Top of the big era

Chapter 2717 Graphics Card

There are many types of chips. To put it simply, all electronic devices related to computing are considered chips.

For example, the sound sensor in a mobile phone can sense external sounds, and then convert the sound signal into a digital signal through calculation and send it to the central processor. This kind of thing is also called a chip, called a sensing chip.

Another example is power consumption.

How much power is needed to run a program? This also requires calculation. The device used to calculate this unit is called a power chip.

Another example is Bluetooth, WiFi, USB interface, etc. These connection processes require calculation. If the connection is done through calculation, there will be corresponding communication chips for the connection.

However, most of these chips have low technical content, a small amount of calculation, and are not considered core technologies.

The most well-known chip is the processor chip, also known as the CPU. In a product, the CPU is equivalent to the human brain, responsible for almost all calculations, including a large number of complex calculations.

Artificial intelligence also requires a lot of computing.

Therefore, the industry’s understanding of artificial intelligence in the past also followed the same logic. When processing a large amount of calculations for artificial intelligence, various AI algorithms are also calculated through the CPU.

This is industry common sense.

But it wasn't until this month...actually just a few days ago that the entire artificial intelligence industry finally woke up and developed a brand-new industry understanding.

When dealing with artificial intelligence calculations, CPU chips should not be used!

Instead use a GPU chip!

GPU, also known as graphics card.

The reason why such a change can be made is the "human-machine war" launched by Rokid-go, which shocked the world in the Go industry last week!

The huge data processing behind Rokid-go relies on the operation of a large number of GPUs.

For CPUs and GPUs with the same specifications, when processing AI calculations, the computing power of the GPU can exceed 100 times that of the CPU! And the energy consumption is less than 5% of the CPU!

After the great success of Rokid-go, seven papers in this area have been published recently. The practice of "man-machine war" has also proved that Ziweixing has led the global artificial intelligence computing industry's major turn.

Artificial intelligence, under the leadership of Ziweixing, will truly enter the GPU era!

And this has further raised the threshold for doing artificial intelligence research.

Because the past model of relying on violently accumulating a large number of CPUs to increase computing power is completely unworkable in the field of artificial intelligence.

Just like a supercomputer.

Supercomputers are computers with super computing capabilities. In this field, China has very strong capabilities, not much worse than the United States.

Why?

Because of the superposition properties of the CPU.

The vast majority of CPU calculations are linear, just like workers on an assembly line, executing tasks one by one. The United States has the best CPU chips. At most, the work level of the assembly line workers is high and the individual work ability is relatively strong.

Domestic CPU chips are relatively poor, but they can use violent accumulation methods to solve computing power deficiencies.

The worker's individual ability is poor, but that doesn't matter. Just arrange more workers on the assembly line. The American supercomputer uses 1,000 chips, and ours uses 10,000 chips. Can we always achieve the same computing power as them?

It's nothing more than wasting more electricity.

In the face of national strategy, how much does electricity cost?

For a supercomputer, the mainframe can fill a whole floor of a building. There is no space limit, and CPUs with low computing power can be piled up infinitely. By relying on the accumulation of countless CPU chips, supercomputing effects can be achieved.

Therefore, Lenovo, Sugon, Inspur, including some universities and the military, can all develop supercomputers with self-developed intellectual property rights and break technical barriers in this field.

However, the GPU cannot.

The reason why GPU can replace CPU as the main task of artificial intelligence computing is because GPU mainly handles parallel computing.

The neural network algorithm trend of artificial intelligence requires parallel computing, which just matches the GPU.

Once it is parallel computing, it is no longer possible to increase productivity by stacking quantities like assembly line workers. This has to rely on real hard power.

For example, a doctoral student and a primary school student can do math problems, but the primary school student cannot do it at all. Even if 10,000 or 100,000 primary school students were combined, they would not be able to compete with a doctoral student.

This is Qualcomm and Intel, companies that mainly sell CPUs. Their highest market value is hundreds of billions of dollars. However, NVIDIA, which mainly sells GPUs, has a market value of US$500 billion, US$1 trillion, and US$2 trillion. The reason why the dollar is even higher.

No matter how CPU is the core technology, it is also fungible.

The GPU cannot be replaced by stacking.

Without top-level GPUs, it is impossible to provide top-level computing power, and it is impossible to become a top-level artificial intelligence company.

Just like the current situation of domestic artificial intelligence in 2023 ten years later.

The best GPU chip in the world is Nvidia's H100, which sells for $40,000 a piece. Next is Nvidia’s A100, which sells for $10,000 a piece.

No matter how many A100s are piled up, there will be no quantitative change or qualitative change to reach the AI ​​computing power of H100. This is determined by the character length of the chip structure.

The length of 32 bits is out of reach compared to the length of 16 bits.

Just like a person's logical level can reach 32 levels, if 1,000 or 10,000 people with 16-level thinking logic are gathered together, their thinking level can only be limited to 16 levels.

At the same time, the domestic Huawei Ascend 910B sold for US$40,000 per unit... Its ability to handle simple calculations and linear calculations is comparable to Nvidia's A100. But linear computing is meaningless. What artificial intelligence requires is deep parallel computing capabilities, which is far worse than the A100, let alone compared with the H100.

And this also determines the general trend of the domestic AI industry.

Artificial intelligence requires a lot of computing power, which needs to be provided by cloud computing platforms. Alibaba Cloud has a stronger international background, so it can purchase more NVIDIA A100 and H100 chips to provide users with the top AI computing power in the country.

Therefore, most domestic companies that are serious about artificial intelligence products have either moved closer to Alibaba Cloud, or are on the way to move closer to Alibaba Cloud.

Baidu Cloud and Penguin Cloud are not as international as Alibaba, and the number of NVIDIA graphics cards purchased is not as many as Alibaba, so they will gradually be left behind in this field.

Those who do not have international resources and cannot buy NVIDIA graphics cards can only support domestically produced cloud computing platforms in the name of marketing and patriotism. They can only play marketing and use marketing to please the authorities and the public. Go and start a business like Government Affairs Cloud that relies on political and business relationships to survive.

It’s only 2013, so it’s still early.

Let alone Zhou Buqi discussing artificial intelligence computing solutions with an outsider like Tang Tianhua, even many top scientists in the industry may not be able to turn this corner in a short time.

Tang Tianhua asked with a confused look: "What GPU? Isn't the GPU a graphics card?"

Zhou Buqi said: "Use the graphics card to run the AI ​​software."

“Isn’t the graphics card responsible for rendering images?”

"Well……"

"Okay..." Tang Tianhua didn't know what was going on at all, but he knew that Mr. Zhou was a leading entrepreneur in the technology industry. He couldn't be untargeted, so he shook his head, "I don't know the specifics. I just know Use this software.”

It's true that Lucasfilm is not a software company.

The software products they develop are all professional-grade and the threshold is very high. After studying for half a year, it would be considered good if I could learn how to use their software.

Looking at this situation, Zhou Buqi basically made a judgment. Lucasfilm's AI software that intelligently changes the mouth shape of animated characters must use traditional computing logic, and is executed by the CPU behind it.

Just like Tang Tianhua said, several computers were burned out.

What should be burned is the CPU.

Zhou Buqi said: "This kind of movie-level image requires a lot of calculations, but it also needs to be divided into objects. It is really difficult to get your own computers and servers to provide computing power. It will burn out several computers and several computers." The servers are all normal. But if Ziweiyun handles it, it’s nothing. It saves money and the processing speed is fast."

"Ah? Zi Weiyun?"

Tang Tianhua was stunned again.

Faced with such a layman, Zhou Buqi had nothing to say. If you want to do this, you have to notify Lucasfilm to make adjustments.

The first is to overhaul their software, changing from calling the CPU to calling the GPU. Then, this software is combined with the Ziwei Cloud platform, and cloud computing provides more efficient, faster, and more powerful computing power.

In this way, it can be better applied in the film industry.

In the previous life of the film industry, without the participation of technology industry leaders like Zhou Buqi, it could only rely on great directors such as George Lucas and James Cameron to promote it.

With Boss Zhou in this life, he is committed to promoting the development trend of "technology + movies". Through the empowerment of technology, he can make movies more dreamy and audio-visual and audio-visual more consistent, thereby achieving a better movie-watching experience.

Zhou Buqi asked: "How long will it take to complete your project?"

Tang Tianhua said: "There are still about 10-12 months."

"Yeah." Zhou Buqi nodded, "That's a little too late. You should follow the established plan first. In the future... maybe there will be a better way to deal with it in the future."

Time is tight.

With the efficiency of the people at Lucasfilm, even if they were urged to change it immediately, it would not be possible to implement it within a year even if they had to change the underlying software and integrate it with Ziweiyun.

It seems that the project of "Kung Fu Panda 3" cannot apply better technical solutions.

Zhou Buqi decided to go to the United States this time to promote this matter.

Then, he walked around the Jieyu Media Animation Center. There were only 30 employees in total, shaking hands with everyone one by one, until he saw a beautiful female artist in her twenties.

Just a little dazed.

Depend on!

An acquaintance?

An acquaintance from his previous life, one of his most impressive lovers—Ji Baizhu.