:01:13But you break it up.
01:01:14Into small.
01:01:15Things.
01:01:15That you could.
01:01:15Then deploy.
01:01:16CICD.
01:01:17Over time.
01:01:18This is an example.
01:01:19Of Cisco's.
01:01:20Now.
01:01:20The question is.
01:01:21How do you now.
01:01:22Deploy this.
01:01:23Because as I mentioned.
01:01:24Earlier.
01:01:24There.
01:01:25Public clouds.
01:01:26Where Nvidia's.
01:01:27Compute is.
01:01:28There are.
01:01:29Regional clouds.
01:01:30We call them.
01:01:31NCPs.
01:01:32Here.
01:01:32For example.
01:01:33Mistral.
01:01:34You might have something.
01:01:35That is.
01:01:36Private cloud.
01:01:38Because of your security.
01:01:39Requirements.
01:01:40And your data.
01:01:40Data privacy.
01:01:41Requirements.
01:01:42You might even decide.
01:01:43That you have something.
01:01:44On your desk.
01:01:45And so the question is.
01:01:46How do you run.
01:01:47All of these.
01:01:48And sometimes.
01:01:49They were on.
01:01:49Different places.
01:01:50Because these are all.
01:01:51Microservices.
01:01:52These are AIs.
01:01:53That could talk to each other.
01:01:54They could obviously.
01:01:54Talk to each other.
01:01:55Over networking.
01:01:56And so.
01:01:57How do you deploy.
01:01:58All of these microservices.
01:01:59Well.
01:02:00We now have.
01:02:01A great system.
01:02:02I'm so happy.
01:02:03To announce this for you.
01:02:04This is called.
01:02:05Our DGX.
01:02:06Lepton.
01:02:08DGX.
01:02:08Lepton.
01:02:08What you're looking at here.
01:02:10Is.
01:02:10A whole bunch.
01:02:11Of different clouds.
01:02:12Here's the lambda cloud.
01:02:13The AWS cloud.
01:02:15You know.
01:02:15Here's your own.
01:02:17Developers machine.
01:02:18Your own system.
01:02:19Could be a DGX station.
01:02:20NBS.
01:02:21Yoda.
01:02:21And scale.
01:02:22It could be.
01:02:22AWS.
01:02:23It could be.
01:02:24GCP.
01:02:24NVIDIA's architecture.
01:02:26Is everywhere.
01:02:27And so.
01:02:27You could decide.
01:02:29Where you would like.
01:02:30To run your models.
01:02:31You deploy it.
01:02:32Using.
01:02:32One.
01:02:33Super cloud.
01:02:34So.
01:02:35It's a cloud of clouds.
01:02:36Once you get it to work.
01:02:38Once you get this NIMS.
01:02:39Deployed into Lepton.
01:02:41It will go.
01:02:43And.
01:02:44Be hosted.
01:02:46And run.
01:02:47On.
01:02:47The various clouds.
01:02:48That you decide.
01:02:49One.
01:02:50Model.
01:02:51Architecture.
01:02:52One.
01:02:53Deployment.
01:02:54Everywhere.
01:02:55You can even run it.
01:02:56On this little.
01:02:56Tiny machine here.
01:02:58You know.
01:02:59This.
01:03:00This.
01:03:02DGX.
01:03:03Spark.
01:03:05It's.
01:03:06It's.
01:03:10It's.
01:03:11Is.
01:03:11Is.
01:03:12Is it a.
01:03:12Cafe time.
01:03:16Look at this.
01:03:26This lift.
01:03:27Is 2,000 horsepower.
01:03:30This is my favorite little machine.
01:03:33DGX.
01:03:34Spark.
01:03:35The first.
01:03:36The AI supercomputer.
01:03:37We built an AI supercomputer in 2016.
01:03:40It's called the DGX one.
01:03:41It was the first version of everything that I've been talking about.
01:03:46Eight.
01:03:46Volta.
01:03:47GPUs.
01:03:48Connected with.
01:03:49Envy link.
01:03:50It took us.
01:03:52Billions of dollars to build.
01:03:54And.
01:03:54On the day we announced it.
01:03:56DGX one.
01:03:57There were no customers.
01:03:58No interest.
01:04:00No applause.
01:04:01A hundred percent confusion.
01:04:04Why would somebody build a computer like that?
01:04:07Does it run Windows?
01:04:08Nope.
01:04:13And so we built it anyways.
01:04:15Well thank.
01:04:16A young company.
01:04:18A startup.
01:04:19A non-profit startup.
01:04:20In San Francisco.
01:04:21Was so delighted to see the computer.
01:04:23They said.
01:04:24Can we have one?
01:04:25And I thought.
01:04:25Oh my gosh.
01:04:26We sold one.
01:04:27But then I discovered it was a non-profit.
01:04:32But I put a computer.
01:04:34It put a DGX one in my car.
01:04:36And I drove it up to San Francisco.
01:04:38And the name of that company is open AI.
01:04:48I don't know what the life lesson is there.
01:04:51There are a lot of non-profits.
01:04:54You know.
01:04:55So next time.
01:04:56Next time.
01:04:56But maybe the lesson is this.
01:04:58If a developer reaches out to you.
01:04:59Need a GPU.
01:05:01The answer is yes.
01:05:03And so.
01:05:03So.
01:05:04That's right.
01:05:09So imagine you have Lepton.
01:05:11It's in your browser.
01:05:12And you have.
01:05:13You have this.
01:05:14This helm chart.
01:05:16An AI agent that you've developed.
01:05:17And you want to run it here.
01:05:19And parts of it.
01:05:20You want to run in AWS.
01:05:21And parts of it.
01:05:22You want to run.
01:05:23You know.
01:05:23In a regional cloud somewhere.
01:05:24You use Lepton.
01:05:25You deploy your helm chart.
01:05:27And it magically shows up here.
01:05:29Okay.
01:05:30And so.
01:05:30If you would like to run it here.
01:05:32Until you're done.
01:05:32And ready to deploy it.
01:05:34And then deploy it into the cloud.
01:05:35Terrific.
01:05:36But the beautiful thing is.
01:05:37This architecture.
01:05:38Is based on Grace Blackwell.
01:05:41GB10 versus GB200.
01:05:43Versus GB300.
01:05:45And all of these different versions of.
01:05:47But this architecture is exactly.
01:05:49Grace Blackwell.
01:05:50Now this is amazing.
01:05:52So.
01:05:52We're doing this.
01:05:54For Lepton.
01:05:55But.
01:05:56Next.
01:05:58Hugging Face.
01:05:59And NVIDIA has connected.
01:06:02Lepton together.
01:06:03And so.
01:06:04Whenever you're training a model.
01:06:05On Hugging Face.
01:06:06If you would like to deploy it into Lepton.
01:06:08And directly into Spark.
01:06:10No problem.
01:06:11It's just one click.
01:06:13So whether you're training.
01:06:14Or inferencing.
01:06:15We're now.
01:06:16Connected.
01:06:17To Hugging Face.
01:06:19And Lepton.
01:06:20Will help you.
01:06:21Decide where you want to deploy it.
01:06:22Let's take a look at it.
01:06:27Developers need easy and reliable access to compute.
01:06:30That keeps up with their work.
01:06:32Wherever they are.
01:06:33Whatever they're building.
01:06:35DGX Cloud Lepton.
01:06:37Provides on-demand access to a global network of GPUs.
01:06:41Across clouds.
01:06:42Regions.
01:06:43And partners.
01:06:44Like Yoda and Nebius.
01:06:47Multi-cloud GPU clusters.
01:06:49Are managed through a single unified interface.
01:06:55Provisioning is fast.
01:06:57Developers can scale up the number of nodes quickly.
01:07:00Without complex setups.
01:07:01And start training right away.
01:07:03With pre-integrated tools.
01:07:05And training ready infrastructure.
01:07:07Progress is monitored in real time.
01:07:10GPU performance.
01:07:11Convergence.
01:07:12And throughput.
01:07:13Are at your fingertips.
01:07:15You can test your fine-tuned models.
01:07:17Right within the console.
01:07:18DGX Cloud Lepton.
01:07:21Can deploy NIM endpoints.
01:07:23Or your models in multiple clouds or regions.
01:07:25For fast distributed inference.
01:07:29Just like ride sharing apps connect riders to drivers.
01:07:32DGX Cloud Lepton.
01:07:34Connects developers to GPU compute.
01:07:37Powering a virtual global AI factory.
01:07:43DGX Cloud Lepton.
01:07:45DGX Cloud Lepton.
01:07:48Okay, so that's Cisco.
01:07:49This is the way SAP, they're building an AI platform in NVIDIA.
01:07:53SANA is building an AI business application automation on NVIDIA.
01:07:58DeepL is building their language framework and platform on NVIDIA AI.
01:08:05PhotoRoom.
01:08:06A video editing and AI editing platform.
01:08:10Building their platform on NVIDIA.
01:08:13And this is Codo.
01:08:14Used to be I think Codium.
01:08:16Incredible coding agent.
01:08:18Built on NVIDIA.
01:08:19And this is Iola.
01:08:20A voice platform.
01:08:22Built on NVIDIA.
01:08:23And this one is a clinical trial platform.
01:08:27The world's largest automation platform for clinical trials.
01:08:31Built on NVIDIA.
01:08:33And so all of these.
01:08:34All of these basically builds on the same idea.
01:08:39NIMS.
01:08:39That encapsulates and packages up.
01:08:43In a virtual container that you could deploy anywhere.
01:08:46The NemoTron large language model.
01:08:49Or other large language models like Mistral or others.
01:08:52We then integrate libraries.
01:08:54That basically covers the entire life cycle.
01:08:58Of an AI.
01:08:59An AI agent.
01:09:01The way you treat an AI agent.
01:09:02It's a little bit like a digital employee.
01:09:04So your IT department.
01:09:05Would have to onboard them.
01:09:07Fine tune them.
01:09:08Train them.
01:09:09Evaluate them.
01:09:10Keep them guard railed.
01:09:12You know.
01:09:13Keep them secure.
01:09:14And continuously improve them.
01:09:15And that entire framework.
01:09:17Platform is called Nemo.
01:09:19And all of that is now being integrated.
01:09:21Into one application framework.
01:09:23After another.
01:09:24All over the world.
01:09:24This is just an example of a few of them.
01:09:26And then now we make it possible.
01:09:28For you to deploy them.
01:09:29Anywhere.
01:09:30If you want to deploy it in the cloud.
01:09:32You got DGX.
01:09:33You got GB 200s in the cloud.
01:09:35If you want to deploy it on-prem.
01:09:36Because you've got.
01:09:38Uh.
01:09:39Uh.
01:09:40Or Red Hat Linux.
01:09:41Or.
01:09:43Nutanix.
01:09:43And you want to deploy it in your virtual machines.
01:09:45On-prem.
01:09:46You can do that.
01:09:46If you want to deploy it as a private cloud.
01:09:48You could do that.
01:09:49You can deploy it all the way on your DGX Spark.
01:09:52Or DGX Station.
01:09:53No problem.
01:09:54And so.
01:09:55Lepton.
01:09:56Will help you.
01:09:56Do all of that.
01:09:58Let's talk about.
01:10:00Industrial AI.
01:10:01This is one of my favorite moments.
01:10:05This is Roland Bush.
01:10:08He just.
01:10:08This is a really fun moment.
01:10:10He wanted to remind me.
01:10:12That neural computers.
01:10:15Neural network computers.
01:10:16Were invented in Europe.
01:10:19That's this whole slide.
01:10:22I just.
01:10:23It is.
01:10:24It was such a great moment.
01:10:26This is the Synapse One.
01:10:28This is incredible you guys.
01:10:31Synapse One.
01:10:32This is Synapse One.
01:10:341992.
01:10:35It runs neural networks.
01:10:378,000 times faster than CPUs at that time.
01:10:41Isn't it incredible?
01:10:42So this is the world's AI computer.
01:10:48And.
01:10:49And Roland just wants to.
01:10:50Just makes.
01:10:51Never forget that Jensen.
01:10:54Never ever forget that.
01:10:55I said okay.
01:10:56All right.
01:10:56I'll tell.
01:10:57And I'll even tell everybody.
01:10:58Siemens 1992.
01:11:02Siemens 1992.
01:11:03We have a great partnership with Siemens.
01:11:05Siemens.
01:11:06And.
01:11:06Roland Bush.
01:11:07The CEO.
01:11:08Is.
01:11:09Supercharging the company.
01:11:11So that they could.
01:11:12Leap.
01:11:13Completely leap.
01:11:15The last.
01:11:16IT.
01:11:17Industrial Revolution.
01:11:18And fuse.
01:11:20The industrial capabilities of Europe.
01:11:22The industrial capabilities of might of Siemens.
01:11:25With artificial intelligence.
01:11:27And create what is called.
01:11:28The industrial AI revolution.
01:11:31We're partnering with Siemens on so many different fronts.
01:11:33Everything from design.
01:11:35To simulation.
01:11:38To digital twins of factories.
01:11:41To operations of the AI's in the factories.
01:11:45Everything from end to end.
01:11:47And it just reminds us.
01:11:49It reminds me.
01:11:50How.
01:11:51How incredible.
01:11:53The industrial capabilities of Europe is.
01:11:57And what an extraordinary opportunity this is for you.
01:11:59What an extraordinary opportunity.
01:12:02Because AI.
01:12:03Is unlike software.
01:12:05AI.
01:12:05Is really.
01:12:07Really smart software.
01:12:08And this smart software.
01:12:10Can finally do something.
01:12:12That can revolutionize.
01:12:13The very industries that you serve.
01:12:15And so we made.
01:12:16Made a.
01:12:17A love letter.
01:12:18Video.
01:12:19If you will.
01:12:19Let's play it.
01:12:24It began.
01:12:25Here.
01:12:27The first industrial revolution.
01:12:29What's steam engine.
01:12:32And the mechanized loom.
01:12:34Introduced automation.
01:12:37And the advent of factories.
01:12:40And industry.
01:12:41Was born.
01:12:43The age.
01:12:45Of electricity.
01:12:47Ampere.
01:12:49Unraveled electromagnetism.
01:12:52Faraday.
01:12:53Built the first electric generator.
01:12:55And Maxwell laid the foundations.
01:12:58For modern.
01:13:00Electrical engineering.
01:13:02Siemens and Wheatstone's dynamo.
01:13:05The engine.
01:13:07Of electricity.
01:13:09Bringing machines.
01:13:11Trains.
01:13:12Factories.
01:13:14And cities to life.
01:13:16Electrifying the planet.
01:13:19Igniting.
01:13:21Modern.
01:13:21Manufacturing.
01:13:25And today.
01:13:27Born out of the computing and information age.
01:13:30The fourth industrial revolution.
01:13:33The age of AI.
01:13:36Reimagining every part of industry.
01:13:39Across the continent.
01:13:41Industrial AI is taking hold.
01:13:44From design to engineering.
01:13:46You're blazing new trails.
01:13:49Toward understanding.
01:13:51And reinvention.
01:13:54You brought the physical world.
01:13:57Into the virtual.
01:13:59To plan and optimize.
01:14:01The world's modern factories.
01:14:04You're building the next frontier.
01:14:07Where everything that moves is robotic.
01:14:12Every car.
01:14:13An intelligent autonomous agent.
01:14:18And a new collaborative workforce.
01:14:21To help close the global labor shortage gap.
01:14:27Developers across the continent.
01:14:29Are building every type of robot.
01:14:32Teaching them new skills.
01:14:34In digital twin worlds.
01:14:36And robots.
01:14:40Preparing them to work alongside us.
01:14:43In our factories.
01:14:47Warehouses.
01:14:49The operating room.
01:14:51And at home.
01:14:54The fourth industrial revolution.
01:14:57Is here.
01:14:58Right where the first began.
01:15:03What do you think?
01:15:11I love that video.
01:15:12You made it.
01:15:14That's so great.
01:15:15You made it.
01:15:17Well.
01:15:17We're.
01:15:18Working on industrial AI.
01:15:20With.
01:15:21One company after another.
01:15:22This is a BMW.
01:15:24Doing.
01:15:25Building their next generation factory.
01:15:27In Omniverse.
01:15:31In Omniverse.
01:15:31This is a.
01:15:33I don't know how to say it.
01:15:34Can somebody teach me?
01:15:36Boo Jess.
01:15:38Sounds good.
01:15:40Exactly.
01:15:41That's exactly right.
01:15:42Good job.
01:15:43Good job.
01:15:44That's exactly right.
01:15:45They're building of course.
01:15:47Their plants.
01:15:48Digital twins.
01:15:49In Omniverse.
01:15:50This is a key on their.
01:15:51Uh.
01:15:53Uh.
01:15:53Their digital twin for.
01:15:55Uh.
01:15:55Warehouse logistics.
01:15:56This is a.
01:15:58Mercedes Benz.
01:15:59And their digital twins.
01:16:00Of their factories.
01:16:02Built in Omniverse.
01:16:03This is Schaeffler.
01:16:04And their digital twin.
01:16:06Of their.
01:16:07Warehouse.
01:16:08Built in Omniverse.
01:16:10This is your.
01:16:11Train station.
01:16:13Here in France.
01:16:14Building a digital twin.
01:16:16Of their train stations.
01:16:17In Omniverse.
01:16:19In Omniverse.
01:16:19And this is Toyota.
01:16:20Building a digital twin.
01:16:22Of their.
01:16:23Warehouse.
01:16:23In Omniverse.
01:16:24And when.
01:16:25When you build these warehouses.
01:16:26And these factories.
01:16:27In Omniverse.
01:16:28Then you could.
01:16:29You could.
01:16:29You could design it.
01:16:31You can plan it.
01:16:31You can change it.
01:16:33In Greenfield.
01:16:34Is wonderful.
01:16:35In Brownfield.
01:16:36Is wonderful.
01:16:36You could simulate.
01:16:37Its effectiveness.
01:16:38Before you go.
01:16:39And physically lift.
01:16:40And move things around.
01:16:41To discover.
01:16:42It wasn't optimal.
01:16:43And so the ability.
01:16:45To do everything digitally.
01:16:46In a digital twin.
01:16:47Is incredible.
01:16:48But.
01:16:49The question is.
01:16:50Why does the digital twin.
01:16:52Has to look photo real.
01:16:53And why does it.
01:16:54Has to obey.
01:16:55The laws of physics.
01:16:56The reason for that.
01:16:57Is because.
01:16:58We wanted.
01:16:59Ultimately.
01:17:00To be a digital twin.
01:17:01Where a robot.
01:17:02Can learn.
01:17:03How to operate.
01:17:04As a robot.
01:17:05And robots.
01:17:06Rely.
01:17:07On photons.
01:17:08For their perception system.
01:17:10And those photons.
01:17:11Are generated.
01:17:12Through Omniverse.
01:17:13And.
01:17:14Robots.
01:17:15Needs to interact.
01:17:16With the physical world.
01:17:17So that it could.
01:17:18Know whether it's doing.
01:17:19The right things.
01:17:20And do.
01:17:20Learn how to do it properly.
01:17:22And so.
01:17:22These digital twins.
01:17:23Have to look real.
01:17:25And behave.
01:17:26Realistically.
01:17:27Okay.
01:17:27So.
01:17:27That's the reason.
01:17:28Why Omniverse.
01:17:29Was built.
01:17:29This is.
01:17:30This is fantastic.
01:17:32This is a fusion reactor.
01:17:33Digital twin.
01:17:34Incredibly.
01:17:35Complicated.
01:17:36Piece of instrument.
01:17:37As you know.
01:17:37And without AI.
01:17:38The next generation.
01:17:39Fusion reactor.
01:17:40Would not be possible.
01:17:41Well.
01:17:42We're.
01:17:42Announcing today.
01:17:43That we are going to build.
01:17:46The world's first.
01:17:48Industrial.
01:17:48AI Cloud.
01:17:49Here in Europe.
01:17:50Applaudissements
Commentaires