news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

New Intelligence Report

Editor: Editorial Department

【New Wisdom Introduction】Just now, Huang revealed at the SIGGRAPH conference that Blackwell's engineering samples have been officially shipped to the world this week! Afterwards, Huang and Zuckerberg had a fireside chat and changed clothes intimately. Zuckerberg was so excited that he even swore.

Shocking news is coming!

Just now, Huang revealed at the SIGGRAPH computer graphics conference: This week, NVIDIA has begun sending Blackwell engineering samples to the world!

Then, the host Lauren Goode joked: That's right, everyone lower your heads, it's under the stool.

It is worth mentioning that behind this most powerful AI chip today, AI is also indispensable——

Without AI, Hopper would not have been possible; without AI, Blackwell would not have been possible.

During a fireside chat between him and Zuckerberg, Zuckerberg even couldn't help but swear at one point when he became emotional.

The effect of the two bosses' changing of clothes was really hard to describe.

This time, Zuckerberg specially gave Huang a customized "black leather style" cotton jacket.

After putting it on, the effect is indeed outstanding!

Of course, Zuckerberg also wore a second-hand leather jacket that Huang had used for only 2 hours. (This is much more valuable than a brand new one.)

Nvidia's digital "copy" world

At the conference, Huang announced that NVIDIA has built the world’s first generative AI model that can understand OpenUSD (language, geometry, materials, physics, and space).

What is OpenUSD? It refers to Universal Scene Description, which can be understood as a universal scene description.

Huang said that what is more exciting than what AI can do with text is that we can do the same with images.

For example, the Edify AI model created by NVIDIA is a basic text-to-2D model.

For brands, it can create Coca-Cola, cars, luxury goods and so on, but controlling cues is a difficult thing.

This is because words have a very low latitude, they are extremely compressed in content, but at the same time very imprecise.

To do this, Nvidia has created a method - creating another model that can be controlled and adjusted to align with more conditions.

With Omniverse, you can combine all of this multimodal data and content, whether it's 3D, AI, animation, or materials.

We can change its posture, position, in short, anything we want.

Using conditional prompts in Omniverse, just like search enhancement generation, can be understood as a kind of 3D enhancement generation.

This way, we can generate images the way we like.

Next, the work that WPP completed using Shutterstock and world-renowned brands directly shocked the audience.

Build me a table in an empty room, with chairs around it, in a busy restaurant.

Build me a table with tacos and a bowl of salsa in the morning light.

Build me a car on an empty road, surrounded by trees, near a modern house.

Build me a tree in an empty field.

Build me hundreds of these trees in all directions.

Let this wood have bushes and vines hanging in it.

Build me a giant rain forest with exotic flowers and rays of sunlight.

Omniverse now understands text to USD conversion. It understands text and has a semantic database so all 3D objects are searchable.

So the little girl can depict how she wants to populate the 3D tree, and once she’s done, the 3D scene goes into a generative AI model that turns it into a photorealistic model.

From here, more and more generative AI will appear in Omniverse to help people create these simulations, or digital twins.

For example, the digital AI below will enable every company to have customer service.

Today, customer service is done by humans, but in the future, AI will be involved.

Customer service will be connected to a digital human front end, an IO, which can speak and make eye contact with us.

All kinds of AI can be connected to this digital human, and even the digital human can be connected to NVIDIA's retrieval-enhanced generative customer service AI.

NIM Services

At this conference, NVIDIA launched a new set of NIM microservices.

NIM is tailored for different workloads, including OpenUSD, 3D modeling, physics, materials, robotics, industrial digital twins, and physical AI.

In the area of ​​AI and graphics, NVIDIA launched a new OpenUSD NIM microservice designed specifically for building physics AI applications.

This workflow includes new NIM microservices for robotics simulation and more to accelerate the development of humanoid robots.

"Three Body" creates robots

Lao Huang predicts that the next wave of AI will be physical AI.

If robotics is to advance, it will require advanced AI and realistic virtual worlds; and before we can deploy the next generation of humanoid robots, we need to train AI.

Robotics requires three computers: one to train the AI, one to test the AI ​​in a physically accurate simulation, and one inside the robot itself that can learn how to optimize the robot.

That is, the third AI is the computer that actually runs the AI.

Nvidia created three computers for this purpose.

Without AI, there would be no H100/H200 and B100

Throughout Nvidia's history, which began in the 1990s, its real DNA lies in computer graphics.

Computer graphics has also brought NVIDIA to its current position.

This picture shows some important milestones in the computer industry, including the IMB 360 system, the Utah teapot, ray tracing, programmable shading, etc.

NVIDIA was founded in 1993. Eight years later, they invented the first programmable shading GPU, which largely promoted the development of NVIDIA.

It can be said that the core behind everything NVIDIA does is accelerated computing. They firmly believe that if a computing model is created to enhance general computing, problems that ordinary computers cannot solve can be solved.

The first choice was computer graphics. They made the right bet.

The application of computer graphics to the then-non-mainstream field - 3D graphics video games - directly drove Nvidia's flywheel.

After that, they spent a long time making CUDA ubiquitous, and then in 2012, like something out of Star Trek, Nvidia got its first exposure to AlexNet.

In 2012, it was a breakout moment when AlexNet made an amazing breakthrough in computer vision. At its core, deep learning was so profound that engineers no longer needed to provide input and imagine what the output would look like.

In 2016, Nvidia launched the first computer built for deep learning, DGX-1, which was favored by Musk and then delivered to the then-unknown OpenAI.

Subsequently, RTX and DLSS were invented.

Then, ChatGPT was born.

In the future, everyone will have an AI assistant

Today, we have learned to use AI to learn everything, not just words, but also images, videos, 3D, chemicals, proteins, physics, thermodynamics, fluid dynamics, particle physics, and more.

We understand the meaning of all these different modalities.

In Huang's view, the generative AI revolution based on visual computing is enhancing human creativity.

We are truly at a revolutionary moment, moving towards the era of software 3.0 - no industry can escape the impact of AI!

Lao Huang predicts: Everyone will have an AI assistant, and every company and every job within the company will be assisted by AI.

Accelerated computing solves energy problems

Although generative AI has the potential to improve human productivity, the energy consumption of AI infrastructure is a major problem that plagues the entire planet.

One search on ChatGPT is equivalent to the power of 10 Google searches.

Data centers consume 1% to 2% of the world's total energy, and may reach 6% within a decade.

What to do? Lao Huang has the solution.

He said that accelerated computing technology is expected to make computing more energy-efficient.

“Accelerated computing can help us save a lot of energy. It can save 20 times, 50 times, and perform the same processing,” Huang said.

“The first thing we need to do as a society is accelerate every application we can: This reduces energy usage around the world.”

That's why Blackwell is so anticipated, because it uses the same amount of energy but significantly speeds up applications.

And it will get cheaper.

Huang stressed: Remember, the goal of generative AI is not training, but reasoning. Ideally, reasoning can create new models for us to predict weather, predict new materials, optimize supply chains, and so on.

You know, data centers aren’t the only places that consume energy. Global data centers only account for 40% of total computing, and 60% of energy consumption is spent online, moving electrons, bits, and bytes.

Therefore, generative AI will reduce energy consumption on the Internet because instead of having to retrieve information, we can generate it directly on the spot.

And just now, NVIDIA deployed GPUs in GCP to run Pandas.

The world's most advanced data science platform directly increased the speed from 50 to 100 times, exceeding general computing.

Over the last 10 to 12 years, we’ve made deep learning a million times faster and a million times cheaper and more energy efficient, which is why LLM was born.

However, Nvidia will also bring new innovations to AI by designing new processors, new systems, Tensor core GPUs, and NVLink switch structures.

Fireside chat between Huang and Zuckerberg

Many people have been looking forward to the fireside chat between the two CEOs at SIGGRAPH this year. In Zuckerberg's own words, what kind of sparks will the collision of "two of the most senior founders in the industry" produce?

The next wave

Not surprisingly, these two heroes of "Childhood Plum Drinking Wine" each shared their own predictions and talked about future technological development trends, from GenAI, to Agent, and then to the "Metaverse" that Zuckerberg has always been thinking about.

Huang said he was also shocked by the technical power of GenAI. "I can't remember any technology that has affected consumers, businesses, industry and academia at such a rapid speed, and across all different fields from climate technology to biotechnology to physical science."

Zuckerberg also said that GenAI is likely to reshape Meta's various social media software.

Once upon a time, the core of these products - the recommendation system, was simply to push content of interest to users.

But GenAI will no longer be limited to existing content. It will not only assist creators, but also create instant content for users, or generate it by integrating existing content.

The two seem to have similar views on the development of Agent.

In a previous speech, Huang made it clear that "everyone will have their own AI assistant in the future."

In this conversation, Zuckerberg also expressed a similar vision. He is planning AI assistant and AI Studio products for Meta, allowing everyone to create their own Agent for different purposes.

In the future, every business will have its own AI, just as every company has its own social media and email accounts today.

To what extent should the "AI assistants" they talk about be "intelligent"?

The Llama 3 we have seen so far is just a "chatbot"-like language model that can only respond to human questions. But Zuckerberg hopes to give AI "intent".

Huang described it as "planning ability", which can form a "decision tree" in the mind like humans to guide behavior.

He even boldly predicted that this AI assistant would cost only $10 per hour and could greatly improve engineers' work performance. "If you haven't hired AI yet, take action now!"

For Meta's most core and unique AR/VR technology, Zuckerberg's blueprint is also quite precise, fully reflecting his obsessive-compulsive personality.

(According to Lao Huang, Xiao Zha cuts tomatoes with millimeter-level precision, and each slice of tomato cannot touch each other.)

Last September, Meta and Ray-Ban launched a new type of smart glasses equipped with audio equipment and cameras, allowing users to take pictures directly from the perspective of both eyes, or live broadcast the view seen in the glasses directly to Facebook or Instagram, and integrated the conversational assistant Meta AI.

Zuckerberg said that based on the current situation of Ray-Ban glasses, a $300 display-free AI glasses will become a very popular product.

According to his prediction, smart glasses will become a device similar to mobile phones in the future, and everyone who wears glasses will wear smart glasses (more than 1 billion people worldwide).

In the next few years, Meta will also launch glasses with holographic AR functions. Although the cost will still be high, it will be a viable product.

Unlike smart glasses, mixed reality headsets are more like workstations or game consoles. They are not easy to carry but have stronger computing power and can provide users with a more immersive experience.

Moreover, with the development of holographic AR technology, "virtual meetings" will soon become a reality.

Unlike avatars or videos on the Zoom platform, everyone will have their own holographic image, allowing the "virtual people" created by the hologram to collaborate and interact in the same space even if they are in different physical spaces.

Open source is the way forward

When it comes to Meta, their consistent "open source" strategy is also a point that cannot be ignored.

Huang highly appreciates this strategy. He said that Llama 2 may be the most important event in the field of AI last year; together with PyTorch and the newly released Llama 3.1, Meta has built an entire ecosystem.

But Zuckerberg said that their decision to go open source was also a "flexible response."

In many areas, especially distributed computing systems and data centers, Meta's starting point was actually behind other companies, so the team thought of open source, especially open computing.

Unexpectedly, this stopgap measure has become a key strategy for "overtaking on a curve".

It is open source that makes the products released by Meta become industry standards, and the entire supply chain is built around it. By making the project open source, Meta has even saved billions of dollars.

For example, Meta entered the GPU field later than most companies, but the GPU supercomputing clusters they currently operate are larger than almost all of their competitors.

Of course, Huang’s strong support is indispensable behind this. After all, Meta’s 600,000 GPUs are all produced by NVIDIA.

Although open source can promote the progress of this community and industry, Zuckerberg also honestly stated that open source is not charity, and we chose this strategy not because we have a selfless heart.

The more important purpose is to make the product we are building reach its peak and become the best.

PyTorch is the most typical example. Developers from all over the world, including two or three hundred engineers from NVIDIA, are helping to find bugs and optimize this open source framework, forming what Huang calls the "PyTorch engineering mountain."

Although Zuckerberg himself admitted that open source is selfish, he still couldn't control his emotions when talking about "closed" platforms. The only swear word in the audience came from this topic.

Although Meta owns a number of top social software, these applications need to be distributed through competitors' platforms, especially Apple's App Store and Google's Android system.

What annoys Zuckerberg is that he once had many product ideas, but they were ultimately not realized due to the various limitations of these mobile platforms.

This extreme dependence on platforms in the mobile Internet era is completely different from the openness of the PC era, which makes Zuckerberg miss the web version of Facebook.

Therefore, he expressed confidence that we are shaping the next generation of computing platforms, namely mixed reality technology, in which open source software will once again have greater value.

The next generation of platforms and ecosystems will be more open and inclusive, similar to the previous Windows or Android ecosystems, rather than the completely closed Apple.

This ambition to "make open source great again" reminds people of his metaphor when Llama 3.1 was released - Llama 3.1 is the Linux of this era.

It's Not Easy to Be a CEO

Throughout the conversation, the two men had a sense of mutual admiration and often talked about the difficulties of the CEO profession.

Jensen, 61 years old at the time and wearing a leather jacket, even compared himself to a delicate flower with a serious face: "We are CEOs, like delicate flowers, we need a lot of support."

Zuckerberg even followed up with, "We are quite exhausted now."

This feeling may come from the ups and downs that the two senior founders have experienced with the company.

In Zuckerberg's view, Huang insisted on making computers into "super behemoths" despite the pressure of being looked down upon, making Nvidia a legend in the industry.

In Huang's view, Zuckerberg has led Meta through many transformations, from PC to mobile, from social media to VR/AR and AI research.

At the end of the conversation, Huang spoke frankly about what they had in common, "I know how hard it is to do that (transformation). We have both been hit hard before, but that is what it takes to be a pioneer and an innovator."

References:

https://www.youtube.com/watch?v=H0WxJ7caZQU

https://www.youtube.com/watch?v=w-cmMcMZoZ4