NVIDIA CEO Jensen Huang Keynote at COMPUTEX 2024
NVIDIA2024-06-02
computex#nvidia#nvidia blackwell#robotics#nvidia spectrum-x#ai
328K views|3 months ago
💫 Short Summary
The video discusses the evolution of computing over the past 60 years, emphasizing the shift towards accelerated computing to address data processing requirements. It highlights the impact of Nvidia's advancements in AI, generative AI, and the development of new platforms in telecom and quantum computing. The segment also explores the future of AI technology, including the creation of digital humans and physically based AI systems. Nvidia's innovative platforms like Blackwell aim to enhance AI capabilities and drive down costs for global adoption, with a focus on robotics and automation in various industries.
✨ Highlights
📊 Transcript
✦
The importance of taking risks and committing fully to one's goals.
03:44Emphasizes the idea of going all in, risking everything, and persevering despite challenges.
Encourages resilience, determination, and unwavering belief in oneself to achieve success.
Conveys a sense of empowerment, motivation, and the mindset of never giving up.
Highlights the message of facing fear and uncertainty with determination and perseverance.
✦
The segment emphasizes perseverance, strength, and resilience in the face of challenges.
10:36The lyrics convey a message of determination and overcoming obstacles.
The theme of empowerment and self-belief is highlighted through phrases like 'I got the power' and 'spread my wings never coming down'.
The overall tone is motivational and uplifting, inspiring listeners to keep moving forward with confidence and positivity.
✦
Importance of Taiwan in Nvidia's AI infrastructure.
21:54Taiwan is home to Nvidia's partners and the origin of their AI infrastructure.
Generative AI is discussed and its impact on various industries is outlined.
Significance of accelerated computing and AI within the Omniverse is emphasized.
Reference to the IBM System 360's introduction in 1964 as a pivotal moment in computing history.
✦
Evolution of Computing over the past 60 years.
22:56Major technology shifts and impact of CPU performance scaling slowing down.
Introduction of accelerated Computing to address exponential growth in data processing requirements.
Offloading and accelerating tasks to specialized processors like GPUs to enhance performance and efficiency in processing-intensive applications.
Application of accelerated Computing in artificial intelligence and deep learning algorithms.
✦
Benefits of accelerated computing in terms of performance, power efficiency, and cost-effectiveness.
27:12GPUs can significantly improve performance in PCs and data centers.
Companies can save hundreds of millions of dollars by processing data faster with accelerated computing.
Implementing accelerated computing is challenging due to the need to rewrite software for parallel processing.
Advancements in deep learning libraries have made accelerated computing more accessible, enabling applications in AI, physics, and telecommunications.
✦
The development of new platforms in telecom through computational lithography.
31:11TSMC's use of K Litho to save energy and costs in chip manufacturing processes.
Introduction of C Quantum as an emulation system for designing quantum computers and algorithms.
Emphasis on domain-specific libraries like QDF for accelerating data processing and enabling accelerated computing.
Significance of these advancements in various industries.
✦
Pandas in C collab on Google's cloud data platform accelerate data processing with CDF.
36:57Cuda breaks traditional computing architecture cycles, attracting 5 million developers globally.
Reduction in computing costs leads to advancements in algorithms and new computer utilization methods.
Decreasing marginal computing costs enable training of large language models and drive artificial intelligence emergence.
✦
The impact of making computing cheaper has led to increased demand and innovation in the field.
40:09The creation of Cuda has resulted in advancements in AI, particularly generative AI.
The concept of Earth two, a digital twin of Earth, is being used for predicting climate change.
Nvidia's first contact with AI in 2012 has led to breakthroughs in deep learning and computer vision.
The video emphasizes the importance of continuous weather prediction and the energy efficiency of AI.
✦
Nvidia's advancements in architecture post-2012 led to the creation of the world's first AI supercomputer delivered to OpenAI in 2016.
48:00The company continued to develop larger supercomputers, resulting in the creation of GPT-3 in 2022.
GPT-3 quickly became popular, gaining a million users in just five days and a hundred million users after two months.
The success of GPT-3 was attributed to its ease of use and engaging interaction with users.
The speaker also shared personal memories of visiting a night market in their childhood, specifically mentioning the presence of a fruit lady.
✦
The transformative impact of generative AI on the industry.
53:05The creation of a new commodity with significant market opportunities.
The AI Factory's ability to generate tokens for various industries, heralding a new Industrial Revolution.
The substantial impact of generative AI on the IT industry, potentially tapping into a hundred trillion dollar market.
The evolution from accelerated computing to AI to generative AI, showcasing the significant technological progression.
✦
Shift from CPUs to GPU accelerated computing in the evolution of computing.
55:23Future computers will generate data instead of retrieving it, consuming less energy and being more contextually relevant.
This shift will result in computers generating skills and performing tasks rather than just being tools.
AI models and computing stacks are becoming more complex, with large parameters requiring distribution across multiple GPUs for faster processing.
Data center throughput utilization is now crucial, impacting financial performance and leading to the creation of AI containers with various software components.
✦
Advancements in AI technology with focus on Nim software.
59:51Nim enables chat-like communication with AI through common and standard APIs.
Pre-trained models in various domains like physics, language, vision, and healthcare are integrated into Nim.
Emphasis on AI augmenting capabilities of customer service agents in different industries.
Nim consists of reasoning agents for tasks like information retrieval and problem-solving.
✦
The future of applications involves assembling teams of experts known as Nims to work collaboratively without detailed instructions.
01:05:30Digital humans, created using AI technology, have the potential to revolutionize industries like customer service, advertising, and healthcare.
These digital humans can assist with interior design suggestions, customer service tasks, healthcare monitoring, and marketing strategies.
The development of digital humans involves advanced AI models that can understand and interact with humans in a natural and empathetic way.
This technology bridges the gap between technology and human-like interactions.
✦
Nvidia Ace utilizes generative AI for realistic 3D models with lifelike skin appearance.
01:08:11Ace Nims are digital human technologies that developers can easily integrate.
Nvidia is shipping AI GPUs with tensor core processing to create an AI platform for PCs.
The future of PCs will involve constant AI assistance in applications like photo editing and writing tools.
Transformers enable unsupervised learning on large datasets to allow AI to independently find patterns and relationships.
✦
Advancements in AI require physically based AIS to understand physics.
01:13:06Methods include learning from video, synthetic data simulation, and using computers to learn.
The development of AIS through self-play and reinforcement learning is compared to AlphaZero.
The introduction of Blackwell, a second generation GPU designed for AI, features technologies like secure AI, MV link for connecting multiple GPUs, and a reliability and availability engine for testing chip reliability.
Enhancements in reliability and data processing speeds are crucial for supercomputers running models over long periods.
✦
Blackwell's high-performance computer has incredible computational capabilities and energy efficiency.
01:17:43The computational power has greatly increased, now only requiring 3 gwatt hours compared to the previous 1,000 gwatt hours.
The advancement has led to a drastic reduction in energy consumption during token generation.
Energy needed for generating tokens has significantly decreased, allowing for faster token creation with minimal energy consumption.
✦
Introduction of DGX Blackwell system with advanced cooling mechanisms and x86 support.
01:23:45Introduction of MGX modular system for liquid cooling and improved performance.
Highlight of MV link switch technology connecting multiple GPUs for a powerful 72 GPU Blackwell system.
Emphasis on MV link chip for large language models due to high bandwidth and deep learning capabilities.
✦
Discussion on advanced GPU technology and MV link spine innovation.
01:26:10MV link spine utilizes 5,000 wires connecting 702 GPUs for high-speed networking, saving 20 kilowatts per rack in Copper.
Integration of InfiniBand capabilities into Ethernet architecture for AI factories, addressing challenges in managing InfiniBand switches.
Emphasis on network-level RDMA for Ethernet, congestion control, and telemetry for efficient data communication in GPU-centric environments.
✦
Importance of adaptive routing in Ethernet for preventing congestion and optimizing data transmission.
01:30:32Noise isolation in data centers is crucial to avoid jitter and delays in training models.
Network utilization cost impacts training time and efficiency significantly.
Introduction of Spectrum X technology and upcoming models like x800 Ultra and x1600 designed for different GPU capacities.
Future trend towards larger GPU data centers driven by the need for generative AI in cloud computing, enabling constant interactions with AI-generated content.
✦
Nvidia introduces Blackwell platform for generative AI era.
01:36:54Blackwell aims to enhance performance, drive down costs, and scale AI capabilities for global adoption.
The platform integrates CPU, GPU, MV link, Nic, and switch for seamless connectivity.
Nvidia's strategy involves building and offering entire platforms as AI factories to foster innovation and customization for different data centers and customers.
Blackwell's success story follows the company's philosophy of one-year rhythm, focusing on pushing technological limits in process, packaging, memory, and optics.
✦
The rise of physical AI and its impact on the future of robotics.
01:42:18Robots are being powered by models that enable them to understand instructions and carry out tasks independently.
Automation in factories is increasing, with robots working together to build products.
Technology breakthroughs are crucial for the advancement of robotic automation.
Companies globally are developing autonomous robots with physical AI capabilities to collaborate with humans in various industries.
✦
Breakthrough technologies like multimodal LLMs enable robots to learn, perceive, understand, and plan actions from human demonstrations.
01:44:12Reinforcement learning plays a crucial role in robots learning specific skills, especially in simulated environments with physics feedback.
Nvidia Omniverse provides a platform for virtual world simulation, allowing robots to autonomously manipulate objects, refine skills, and learn complex tasks.
The next wave of AI involves robotics powered by physical AI, which is revolutionizing various industries.
Different platforms are being developed for different types of robotic systems, integrating acceleration libraries, pre-trained models, and testing within Omniverse.
✦
Foxconn uses robotic factories with AI to meet demand for NVIDIA accelerated computing.
01:49:57Omniverse is utilized to integrate data, optimize layouts, and simulate operations in the factories.
The digital twin created by Foxconn saves costs and aids in communication.
Foxconn developers train AI applications for robotic perception and manipulation in the factories.
The factories are designed with three computer systems for AI training, robot operation, and simulation in Omniverse.
✦
Integration of AI models in robotics systems by companies like Nvidia, Siemens, and ABB.
01:53:09The goal is to enhance manufacturing efficiencies and advance robot capabilities.
Focus on integrating Isaac manipulator and perceptor into different AI robots for improved object recognition and motion tracking.
Development of humanoid robots with improved cognitive and world understanding capabilities.
Potential for significant progress in adapting robots to human environments.
✦
Future of robotics involves building walking and rolling computers with similarities to existing technology.
01:57:46Advanced AI and robotics represent a significant milestone in technology development.
Speaker expresses excitement and gratitude for future possibilities in computer technology.
Segment includes music performances and expressions of appreciation from the speaker to the audience.
00:01[Music]
00:26[Music]
00:30oh
00:32[Music]
00:41[Music]
00:53[Music]
01:00w
01:08[Music]
01:13[Music]
01:19[Music]
01:27[Music]
01:35[Music]
01:55oh
02:00oh
02:03[Music]
02:27[Music]
02:34[Applause]
02:36[Music]
02:59w o
03:04[Applause]
03:07[Music]
03:17[Music]
03:32[Music]
03:44there comes a time when you
03:49realize that everything you've done
03:51leads to this moment and I know you're
03:55scared it's hard to believe
03:58it you better got to make that jump
04:01before you
04:02g before you gone cuz I'm going all in
04:07I'm taking all my cards and put them in
04:10the middle this is just what I want I
04:14will risk it all I will risk it all for
04:18me I will do it all I will do it all for
04:23me I will risk it all I will risk it all
04:27for me I will do it all
04:30I will do it all for
04:38[Music]
04:49me when you have the chance don't think
04:53about it because the push and pull will
04:57drive you insane you can come up with a
05:01million reasons not to but all it takes
05:05is one to know for
05:09sure so I'm going all in I'm taking all
05:14my cars and put them in the middle this
05:18is just what I want I will risk it all I
05:22will risk it all for me I will do it all
05:27I will do it all for me I will risk it
05:31all I will risk it all for me I will do
05:35it all I will do it all me
05:40[Music]
06:03[Music]
06:13[Music]
06:18never give
06:19[Music]
06:23up give up
06:26[Music]
06:30sometimes I just feel
06:34Unstoppable other times I feel I want to
06:38go to
06:40go me something long
06:45ago always there when I need it
06:49the never up when up your hand C the I
06:55always when you had enough just try to
06:58hold on when you want to get
07:05just never give
07:08[Music]
07:14up I
07:17[Music]
07:18it I
07:20always on theak the
07:24breaks when I'm down I think of what you
07:30it always helps to get me out
07:33my
07:38up give
07:42[Music]
07:44up give
07:47[Music]
07:56up never give up when
08:00Just Hold Your Hand high till you reach
08:02the top know it's when you enough just
08:05try to hold on when you want to give
08:12up just never give
08:15[Music]
08:18up up
08:20when hold
08:22your to I know it's tough when you
08:25enough just try to hold on when you want
08:27to give up oh
08:34just never give
08:37[Music]
08:47up the clock keeps taking I just close
08:50my eyes and
08:52manifest the braise is almost finished I
08:56just want to see the end I keep keep
08:59lifting when the going get St I'mma keep
09:03going nothing slowing me down no one
09:06stopping me now and when my faith is on
09:09the ledge I won't ever B when they try
09:14to push me I Stand
09:17Tall I'm not going to
09:20fall holding on to love I'm against the
09:25wall but I'm not giving
09:28up oh
09:32stand strong no one's going to break me
09:35down no one's going to break me down no
09:37one's going to break me now I'm
09:42[Music]
09:46breakable I keep dreaming even when I'm
09:49wide awake I'm Living Like There's No
09:53Tomorrow Never looking back just down
09:55the road ahead I keep pushing never stop
10:00fighting never stop
10:02laughing nothing slowing me down no one
10:04stopping me now stopping me now my Fai
10:08is on the ledge I won't ever B when they
10:12try to push me I
10:16stand I'm not going to
10:18fall holding on to
10:21love up against the war but I'm not
10:25giving up I'm holding on
10:30standing strong no one's going to break
10:33me down no one's going to break me down
10:36no one's going to break me now I'm not
10:38going to
10:40fall holding on to love up against the
10:45wall but I'm not giving
10:48up hold
10:50it stand strong no one's going to break
10:54me down no one's going to break me down
10:57no one's going to break me down I'm
11:01[Music]
11:06unbreakable used to feel locked up
11:09inside my
11:11head you say I close the door now I know
11:16that I'm the only one to open up the
11:22window I can't stop got to let it out I
11:28got the power
11:29I got the power spreading my wings never
11:33coming down there's so much power
11:37there's so much power in
11:42me I got the power I got the power in
11:51me I got the power I got the power now I
11:56see what I have always been I Turn To
12:01Face my
12:03future now I know that I'm the only one
12:08to open up the window I can't stop got
12:13to let it out I got the power I got the
12:18power spread my wings never coming down
12:24there's so much power there's so much
12:27power in me
12:30[Applause]
12:30[Music]
12:32I got the power I got the power in
12:39me I got the power I got the
12:47power I got the power I got the power in
12:55me I got the power I got the power in
13:18[Music]
13:19me all right let's get
13:24started good take good
13:27take oh oh
13:30je okay you guys ready yeah yeah
13:33everybody thinks we make
13:38GPS I video is so much more than
13:42[Music]
13:44that this whole keynote is going to be
13:46about that okay so we'll start at the
13:48top examples of the use cases and then
13:50seeing it in action that's kind of the
13:52flow in such your compelling story I'm
13:55super nervous about this just got to get
13:57in the Rhythm we're 2 weeks away from
13:59you guys can go really really make it
14:01bre we should take a swing at it yeah
14:03that's the plan we need to get daily
14:05status on that animation can you mute
14:07cuz I hear myself sorry what's the drop
14:10date for all the videos it needs to be
14:12done on the
14:1428th did you get all
14:19that safe travels everybody super
14:23excited to see everyone see you guys
14:25soon okay bye
14:29we're basically moving as fast as the
14:32world can absorb technology so we've got
14:34to LEAP from ourselves
14:36[Music]
14:40in now the spine you just have to figure
14:42out a way to make it pop you know what
14:44I'm
14:46[Music]
14:47saying you know what I'm saying you want
14:51to yeah that kind of thing
14:54[Music]
15:08okay thank you I'm super late let's
15:18[Music]
15:27go please welcome welcome to the stage
15:29Nvidia founder and CEO Jensen
15:32[Music]
15:50W I am very happy to be
15:53back thank you NTU for letting us use
15:56your Stadium
15:59the last time I was
16:01here I received a degree from
16:09[Applause]
16:14n and I gave the Run don't walk
16:19speech and today we have a lot to cover
16:23so I cannot walk I must
16:25run we have a lot to cover I have many
16:29things to tell you I'm very happy to be
16:32here in
16:32Taiwan Taiwan is the home of our
16:37treasured
16:39Partners this is in
16:41fact where everything Nvidia does
16:44begins our partners and ourselves take
16:48it to the
16:49world
16:51Taiwan and our
16:54partnership has created the world's AI
16:58infrastructure
17:01today I want to talk to you about
17:03several
17:05things
17:08one what is happening and the meaning of
17:13the work that we do together what is
17:15generative
17:16AI what is its impact on our industry
17:20and on every
17:24industry a blueprint for how we will go
17:27forward and engage this incredible
17:33opportunity and what's coming
17:37next generative Ai and its impact our
17:42blueprint and what comes
17:44next these are really really exciting
17:48times a
17:50restart of our computer industry an
17:53industry
17:55that you have forged an IND industry
17:59that you have
18:00created and now you're
18:02prepared for the next major Journey but
18:06before we
18:09start Nvidia lives at the
18:12intersection of computer
18:15graphics
18:17simulations and artificial
18:20intelligence this is our
18:23soul everything that I show you
18:26today is simulation
18:29it's math it's science it's computer
18:33science it's amazing computer
18:36architecture none of it's
18:39animated and it's all
18:41homemade this is invidious soul and we
18:45put it all into this virtual world we
18:48called Omniverse please enjoy
19:04[Music]
19:11[Music]
19:27[Music]
19:33w
19:36[Music]
19:42[Music]
20:05[Music]
20:17[Music]
20:33[Music]
20:51[Applause]
20:57when doubt
21:03[Applause]
21:09I want to speak to you in Chinese but I
21:12have so much to tell you I have to think
21:15too
21:16hard to speak
21:18Chinese so I have to speak to you in
21:21English at the foundation of everything
21:24that you saw was two fundamental
21:27Technologies
21:29accelerated Computing and artificial
21:31intelligence running inside the
21:37Omniverse those two technologies those
21:40two fundamental forces of
21:43computing are going to reshape the
21:46computer
21:48industry the computer
21:51industry is now some 60 years
21:54old in a lot of ways everything that we
21:57do today was invented the year after my
22:00birth in
22:021964 the IBM system 360 introduced
22:06central processing units general purpose
22:09Computing the separation of hardware and
22:13software through an operating
22:16system multitasking IO subsystems
22:20dma all kinds of technologies that we
22:23use today architectural
22:26compatibility backwards compa ibility
22:29family compatibility all of the things
22:31that we know today about Computing
22:33largely described in
22:361964 of course the PC Revolution
22:39democratized Computing and put it in the
22:42hands and the houses of everybody and
22:44then off in 2007 the iPhone introduced
22:48mobile Computing and put the computer in
22:50our
22:51pocket ever since everything is
22:54connected and running all the time
22:56through the mobile cloud
22:59this last 60 years we saw
23:03several just several not that many
23:06actually two or three major technology
23:10shifts two or
23:13three tectonic shifts in Computing where
23:17everything changed and we're about to
23:20see that happen again there are two
23:22fundamental things that are
23:24happening the first is that the
23:27processor the engine by which the
23:29computer industry runs on the central
23:32processing unit the performance scaling
23:35has slowed
23:37tremendously and yet the amount of
23:40computation we have to do is still
23:43doubling very quickly
23:46exponentially if processing requirement
23:50if the data that is that we need to
23:51process continues to scale exponentially
23:54but performance does not we will
23:57experience computation inflation and in
24:01fact we're seeing that right now as we
24:02speak the amount of data center power
24:05that's used all over the world is
24:06growing quite
24:08substantially the cost of computing is
24:11growing we are seeing computation
24:15inflation this of course cannot
24:18continue the data is going to continue
24:20to increase exponentially and CPU
24:25performance scaling will never return
24:28there is a better way for almost two
24:30decades now we've been working on
24:32accelerated
24:33Computing Cuda augments a CPU offloads
24:37and accelerates the work that A
24:40specialized processor can do much much
24:43better in fact the performance is so
24:45extraordinary that it is very clear now
24:49as CPU scaling has slowed and event
24:52substantially stopped we should
24:55accelerate everything I predict that
24:58every application that is processing
25:00intensive will be accelerated and surely
25:03every data center will be accelerated
25:06the near future now accelerated
25:08Computing is very sensible it's very
25:10common sense if you take a look at an
25:13application and here the 100t means 100
25:17units of time it could be 100 seconds it
25:20could be 100 hours and in many cases as
25:23you know we're now working on artificial
25:25intelligence applications that run for a
25:27100 days
25:31the ont is code that is requires
25:35sequential processing where
25:36single-threaded CPUs are really quite
25:39essential operating systems control
25:42logic really essential to have one
25:45instruction executed after another
25:47instruction however there are many
25:49algorithms computer Graphics is one that
25:52you can operate completely in parallel
25:55computer Graphics image processing
25:57physics
25:58simulations combinatorial
26:01optimizations graph processing database
26:04processing and of course the very famous
26:07linear algebra of deep learning there
26:09are many types of algorithms that are
26:11very conducive to
26:14acceleration through parallel processing
26:17so we invented an architecture to to do
26:20that by adding the GPU to the
26:25CPU the specialized processor can take
26:29something that takes a great deal of
26:30time and accelerate it down to something
26:33that is in incredibly fast and because
26:36the two processors can work side by side
26:39they're both autonomous and they're both
26:40separate independent that is we could
26:43accelerate what used to take a 100 units
26:45of time down to one unit of time well
26:48the speed up is incredible it almost
26:52sounds
26:54unbelievable it almost sounds
26:56unbelievable but today I'll demonstrate
26:58many examples for
26:59you the benefit is quite extraordinary a
27:03100 times speed up but you only increase
27:06the power by about a factor of three and
27:09you increase the cost by only about
27:1250% we do this all the time in the PC
27:15industry we add a GPU a $500 GPU gForce
27:20GPU to a $1,000 PC and the performance
27:23increases tremendously we do this in a
27:26data center a billion dooll data center
27:30we add $500 million worth of gpus and
27:33all of a sudden it becomes an AI
27:36Factory this is happening all over the
27:39world today well the savings are quite
27:45extraordinary you're getting 60 times
27:48performance per dollar a 100 times speed
27:52up you only increase your power by 3x
27:55100 times speed up you only increase
27:58inre your cost by 1.5x the savings are
28:03incredible the savings are measured in
28:06dollars it is very clear that many many
28:10companies spend hundreds of millions of
28:13dollars processing data in the cloud if
28:15it was accelerated it is not unexpected
28:20that you could save hundreds of millions
28:22of dollars now why is that well the
28:25reason for that is very clear we've been
28:28experiencing inflation for so long in
28:31general purpose Computing now that we
28:34finally came to we finally determined to
28:38accelerate there's an enormous amount of
28:41captured loss that we can now regain a
28:45great deal of captured retained waste
28:48that we can now relieve out of the
28:50system and that will translate into
28:52savings Savings in money Savings in
28:55energy and that's the reason why you've
28:58heard me
29:00say the more you buy the more you
29:10save and now I've shown you the
29:13mathematics it is not accurate but it is
29:17correct okay that's called CEO
29:20math CEO math is not accurate but it is
29:23correct the more you buy the more you
29:25save well accelerated computing does
29:28deliver extraordinary results but it is
29:30not easy why is it that it saves so much
29:34money but people haven't done it for so
29:35long the reason for that is because it's
29:38incredibly
29:39hard there is no such thing as a
29:42software that you can just run through a
29:44c compiler and all of a
29:46sudden that application runs a 100 times
29:49faster that is not even logical if it
29:53was possible to do that they would have
29:55just changed the CPU to do that you in
29:57fact have to rewrite the software that's
30:00the hard part the software has to be
30:03completely Rewritten so that you could
30:06reactor
30:09reexpress the algorithms that was
30:11written on a CPU so that it could be
30:14accelerated offloaded accelerated and
30:16run in parallel that computer science
30:20exercise is insanely hard well we've
30:23made it easy for the world over the last
30:2520 years of course the very famous cdnn
30:29the Deep learning library that processes
30:31neural networks we have a library for AI
30:35physics that you could use for fluid
30:37dynamics and many other applications
30:39where the neural network has to obey the
30:41laws of physics we have a great new
30:43library called aerial that is a Cuda
30:46accelerated 5G radio so that we can
30:50software Define and accelerate the
30:53Telecommunications networks the way that
30:55we've soft software defined the world
30:58world's networking um uh internet and so
31:02the ability for us to accelerate that
31:04allows us to turn all of Telecom into
31:07essentially the same type of platform a
31:09Computing platform just like we have in
31:11the cloud kitho is a comp computational
31:14lithography platform that allows us to
31:18uh process the most computationally
31:20intensive parts of Chip manufacturing
31:23making the mask tsmc is in the process
31:26of going to production with k litho
31:28saving enormous amounts of energy and
31:30more enormous amounts of money but the
31:33goal for tsmc is to accelerate their
31:36stack so that they're prepared for even
31:38further advances in algorithm and more
31:41computation for deeper and deeper uh
31:45narrow and narrow transistors parab
31:47brakes is our Gene sequencing Library it
31:50is the highest throughput library in the
31:51world for Gene sequencing coopt is an
31:54incredible library for combinatorial
31:56optimization
31:58route planning optimization the
32:01traveling salesman problem incredibly
32:03complicated people do people have well
32:07scientists have largely concluded that
32:09you needed a quantum computer to do that
32:11we created an algorithm that runs on
32:12accelerated Computing that runs
32:14Lightning Fast 23 World Records we hold
32:18every single major world record
32:20today C Quantum is an emulation system
32:24for a quantum computer if you want to
32:26design a quantum computer comp you need
32:28a simulator to do so if you want to
32:30design Quantum algorithms you need a
32:31Quantum emulator to do so how would you
32:34do that how would you design these
32:35quantum computers create these Quantum
32:38algorithms if the quantum computer
32:40doesn't exist while you use the fastest
32:42computer in the world that exist today
32:44and we call it of course Nvidia Cuda and
32:47on that we have an emulator that
32:50simulates quantum computers it is used
32:53by several hundred thousand researchers
32:56around the world it is integrated into
32:59all the leading M leading Frameworks for
33:01Quantum Computing and is used in
33:03scientific super super Computing centers
33:05all over the world kdf is an
33:09unbelievable library for data
33:12processing data processing consumes the
33:15vast majority of cloud spend today all
33:18of it should be accelerated qdf
33:21accelerates the major libraries used in
33:23the world spark many of you probably use
33:26spark in your companies
33:30pandas um a new one called polar and of
33:34course uh networkx which is a graph
33:36processing graph processing uh database
33:39um library and so these are just some
33:41examples there are so many more each one
33:44of them had to be created so that we can
33:47enable the ecosystem to take advantage
33:49of accelerated Computing if we hadn't
33:51created cudnn Cuda alone wouldn't have
33:55been able wouldn't have been possible
33:57for of the deep learning scientists
33:59around the world to use because
34:01Cuda and the algorithms that are used in
34:04tensor flow and Pi torch the Deep
34:06learning algorithms the separation is
34:09too far apart it's almost like trying to
34:12do computer Graphics without open G it's
34:14almost like doing data processing
34:16without
34:18SQL these domain specific libraries are
34:22really The Treasure of our company we
34:23have 350 of
34:26them the these libraries Is What It
34:29Takes and what has made it possible for
34:31us to have such open so many markets
34:35I'll show you some other examples today
34:37well just last week Google announced
34:40that theyve put CDF in the cloud and
34:43accelerate pandas pandas is the most
34:46popular data science library in the
34:48world many of you in here probably
34:50already use pandas it's used by 10
34:52million data scientists in the world
34:54downloaded 170 million times each month
34:59it is the Excel that is the spreadsheet
35:01of data
35:02scientists well with just one click you
35:05can now use pandas in C collab which is
35:09Google's cloud data centers platform
35:12accelerated by CDF the speed up is
35:14really incredible let's take a look
35:17[Music]
35:29that was a great demo right didn't take
35:34[Applause]
35:38long when you accelerate data processing
35:42that fast demos don't take
35:45long okay well Cuda has now achieved
35:50what people call a Tipping Point but
35:52it's even better than that Cuda has now
35:54achieved a virtuous cycle this rarely
35:57happens if you look at history and all
36:00the Computing architecture Computing
36:03Platforms in the case of microprocessor
36:06CPUs it has been here for 60
36:09years it has not been changed for 60
36:13years at this level this way of doing
36:17Computing accelerated Computing has been
36:19around
36:21has creating a new platform is extremely
36:23hard because it's a chicken and egg
36:26problem if there are no
36:28developers that use your platform then
36:31of course there will be no users but if
36:33there are no users there are no install
36:35base if they're no install base
36:36developers aren't interested in it
36:39developers want to write software for a
36:41large install base but a large install
36:43base requires a lot of applications so
36:46that users would create that install
36:49base this chicken or the egg problem has
36:52rarely been broken and has taken us now
36:5520 years one domain Library after
36:57another one acceleration Library another
37:00and now we have 5 million developers
37:02around the
37:03world we serve every single industry
37:07from Health Care Financial Services of
37:09course the computer industry automotive
37:10industry just about every major industry
37:12in the world just about every field of
37:16science because there are so many
37:19customers for our architecture oems and
37:22cloud service providers are interested
37:24in building our
37:25systems system makers am amazing system
37:28makers like the ones here in Taiwan are
37:29interested in building our systems which
37:31then takes and offers more systems to
37:34the
37:35market which of course creates greater
37:38opportunity for us which allows us to
37:40increase our scale R&D scale which
37:43speeds up the application even more well
37:46every single time we speed up the
37:48application the cost of computing goes
37:51down this is that slide I was showing
37:54you earlier 100x speed up translates to
37:5897 96% 98% savings and so when we go
38:02from 100x speed up to 200x speed up to
38:05a000 X speed up the savings the marginal
38:08cost of computing continues to
38:10fall well of
38:13course we believe that by reducing the
38:17cost of computing
38:19incredibly the market developers
38:23scientists inventors will continue to
38:25discover new algorithms that cons
38:27consume more and more and more Computing
38:30so that one
38:33day something
38:35happens that a pH shift happens that the
38:38marginal cost of computing is so low
38:41that a new way of using computers
38:44emerge in fact that's what we're seeing
38:47now over the years we have driven down
38:50the marginal cost of computing in the
38:52last 10 years in one particular
38:54algorithm by a million times well as a
38:58result it is now very
39:01logical and very common
39:04sense to train large language models
39:08with all of the data on the internet
39:10nobody thinks
39:12twice this idea that you could create a
39:16computer that could process so much data
39:19to write its own software the emergence
39:22of artificial intelligence was made
39:24possible because of this complete belief
39:27that if we made Computing cheaper and
39:29cheaper and cheaper somebody's going to
39:31find a great use well today Cuda has
39:34achieved the virtual cycle install base
39:37is growing Computing cost is coming down
39:41which causes more developers to come up
39:43with more
39:44ideas which drives more
39:47demand and now we're on in the beginning
39:50of something very very important but
39:52before I show you that I want to show
39:54you what is not possible if not for the
39:57fact that we created Cuda that we
40:01created the modern version of General
40:04the modern Big Bang of AI generative AI
40:07what I'm about to show you would not be
40:09possible this is Earth two the idea that
40:14we would create a digital twin of the
40:18earth that we would go and simulate the
40:21Earth so that we could predict the
40:24future of our planet to better Aver
40:29disasters or better understand the
40:31impact of climate change so that we can
40:33adapt better so that we could change our
40:35habits now this digital twin of Earth is
40:40probably one of the most ambitious
40:41projects that the world's ever
40:43undertaken and we're taking step large
40:46steps every single year and I'll show
40:47you results every single year but this
40:49year we made some great breakthroughs
40:50let's take a look
41:02on Monday the storm will Veer North
41:04again and approach Taiwan there are big
41:06uncertainties regarding its path
41:09different paths will have different
41:10levels of impact on
41:26Taiwan
41:35forvia Earth
41:37[Music]
41:45[Music]
41:562f
42:20[Music]
42:26for
42:38fore cordi AI
43:08[Music]
43:12Nvidia
43:18[Music]
43:24[Music]
43:41[Applause]
43:48someday in the near future we will have
43:51continuous weather
43:54prediction at every at every Square
43:57kilometer on the
43:58planet you will always know what the
44:01climate's going to be you will always
44:02know and this will run continuously
44:05because we trained the
44:07AI and the AI requires so little
44:11energy and so this is just an incredible
44:13achievement I hope you you enjoyed it
44:15and very
44:17importantly uh what
44:23[Applause]
44:30the truth is that was
44:34a Jensen AI that was not
44:38me I I wrote I wrote it but an AI Jens
44:43AI had to say
44:49it because of our dedication to
44:51continuously improve the performance of
44:54Drive the cost down researchers discover
44:57discovered AI researchers discovered
44:58Cuda in
44:592012 that was
45:02nvidia's first contact with
45:06AI this was a very important day we had
45:09the good
45:10wisdom to work with the the scientists
45:14to make it possible for a deep learning
45:16to happen and Alex net achieved of
45:18course a tremendous computer vision
45:21breakthrough but the great wisdom was to
45:24take a step back and understanding what
45:26was the
45:28background what is the foundation of
45:30deep learning what is its long-term
45:32impact what is its potential and we
45:36realized that this technology has great
45:38potential to scale an algorithm that was
45:40invented and discovered decades ago all
45:45of a sudden because of more
45:47data larger networks and very
45:50importantly a lot more
45:52compute all of a sudden deep learning
45:55was able to achieve what no human
45:58algorithm was able to now imagine if we
46:02were to scale up the architecture even
46:04more larger networks more data and more
46:07compute what could be possible so we
46:09dedicated ourselves to reinvent
46:13everything after 2012 we changed the
46:16architecture of our GPU to add tensor
46:18course we invented MV link that was 10
46:22years ago
46:24now CNN
46:27tensor RT nickel we bought
46:32melanox tensor RT LM the Triton
46:35inference server and all of it came
46:38together on a brand new computer nobody
46:42understood nobody asked for it nobody
46:45understood it and in fact I was certain
46:48nobody wanted to buy it and so we
46:50announced it at
46:52GTC and open AI a small company in San
46:55Francisco saw it and they asked me to
46:58deliver one to them I delivered the
47:01first dgx the F the world's first AI
47:04supercomputer to open
47:07AI in
47:102016 well after that we continued to
47:13scale from one AI supercomputer one AI
47:18Appliance we scaled it up to large
47:20supercomputers even larger by
47:242017 the world discovered Transformers
47:27so that we could train enormous amounts
47:30of data and recognize and learn patterns
47:33that are sequential over large spans of
47:36time it is now possible for for us to
47:39train these large language models to
47:41understand and achieve a breakthrough in
47:43natural language
47:46understanding and we kept going after
47:48that we built even larger ones and then
47:51in November
47:532022 trained on thousands
47:57tens of thousands of Nvidia gpus in a
48:00very large AI supercomputer open AI
48:03announced chat GPT 1 million users after
48:07five
48:09days 1 million after five days a 100
48:13million after two months the fastest
48:15growing application in history and the
48:17reason for that is very simple it is
48:21just so easy to use and it was so
48:23magical to use to be able to interact
48:26with computer like it's
48:28Human Instead of being clear about what
48:31you want it's like the computer
48:34understands your meaning it understands
48:36your
48:38intention oh I think here it it um asked
48:41the the closest Night Market night as
48:43you know U the night market is very
48:46important to
48:49me so when I was young uh I was I think
48:53I was four and a half years old I used
48:55to love going to the night mark Market
48:57because I I just love watching people
48:59and and uh uh and so we went my parents
49:02used to take us to the night market
49:11uh and and um uh and I love I love U I
49:16love going and one day uh my face you
49:19guys might might see that I have a large
49:21scar on my face my face was cut because
49:23somebody was washing their knife and I
49:25was a little kid um but my memories of
49:29of the Night Market is uh so deep
49:32because of that and I used to love I I
49:34just I still love going to the night
49:35market and I just need to tell you guys
49:37this the the Tona Night Market is is
49:41really good because there's a lady uh
49:44she's been working there for 43
49:47years she's the fruit lady and it's in
49:50the middle of the stre the middle
49:52between the two go find her okay she
49:59[Applause]
50:04she's really
50:05terrific I think it would be funny after
50:08this all of you go to see
50:10her she every year she's doing better
50:13and better her cart has improved and
50:16yeah I just love watching her succeed
50:18anyways anyways uh Chad GPT came along
50:22and um and something is very important
50:25in this slide
50:27here let me show you
50:31something this
50:33slide okay and this
50:38Slide the fundamental difference is
50:42this until chat
50:45GPT revealed it to the
50:48world AI was all about
50:52perception natural language
50:55understanding Compu computer vision
50:58speech
50:59recognition it's all about perception
51:03and
51:04detection this was the first time the
51:06world saw a generative
51:10AI It produced
51:12tokens one token at a time and those
51:16tokens were words some of the tokens of
51:19course could now be images or charts or
51:23tables songs words speech
51:28videos those tokens could be anything
51:31they anything that that you can learn
51:33the meaning of it could be tokens of
51:37chemicals tokens of proteins
51:40genes you saw earlier in Earth 2 we were
51:44generating
51:45tokens of the
51:47weather we can we can learn physics if
51:51you can learn physics you could teach an
51:53AI model physics the AI model could
51:55learn the meaning of physics and it can
51:58generate physics we were scaling down to
52:011 kilometer not by using filtering it
52:04was
52:06generating and so we can use this method
52:11to generate tokens for almost
52:13anything almost anything of value we can
52:17generate steering wheel control for a
52:21car we can generate articulation for a
52:24robotic arm
52:27everything that we can learn we can now
52:31generate we have now arrived not at the
52:34AI era but a generative AI era but
52:38what's really important is
52:42this this computer that started out as a
52:46supercomputer has now evolved into a
52:49Data Center and it
52:52produces one thing it produces tokens
52:57it's an AI
52:59Factory this AI Factory is generating
53:03creating producing something of Great
53:05Value a new
53:08commodity in the late
53:101890s Nicola Tesla invented an AC
53:15generator we invented an AI
53:19generator the AC generator generated
53:22electrons nvidia's AI generator
53:25generates tokens
53:27both of these things have large Market
53:31opportunities it's completely fungible
53:33in almost every
53:34industry and that's why it's a new
53:38Industrial
53:39Revolution we have now a new Factory
53:43producing a new commodity for every
53:46industry that is of extraordinary value
53:49and the methodology for doing this is
53:51quite scalable and the methodology of
53:53doing this is quite repeatable
53:56notice how quickly so many different AI
53:59models generative AI models are being
54:01invented literally daily every single
54:05industry is now piling
54:07on for the very first
54:10time the IT industry which is $3
54:14trillion $3 trillion IT industry is
54:18about to create something that can
54:21directly serve a hundred trillion dollar
54:24of Industry no longer just an instrument
54:28for information storage or data
54:32processing but a factory for generating
54:36intelligence for every industry this is
54:39going to be a manufacturing industry not
54:43a manufacturing industry of computers
54:45but using the computers in
54:48manufacturing this has never happened
54:50before quite an extraordinary thing what
54:54led started with accelerated comp
54:56Computing led to AI led to generative Ai
55:00and now an industrial revolution now the
55:04impact to our industry is also quite
55:10significant of course we could create a
55:12new commodity a new product we call
55:15tokens for many Industries but the
55:18impact of ours is also quite profound
55:21for the very first time as I was saying
55:23earlier in 60 years every single layer
55:26of computing has been changed from CPUs
55:29general purpose Computing to accelerated
55:31GPU
55:32Computing where the computer needs
55:36instructions now computers process llms
55:39large language models AI models and
55:43whereas the Computing model of the past
55:45is retrieval
55:47based almost every time you touch your
55:49phone some pre-recorded text or
55:53pre-recorded image or pre-recorded video
55:56is retrieved for you and recomposed
55:59based on a recommender system to present
56:02it to you based on your habits but in
56:05the
56:06future your computer will generate as
56:09much as possible retrieve only what's
56:13necessary and the reason for that is
56:15because generated generated data
56:17requires less energy to go fetch
56:20information generated data also is more
56:22contextually relevant it will encode
56:25knowledge it will your understanding of
56:27you and instead
56:30of get that inform information for me or
56:33get that file for me you just
56:35say ask me for an answer and instead of
56:40a
56:40tool instead of your computer being a
56:43tool that we use the computer will now
56:47generate
56:48skills it performs tasks and instead of
56:52an industry that is producing software
56:55was which was a revolu evolutionary idea
56:57in the early
56:5890s
56:59remember the idea that Microsoft created
57:03for packaging software revolutionize the
57:05PC industry without packaged software
57:08what would we use the PC to
57:11do it drove this industry and now we
57:16have a new Factory a new computer and
57:19what we will run on top of this is a new
57:22type of software and we call it Nims
57:25Nvidia inference
57:28microservices now what what happens is
57:31the Nim runs inside this Factory and
57:33this Nim is a pre-trained model it's an
57:38AI well this AI is of course quite
57:42complex in itself but the the Computing
57:45stack that runs AIS are insanely complex
57:49when you go and use chat GPT underneath
57:52their stack is a whole bunch of software
57:55underneath that prompt is a ton of
57:57software and it's incredibly complex
58:00because the models are large billions to
58:02trillions of parameters it doesn't run
58:04on just one computer it runs on multiple
58:06computers it has to distribute the
58:08workload across multiple gpus tensor
58:10parallelism pipeline parallelism data
58:13parallel all kinds of
58:16parallelism expert parallelism all kinds
58:18of parallelism Distributing the workload
58:21across multiple gpus processing it as
58:24fast as possible because if you are in a
58:27factory if you run a factory your
58:30throughput directly correlates to your
58:33revenues your throughput directly
58:35correlates to quality of service and
58:37your throughput directly correlates to
58:39number of people who can use your
58:40service we are now in a world where data
58:43center throughput
58:45utilization is vitally important it was
58:48important in the past but not Vally
58:50important it was important in the past
58:52but people don't measure it today every
58:55par parameter is measured start time
58:58uptime utilization throughput idle time
59:01you name it because it's a factory when
59:05something is a factory its operations
59:08directly correlate to the financial
59:11performance of the company and so we
59:14realized that this is incredibly complex
59:16for most companies to do so what we did
59:19was we created this AI in a box and it
59:23containers an incredible amounts of
59:25software
59:27inside this container is Cuda cudnn
59:31tensor RT Triton for inference Services
59:35it is cloud native so that you could
59:38Auto scale in a kubernetes environment
59:40it has Management Services and hooks so
59:42that you can monitor your AIS it has
59:45common apis standard apis so that you
59:48could literally chat with this
59:51box you download this Nim and you can
59:55talk to it so long as you have Cuda on
59:58your
59:59computer which is now of course
01:00:01everywhere it's in every cloud available
01:00:03from every computer maker it is
01:00:05available in hundreds of millions of PCS
01:00:08when you download this you have an AI
01:00:11and you can chat with it like chat GPT
01:00:13all of the software is now integrated
01:00:15400 dependencies all integrated into one
01:00:19we tested this Nim each one of these
01:00:22pre-trained models against all kind are
01:00:26Tire install base that's in the cloud
01:00:28all the different versions of Pascal and
01:00:30ampers and Hoppers
01:00:35and all kinds of different versions I
01:00:37even forget
01:00:41some Nims incredible invention this is
01:00:45one of my favorites and of course as you
01:00:49know we now have the ability to create
01:00:52large language models and pre-trained
01:00:53models of all kinds and we we have all
01:00:56of these various versions whether it's
01:00:59language based or Vision based or
01:01:00Imaging based or we have versions that
01:01:03are available for healthc care digital
01:01:06biology we have versions that are
01:01:09digital humans that I'll talk to you
01:01:11about and the way you use this just come
01:01:13to ai. nvidia.com and today we uh just
01:01:18posted up in hugging face the Llama 3
01:01:22Nim fully optimized it's available there
01:01:25for you to try and you can even take it
01:01:28with you it's available to you for free
01:01:30and so you could run it in the cloud run
01:01:33it in any Cloud you could download this
01:01:35container put it into your own Data
01:01:37Center and you could host it make it
01:01:39available for your customers we have as
01:01:42I mentioned all kinds of different
01:01:44domains physics some of it is for
01:01:48semantic retrieval called Rags Vision
01:01:51languages all kinds of different
01:01:53languages and the way that you use it
01:01:57is connecting these microservices into
01:02:00large applications one of the most
01:02:02important applications in the coming
01:02:04future of course is customer service
01:02:07agents customer service agents are
01:02:09necessary in just about every single
01:02:11industry it represents trillions of
01:02:14dollars of of customer service around
01:02:17the world nurses are customer service
01:02:20agents um in some ways some of them are
01:02:23nonprescription or or non Diagnostics um
01:02:27based nurses are essentially customer
01:02:30service uh customer service for retail
01:02:32for uh Quick Service Foods Financial
01:02:35Services Insurance just tens and tens of
01:02:38millions of customer service can now be
01:02:41augmented by language models and
01:02:45augumented by Ai and so these one these
01:02:47boxes that you see are basically Nims
01:02:50some of the NIMS are reasoning agents
01:02:53given a task figure out what the miss is
01:02:56break it down into a plan some of the
01:02:59NIMS retrieve information some of the
01:03:01NIMS might uh uh uh go and do search
01:03:05some of the NIMS uh might use a tool
01:03:08like kuop that I was talking about
01:03:09earlier they could use a tool that uh
01:03:12could be running on sap and so it has to
01:03:15learn a particular uh language called
01:03:18abap maybe some Nims have to uh uh do
01:03:21SQL queries and so all of these Nims are
01:03:25experts
01:03:26that are now assembled as a
01:03:28team so what's
01:03:31happening the application layer has been
01:03:34changed what used to be applications
01:03:37written with
01:03:38instructions are now
01:03:41applications that are assembling teams
01:03:44assembling teams of
01:03:46AIS very few people know how to write
01:03:48programs almost everybody knows how to
01:03:50break down a problem and assemble teams
01:03:53very every company I believe in the
01:03:55future will have a large collection of
01:03:59Nims and you would bring down the
01:04:01experts that you want you connect them
01:04:04into a team and you you don't even have
01:04:08to figure out
01:04:09exactly how to connect
01:04:11them you just give the mission to an
01:04:16agent to a Nim to figure out who to
01:04:20break the task down and who to give it
01:04:22to and they that a that Central
01:04:26the leader of the of the application if
01:04:28you will the leader of the team would
01:04:30break down the task and give it to the
01:04:32various team members the team members
01:04:35would do their perform their task bring
01:04:37it back to the team leader the team
01:04:39leader would reason about that and
01:04:41present an information back to you just
01:04:44like humans this is in our near future
01:04:48this is the way applications are going
01:04:50to look now of
01:04:52course we could interact with these
01:04:54large these AI services with text
01:04:58prompts and speech
01:05:00prompts however there are many
01:05:02applications where we would like to
01:05:04interact with what what is otherwise a
01:05:07humanlike form we call them digital
01:05:09humans Nvidia has been working on
01:05:11digital human technology for some time
01:05:14let me show it to you and well before I
01:05:16do that hang on a second before I do
01:05:19that okay digital humans has the
01:05:22potential of being a great interact
01:05:25Interactive with you they make much more
01:05:28engaging they could be much more
01:05:30empathetic and of course um we have to
01:05:35uh uh cross this incredible Chasm this
01:05:40uncanny Chasm of realism so that the
01:05:43digital humans would appear much more
01:05:46natural this is of course our vision
01:05:49this is a vision of where we love to go
01:05:51uh but let me show you where we are
01:05:56great to be in Taiwan before I head out
01:05:59to the night market let's dive into some
01:06:01exciting frontiers of digital
01:06:03humans imagine a future where computers
01:06:06interact with us just like humans can hi
01:06:09my name is Sophie and I am a digital
01:06:11human brand ambassador for Unique this
01:06:14is the incredible reality of digital
01:06:17humans digital humans will revolutionize
01:06:20industries from customer service to
01:06:24advertising and ging the possibilities
01:06:27for digital humans are endless using the
01:06:30scans you took of your current kitchen
01:06:32with your phone they will be AI interior
01:06:34designers helping generate beautiful
01:06:37photorealistic suggestions and sourcing
01:06:39the materials and Furniture we have
01:06:41generated several design options for you
01:06:43to choose from they'll also be AI
01:06:46customer service agents making the
01:06:48interaction more engaging and
01:06:49personalized or digital healthcare
01:06:51workers who will check on patients
01:06:54providing Timely personalized care um I
01:06:57did forget to mention to the doctor that
01:06:59I am allergic to penicillin is it still
01:07:01okay to take the medications the
01:07:03antibiotics you've been prescribed cicin
01:07:06and metronidazol don't contain
01:07:08penicillin so it's perfectly safe for
01:07:10you to take them and they'll even be AI
01:07:13brand ambassadors setting the next
01:07:15marketing and advertising Trends hi I'm
01:07:18EMA Japan's first virtual
01:07:21model new breakthroughs in generative Ai
01:07:25and computer Graphics let digital humans
01:07:28see understand and interact with us in
01:07:31humanlike
01:07:33ways H from what I can see it looks like
01:07:37you're in some kind of recording or
01:07:39production setup the foundation of
01:07:41digital humans are AI models built on
01:07:44multilingual speech recognition and
01:07:47synthesis and llms that understand and
01:07:49generate conversation
01:08:00the AIS connect to another generative AI
01:08:03to dynamically animate a lifelike 3D
01:08:06mesh of a face and finally AI models
01:08:11that reproduce lifelike appearances
01:08:14enabling real-time path traced
01:08:16subsurface scattering to simulate the
01:08:18way light penetrates the skin scatters
01:08:21and exits at various points giving skin
01:08:24its soft and translucent
01:08:26appearance Nvidia Ace is a suite of
01:08:29digital human Technologies packaged as
01:08:32easy to deploy fully optimized
01:08:35microservices or Nims developers can
01:08:38integrate Ace Nims into their existing
01:08:40Frameworks engines and digital human
01:08:43experiences neotron slm and llm Nims to
01:08:48understand our intent and orchestrate
01:08:50other models Reva speech Nims for
01:08:52interactive speech and translation audio
01:08:55to face and gesture Nims for facial and
01:08:58body animation an Omniverse RTX with
01:09:01dlss for neuro rendering of skin and
01:09:03hair Ace Nims run on Nvidia gdn a Global
01:09:08Network of Nvidia accelerated
01:09:10infrastructure that delivers low latency
01:09:12digital human processing to over 100
01:09:19regions Kaa
01:09:22[Applause]
01:09:28pretty incredible well those those Ace
01:09:32runs in a cloud but it also runs on PCS
01:09:35we had the good wisdom of including
01:09:37tensor core gpus in all of RTX so we've
01:09:41been shipping AI gpus for some time
01:09:46preparing ourselves for this day the
01:09:49reason for that is very simple we always
01:09:51knew that in order to create a new
01:09:52Computing platform you need to install
01:09:54Bas first
01:09:55eventually the application will come if
01:09:58you don't create the install base how
01:10:00could the application come and so if you
01:10:03build it they might not
01:10:05come but if you build it if you don't
01:10:07build it they cannot come and so we
01:10:10installed every single RTX GPU with
01:10:13tensor core G tensor core processing and
01:10:16now we have 100 million GeForce RTX AIP
01:10:20PCS in the world and we're shipping 200
01:10:23and this this compy Tex wor featuring
01:10:25four new amazing
01:10:28laptops all of them are able to run AI
01:10:32your future laptop your future PC will
01:10:34become an AI it'll be constantly helping
01:10:37you assisting you in the background the
01:10:41PC will also run applications that are
01:10:44enhanced by AI of course all your photo
01:10:47editing and your writing and your tools
01:10:49all the things that you use will all be
01:10:51enhanced by Ai and your PC will also
01:10:56host applications with digital humans
01:11:00that are AIS and so there are different
01:11:02ways that AIS will manifest themselves
01:11:05and become used in PCS but PCS will
01:11:08become very important AI platform and so
01:11:10where do we go from
01:11:12here I spoke earlier about the scaling
01:11:17of our data centers and every single
01:11:19time we scaled we found a new phase
01:11:23change when we scaled from dgx into
01:11:26large AI supercomputers we enabled
01:11:29Transformers to be able to train on
01:11:31enormously large data data sets well
01:11:35what happened was in the beginning the
01:11:38data was human
01:11:41supervised it required human labeling to
01:11:44train AIS unfortunately there is only so
01:11:47much you can human label Transformers
01:11:50made it possible for unsupervised
01:11:53learning to happen
01:11:55now Transformers just look at an
01:11:57enormous amount of data or look at
01:11:59enormous amount of video or look at
01:12:01enormous amount of uh images and it can
01:12:04learn from studying an enormous amount
01:12:06of data find the patterns and
01:12:07relationships
01:12:09itself while the next generation of AI
01:12:13needs to be physically based most of the
01:12:16AIS today uh don't understand the laws
01:12:18of physics it's not grounded in the
01:12:20physical world in order for us to
01:12:24generate uh uh images and videos and 3D
01:12:29graphics and many physics phenomenons we
01:12:33need AIS that are physically based and
01:12:36understand the laws of physics well the
01:12:38way that you could do that is of course
01:12:40learning from video is One Source
01:12:42another way is synthetic data simulation
01:12:44data and another way is using computers
01:12:47to learn with each other this is really
01:12:50no different than using alphao having
01:12:52alphao play itself selfplay
01:12:55and between the two
01:12:58capabilities same capabilities playing
01:13:00each other for a very long period of
01:13:02time they emerge even smarter and so
01:13:06you're going to start to see this type
01:13:08of AI emerging well if the AI data is
01:13:13synthetically generated and using
01:13:15reinforcement learning it stands to
01:13:17reason that the rate of data generation
01:13:20will continue to advance and every
01:13:22single time data generation grows the
01:13:25amount of computation that we have to
01:13:27offer needs to grow with it we are about
01:13:29to enter a phase where AIS can learn the
01:13:32laws of physics and understand and be
01:13:34grounded in physical world data and so
01:13:36we expect that models will continue to
01:13:38grow and we need larger gpus well
01:13:41Blackwell was designed for this
01:13:43generation this is Blackwell and has
01:13:46several very important Technologies uh
01:13:49one of course is just the size of the
01:13:51chip we took two of the largest a chip
01:13:55that is as large as you can make it at
01:13:57tsmc and we connected two of them
01:14:00together with a 10 terabytes per second
01:14:03link between the world's most advanced
01:14:05cies connecting these two together we
01:14:08then put two of them on a computer node
01:14:11connected with a gray CPU the gray CPU
01:14:14could be used for several things in the
01:14:17training situation it could use it could
01:14:19be used for fast checkpoint and restart
01:14:22in the case of inference and generation
01:14:25could be used for storing context memory
01:14:27so that the AI has memory and
01:14:31understands uh the context of the the
01:14:33conversation you would like to have it's
01:14:34our second generation Transformer engine
01:14:37Transformer engine allows us to adapt
01:14:40dynamically to a lower precision based
01:14:43on the Precision and the range necessary
01:14:46for that layer of computation uh this is
01:14:49our second generation GPU that has
01:14:51secure AI so that you could you could
01:14:53ask your service provider
01:14:55to protect your AI from being either
01:14:58stolen from theft or tampering this is
01:15:02our fifth generation MV link MV link
01:15:05allows us to connect multiple gpus
01:15:07together and I'll show you more of that
01:15:08in a second and this is also our first
01:15:11generation with a reliability and
01:15:14availability engine this system this Ras
01:15:17system allows us to test every single
01:15:20transistor flip-flop memory on chip
01:15:23memory off chip so that we
01:15:26can in the field determine whether a
01:15:29particular chip is uh failing the mtbf
01:15:33the meantime between failure of a
01:15:36supercomputer with 10,000
01:15:38gpus is a measured in
01:15:41hours the meantime between failure of a
01:15:45supercomputer with 100,000 gpus is
01:15:48measured in minutes and so the ability
01:15:52for a supercomputer to run for a long
01:15:55period of time and train a model that
01:15:57could last for several months is
01:15:59practically impossible if we don't
01:16:01invent Technologies to enhance its
01:16:04reliability reliability would of course
01:16:06enhance its uptime which directly
01:16:08affects the cost and then lastly
01:16:10decompression engine data processing is
01:16:12one of the most important things we have
01:16:14to do we added a data compression engine
01:16:17decompression engine so that we can pull
01:16:18data out of storage 20 times faster than
01:16:22what's possible today well all of this
01:16:24repres presents Blackwell and I think we
01:16:26have one here that's in production
01:16:28during GTC I showed you Blackwell in a
01:16:31prototype State um the other
01:16:38side this is why we practice
01:16:53[Laughter]
01:16:55ladies and gentlemen this is
01:17:03Blackwell black Blackwell is in
01:17:07production incredible amounts of
01:17:12Technology this is our production board
01:17:16this is the most complex highest
01:17:18performance computer the world's ever
01:17:22made this is the great CPU
01:17:26and these are you could see each one of
01:17:28these blackw dieses two of them
01:17:30connected together you see that it is
01:17:33the largest die the largest chip the
01:17:36world makes and then we connect two of
01:17:38them together with a 10 terabyte per
01:17:40second
01:17:43link okay and that makes the Blackwell
01:17:46computer and the performance is
01:17:49incredible take a look at
01:17:52this so um
01:17:58you see you see our uh the the
01:18:02computational the flops the AI flops uh
01:18:06for each generation has increased by a
01:18:10thousand times in eight years Mo's law
01:18:13in eight
01:18:15years is something along the lines of oh
01:18:19I don't know maybe 40
01:18:2360 and in the last eight years Moors law
01:18:27has gone a lot lot less and so just to
01:18:31compare even Moors law at its best of
01:18:34times compared to what Blackwell could
01:18:37do so the amount of computations is
01:18:40incredible and when whenever we bring
01:18:42the computation High the thing that
01:18:45happens is the cost goes down and I'll
01:18:47show you what we've done is we've
01:18:50increased through its computational
01:18:52capability the energy
01:18:55used to train a
01:18:59gp4 2 trillion parameter 8 trillion
01:19:03Tokens The amount of energy that is used
01:19:07has gone down by 350 times Well Pascal
01:19:12would have taken
01:19:141,000 gwatt hours 1,000 gwatt hours
01:19:18means that it would take a gigawatt data
01:19:21center the world doesn't have a gigawatt
01:19:23data center but if you had a gigawatt
01:19:24data center it would take a month if you
01:19:28had 100 watt 100 megawatt data center it
01:19:30would take about a year and so nobody
01:19:34would of course um uh create such a
01:19:38thing and um and that's the reason why
01:19:40these large language models chat GPT was
01:19:42impossible only eight years ago by us
01:19:45driving down the increasing the
01:19:47performance the energy efficent while
01:19:49keeping and improving energy efficent
01:19:51efficiency along the way we've now taken
01:19:54with Blackwell what used to be 1,000
01:19:57gwatt hours to three an incredible
01:20:00Advance uh 3 gwatt hours if it's a if
01:20:05it's a um uh uh a 10,000 gpus for
01:20:09example it would only take a coule
01:20:1110,000 gpus I guess it would take a few
01:20:13days 10 days or so so the amount of
01:20:17advance in just eight years is
01:20:20incredible well this is for inference
01:20:23this is for token generation Our token
01:20:27generation performance has made it
01:20:30possible for us to drive the energy down
01:20:32by three 45,000
01:20:35times 177,000 Jewels per token that was
01:20:39Pascal 177,000 Jewels it's kind of like
01:20:43two light bulbs running for two days it
01:20:47would take two light bulbs running for
01:20:49two days amounts of energy 200 Watts
01:20:52running for two days to generate one
01:20:54token of
01:20:57gp4 it takes about three tokens to
01:21:00generate one
01:21:01word and so the amount of energy used
01:21:04necessary for Pascal to generate gp4 and
01:21:08have a chat GPT experience with you was
01:21:10practically impossible but now we only
01:21:13use 0.4 Jew per token and we can
01:21:17generate tokens at incredible rates and
01:21:20very little energy okay so Blackwell is
01:21:23just an enormous lead
01:21:25well even so it's not big enough and so
01:21:29we have to build even larger
01:21:31machines and so the way that we build it
01:21:33is called dgx so this is this is our
01:21:36Blackwell chips and it goes into dgx
01:21:47systems that's why we should
01:21:53practice so this is a dgx Blackwell this
01:21:58has this is air cooled has eight of
01:22:01these gpus inside look at the size of
01:22:04the heat sinks on these
01:22:08gpus about 15 kilowatt 15,000 watts and
01:22:14completely air cooled this version
01:22:16supports x86 and it's it goes into the
01:22:20infrastructure that we've been shipping
01:22:22Hoppers into however if you you would
01:22:24like to have liquid cooling we have a
01:22:26new
01:22:27system and this new
01:22:31system is based on this board and we
01:22:34call it mgx for
01:22:36modular and this modular
01:22:39system you won't be able to see this
01:22:42this can they see this can you see this
01:22:47you can the are
01:22:50you okay
01:22:55I
01:22:57say and so this is the mgx system and
01:23:00here's the two uh black Blackwell boards
01:23:03so this one node has four Blackwell
01:23:06chips these four Blackwell chips this is
01:23:09liquid
01:23:12cooled nine of
01:23:15them nine of them uh well 72 of these 72
01:23:20of these gpus 72 of these G gpus are
01:23:24then connected together with a new MV
01:23:27link this is MV link
01:23:29switch fifth
01:23:31generation and the mvlink switch is a
01:23:34technology Miracle this is the most
01:23:36advanced switch the world's ever made
01:23:38the data rate is
01:23:39insane and these switches connect every
01:23:42single one of these black Wells to each
01:23:45other so that we have one giant 72 GPU
01:23:50Blackwell well the benefit
01:23:55the benefit of this is that in one
01:23:58domain one GPU domain this now looks
01:24:01like one GPU this one GPU has 72 versus
01:24:05the last generation of eight so we
01:24:07increased it by nine times the amount of
01:24:09bandwidth we've increased by 18 times
01:24:11the AI flops we've increased by 45 times
01:24:15and yet the amount of power is only 10
01:24:17times this is 100 kilowatt and that is
01:24:2110 kilow and that's for one now of
01:24:25course well you could always connect
01:24:27more of these together and I'll show you
01:24:29how to do that in a second but what's
01:24:30the miracle is this chip this MV link
01:24:33chip people are starting to awaken to
01:24:35the importance of this mvlink chip as it
01:24:38connects all these different gpus
01:24:40together because the large language
01:24:41models are so large it doesn't fit on
01:24:44just one GPU doesn't fit on one just
01:24:46just one node it's going to take the
01:24:48entire rack of gpus like this new dgx
01:24:52that I that I was just standing next to
01:24:54uh to to hold a large language model
01:24:56that are tens of trillions of parameters
01:24:58large mvlink which in itself is a
01:25:01technology Miracle is 50 billion
01:25:02transistors 74 ports at 400 gbits each
01:25:06four
01:25:07lengths cross-sectional bandwidth of 7
01:25:107.2 terabytes per second but one of the
01:25:12important things is that it has
01:25:13mathematics inside the switch so that we
01:25:16can do reductions which is really
01:25:18important in deep learning right on the
01:25:20Chip And so this is what this is what um
01:25:25a dgx looks like now and a lot of people
01:25:28ask us
01:25:30um you know they say and there's this
01:25:33there's this confusion about what Nvidia
01:25:35does and and um how is it
01:25:40possible that that Nvidia became so big
01:25:43building
01:25:45gpus and so there's an impression that
01:25:48this is what a GPU looks like now this
01:25:50is a GPU this is one of the most
01:25:52advanced gpus in the world but this is a
01:25:54gamer GPU but you and I know that this
01:25:57is what a GPU looks like this is one
01:26:01GPU ladies and gentlemen dgx
01:26:05GPU you
01:26:10know the back of this GPU is the MV link
01:26:15spine the MV link spine is 5,000
01:26:20wires two miles
01:26:26and it's right
01:26:28here
01:26:31this is an mvlink
01:26:34spine and it connects
01:26:37702 gpus to each
01:26:40other this is a electrical mechanical
01:26:45Miracle the transceivers makes it
01:26:48possible for us to drive the entire
01:26:49length in Copper and as a result this
01:26:53switch the mvy switch mvlink switch
01:26:56driving the mvlink spine in Copper makes
01:26:59it possible for us to save 20 kilow in
01:27:02one rack 20 kilowatts could now be used
01:27:05for processing just an incredible
01:27:08achievement so this is the the MV links
01:27:13[Applause]
01:27:21spine wow
01:27:28I went down
01:27:31today and if you and even this is not
01:27:34big
01:27:35enough even this is not big enough for
01:27:37AI factories so we have to connect it
01:27:39all together with very high-speed
01:27:41networking well we have two types of
01:27:44networking we have infin band which has
01:27:45been used uh in supercomputing and AI
01:27:48factories all over the world and it is
01:27:52growing incredibly fast for us however
01:27:54not every data center can handle
01:27:56infiniband because they've already
01:27:58invested their ecosystem in Ethernet for
01:28:00too long and it does take some specialty
01:28:03and some expertise to manage infiniband
01:28:06switches and infiniband networks and so
01:28:08what we've done is we've brought the
01:28:10capabilities of infiniband to the
01:28:13ethernet architecture which is
01:28:15incredibly hard and the reason for that
01:28:17is
01:28:18this ethernet was designed for high
01:28:22average throughput
01:28:25because every single note every single
01:28:28computer is connected to a different
01:28:29person on the internet and most of the
01:28:31communications is the data center with
01:28:33somebody on the other side of the
01:28:35internet
01:28:37however deep learning in AI
01:28:40factories the gpus are not communicating
01:28:43with people on the internet mostly it's
01:28:45communicating with each
01:28:47other they're communicating with each
01:28:49other because they're all they're
01:28:51collecting partial products and they
01:28:54have to reduce it and then redistribute
01:28:56it chunks of partial products reduction
01:29:00redistribution that traffic is
01:29:02incredibly
01:29:03bursty and it is not the average
01:29:06throughput that matters it's the last
01:29:10arrival that matters because if you're
01:29:13reducing collecting partial products
01:29:15from everybody if I'm trying to take all
01:29:18of your
01:29:27so it's not the average throughput it's
01:29:30whoever gives me the answer
01:29:33last okay ethernet has no provision for
01:29:36that and so there are several things
01:29:39that we had to create we created an
01:29:41endtoend architecture so that the the
01:29:43Nick and the switch can communicate and
01:29:46we applied four different Technologies
01:29:48to make this possible number one Nvidia
01:29:50has the world's most advanced RDMA and
01:29:53so now we have the ability to have a
01:29:55network level RDMA for ethernet that is
01:29:58incredibly great number two we have
01:30:00congestion control the switch does
01:30:03Telemetry at all times incredibly fast
01:30:06and whenever the uh the uh the the gpus
01:30:10or the the Nyx are sending too much
01:30:12information we can tell them to back off
01:30:14so that it doesn't create hotspots
01:30:16number three adaptive
01:30:18routing ethernet needs to transmit and
01:30:22receive in order
01:30:24we see
01:30:27congestions or we see uh ports that are
01:30:30not currently being used irrespective of
01:30:32the ordering we will send it to the
01:30:34available ports and Bluefield on the
01:30:37other end reorders it so that it comes
01:30:40back in order that adaptive routing
01:30:43incredibly powerful and then lastly
01:30:45noise
01:30:46isolation there's more than one model
01:30:49being trained or something happening in
01:30:51the data center at all times and their
01:30:53noise and their traffic could get into
01:30:55each other and causes Jitter and so when
01:30:58it when the noise of one training model
01:31:01one model training causes the last
01:31:03arrival to end up too late it really
01:31:07slows down to training well overall
01:31:09remember you have you have built A5
01:31:13billion dollar or3 billion doll Data
01:31:15Center and you're using this for
01:31:17training if the
01:31:19utilization Network
01:31:22utilization was
01:31:2540% lower and as a result the training
01:31:28time was 20%
01:31:30longer
01:31:31the5 billion data center is effectively
01:31:35like a$6 billion doll data center so the
01:31:37cost is incredible the cost impact is
01:31:40quite
01:31:41High ethernet with Spectrum X basically
01:31:46allows us to improve the performance so
01:31:48much that the network is basically free
01:31:50and so this is really quite an
01:31:52achievement we're very we have a whole
01:31:55pipeline of ethernet products behind us
01:31:56this is Spectrum
01:31:58x800 it is uh 51.2 terabits per second
01:32:03and 256
01:32:05radic the next one coming is 512 radic
01:32:08is one year from now 512 Radix and
01:32:11that's called Spectrum x800 Ultra and
01:32:13the one after that is x1600 but the
01:32:17important idea is this x800 is designed
01:32:20for tens of
01:32:22thousands tens T of thousands of gpus
01:32:26x800 ultra is designed for hundreds of
01:32:29thousands of gpus and x1600 is designed
01:32:32for millions of
01:32:34gpus the days of millions of GPU data
01:32:38centers are coming and the reason for
01:32:40that is very simple of course we want to
01:32:43train much larger models but very
01:32:45importantly in the future almost every
01:32:48interaction you have with the Internet
01:32:50or with a computer will likely have a
01:32:52generative AI running in the cloud
01:32:55somewhere and that generative AI is
01:32:58working with you interacting with you
01:33:00generating videos or images or text or
01:33:04maybe a digital human and so you're
01:33:06interacting with your computer almost
01:33:08all the time and there's always a
01:33:10generative AI connected to that some of
01:33:13it is on Prem some of it is on your
01:33:15device and a lot of it could be in the
01:33:17cloud these generative AIS will also do
01:33:20a lot of reasoning capability instead of
01:33:22just one shot answer
01:33:24they might iterate on answers so that it
01:33:26improve the quality of the answer before
01:33:28they give it to you and so the amount of
01:33:30generation we're going to do in the
01:33:32future is going to be extraordinary
01:33:34let's take a look at all of this put
01:33:36together now tonight this is our first
01:33:38nighttime
01:33:41keyote I want to
01:33:50thank I want to thank all of you for
01:33:52coming out tonight at 7:00 and and so
01:33:56what I'm about to show you has a new
01:33:59Vibe okay there's a new Vibe this is
01:34:02kind of the nighttime keynote Vibe so
01:34:05enjoy
01:34:13this black
01:34:21[Music]
01:34:25let's go go go go
01:34:31go okay
01:34:38[Music]
01:34:55[Music]
01:35:06[Music]
01:35:08come on yeah yeah yeah
01:35:14yeah get it y get it
01:35:21y let's go
01:35:26[Music]
01:35:26[Applause]
01:35:33the more you back the more you sa with
01:35:35top ey Tor made that's to speed a
01:35:41[Music]
01:35:47[Applause]
01:35:57now you can't do that on a morning
01:36:03keynote I think that style of keynote
01:36:06has never been done in compy Tex
01:36:09ever might be the
01:36:17last only Nvidia can pull off
01:36:20that only I could do that
01:36:23[Applause]
01:36:29Blackwell of course uh is the first
01:36:32generation of Nvidia platforms that was
01:36:35launched at the beginning at the right
01:36:37as the world knows the generative AI era
01:36:41is here just as the world realized the
01:36:44importance of AI factories just as the
01:36:46beginning of this new Industrial
01:36:48Revolution uh we have so much support
01:36:51nearly every OEM every computer maker
01:36:54every CSP every GPU Cloud Sovereign
01:36:58clouds even telecommunication
01:37:01companies Enterprises all over the world
01:37:05the amount of success the amount of
01:37:07adoption the amount of of uh enthusiasm
01:37:10for Blackwell is just really off the
01:37:11charts and I want to thank everybody for
01:37:19that we're not stopping there uh during
01:37:23this during the time of this incredible
01:37:25growth we want to make sure that we
01:37:27continue to enhance performance continue
01:37:30to drive down cost cost of training cost
01:37:33of inference and continue to scale out
01:37:36AI capabilities for every company to
01:37:39embrace the further we the further
01:37:42performance we drive up the Greater the
01:37:44cost decline Hopper platform of course
01:37:47was the most successful data center
01:37:49processor probably in history and this
01:37:52is just an incredible incredible success
01:37:54story however Blackwell is here and
01:37:59every single platform as you'll notice
01:38:01are several things you got the CPU you
01:38:03have the GPU you have MV link you have
01:38:04the Nick and you have the switch the the
01:38:07mvlink switch the uh connects all of the
01:38:10gpus together as large of a domain as we
01:38:13can and whatever we can do we connect it
01:38:15with large um very large and very
01:38:17highspeed switches every single
01:38:19generation as you'll see is not just a
01:38:21GPU but it's entire platform
01:38:24we build the entire platform we
01:38:26integrate the entire platform into an AI
01:38:28Factory supercomputer however then we
01:38:31this aggregate it and offer it to the
01:38:33world and the reason for that is because
01:38:35all of you could create interesting and
01:38:38Innovative configurations and and and
01:38:42all kinds of different uh uh Styles and
01:38:44and fit different uh data centers and
01:38:47different customers in different places
01:38:49some of it for Edge some of it for Telco
01:38:51and all of the different Innovation are
01:38:53POS possible if it would we made the
01:38:55systems open and make it possible for
01:38:57you to innovate and so we design it
01:39:00integrated but we offer it to you
01:39:02disintegrated so that you could create
01:39:04modular systems the Blackwell platform
01:39:07is
01:39:07here our company is on a one-year
01:39:10Rhythm we our basic philosophy is very
01:39:14simple one build the entire data center
01:39:16scale disaggregated and sell it to you
01:39:19in Parts on a one-year Rhythm and we
01:39:22push everything the technology limits
01:39:25whatever tsmc process technology will
01:39:29push it to the absolute limits whatever
01:39:30packaging technology push it to the
01:39:32absolute limits whatever memory
01:39:33technology push it to Absolute limits
01:39:35series technology Optics technology
01:39:38everything is pushed to the
01:39:40Limit well and then after that do
01:39:43everything in such a way so that all of
01:39:45our software runs on this entire install
01:39:48base software inertia is the single most
01:39:51important thing in computers it'll when
01:39:53a computer is backwards compatible and
01:39:55it's architecturally compatible with all
01:39:57the software that has already been
01:39:59created your ability to go to market is
01:40:01so much faster and so the velocity is
01:40:06incredible when we can take advantage of
01:40:08the entire installed base of software
01:40:09has already been created well Blackwell
01:40:11is here next year is Blackwell Ultra
01:40:16just as we had h100 and h200 you'll
01:40:19probably see some pretty exciting New
01:40:21Generation from us for black Blackwell
01:40:24Ultra again again push to the limits and
01:40:28the Next Generation Spectrum switches I
01:40:30mentioned well this is the very first
01:40:33time that this next click has been
01:40:38made and I'm not sure yet whe I'm going
01:40:40to regret this or
01:40:45[Applause]
01:40:49not we have code names in our company
01:40:52and uh we try to keep them very secret
01:40:55uh often times uh most of the employees
01:40:57don't even know but our next Generation
01:40:59platform is called Reuben the Reuben
01:41:01platform the Reuben platform um I'm I'm
01:41:05not going to spend much time on it uh I
01:41:07know what's going to happen you're going
01:41:08to take pictures of it and you're going
01:41:09to go look at the fine prints uh and
01:41:11feel free to do that so we have the
01:41:13Reuben platform and one year later we'
01:41:15have the Reuben um Ultra platform all of
01:41:18these chips that I'm showing you here
01:41:19are all in full development 100% of them
01:41:22and the rhythm is one year at the limits
01:41:25of Technology all 100% architecturally
01:41:28compatible so this is this is basically
01:41:31what Nvidia is building and all of the
01:41:33riches of software on top of it so in a
01:41:35lot of ways the last 12
01:41:37years from that moment of imag net and
01:41:42US realizing that the future of
01:41:44computing was going to radically change
01:41:46to today is really exactly as I was
01:41:49holding up earlier GeForce pre 20102 and
01:41:53Nvidia
01:41:55today the company has really transformed
01:41:58tremendously and I want to thank all of
01:41:59our partners here for supporting us
01:42:01every step along the way this is the
01:42:04Nvidia black wall
01:42:14platform let me talk about what's
01:42:18next the next wave of AI is physical AI
01:42:23AI that understands the laws of physics
01:42:25AI that can work among us and so they
01:42:30have to understand the world model so
01:42:34that they understand how to interpret
01:42:36the world how to perceive the world they
01:42:38have to of course have excellent
01:42:40cognitive capabilities so they can
01:42:42understand us understand what we asked
01:42:45and perform the
01:42:46tasks in the
01:42:49future robotics is a much more p
01:42:53pervasive idea of course when I say
01:42:55robotics there's a humanoid robotics
01:42:58that's usually the representation of
01:43:00that but that's not at all true
01:43:03everything is going to be robotic all of
01:43:05the factories will be robotic the
01:43:08factories will orchestrate robots and
01:43:11those robots will be building products
01:43:13that are
01:43:15robotic robots interacting with robots
01:43:19building products that are robotic well
01:43:23in order for us to do that we need to
01:43:24make some breakthroughs and let me show
01:43:26you the
01:43:33video the era of Robotics has
01:43:37arrived one day everything that moves
01:43:40will be
01:43:42autonomous researchers and companies
01:43:44around the world are developing robots
01:43:47powered by physical
01:43:49AI physical AIS are models that can
01:43:53understand instructions and autonomously
01:43:56perform complex tasks in the real
01:43:59world multimodal llms are breakthroughs
01:44:03that enable robots to learn perceive and
01:44:07understand the world around them and
01:44:08plan how they'll act and from Human
01:44:12demonstrations robots can now learn the
01:44:14skills required to interact with the
01:44:16world using gross and fine motor
01:44:19skills one of the integral Technologies
01:44:22for advancing robotics
01:44:23is reinforcement learning just as llms
01:44:26need rhf or reinforcement learning from
01:44:29Human feedback to learn particular
01:44:32skills generative physical AI can learn
01:44:35skills using reinforcement learning from
01:44:37physics feedback in a simulated World
01:44:40these simulation environments are where
01:44:42robots learn to make decisions by
01:44:45performing actions in a virtual world
01:44:48that obeys the laws of
01:44:50physics in these robot gyms
01:44:53a robot can learn to perform complex and
01:44:56dynamic tasks safely and quickly
01:44:59refining their skills through millions
01:45:02of Acts of trial and error we built
01:45:05Nvidia Omniverse as the operating system
01:45:08where physical a can be
01:45:10created Omniverse is a development
01:45:13platform for virtual world simulation
01:45:16combining realtime physically based
01:45:19rendering physics
01:45:21simulation and generative AI
01:45:25Technologies in Omniverse robots can
01:45:27learn how to be robots they learn how to
01:45:30autonomously manipulate objects with
01:45:33Precision such as grasping and handling
01:45:36objects or navigate environments
01:45:39autonomously finding optimal paths while
01:45:42avoiding obstacles and
01:45:45Hazards learning in Omniverse minimizes
01:45:48the Sim to real Gap and maximizes the
01:45:51transfer of learned beh
01:45:53Behavior building robots with generative
01:45:56physical AI requires three
01:45:58computers Nvidia AI supercomputers to
01:46:02train the models Nvidia Jetson Orin and
01:46:05Next Generation Jetson Thor robotic
01:46:08supercomputer to run the
01:46:10models an Nvidia Omniverse where robots
01:46:13can learn and refine their skills in
01:46:15simulated
01:46:17worlds we build the platforms
01:46:20acceleration libraries and AI models
01:46:23needed by developers and companies and
01:46:26allow them to use any or all of the
01:46:30stacks that suit them best the next wave
01:46:33of AI is here robotics powered by
01:46:36physical AI will revolutionize
01:46:39[Music]
01:46:49Industries this isn't the future this is
01:46:52happening
01:46:54now there are several ways that we're
01:46:57going to serve the market the first
01:46:58we're going to create platforms for each
01:47:01type of robotic systems one for robotic
01:47:03factories and warehouses one for robots
01:47:07that manipulate things one for robots
01:47:10that move and one for uh robots that are
01:47:14humanoid and so each one of these
01:47:16robotic robotics platform is like almost
01:47:20everything else we do a computer acceler
01:47:23libraries and pre-train models computers
01:47:25acceleration libraries pre-train models
01:47:28and we test everything we train
01:47:30everything and integrate everything
01:47:32inside Omniverse where Omniverse is as
01:47:35the video was saying where robots learn
01:47:37how to be robots now of course the
01:47:40ecosystem of robotic warehouses is
01:47:42really really complex it takes a lot of
01:47:45companies a lot of tools a lot of
01:47:47technology to build a modern warehouse
01:47:50and warehouses are increasingly robotic
01:47:52one of these things days will be fully
01:47:53robotic and so in each one of these
01:47:57ecosystems we have sdks and apis that
01:48:00are connected into the software industry
01:48:04sdks and apis connected into Edge AI
01:48:07industry um and companies and then also
01:48:10of course uh systems that are designed
01:48:12for plc's and robotic systems for the
01:48:14odms it's then integrated by integrators
01:48:18uh created for ultimately uh building
01:48:20warehouses uh for customers here we have
01:48:23an example of kenmac building a robotic
01:48:26warehouse for giant giant
01:48:36group okay and then here now let's talk
01:48:39about factories factories has a
01:48:41completely different ecosystem and
01:48:43foxcon is building some of the world's
01:48:45most advanced factories their ecosystem
01:48:48again Edge Computers and Robotics
01:48:51software for designing the factories the
01:48:55workflows programming the robots and of
01:48:58course PLC computers that orchestrate uh
01:49:01the digital factories and the AI
01:49:02factories we have sdks uh that are
01:49:05connected into each one of these
01:49:07ecosystems as well this is happening all
01:49:10over Taiwan foxcon has buil is building
01:49:14digital twins of their
01:49:16factories Delta is building digital
01:49:19twins of their
01:49:20factories by the way half is real half
01:49:22half is digital half is
01:49:25Omniverse pegatron is building digital
01:49:28twins of their robotic
01:49:31factories wistron is building digital
01:49:34twins of their robotic
01:49:37factories and this is really cool this
01:49:39is a video of foxc con's new Factory
01:49:42let's take a
01:49:47look demand for NVIDIA accelerated
01:49:49Computing is skyrocketing as the World
01:49:52modernizes traditional data centers into
01:49:55generative AI
01:49:57factories foxcon the world's largest
01:50:01electronics manufacturer is gearing up
01:50:03to meet this Demand by building robotic
01:50:05factories with Nvidia Omniverse and
01:50:08AI Factory planners use Omniverse to
01:50:11integrate facility and Equipment data
01:50:13from leading industry applications like
01:50:15seamen team Center X and Autodesk
01:50:19Revit in the digital twin they optimize
01:50:22floor layout and line configurations and
01:50:25locate optimal camera placements to
01:50:27monitor future operations with Nvidia
01:50:29Metropolis powered Vision
01:50:33AI virtual integration saves planners on
01:50:36the enormous cost of physical change
01:50:40orders during construction the foxcon
01:50:43teams use the digital twin as the source
01:50:45of Truth to communicate and validate
01:50:48accurate equipment
01:50:50layout the Omniverse digital twin is
01:50:53also the robot gym where foxcon
01:50:56developers train and test Nvidia ISAC AI
01:50:59applications for robotic perception and
01:51:02manipulation and Metropolis AI
01:51:04applications for Sensor
01:51:07Fusion in Omniverse foxcon simulates two
01:51:10robot AIS before deploying run times to
01:51:13Jets and
01:51:14computers on the assembly line they
01:51:17simulate Isaac manipulator libraries and
01:51:19AI models for automated Optical
01:51:22inspection for object identification
01:51:24defect detection and trajectory
01:51:28planning to transfer hgx systems to the
01:51:31test pods they simulate Isaac perceptor
01:51:34powered fobot amrs as they perceive and
01:51:37move about their environment with 3D
01:51:39mapping and
01:51:41reconstruction with Omniverse foxcon
01:51:44builds their robotic factories that
01:51:46orchestrate robots running on Nvidia
01:51:48ISAC to build Nvidia AI supercomputers
01:51:52which which in turn train boxon
01:51:56[Music]
01:52:04[Applause]
01:52:09robots so a robotic Factory is designed
01:52:13with three computers train the AI on
01:52:15Nvidia AI you have the robot running on
01:52:19the PLC systems uh for orchestrating the
01:52:22the the uh the factories and then you of
01:52:24course simulate everything inside
01:52:26Omniverse well the robotic arm and the
01:52:29robotic amrs are also the same way three
01:52:31computer systems the difference is the
01:52:34two omniverses will come together so
01:52:37they'll share one virtual space when
01:52:39they share one virtual space that
01:52:42robotic arm will become
01:52:44inside the robotic
01:52:46Factory and again three three uh three
01:52:50computers and we provide the computer
01:52:53the acceleration layers and pre-train uh
01:52:56pre-train AI models we've connected
01:52:58Nvidia manipulator and Nvidia Omniverse
01:53:02with Seamans the world's leading
01:53:04Industrial Automation software and
01:53:05systems company this is really a
01:53:07fantastic partnership and they're
01:53:09working on factories all over the world
01:53:11semantic pick aai now integrates Isaac
01:53:15manipulator and sematic pick AI uh runs
01:53:19operates ABB CA yasawa uh Fook uh
01:53:23Universal robotics um and techman and so
01:53:27semmens is a fantastic integration we
01:53:30have all kinds of other Integrations
01:53:32let's take a
01:53:35look arcbest is integrating Isaac
01:53:38perceptor into Vox smart autonomy robots
01:53:41for enhanced object recognition and
01:53:44human motion tracking in Material
01:53:47Handling byd Electronics is integrating
01:53:50Isaac manipulator and percep ctor into
01:53:53their AI robots to enhance manufacturing
01:53:56efficiencies for Global
01:53:58customers ideal works is building Isaac
01:54:01perceptor into their iwos software for
01:54:04AI robots in Factory
01:54:07Logistics intrinsic an alphabet company
01:54:10is adopting Isaac manipulator into their
01:54:12Flow State platform to advance robot
01:54:15grasping Gideon is integrating Isaac
01:54:18perceptor into Trey AI powered forklifts
01:54:21to advance AI enabled Logistics Argo
01:54:25robotics is adopting Isaac perceptor
01:54:27into perception engine for Advanced
01:54:29Vision based amrs Solomon is using Isaac
01:54:33manipulator AI models in their acup piic
01:54:363D software for industrial
01:54:39manipulation techman robot is adopting
01:54:41Isaac Sim and manipulator into TM flow
01:54:44accelerating automated Optical
01:54:47inspection pteridine robotics is
01:54:49integrating Isaac manipulator into po
01:54:52scope X for cobots and Isaac perceptor
01:54:55into mere
01:54:57amrs vention is integrating Isaac
01:55:00manipulator into machine Logic for AI
01:55:03manipulation
01:55:04[Music]
01:55:08robots robotics is here physical AI is
01:55:12here this is not science fiction and
01:55:14it's being used all over Taiwan and just
01:55:17really really exciting and that's the
01:55:20factory the robots inside and of course
01:55:22all the products are going to be
01:55:23robotics there are two very high volume
01:55:26robotics products one of course is the
01:55:29self-driving car or cars that have a
01:55:31great deal of autonomous capability
01:55:33Nvidia again builds the entire stack
01:55:35next year we're going to go to
01:55:37production with the Mercedes Fleet and
01:55:38after that in 2026 the jlr fleet uh we
01:55:43offer the full stack to the world
01:55:45however you're welcome to take whichever
01:55:47Parts uh which whichever layer of our
01:55:50stack just as the entire the entire uh
01:55:53uh Drive stack is open the next high
01:55:56volume robotics product that's going to
01:55:58be manufactured by robotic factories
01:56:01with robots inside will likely be
01:56:03humanoid robots and this has great
01:56:07progress in recent years in both the
01:56:10cognitive capability because of
01:56:12foundation models and also the world
01:56:14understanding capability that we're in
01:56:16the process of developing I'm really
01:56:18excited about this area because
01:56:20obviously the easiest robot to adapt
01:56:22into the world are human or robots
01:56:24because we built the world for us we
01:56:26also have the vast the most amount of
01:56:28data to train these robots than other
01:56:31types of robots because we have the same
01:56:33uh physique and so the amount of
01:56:35training data we can provide through
01:56:37demonstration capabilities and video
01:56:39capabilities is going to be really great
01:56:40and so we're going to see a lot of
01:56:42progress in this area well I think we
01:56:45have um some robots that we' like to uh
01:56:48welcome
01:56:53there we go about my
01:56:59[Applause]
01:57:02size and we have we have some friends to
01:57:05join us so the fut the future of robot
01:57:08robotics is here the next wave of AI and
01:57:12and of course you know Taiwan
01:57:16builds computers with
01:57:18keyboards you build computers for your
01:57:21pocket you build computers for data
01:57:24centers in the cloud in the future
01:57:27you're going to build computers that
01:57:28walk and computers that roll you know
01:57:33around and um so these are all just
01:57:36computers and as as it turns out uh the
01:57:38technology is very similar to the
01:57:40technology of building um all of the
01:57:43other computers that you already build
01:57:44today so this is going to be a really
01:57:46extraordinary uh Journey for
01:57:49us well uh I want to thank I want
01:57:58to I want to thank I want to I have I've
01:58:00I've made one last video if you don't
01:58:02mind uh something that that uh uh uh we
01:58:06we really enjoyed making um and if you
01:58:10let's run
01:58:21it
01:58:23[Music]
01:58:35[Music]
01:58:51for
01:59:06[Music]
01:59:21for for
01:59:24[Music]
01:59:31[Music]
01:59:43[Music]
01:59:54[Music]
01:59:56[Applause]
02:00:01[Music]
02:00:06[Music]
02:00:15[Applause]
02:00:20thank you
02:00:23I love you guys thank
02:00:24[Applause]
02:00:32you thank you all for coming have a
02:00:34great comput
02:00:40text thank
02:00:41[Applause]
02:00:46you I will make you like me yeah singing
02:00:51with me
02:00:54[Music]
02:01:10once again a come T we will show you all
02:01:15about best yeah digital dreams and
02:01:20me with
02:01:24to all of in
02:01:26Tai we hope that you like this song yeah
02:01:31good
02:01:33me and
02:01:37[Music]
02:01:47[Music]
02:02:04once
02:02:07get thank you for joining us yeah digal
02:02:12Twins and
02:02:14me we been
02:02:22I
02:02:28[Music]
02:02:37feeled
02:02:39hey I am awake for the new day feeling
02:02:43the sun upon my face I'm all the bad
02:02:46vibes away look to the sky I have this
02:02:50moment inside the glove I am the Sun up
02:02:53so high above I am the star you are
02:02:55thinking know take a minute now take a
02:02:58minute to see take a minute now we are
02:03:01put in the SE this is everything that I
02:03:03could ever KN on the top again in the
02:03:05land of the free
02:03:20and
02:03:50e
02:04:20e e
🎥 Related Videos![What vaccinating vampire bats can teach us about pandemics | Daniel Streicker](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FlSPxeA6Z_m4%2Fhqdefault.jpg&w=3840&q=75)
![a16z Podcast | Things Come Together -- Truths about Tech in Africa](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fjbl5I7kYWHM%2Fhqdefault.jpg&w=3840&q=75)
![2024 TSCRS Applications of anterior segments diagnostic instruments in cataract surgery](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FWnUDk5vKuqE%2Fhqdefault.jpg&w=3840&q=75)
![a16z Podcast | The Infrastructure of Total Health](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FoLGRslHa5FE%2Fhqdefault.jpg&w=3840&q=75)
![The Robot Lawyer Resistance with Joshua Browder of DoNotPay](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FAmVdYPTdw2c%2Fhqdefault.jpg&w=3840&q=75)
![NES Controllers Explained](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnAStgQzPrAQ%2Fhqdefault.jpg&w=3840&q=75)
![What vaccinating vampire bats can teach us about pandemics | Daniel Streicker](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FlSPxeA6Z_m4%2Fhqdefault.jpg&w=3840&q=75)
What vaccinating vampire bats can teach us about pandemics | Daniel Streicker
![a16z Podcast | Things Come Together -- Truths about Tech in Africa](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fjbl5I7kYWHM%2Fhqdefault.jpg&w=3840&q=75)
a16z Podcast | Things Come Together -- Truths about Tech in Africa
![2024 TSCRS Applications of anterior segments diagnostic instruments in cataract surgery](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FWnUDk5vKuqE%2Fhqdefault.jpg&w=3840&q=75)
2024 TSCRS Applications of anterior segments diagnostic instruments in cataract surgery
![a16z Podcast | The Infrastructure of Total Health](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FoLGRslHa5FE%2Fhqdefault.jpg&w=3840&q=75)
a16z Podcast | The Infrastructure of Total Health
![The Robot Lawyer Resistance with Joshua Browder of DoNotPay](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FAmVdYPTdw2c%2Fhqdefault.jpg&w=3840&q=75)
The Robot Lawyer Resistance with Joshua Browder of DoNotPay
![NES Controllers Explained](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnAStgQzPrAQ%2Fhqdefault.jpg&w=3840&q=75)
NES Controllers Explained
🔥 Recently Summarized Examples![The Hitler-Stalin Pact | Reflections Episode 9](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FFiCmS9niS2U%2Fhqdefault.jpg&w=3840&q=75)
![Uncovering Corruption From Health "Experts" | Scott Carney](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FJP76ekb_DxI%2Fhqdefault.jpg&w=3840&q=75)
![The Forgotten Geometry: A New Path to Unification](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FTTSeqsCgxj8%2Fhqdefault.jpg&w=3840&q=75)
![Joe Rogan Experience #2194 - Luis Elizondo](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2F9gLPtRwXgCM%2Fhqdefault.jpg&w=3840&q=75)
![From Tesla to DNA: The Science of Scalar Waves - Dr. Sandra Rose Michael - Think Tank E44](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FzPxOsOsbNo8%2Fhqdefault.jpg&w=3840&q=75)
![Bitcoin Holders...Watch Out for Sept](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FNedUS9J4Er0%2Fhqdefault.jpg&w=3840&q=75)
![The Hitler-Stalin Pact | Reflections Episode 9](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FFiCmS9niS2U%2Fhqdefault.jpg&w=3840&q=75)
The Hitler-Stalin Pact | Reflections Episode 9
![Uncovering Corruption From Health "Experts" | Scott Carney](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FJP76ekb_DxI%2Fhqdefault.jpg&w=3840&q=75)
Uncovering Corruption From Health "Experts" | Scott Carney
![The Forgotten Geometry: A New Path to Unification](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FTTSeqsCgxj8%2Fhqdefault.jpg&w=3840&q=75)
The Forgotten Geometry: A New Path to Unification
![Joe Rogan Experience #2194 - Luis Elizondo](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2F9gLPtRwXgCM%2Fhqdefault.jpg&w=3840&q=75)
Joe Rogan Experience #2194 - Luis Elizondo
![From Tesla to DNA: The Science of Scalar Waves - Dr. Sandra Rose Michael - Think Tank E44](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FzPxOsOsbNo8%2Fhqdefault.jpg&w=3840&q=75)
From Tesla to DNA: The Science of Scalar Waves - Dr. Sandra Rose Michael - Think Tank E44
![Bitcoin Holders...Watch Out for Sept](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FNedUS9J4Er0%2Fhqdefault.jpg&w=3840&q=75)