GTC 2025 科技大会关键亮点：18 分钟速览！Nvidia AI 与 高能 GPU 燃爆全场 | 零度解说

00:18:01
https://www.youtube.com/watch?v=TYDZQeWnpUQ

Sintesi

TLDR在2023年GTC大会上，Nvidia创始人兼首席执行官Jensen Huang展现了最新技术的突破，其中包括GeForce 5090显卡和与通用汽车的合作。他强调了人工智能改变计算机图形的能力，以及数据中心架构的升级，介绍了新操作系统Nvidia Dynamo，旨在优化AI工厂的效率。此外，Huang还分享了在机器人技术和安全性方面的进展，展望未来的发展。

Punti di forza

  • 🎉 Nvidia创始人Jensen Huang在GTC大会上发表演讲！
  • 💻 最新发布的GeForce 5090显卡比前代更小、更高效。
  • 🚗 Nvidia与通用汽车合作开发未来自动驾驶汽车。
  • 🔒 Nvidia首次实现代码的安全性评估，确保透明性和可解释性。
  • 🌐 数据中心架构升级，采用液冷技术。
  • 🎮 Omniverse用于构建物理AI的新操作系统。
  • 🖥️ 新操作系统Nvidia Dynamo将驱动AI工厂。
  • 🤖 机器人技术的进步与新物理引擎Newton的合作。
  • 📈 预计到本世纪末，全球将缺乏5000万劳动力。
  • 🚀 Nvidia的未来展望充满机遇！

Linea temporale

  • 00:00:00 - 00:05:00

    在Nvidia的年度GTC大会上，创始人兼首席执行官詹森·黄分享了公司过去一年的成就，特别是新推出的GeForce 5090显卡，该显卡在体积和能效方面都有显著提升。黄还介绍了人工智能如何在计算图形学中发挥重要作用，并指出Nvidia与通用汽车合作开发自主驾驶技术的计划。

  • 00:05:00 - 00:10:00

    黄强调了汽车安全的重要性，介绍了Nvidia在软件和算法的开发中以安全、透明和可解释性为核心理念的做法。同时，他还展示了Nvidia的最新数据中心架构，采用液冷技术和强大的计算能力，实现了前所未有的计算规模和效率。

  • 00:10:00 - 00:18:01

    Nvidia还在机器人技术方面取得了突破，推出了Omniverse操作系统，并展示了新开发的Cosmos生成模型和与DeepMind及迪士尼研究合作的Newton物理引擎。这将为自动化和智能机器人提供更强大的支持，并应对未来人力资源短缺的问题。

Mappa mentale

Video Domande e Risposte

  • Nvidia最近发布了哪个显卡？

    Nvidia最近发布了GeForce 5090显卡。

  • Nvidia与哪家公司合作开发自动驾驶汽车？

    Nvidia与通用汽车（GM）合作开发自动驾驶汽车。

  • Nvidia Dynamo是什么？

    Nvidia Dynamo是一个新的操作系统，旨在支持AI工厂的运作。

  • Nvidia在数据中心技术方面的进步是什么？

    Nvidia在数据中心技术方面的进步包括从集成MVlink过渡到分布式MVlink，以及液冷技术的应用。

  • Nvidia的Omniverse是什么？

    Omniverse是Nvidia为物理AI构建的操作系统。

Visualizza altre sintesi video

Ottenete l'accesso immediato ai riassunti gratuiti dei video di YouTube grazie all'intelligenza artificiale!
Sottotitoli
en
Scorrimento automatico:
  • 00:00:03
    [Music]
  • 00:00:05
    s
  • 00:00:10
    [Music]
  • 00:00:49
    tell me that wasn't
  • 00:00:54
    amazing welcome to the stage Nvidia
  • 00:00:57
    founder and CEO Jensen Wong
  • 00:01:01
    [Music]
  • 00:01:10
    welcome to
  • 00:01:13
    [Music]
  • 00:01:18
    GTC what an amazing
  • 00:01:21
    year we wanted to do this at
  • 00:01:25
    Nvidia so through the magic of
  • 00:01:27
    artificial
  • 00:01:28
    intelligence we're going to to bring you
  • 00:01:31
    to nvidia's
  • 00:01:35
    headquarters I think I'm bringing you to
  • 00:01:37
    nvidia's
  • 00:01:43
    headquarters what do you
  • 00:01:45
    think this
  • 00:01:47
    is this is where we work this is where
  • 00:01:50
    we work what an amazing year it was and
  • 00:01:53
    we have a lot of incredible things to
  • 00:01:54
    talk about and I just want you to know
  • 00:01:57
    that I'm up here without a net there are
  • 00:02:00
    no scripts there's no teleprompter and
  • 00:02:02
    I've got a lot of things to cover so
  • 00:02:04
    let's get started first of all I want to
  • 00:02:06
    thank all of the sponsors all the
  • 00:02:08
    amazing people who are part of this
  • 00:02:10
    conference just about every single
  • 00:02:12
    industry is
  • 00:02:14
    represented healthc care is here
  • 00:02:16
    Transportation
  • 00:02:18
    retail gosh the computer industry
  • 00:02:22
    everybody in the computer industry is
  • 00:02:24
    here and so it's really really terrific
  • 00:02:27
    to see all of you and thank you for
  • 00:02:28
    sponsoring it
  • 00:02:30
    GTC started with GeForce it all started
  • 00:02:35
    with GeForce and
  • 00:02:37
    today I have here a GeForce
  • 00:02:41
    5090 and
  • 00:02:43
    5090 unbelievably 25 years
  • 00:02:47
    later 25 years after we started working
  • 00:02:50
    on GeForce GeForce is sold out all over
  • 00:02:54
    the world this is the 5090 the Blackwell
  • 00:02:57
    generation and comparing it to the 40 9
  • 00:03:00
    you look how it's 30% smaller in volume
  • 00:03:04
    it's
  • 00:03:05
    30% better at dissipating energy and
  • 00:03:10
    incredible performance hard to even
  • 00:03:12
    compare and the reason for that is
  • 00:03:14
    because of artificial
  • 00:03:16
    intelligence GeForce brought Cuda to the
  • 00:03:19
    world Cuda enabled Ai and AI has now
  • 00:03:23
    come back to revolutionize computer
  • 00:03:26
    Graphics what you're looking at is
  • 00:03:28
    realtime computer Graphics 100% path
  • 00:03:31
    traced for every pixel that's rendered
  • 00:03:36
    artificial intelligence predicts the
  • 00:03:38
    other
  • 00:03:39
    15 think about this for a second for
  • 00:03:42
    every pixel that we mathematically
  • 00:03:44
    rendered artificial intelligence
  • 00:03:47
    inferred the other 15 and it has to do
  • 00:03:50
    so with so much Precision that the image
  • 00:03:53
    looks right and it's temporarily
  • 00:03:56
    accurate meaning that from frame to
  • 00:03:58
    frame to frame going forward backwards
  • 00:04:00
    because it's computer Graphics it has to
  • 00:04:02
    stay temporarily stable incredible
  • 00:04:06
    today I'm super excited to
  • 00:04:09
    announce that GM has selected Nvidia to
  • 00:04:12
    partner with them to build their future
  • 00:04:15
    self-driving car
  • 00:04:22
    Fleet the time for autonomous vehicles
  • 00:04:25
    has arrived and we're work looking
  • 00:04:28
    forward to building with GM M AI in all
  • 00:04:32
    three areas AI for manufacturing so they
  • 00:04:35
    can revolutionize the way they
  • 00:04:36
    manufacture AI for Enterprise so they
  • 00:04:39
    can revolutionize the way they work
  • 00:04:41
    design cars and simulate cars and and
  • 00:04:43
    then also AI for in the car so AI
  • 00:04:47
    infrastructure for GM partnering with GM
  • 00:04:50
    and building with GM their AI so I'm
  • 00:04:53
    super excited about that one of the
  • 00:04:54
    areas that I'm deeply proud of and it
  • 00:04:57
    rarely gets any attention
  • 00:05:00
    is safety Automotive Safety it's called
  • 00:05:06
    halos in our companies call
  • 00:05:08
    Halos safety
  • 00:05:12
    requires technology from Silicon the
  • 00:05:15
    systems the system software the
  • 00:05:20
    algorithms the
  • 00:05:23
    methodologies everything from diversity
  • 00:05:27
    to ensuring diversity
  • 00:05:30
    monitoring and
  • 00:05:32
    transparency
  • 00:05:35
    explainability all of these different
  • 00:05:37
    philosophies have to be deeply ingrained
  • 00:05:39
    into every single part of how you
  • 00:05:41
    develop the system and the software
  • 00:05:44
    we're the first company in the world I
  • 00:05:45
    believe to have every line of code
  • 00:05:48
    safety assessed 7 million lines of code
  • 00:05:51
    safety assessed our chip our system our
  • 00:05:54
    system software and our algorithms are
  • 00:05:57
    safety assessed by Third parties that
  • 00:06:01
    crawl through every line of code to
  • 00:06:04
    ensure that it is designed to ensure
  • 00:06:08
    diversity transparency and
  • 00:06:11
    explainability all right let's talk
  • 00:06:12
    about data
  • 00:06:23
    centers that's not bad huh uh blackw is
  • 00:06:27
    in full production and this is what it
  • 00:06:29
    looks like like it's an incredible
  • 00:06:32
    incredible you know for for people for
  • 00:06:35
    us this is a sight of beauty would you
  • 00:06:38
    agree this
  • 00:06:41
    is how how is this not beautiful how is
  • 00:06:45
    this not beautiful well this is a big
  • 00:06:49
    deal
  • 00:06:50
    because we made a fundamental transition
  • 00:06:55
    in computer architecture and so the
  • 00:06:58
    thing that we had to do was scale up
  • 00:06:59
    first well this is the way we scaled up
  • 00:07:02
    I'm not going to lift this this is this
  • 00:07:03
    is 70 lbs this is the the the last
  • 00:07:06
    generation system architecture is called
  • 00:07:09
    hgx this revolutionize Computing As We
  • 00:07:12
    Know It This revolutionize artificial
  • 00:07:14
    intelligence this is eight
  • 00:07:16
    gpus eight gpus each one of them is kind
  • 00:07:19
    of like this okay this this is two gpus
  • 00:07:24
    two Blackwell gpus in one Blackwell
  • 00:07:28
    package two black wall gpus and one
  • 00:07:30
    black black black wall package and um uh
  • 00:07:33
    there are eight of these underneath
  • 00:07:36
    this okay and this connects into what we
  • 00:07:39
    call MV link
  • 00:07:41
    8 this then connects to a CPU
  • 00:07:45
    shelf like that so there's dual CPUs and
  • 00:07:49
    that sits on top and we connect it over
  • 00:07:51
    PCI Express and then many of these get
  • 00:07:54
    connected with
  • 00:07:56
    infiniband which turns into uh what is
  • 00:07:59
    an AI supercomputer this is the way it
  • 00:08:02
    was in the past we need to disaggregate
  • 00:08:05
    the mvlink system and take it out so
  • 00:08:07
    this is the mvy link system okay this is
  • 00:08:10
    an MV link
  • 00:08:11
    switch this is the most this is the
  • 00:08:13
    highest performance switch the world's
  • 00:08:15
    ever made and this makes it possible for
  • 00:08:17
    every GPU to talk to every GPU at
  • 00:08:21
    exactly the same time at full bandwidth
  • 00:08:24
    okay so this is the mvlink switch we
  • 00:08:26
    disaggregated it we took it out and we
  • 00:08:29
    put put it in the center of the
  • 00:08:32
    chassis so there's all the there 18 of
  • 00:08:35
    these switches in nine different racks
  • 00:08:39
    nine different
  • 00:08:41
    switch trays we call them and then the
  • 00:08:45
    switches are disaggregated the compute
  • 00:08:47
    is now sitting in here this is
  • 00:08:50
    equivalent to these two things in
  • 00:08:52
    compute what's amazing is this is
  • 00:08:55
    completely liquid cooled and by liquid
  • 00:08:57
    cooling it we can compress
  • 00:09:01
    all of these compute nodes into one rack
  • 00:09:04
    this is the big change of the entire
  • 00:09:06
    industry all of you in the audience I
  • 00:09:08
    know how many of you are here I want to
  • 00:09:10
    thank thank you for making this
  • 00:09:12
    fundamental shift from integrated MV
  • 00:09:16
    link to disaggregated MV link from air
  • 00:09:22
    cooled to liquid
  • 00:09:25
    cooled from 60,000 components
  • 00:09:30
    per computer or so to 600,000 components
  • 00:09:35
    per
  • 00:09:39
    rack
  • 00:09:40
    120
  • 00:09:42
    kilow fully liquid cooled and as a
  • 00:09:45
    result we have a one Exel flops computer
  • 00:09:50
    in one rack isn't it
  • 00:09:55
    incredible the way to solve this problem
  • 00:09:57
    is to disaggregate it as described into
  • 00:10:01
    the grace Blackwell MV link
  • 00:10:04
    72 rack but as a result we have done the
  • 00:10:08
    ultimate scale up this is the most
  • 00:10:12
    extreme scale up the world has ever done
  • 00:10:15
    the amount of computation that's
  • 00:10:17
    possible here the memory bandwidth 570
  • 00:10:22
    terabytes per second everything is
  • 00:10:25
    everything in this machine is now in
  • 00:10:27
    te's everything's a trillion and you
  • 00:10:31
    have uh an exit flops which is a million
  • 00:10:35
    trillion floating Point operations per
  • 00:10:38
    second today we're announcing the Nvidia
  • 00:10:47
    Dynamo Nvidia Dynamo does all that it is
  • 00:10:50
    essentially the operating system of an
  • 00:10:53
    AI
  • 00:10:54
    Factory whereas in the past in the way
  • 00:10:57
    that we ran data centers our operating
  • 00:10:59
    system would be something like VMware
  • 00:11:01
    and we would orchestrate a and we still
  • 00:11:04
    do um you know we're big user
  • 00:11:06
    orchestrate a whole bunch of different
  • 00:11:07
    Enterprise applications running on top
  • 00:11:10
    of our Enterprise
  • 00:11:13
    it but in the future the application is
  • 00:11:17
    not Enterprise it it's agents and the
  • 00:11:20
    operating system is not something like
  • 00:11:22
    VMware it's something like Dynamo and
  • 00:11:25
    this operating system is running on top
  • 00:11:27
    of not a data center but on top of an a
  • 00:11:30
    Factory now we call it Dynamo for a good
  • 00:11:33
    reason as you know the Dynamo was the
  • 00:11:36
    first instrument that started the last
  • 00:11:39
    Industrial Revolution the industrial
  • 00:11:41
    revolution of energy water comes in
  • 00:11:44
    electricity comes out it's pretty
  • 00:11:46
    fantastic you know water comes in you
  • 00:11:48
    light it on fire turned to steam and it
  • 00:11:51
    what comes out this this invisible thing
  • 00:11:53
    that's incredibly valuable it took
  • 00:11:55
    another 80 years to go to alternate and
  • 00:11:57
    current but Dynamo Dynamo is the where
  • 00:12:00
    it all started okay so we decided to
  • 00:12:02
    call this operating system this piece of
  • 00:12:04
    software insanely complicated software
  • 00:12:07
    the Nvidia Dynamo this is what a PC
  • 00:12:09
    should look
  • 00:12:11
    like 20 pedop
  • 00:12:14
    flops unbelievable 72 CPU
  • 00:12:17
    cores chipto chip interface hbm memory
  • 00:12:22
    and just just in case some PCI Express
  • 00:12:26
    slots for your uh GeForce
  • 00:12:30
    okay so so this uh is called djx station
  • 00:12:33
    djx spark and djx station are going to
  • 00:12:35
    be available by all of the oems HP Dell
  • 00:12:40
    Lenovo Asus uh it's going to be
  • 00:12:42
    manufactured uh for data scientists and
  • 00:12:45
    researchers all over the world this is
  • 00:12:47
    the computer of the age of
  • 00:12:50
    AI this is what computers should look
  • 00:12:52
    like and this is what computers will run
  • 00:12:54
    in the future and we have a whole lineup
  • 00:12:56
    for Enterprise now from little tiny ones
  • 00:12:59
    to workstation ones the server ones to
  • 00:13:03
    uh supercomputer ones and these will be
  • 00:13:04
    available by all of our partners let's
  • 00:13:07
    go talk about robotics shall
  • 00:13:09
    [Music]
  • 00:13:09
    [Applause]
  • 00:13:11
    we let's talk about
  • 00:13:15
    robots well the time has come the time
  • 00:13:18
    have has come for
  • 00:13:20
    robots uh robots have the benefit the
  • 00:13:23
    benefit of being able to interact with
  • 00:13:25
    the physical world and do things that
  • 00:13:26
    otherwise digital information cannot uh
  • 00:13:29
    we know very clearly that the world is
  • 00:13:32
    has severe shortage of of human labor
  • 00:13:35
    human workers by the end of this decade
  • 00:13:37
    the world is going to be at least 50
  • 00:13:40
    million workers short we'd be more than
  • 00:13:42
    delighted to pay them each $50,000 to
  • 00:13:45
    come to work we're probably going to
  • 00:13:47
    have to pay robots $50,000 a year to
  • 00:13:49
    come to work and so this is going to be
  • 00:13:51
    a very very large industry we created a
  • 00:13:53
    system called Omniverse it's our
  • 00:13:55
    operating system for physical AIS you've
  • 00:13:57
    heard me talk about Omniverse for long
  • 00:13:59
    time we added two technologies to it
  • 00:14:02
    today I'm going to show you two things
  • 00:14:05
    one of them is so that we could scale AI
  • 00:14:09
    with generative capabilities and
  • 00:14:12
    generative model that understand the
  • 00:14:14
    physical world we call it Cosmos using
  • 00:14:18
    Omniverse to condition Cosmos and using
  • 00:14:22
    Cosmos to generate an infinite number of
  • 00:14:27
    environments allows us to create data
  • 00:14:30
    that is grounded grounded controlled by
  • 00:14:35
    us and yet be systematically infinite at
  • 00:14:39
    the same time okay so you see Omniverse
  • 00:14:42
    we use candy colors to give you an
  • 00:14:45
    example of us controlling the robot in
  • 00:14:48
    the scenario perfectly and yet o Cosmos
  • 00:14:53
    can create all these virtual
  • 00:14:54
    environments the second thing just as we
  • 00:14:57
    were talking about earlier one of the
  • 00:15:00
    incredible scaling capabilities of
  • 00:15:01
    language models today is reinforcement
  • 00:15:04
    learning verifiable rewards the question
  • 00:15:08
    is what's the verifiable rewards in
  • 00:15:11
    robotics and as we know very well is the
  • 00:15:14
    laws of physics verifiable physics
  • 00:15:18
    rewards and so we need an incredible
  • 00:15:20
    physics engine well most physics engines
  • 00:15:23
    have been designed for a variety of
  • 00:15:25
    reasons they could be designed because
  • 00:15:26
    we want to use it for large Machin iies
  • 00:15:29
    or maybe we design it for uh virtual
  • 00:15:32
    worlds video games and such but we need
  • 00:15:34
    a physics engine that is designed for
  • 00:15:38
    very fine
  • 00:15:39
    grain rigid and soft bodies designed for
  • 00:15:44
    being able to train tactile feedback and
  • 00:15:48
    fine motor skills and actuator controls
  • 00:15:52
    we needed to be GPU accelerated so that
  • 00:15:55
    we these Virtual Worlds could live in
  • 00:15:58
    super linear time super real time and
  • 00:16:02
    train these AI models incredibly fast
  • 00:16:05
    and we needed to be integrated
  • 00:16:07
    harmoniously into a framework that is
  • 00:16:10
    used by roboticist all over the world
  • 00:16:13
    Moko and so today we're announcing
  • 00:16:16
    something really really special it is a
  • 00:16:19
    partnership of three
  • 00:16:21
    companies Deep
  • 00:16:23
    Mind Disney research and Nvidia and we
  • 00:16:27
    call it Newton
  • 00:16:33
    let's let's take a look at
  • 00:16:42
    Newton tell me that wasn't
  • 00:16:46
    amazing hey
  • 00:16:48
    blue how you
  • 00:16:50
    doing how do you like how do you like
  • 00:16:53
    your new physics engine you like it
  • 00:16:55
    huh yeah I bet I know tactile
  • 00:16:59
    feedback rigid body soft body
  • 00:17:03
    simulation super real
  • 00:17:06
    time can you imagine just now what you
  • 00:17:08
    were looking at is complete realtime
  • 00:17:11
    simulation this is how we're going to
  • 00:17:12
    train robots in a
  • 00:17:14
    future uh just so you know blue has uh
  • 00:17:17
    two computers two Nvidia computers
  • 00:17:23
    inside look how smart you
  • 00:17:26
    are yes you're smart
  • 00:17:30
    okay all right hey blue listen how about
  • 00:17:34
    let's take him home let's finish this
  • 00:17:35
    keynote our robotics has been making
  • 00:17:37
    enormous progress and today we're
  • 00:17:40
    announcing that group
  • 00:17:43
    N1 is open
  • 00:17:46
    [Music]
  • 00:17:47
    sourced well have a great GTC thank
  • 00:17:53
    you hey
  • 00:17:55
    blue let's go home good job
Tag
  • Nvidia
  • 人工智能
  • GeForce 5090
  • 自动驾驶
  • Dynamo
  • Omniverse
  • 数据中心
  • 机器人技术
  • 安全性
  • GTC大会