#黃仁勳 驚喜為 #美超微 #梁見後 #COMPUTEX 主題演講站台｜完整精華

Summary

在这段视频中，Nvidia的CEO Jensen Huang分享了他对人工智能和加速计算的远见。他强调了数据量指数级增长与CPU扩展速度放缓之间的矛盾，以及加速计算如何帮助释放数据中心的浪费能量，并加速每个应用。Jensen还介绍了生成式AI（generative AI）的概念，它将影响全球所有数据中心，推动其现代化。此外，他还提到了液冷技术在降低功耗方面的作用，以及如何通过提高吞吐量、利用率和启动速度来增加收入。最后，他强调了安全性、技术进步和良好政策的重要性，以确保AI的健康发展。

Takeaways

• 🧠 AI正在改变最小化计算，因为数据量呈指数级增长，而CPU的扩展速度已经放缓多年。
• 🌿 绿色计算（Green Computing）即能效计算，是Nvidia所关注的，旨在减少数据中心的能源浪费。
• 🚀 加速计算（Accelerated Computing）时代已经到来，它将释放数据中心的潜能，提高应用和数据中心的效率。
• 🤖 生成式AI（Generative AI）是一种新兴的计算方式，它涉及文本、图像、视频的生成，将影响全球所有的数据中心。
• 💰 数据中心需要现代化改造以适应生成式AI，这将带来巨大的需求和市场潜力。
• 🔧 Nvidia和Supermicro合作，提供多种新产品和解决方案，以支持加速计算和数据中心的现代化。
• 💡 液冷技术（DLC）正在被用于降低数据中心的功耗，提高AI芯片的制造能力。
• 🔗 高速互连技术（如NVLink）使得GPU和CPU之间能够高效地传输数据，这对于AI训练和推理至关重要。
• 🛠️ Nvidia的软件堆栈，包括CUDA和网络计算，为高性能计算提供了强大的支持。
• 🔐 安全技术的发展同样重要，需要为AI制定安全政策和实践，以确保其可靠性和安全性。

Q & A

Nvidia的CEO Jensen在演讲中提到了哪些关于AI和计算的变革？

Jensen在演讲中提到了加速计算(accelerated computing)和绿色计算(Green Computing)的到来，强调了由于数据量指数级增长和CPU扩展放缓，加速计算能够释放数据中心中浪费的能源和成本。他还提到了生成式AI(generative AI)的崛起，这将影响全球所有的数据中心。

Jensen提到了加速计算能带来哪些好处？

加速计算能够节省数据中心中浪费的能源和成本，提高应用和数据中心的效率。通过加速计算，可以释放之前浪费的能源，用于新的目的，比如加速每个应用和数据中心。

什么是生成式AI，它将如何影响数据中心？

生成式AI是一种能够生成文本、图像、视频等内容的人工智能技术。Jensen认为，生成式AI的转变将影响世界上每一个数据中心，因为到2030年，全球数据的价值可能达到3万亿美元，这要求我们现代化所有的数据中心，以适应这些先进的系统。

Jensen如何描述Super Micro和Nvidia的合作关系？

Jensen将Super Micro团队描述为准备好接受订单的合作伙伴，他们准备好提供所需的产品和服务，以帮助现代化数据中心。他还开玩笑说自己是Super Micro的最佳销售人员，并且他们之间有良好的合作关系。

在演讲中提到的数据中心现代化涉及到哪些方面？

数据中心现代化涉及到采用加速计算和生成式AI等先进技术，以及采用液冷系统(DLC)来降低功耗，提高性能和成本效率。此外，还需要软件的兼容性和整个系统的整合，以确保快速启动、高利用率和高吞吐量。

为什么Jensen强调了数据中心的能源效率？

Jensen强调能源效率是因为在加速计算中，能源效率直接转化为更高的性能。通过节省能源，可以将这些能量用于更多的计算任务，从而提高数据中心的整体性能和产出。

Nvidia在数据中心现代化中扮演了什么角色？

Nvidia在数据中心现代化中扮演了关键角色，提供了加速计算平台和生成式AI技术。Nvidia的GPU和其他技术是构建现代AI工厂的基础，这些AI工厂能够直接产生收入。

Jensen提到了哪些Nvidia的新产品或技术？

Jensen提到了Nvidia的多个新产品和技术，包括新的GPU服务器、液冷系统(DLC)以及与Super Micro合作的各种配置，如Blackwell GPU和Grace CPU的结合使用。

在演讲中，Jensen如何描述AI的未来和其在不同行业中的应用？

Jensen认为AI的未来非常重要，他提到生成式AI能够嵌入智能，并且可以在许多不同的行业中使用。他强调了AI的实用性，以及如何通过不同的技术来提高AI的安全性和功能性。

Jensen在演讲中提到了哪些关于AI安全性的观点？

Jensen提到为了确保AI的安全性，我们需要开发出多种技术，比如AI监控AI、人员监控AI以及建立各种安全护栏。他强调了良好的政策、实践和工业标准的重要性，以确保AI的安全和可靠。

Jensen如何描述Super Micro在数据中心现代化中的角色？

Jensen将Super Micro描述为一个准备好立即服务的合作伙伴，他们能够提供所需的硬件和工程支持，以帮助客户快速启动和运行现代化的数据中心。

在演讲中，Jensen提到了哪些关于数据中心的经济效益？

Jensen提到了数据中心的经济效益与吞吐量、利用率和启动时间直接相关。他强调了通过提高这些指标，可以增加数据中心的收入，这也是为什么他们称其为AI工厂的原因。

Jensen在演讲中提到了哪些关于软件在高性能计算中的作用？

Jensen强调了软件在高性能计算中的重要性，他提到了Cuda、网络计算以及分布式计算的软件堆栈，这些都是构建在Nvidia硬件之上的。他还提到了软件的兼容性，这是高性能计算中非常关键的部分。

Outlines

00:00

🤖 AI与加速计算的新时代

本段落介绍了Nvidia CEO Jensen Huang的演讲，他强调了AI如何改变最小化计算，并提出了加速计算和绿色计算的概念。加速计算由于数据量指数级增长和CPU扩展放缓，变得尤为重要。Jensen提到了数据中心的能源和成本浪费问题，并强调通过加速计算可以释放这些资源，用于新的目的。此外，他还提出了生成性AI的概念，这是一种能够生成文本、图像、视频等的AI技术，它将影响全球所有的数据中心，并需要在未来几年内进行现代化改造。

05:01

🚀 数据中心的液体冷却技术革新

在第二段中，讨论了数据中心的液体冷却技术（DLC），这是一种降低功耗并提高AI芯片制造能力的技术。Jensen Huang提到了Supermicro公司每月能够生产1000个液体冷却系统，这将是一个巨大的公司规模。他还提到了GPU服务器的复杂性，强调了除了GPU之外，还有许多其他组件和技术。此外，Jensen还提到了通过液体冷却系统节省下来的能源可以用于未来的计算，这将有助于创建新的商品——计算吞吐量，这将直接影响数据中心的收入。

10:02

🏭 AI工厂的兴起与经济效益

第三段讲述了AI工厂的概念，这些工厂不仅仅是数据存储或文件检索的地方，而是直接产生收入的设施。Jensen Huang强调了AI工厂的启动时间、吞吐量和利用率对收入的影响，以及如何通过整合系统到机架规模来提高这些指标。他还提到了Supermicro公司如何准备好为这些AI工厂提供支持，并且所有这些系统都经过了Nvidia软件的认证。此外，Jensen还讨论了不同类型的配置，以适应不同的使用案例和数据中心需求。

15:03

🌐 网络作为计算结构的重要性

在第四段中，Jensen Huang讨论了网络作为计算结构的重要性，特别是对于分布式计算。他提到了Cuda和DOA这两个软件堆栈，它们使得工作负载能够高效地在网络中分布。此外，他还提到了400和800吉比特每秒的高速互连，以及即将到来的1600吉比特每秒的技术。Jensen强调了在训练阶段和推理阶段，不同类型的CPU如何被用于AI，以及如何通过节能的互连提高性能。

20:05

🛡️ AI的安全性与未来发展

最后一段中，Jensen Huang讨论了AI的安全性和未来发展。他比喻了飞机的自动驾驶技术，强调了为了确保AI的安全性，需要发明许多技术、实践和政策。Jensen提到了AI监控、人为监督以及安全防护措施的重要性。他强调了为了使AI变得“难以置信地好”，需要在技术、安全性、政策和实践方面不断进步。最后，他以幽默的方式结束了演讲，强调了购买更多产品可以提高安全性的观点。

Keywords

💡人工智能

人工智能（AI）是指由人制造出来的能够执行复杂任务的系统，这些任务通常需要人类智能才能完成。在视频中，人工智能被提到是改变数据中心的关键技术之一，并且与加速计算和绿色计算紧密相关。例如，Nvidia的CEO Jensen提到了AI在推动数据中心发展中的作用。

💡加速计算

加速计算是指使用专门的硬件和软件来提高数据处理速度，尤其是在处理大量数据时。视频中提到，随着数据量的指数级增长，CPU的扩展速度已经放缓，因此加速计算变得尤为重要，能够释放数据中心中浪费的能源和成本。

💡绿色计算

绿色计算通常指的是能效高的计算，旨在减少能源消耗和环境影响。在视频中，Nvidia强调了绿色计算的重要性，特别是在加速数据中心的能效和性能方面，以及如何通过绿色计算实现成本和能源的节约。

💡生成式AI

生成式AI是一种人工智能技术，它能够生成新的数据实例，如文本、图像或视频。视频中提到，生成式AI将影响世界上每一个数据中心，并且是推动数据中心现代化的关键因素之一。

💡数据中心现代化

数据中心现代化是指更新和升级数据中心的基础设施，以提高效率、性能和可持续性。视频中提到了全球数据中心的价值，并强调了到2030年，需要对这些数据中心进行现代化改造的紧迫性。

💡液冷系统

液冷系统是一种冷却技术，它使用液体（如水或特殊冷却液）来吸收和传输热量，以降低设备的运行温度。视频中提到了Nvidia正在出货的液冷系统，以降低功耗并提高AI芯片的制造能力。

💡GPU

GPU（图形处理单元）是一种专门设计用于处理图形和图像计算的微处理器。在视频中，GPU被提及为构建复杂系统的核心，这些系统被用来执行高级计算任务，如AI训练和推理。

💡软件栈

软件栈是指一系列相互兼容的软件产品，它们共同工作以提供特定的功能。视频中提到了几个关键的软件栈，包括CUDA和用于网络的软件，这些都是高性能计算的基础。

💡Cuda

CUDA（Compute Unified Device Architecture）是Nvidia推出的一个并行计算平台和编程模型，它允许开发者使用Nvidia的GPU进行通用计算。视频中提到CUDA作为构建软件栈的基础之一。

💡以太网

以太网是一种局域网技术，用于计算机之间的数据传输。在视频中，以太网被提到作为现代计算网络的一部分，不再是仅仅用于发送电子邮件，而是成为了支持分布式计算的计算网络。

💡AI安全

AI安全是指在开发和部署人工智能系统时采取的安全措施，以确保系统的可靠性和安全性。视频中提到了AI安全的重要性，并强调了需要为AI制定安全技术和政策。

Highlights

英伟达CEO Jensen Huang分享了他对AI改变世界的伟大愿景。

加速计算和绿色计算同时到来，标志着数据中心的能源效率和成本效率的显著提升。

数据量呈指数级增长，而CPU扩展速度放缓，导致数据中心存在大量能源和成本浪费。

加速计算可以释放数据中心的潜能，为新目的使用之前浪费的能源。

英伟达宣布了220种新产品，展示了其在加速计算领域的创新能力。

生成性AI（Generative AI）的出现，预示着数据中心的转型。

到2030年，全球数据中心的价值可能达到3万亿美元，需要现代化改造以适应生成性AI。

超微公司（Super Micro）准备接受订单，提供现代化的数据中心解决方案。

英伟达正在出货数据中心液冷技术，以降低功耗并提高AI芯片的制造效率。

超微公司每月出货1000个液冷系统，展现了其在数据中心冷却技术方面的领先地位。

数据中心的现代化改造对于提高能源效率和降低成本至关重要。

英伟达的GPU不仅是芯片，整个系统比单纯的GPU要复杂得多。

超微公司的工程设计使得安装和启动时间极快，提高了数据中心的利用率和吞吐量。

英伟达的软件许可确保了所有产品在数据中心的兼容性和即插即用。

数据中心被视为AI工厂，直接产生收入，强调了可靠性、吞吐量和启动时间的重要性。

英伟达的CUDA和其他软件堆栈为高性能计算提供了基础。

网络现在是计算的织物，而不仅仅是发送电子邮件的工具。

英伟达的Blackwell GPU和Grace CPU通过高速互连连接，优化了训练和推理的性能。

AI技术的发展需要同时关注安全性和效率，确保AI的可靠性和实用性。

英伟达致力于推进AI技术的发展，同时注重安全技术和政策的制定。

Transcripts

I know only

some fortunately we are very lucky again

to invite the AI

genius our common

friend our common friend is very busy

huh Invidia found CEO Jensen to share

his great vision with us

[Applause]

[Music]

[Applause]

[Music]

thank

you hi

everybody now

what that AI is changing minium because

of you what's new

today I have to admit just now when I

was coming to your keynote in the car I

fell asleep

and so right now right now I'm a little

bit groggy so if I say nonsense things

please I let me apologize first no well

let's see um uh Charles we've gone back

a very long ways yeah and and um uh what

are we doing oh I needed some water I

need to spe up okay right my energy

yeah they said I was on this side and

you keep going on my

side this is what happens when we don't

practice you don't need to and you are

no time you you don't need and so so um

I uh what what were we saying um this is

a very important time because we have a

new agent Computing coming there are two

things that are happening at the same

time the first is accelerated Computing

accelerated Computing has arrived at a

time

oh Green Computing

yeah Green computer yeah okay

Computing I think I think when you say

Green Computing you mean energy

efficient Computing right yes Nvidia is

energy efficient Computing yes we have S

we follow you all

right look Green Computing and Green

Computing all right so so um uh

accelerated computing's time has come

because for a very long time the amount

of data processing has been increasing

exponentially yeah and yet CPU scaling

has slowed for many many years so we've

been we have now an enormous amount of

waste wasted energy and wasted cost

trapped inside the data centers so when

we accelerate the data centers the

savings

incredible because it has been sold long

of waste

trapped and so now we can release the

waste and use that energy for a new

purpose number one accelerate every

application accelerate every data center

these

amazing servers here right so many new

products so many new products you have

220 new products unbelievable did he

tell you that already no I W very high

I came to announce super micros products

and so that's the first thing the second

thing is because the Energy Efficiency

and the performance efficiency and the

cost efficiency is so incredibly great

with accelerated Computing a new way of

doing Computing has emerged and it's

called generative AI generative AI is an

incredible thing people say generative

AI inference it's related not the same

inference

recognizing C dog speech inference

generation text Generation image

Generation video generation that's what

we call a generative AI the pressure of

generative AI to not the pressure but

the the transition to generative AI will

affect every single data center in the

world we have a trillion dat a trillion

dollars worth of data centers in the

world that's established $3 trillion

probably by 2030 in another 6 years we

have to to modernize all of them with

these amazing systems yeah that's the

reason why the demand is so great

because all of these data centers has to

be modernized and Charles and the super

micro team is ready to take your

order Json I'm your I'm your best sales

guy thank you I work on commission no

commission we buy more cheaper from you

don't buy more

[Laughter]

chips so

that's Jon sh Michael is now shipping

data center uh liqu cooling DLC R inum

production now to lower the power

consumption so you can manufacture more

AI chip yeah yeah thousand of how here

you see

[Applause]

[Laughter]

I have many American colleagues they

don't understand my Chinese I have many

Chinese colleagues they don't understand

my

Chinese hi

y we are shipping up to 1,000 R per

month now 1,000 R like it is multiply by

ASP yeah you're going to be a gigantic

company yeah thank

you that's why I need a more

CH did you guys all do the

math Millions

times thousands time 52 no no no you

charging me $2 million more than $2

million for d

[Laughter]

are we allowed to do this on TV are we

on

TV I I guess the well is

this so we are shipping about 1,000

that's incredible now this this uh

600,000 Parts this is probably more than

600,000 parts how many pounds oh I don't

know can I move three I think it's 3,000

lb more than 3,000 lb

yeah it's incredible so yeah our goal

this year is to ship more than 10,000

record you know the Charles this is the

thing that's really amazing uh people

think that we're building

gpus you know GP is a

chip there are 72 chips in here and then

there are six

600,000 other

parts

it's 72 chips probably weighs one

pound this is 3, 2,999 other

pounds so the amount of Technology

that's inside one of these RS is really

quite extraordinary this is a technology

Marvel the most most most complex most

advanced computer the world's ever made

yeah exactly the p in the world now yeah

absolutely incredible and the software

that it takes to run this is

unbelievable yeah unbelievable isn't

that right and so I think that that

people now are starting to realize that

when we say GPU server of course the

brain is the GPU yeah but the system is

much much more complex than that and

super micro does amazing engineering

thank

[Laughter]

you huh what I

okay then we there's some Americans this

year we are going to ship hopefully make

when we're together sometimes we speak

Taiwanese sometimes we speak Mandarin

and then when we disagree we speak

[Laughter]

English we try to make a thlc mar share

from 1% to 15%

this year wow Save lot of power for your

TB yeah yeah the Energy Efficiency is so

much better the cost to the data center

is cheaper cheaper that's right people

don't realize this liquid cooled systems

eliminates an enormous amount of cost in

the data center yeah so that you can use

that waste capture that waste and put it

into Computing in the future in the

future Computing throughput is

revenues because it's token generation

and token generation is dollars per

million tokens just like

energy dollars per kilowatt hour we have

now invented a new commodity this is a

very important idea for all of you this

is a new commodity it has value and the

faster you can generate it the higher

throughput the greater utilization the

higher your revenues it is absolutely

true and it's directly measurable that's

why this is a factory not a data center

that's why this is a factory not a file

server it's not a retrieval of files

it's not used for exchanging emails this

is directly generating revenues for

factories that's why we call it AI

factories and so

powerful and only s million

[Laughter]

dollars a

such okay so $3 million and you can

generate who knows how much revenue per

year right uh 3 million 1,000 and every

year have how many

months

12 the the return on the return on large

language model generation token

generation is going to be very very good

yeah be huge and the reason for that is

because the token embeds intelligence

yeah and the int could be used in so

many different Industries and so the

future is very important it's time to

Startup yeah time to

Startup throughput yeah

utilization all matter so

reliability has Revenue implication

throughput has Revenue implication

startup has Revenue implication yeah

that's why it's so important that we

integrate the whole s whole system into

a rack scale get all the software

working connected to all the all the

networking so that and we build all of

our own data centers we build our own

supercomputers so that we know when you

install this when you install super

micro in your factories the startup time

will be extremely fast your utilization

will be extremely high and your

throughput will be extremely high

because your revenues depends on it

Factory output is measured by all of

those factors very complicated yeah and

all of those R are Invidia sofware

license all certified so the sound of

that parking the cable and they can run

and it runs that's right and all of the

Nvidia Nims all of the large language

models it just runs on all these systems

yeah

[Laughter]

[Applause]

we are shipping thousand R

very

yes very

beautiful Charles Charles said that this

is

everything everything in here is NVIDIA

for all the American citizens

there

from to H AI everything all Nvidia sare

all all Green Computing all Green

Computing all green computer all all

support that's

fantastic

good let go through something detail

okay okay

okay H1 H2 B1 for you cooling wow

shipping in B wow and this one your p200

uhhuh fully ready beautiful for your

chip beautiful beautiful this will be

how many time faster than this so we

have we have we have uh uh for Blackwell

Blackwell has air cool

liquor

cooled

x86

Grace MV link 8 MV link 2 MV link 36 MV

link 72 yeah so many different

configurations yeah so that depending on

the type of type of utilization type of

use case you have the type of data

center that you have uh Charles is ready

to serve you immediately right

immediately doesn't need to acheve yeah

one hand we got to acheve second hand we

Shi to C W thank goodness we only need

two hands in two weeks in two

weeks that's incredible and all of it

software compatible this is really this

is really the amazing thing certifi

literally everything here is software

compatible one% yeah and software as we

know is the most complex part of high

performance Computing yeah thank you for

those great offering they are all ready

to service our customer there are three

very important software Stacks that we

have in our company that everything is

built on top of the first of course is

Cuda very famous the second for all of

the networking because networking is

just not networking networking today

networking today is a Computing

fabric networking today is a Computing

fabric not just for sending email to

each

other

4 Mez a gigahertz megahertz this is not

1980s

be

Mez

kilohertz gahz gigahertz yes 400

gigabits per second 800 gigabits per

second and and then of course Next

Generation coming 1600 but the important

thing is all of the software that we

have that runs on the networking for

distributed computing is on top of two

software Stacks one is called DOA for

the nick nickel for the fabric yeah and

it enables us to distribute the workload

across the network very very efficiently

because ethernet was was not designed

for hyperform computing you make our job

easier but still very py because you

have so many

great my job is to help give you

[Laughter]

job we

and because because you do such a good

job it becomes gives me job oh don't

forget that your another

baby yeah yeah yeah

yeah inside

here this this is an incredible

incredible system in fact in

fact in fact um these chips are all

connected together using high-speed

interconnect the world's fastest CIS the

CIS is incredibly fast and very energy

efficient and so we can connect this

great CPU to dual Blackwell gpus and

that's very important because in the

training stage the memory system of

Grace could be used for checkpoint

restart checkpoint and restarting is

very important for high utilization and

high uptime and so checkpoint restart uh

could be stored in the system memory

that system memory is very low energy

very low power and the link between

Blackwell and Grace is very very high

second during inference time as you know

there's a concept called

prompts context in context training

prompting that prompt memory that

context memory is right here this is the

memory memory the thinking memory the

working memory of AI and so this memory

needs to be very high performance very

low energy and so during training we

have good use for gray gray CPU during

inference we have excellent use for gray

CPU and the interconnect is very very

high speed very low power F optimiz and

so the re the benefit is because we

compress so many in one system yeah if

we

save 20 watts 50 Watts on the

interconnect you multiply by the whole

rack then we can take the energy and use

it for computing y so Energy Efficiency

translates to higher performance to

that's right Green

[Applause]

Computing

huh I am a super micro employee

super micro

employee where AI Control

us of course not um we we have to we

have to uh the most important thing of

course at the moment is we have to make

AI work

well right now ai is of course uh

working extremely well and in many

applications AI has become good enough

to good enough to become useful it has

achieved the plateau of good enough very

useful however we want it to be

incredibly good we want it to be very

functional everything from Guard railing

for uh fine-tuning skill learning there

are many different things that we still

have to improve okay so we know that AI

is AI still has long ways to go that's

job number one is Advance the technology

at the same time we have to advanced

Safety technology as you know uh our the

planes that we all flew on to come here

has autopilot and autopilot is automatic

technology in order for planes to be

safe a great deal of Technology had to

be invented to keep the plane safe yeah

also practices to monitor the planes air

traffic control other planes monitor the

planes Pilots monitoring each other many

different ways to keep uh AI uh keep

autopilot safe in the future we'll do

the same thing with AI there will be AIS

that watch AIS there are people that

watch AIS there's gu right guard rails

that keep AI guard rail and so there's

going to be a whole lot of different

Technologies we need to create for

safety technology for safety and then

third of course we need to have good

policies for safety good practices and

good policies for safety talking about

it is very important so that we can all

remind each other that we have to do

good science good engineering good

business practice good policy practice

good industrial practice all of those

things has to advance so perfect

strategy so the conclusion is one the

more you buy the more

safe the more you buy the more you safe

the more you buy the more you safe

yeah thank you Jas thank you so much

good job thank you

everybody thank you okay thank you thank

you thank you thank you all right have a

great

thank you

[Music]