Home > News content

Five years of Chinese AI core "awakening"

via:博客园     time:2019/6/17 8:54:01     readed:1596

Ding Guangsheng

Netease Smart "Enhanced" column

More than a dozen chips have come out, and many mergers and acquisitions have been made, which is the heat of Chinese AI chips in the past 500 days.

Unexpected events once again ignited the academic industry and the entire public opinion field. Under the background of artificial intelligence technology in the forefront of the world, we began to expect it to become the Chinese wedding chip.

The concept of artificial intelligence began to be thrown into the public eye with each new product.

The AI ​​chip refers to an ASIC specifically for the AI ​​algorithm. Unlike the traditional CPU and GPU, the latter can be used to perform AI calculations, but the speed and performance affect the commercial effect.

For example, a chip engineer needs to recognize road pedestrian traffic lights, etc., but if you use the current CPU to calculate, it is estimated that the car has not been found in front of the river.

And if it's a GPU, the speed is of course much faster, but its huge power consumption and high cost are still not ideal for AI scenarios. Not only that, the AI ​​chip is also divided into the cloud and the terminal on the scene side, and the scene of a large amount of data training artificial neural network is completed in the cloud, and the reasoning for pursuing low cost and low power is fixed at the terminal.

Real demand, no shortage of scenes, funds in place, the AI ​​chip companies have gone on the road, but the chip is not the rush, the difficulty of designing, manufacturing, sealing and testing the entire chain has made countless people deterred. And this is still a physical activity that sits on the bench for ten years.

Ups and downs and ups and downs, and the thorns of the sky and the enthusiasm of the heroes, this has gathered into the AI ​​chip entrepreneurs created "core" rdquo; story.

Find the optimal solution

In the fall of 2013, Zhang Jun was the only salesperson of Yunzhisheng in Shenzhen, responsible for opening up Baidian customers.

He often consulted with Li Xiaohan, the co-founder who came on a business trip, how can he help traditional home appliance companies improve their competitiveness? At first, they provided SaaS services in the cloud to provide voice recognition capabilities to customers, but quickly realized that this was not enough.

They found that IoT devices are not the same as mobile phones. Voice recognition on the mobile phone may not be just needed, but in the home life and independent office, the most natural way of human interaction will have its use.

At that time, the Shenzhen office was only 12 square meters. There were two tables and three chairs. Everyone thought about doing something like “U disk”. There is a standard interface, which can be plugged into the home appliance to realize voice interaction.

In August 2015, this product, which was later named Intelligent Voice Module (IVM), was a great success. A hardware version can be standardized for different customers. Model A and Model B can be easily adapted. Gree Air Conditioner and Vantage Hood Smooth shipment.

The Shenzhen team began to move into the big office, and the team recruited.

However, in just a few months, they found a new problem, the board with chips, a variety of things stringed up, each is actually not expensive, but the cost is too high, the customer asked Li Hanhan: If you sell a hundred dollars for this thing, it can only be placed on high-end products, and low-end products sell hundreds of dollars. I can't use it.

How to do? Some people have suggested that they make chips and integrate things into the chip to solve the problem. However, if they find out that they can't be thinned out, it is still expensive.

This matter has been heatedly discussed. Even some investors have complained that they have found that we are not very generous in investing money. You are going to make chips. Everyone knows that the chip has a long investment cycle. Why do you want to do it?

It is also impossible to blame the investors, Zhu Xiaohu, the managing director of Jinshajiang Venture Capital, which is called “Unicorn Catcher”, said that “China VC is not a chip, and we have invested a few before.” "At that time, all investors mentioned high input costs, high thresholds, long periods, and low returns.


In the second half of the year, the team was still making a decision because they found that there is a sharp contradiction in the long-term irreconcilable. Only enough computing power and functions can support the AI ​​application, but the cost is the main limiting factor.

An AI product, with a good underlying support, requires a cost of one or two hundred, but it has to be sold at a lower price, so that you don’t want to reduce your computing power, you can only pass the chip itself. Other solutions.

They judged that this is the problem of the entire IoT direction. Yun Weisheng CEO Huang Wei said, “Do not make a chip is a dead end”.


Fight short-term <Emotions”

A little faster, he was the vice president of Baidu Deep Learning Research Institute, Yu Kai. He left Baidu on June 1, 2015. After a week, he started to build momentum for the new company. "We want to be Intel in the era of artificial intelligence."

The reason for him to resolutely leave and start a fast business is precisely because the variables of AI are not in software, but the processor architecture, the underlying computing efficiency and power consumption are not solved, the application layer will do a lot of useless work.

Because the function of the embedded processor has been extended, the qualitative change has already taken place. He Limin, a professor at Beijing University of Aeronautics and Astronautics, once said that since the era of deep learning of artificial intelligence entering the machine, the hardware acceleration of the original MCU has been unable to meet the high-speed massive numerical calculation. Requirements, and inter-cloud interaction requirements for big data.

In this way, in the embedded field, there are two groups of chips that are fundamentally different between the MCU and the AI ​​chip. The former is known for its control and the latter is known for its calculation.

After finding the right direction, Yu Kai met a good news and bad news.

The smooth start is the establishment of the entrepreneurial team. Basically, it is to find the old friends and old people. After eating a few meals and digging a few laps, a combat force of thirty or fifty people will be pulled. The management troubles encountered after the team has risen to several hundred people. That is a postscript.

The bad news is that after the first round of financing is easily completed under the endorsement of the luxury team, where to find money.

In the autumn of 2015, the people, especially the investors, simply couldn’t understand what this guy was going to do. Yu Kai talked about six or seventy investment institutions in succession.

The vents at that time were shared bikes and P2P.

So that the "human-machine wars" that affected the development of artificial intelligence in the world came, Yu Kai actively participated in and appeared in NetEase's live broadcast room, as a technical guest to explain AlphaGo, and also through this opportunity for what they are doing. Positive name.

In the Go game held in Seoul, South Korea from March 9th to 15th, 2016, AlphaGo Go defeated Li Shishi with a total score of 4 to 1. Yu Kai’s confidence has increased, which means that most people are beginning to understand the beginning of artificial intelligence and AI chips.


Yu Kai's chip & ldquo; troops & rdquo;

From the very beginning of his business, he said that he must adhere to strategic judgments. It is not for the big environment. It has paid a lot of price. What has always puzzled him is that the market impact caused by short-term emotions and competition is magnified in China. This is also why many foreign investors do not make money in China, and they often underestimate the power of “local”.

Fortunately, in the past few months, the Horizon has continuously received funding support from Yuri Milner, Shuanghu Investment, Qingyun Venture Capital and Xiangfeng Investment, and won Intel’s lead in the second year. Over $100 million A+ round.

The grain of winter has finally been prepared.

Yu Kai said that the horizon is doing "anti-consensus", Baidu is a self-driving entrepreneur, only they take different routes, which leads to two extremes when talking, either hit it off or not talk properly, you have to do It is to find people with the same idea.

Giant seed encounter variable

Around 2015, the horizon and Yunzhisheng began to invest heavily in the force. Although the Cambrian was founded the latest, the original accumulation was completed at the Institute of Computing Technology of the Chinese Academy of Sciences.


Chen Yunqi, Chen Tianshi Brothers

In 2014, Chen Yunqi and Chen Tianshi, who are at the Institute of Computing Technology of the Chinese Academy of Sciences, proposed DianNaoYu, a deep learning processor instruction set, which was accepted by ISCA2016, the top international conference in the field of computer architecture, and ranked nearly 300 submissions. First place.

Simulations at that time showed that the Cambrian deep learning processor with the DianNaoYu instruction set had two orders of magnitude performance improvement over the CPU of the x86 instruction set.

Chen Yunqi and Chen Tianshi brothers began to emerge.

Chen Yunqi attended the middle school at the age of 9 and entered the junior college class of the Chinese Academy of Sciences at the age of 14. At the age of 24, he obtained a Ph.D. from the Institute of Computing Technology of the Chinese Academy of Sciences. At the age of 29, he was promoted to a researcher. At the age of 33, he won the China Youth Science and Technology Award and the Young Scientist Award of the Chinese Academy of Sciences. The two brothers stepped forward and followed the same pace until they "received".

In 2008, Chen Yunqi brothers began to jointly create AI chips, using their own scientific research advantages, before this, Chen Yunyi has followed China's "father of Godson" Hu Weiwu for more than ten years. The Loongson 3A3000 processor that Hu Weiwu led the team in 2016 has been used in the Beidou II satellite.

In the same year, Cambrian Technology was founded and launched the first commercially available deep learning processor, Cambrian 1A, which was earlier than the industry.

Chen Tianshi once said that the Cambrian team of only a few hundred people has done a lot of things, the world's first terminal AI processor landing, the world's first multi-core terminal AI processor landing, China's first cloud artificial intelligence chip Landing, three-generation terminal IP products released.

He said that it is hard to imagine that the AI ​​chip field will not give birth to a new giant, which shows the ambition of the Cambrian. But soon they need to face Huawei's own chip, and the outside world refers to “abandoning” the Cambrian question.

Chen Tianshi responded to Huawei's peak performance 16T's rise 310 and Cambrian's MLU100 did not compete, the scene is different, the peak performance is also different, as an open independent chip company, to have a broad vision.

I don't know if this is his heart, but at least the number of variables going to the giant road is increasing.

The AI ​​chip is an opportunity of the times, and it is also a technology demand. In this wave of rising tide, the earliest investment in research and development is still in Hess, but it has always been Huawei's "snow".

As early as 2004, Haisi Semiconductor, a wholly-owned subsidiary of Huawei, gradually entered the historical arena. They initially focused on baseband chips, video coding chips and unicorn chips for their own mobile phones, and the relationship with AI chips began in 2017.

In early September 2017, Huawei released the Kirin 970 in Berlin, claiming that it was “the world's first mobile phone AI chip”, which integrated the IP of the Cambrian 1A processor as its core AI processing unit. In August of the following year, the Kirin 980 was also equipped with an optimized version of Cambrian 1A.

However, since October 2018, the Cambrian has ushered in bad news. In fact, unexpectedly, “Customer” Huawei launched a full-stack full-scenario AI solution and two AI chips, the Soaring 910 and the Soaring 310.


Hai Sicheng's highest growth rate chip design company

Before the survival of the Jedi, Huawei founder Ren Zhengfei planned to invest 400 million US dollars and 20,000 people on the chip. In 23 years, Hess's helmsman He Tingbo went from engineer to president, she went through micron to nano, and now to 5 nanometer technology. .

These accumulations and efforts have become a strong reserve for HiSilicon in the AI ​​chip era, and it is blossoming and solid. The next generation of the 910, the Rising 310 and the Kirin 980 are accelerating.

Creativity and order

In the Cambrian period, the background of the Chinese Academy of Sciences, Haisi has Huawei's huge wheel as a backing, more than a thousand senior semiconductor experts participate in itself is a super barrier.

The construction and running-in of the team is not so easy for startups like Yunzhisheng and Horizon.

Since 2015, it was decided to be an AI chip, and it was not until 2016 that it was officially on the right track. Yunzhisheng formed a chip team that experienced a long and difficult break-in period. Until “Mr. Key”, Mr. Xie Guanchao joined in November 2015 and is responsible for the entire IoT business unit. The most important KPI is the chip.

From Shenzhen to Beijing, Yunzhisheng's solutions began to revolutionize.

With the people who took the lead, the combat troops are still slowly recruiting and polishing, and the most difficult problem for the management at that time is the “interaction” between the algorithm team and the chip team. This is the pain that the AI ​​chip combined with hardware and software must go through. Both sides madly test each other on the edge of the bottom line.

If it is a general-purpose chip, it costs a lot and has a large memory and storage space, but once it reaches the AI ​​chip, especially on the edge side, a little more is a cost pain, then I hope the algorithm team cooperates with the cutting.

At this time, the algorithm team is frying. It is common to shoot a table and blink in the office. I have said that I have done the best, and I am very busy. The two sides are not arguing, they can only wait for the boss to make a decision.

In the eyes of the boss, that is backwards. If the chip is set up, it must meet the resource limitation requirements; the algorithm is the more resources you give me, the better I will give you performance. All that's left to do now is to confirm the performance level, then squeeze out each space and reach a settlement.

In the development of the AI ​​chip, this is what the team will encounter in every engine and every module.

Yu Kai attributed this problem to a conflict of ideas. When recruiting, someone asked him, you do software to do software, what is the combination of software and hardware?

He believes that this is the “wind resistance” that must be experienced in the first place. The Horizon chip team is also not working well for a long time in the initial period. Software development emphasizes rapid iteration, while hardware emphasizes systematic procedural thinking.

One is fast and one slow, one is creative and the other is order. These two people have natural confrontation.

From October 2015, the first chip engineer of the horizon entered the company, the first line of chip code knocked out, began to do front-end development, went through the process of writing code, testing code, drawing layers of physical implementation maps, and then manufacturing The sample was finally lit on August 16, 2017. On December 20 of the same year, the first generation of chips was released along with several typical applications based on chip development.

In fact, at the beginning, the chip team itself doubted, "Can it be formed?" Nothing! ”. But after the project really ran, the team's confidence began to become more determined.

In a key node, in order to ensure the progress of chip development, it is said that Huang Chang, the person in charge of the horizon algorithm, locked himself in a small conference room. He wrote a two-week code and hanged a sign at the door: “Closed development” Do not disturb, rdquo;.

Two weeks later, with the final version of the algorithm architecture, Huang Chang walked out of the small meeting room, dark circles, heavy, with messy hair, eyes flashing, "Get it!" ”

The mature team has reduced the horizon chip development cycle to 22 months. Unlike the rapid trial and error and iteration of the Internet industry, the chip is absolutely impossible. Yu Kai said that the cost of developing an AI chip is more than 50 million US dollars, which is extremely risky and stressful for startups.

Only make weapons without fighting

Two years of core making, in May 2018, Yun Zhisheng handed over the first generation of chips & quoquo; Swift, for smart homes and smart speakers, providing higher computing power at lower cost and power consumption. The chip adopts ARM's architecture and autonomous algorithm, and selects the UMC foundry and a Taiwanese company to complete the beta test.

Did not think, the darkest before dawn, the chip release on May 16th, once let Yunzhisheng decided to cancel, because the chip that was originally ready suddenly did not shine, everyone was shocked by a cold sweat.

Li Xiaohan quickly organized the investigation and found that there were problems in the development of the board map. He immediately revised the version and re-validated it. It took a week, and it was alarming.

After the release of the chip, the voice function test also made the team painstaking. Most of the customers were Southerners. Once, after several days and nights of testing and adjustments, they finally passed the test. However, the last group of people were beaten back, because the client's boss's accent came from a smaller audience.


Huang Weiren Celebrates the First Lighting of Chips

Li Xiaohan attributed the challenge of building AI chips to three reasons. The first is how to achieve the optimal energy consumption ratio in the whole architecture, and how to break through Feng.

Cloud Zhisheng is a dedicated chip for the Internet of Things, and the horizon is a dedicated chip for the calculation of intelligent driving in vehicles and for the IoT scene.

By the same token, their business models are restrained, just as Intel will not become Lenovo and Dell. When others are doing applications vertically up, Yu Kai hopes to stay at the bottom and compete with each other.

One is to earn the money of the whole industry chain, the other is to earn the money of the platform. They choose the latter. Everyone knows that making chips can only survive if they are paved, avoiding war as far as possible, and only acting as the role of transporting ammunition is their way of survival.

Yu Kai used to say that only making weapons and not fighting is this truth.

Similarly, when Yunzhisheng started delivering chips in the second half of 2018, what did they think most about delivering to customers? Although not all chips and applications are arranged by them, they will prepare many ready-made functional applications for customers because of the differences in the field of the Internet of Things.

Team research and development found that the Internet of Things market needs not an AI chip, but an AI function applied to the Internet of Things, such as you come up with a smart light bulb solution, which has a chip is not fake, but the function is what customers value.

Once, Li Xiaohan took AI chips to a lamp manufacturer. The boss looked at him blankly. Why did you give me this? But he also said what other functions the chip has, chips and software together with back-end services, and can achieve such and such functions, the boss invited him into the room and sat down to discuss in detail.

By the end of 2017, Horizon released Journey 1.0 processors and Sunrise 1.0 processors for smart cameras half a year earlier than Cloud Knowledgeable.

Yu Kai compares the chip architecture to the left brain and the right brain, responsible for rational computing or using the ARM architecture. The perceptual part is called BPU (Brain Processing Unit).

Journey 1.0 and Sunrise 1.0 chips use the chip Horizon first generation BPU architecture

For more than four years, Yu Kai is still confused by the enthusiasm and investment in software algorithms. He has always believed that the threshold of algorithms will be lower and lower. In fact, we can also see that many pure AI algorithm companies are adjusting their course.

The excitement is theirs.

This wave of AI chip craze, the company making intelligent voice is the most lively, the company making computer vision is the lowest key.

How do you say that? Wu Lei, a senior analyst, argues that voice control modules for both robots and air conditioners are new things. This is an opportunity in the voice field, so they prefer to try AI chips.

Shang Tang and Yun have not created new devices from the business of such computer vision companies. Web cameras or professional cameras are already there. They are currently trying to solve the problem that data flow to servers and are not effectively utilized.

Yunfu Technology is an artificial intelligence company that focuses on computer vision. Tang Libin, head of engineering technology, encountered a problem at the end of 2017. They had previously put the algorithm on the PC side.

They urgently need to find a way to plug the algorithm into the camera.

The first step is the chip selection, which lasts about three months. The product department and the algorithm department coordinate to get all the chips that meet the needs of the scene on the market to be tested one by one.

If it is not suitable, we will also discuss customization with chip suppliers. For example, some chips are not specially designed for terminal devices, and the compact environment will not be considered thoroughly for PC design.

Moreover, unless a general purpose chip is chosen, the design of the product and the selection of the chip will proceed synchronously, with a slight lag behind the chip. Just in case the chip doesn't work out, the product won't have to be designed. We will stick to the same time point to finalize the product design.

Tang Libin also found that domestic and foreign chip companies are excellent, but the support speed will be slightly different. The support speed of international manufacturers is one third of that of domestic companies of the same type. One reason is that international manufacturers will feed back technology requirements to foreign headquarters. The overall decision-making process is longer, and domestic companies can quickly customize the chip according to customer needs.

March 2018, Cloud from Technology's first Generation AI Smart camera

A few months later, the second generation of AI smart cameras came on the agenda. Huawei Haisi's chip became the main choice. Although the maturity was still polished, Tang Libin was assured that the performance improvement was obvious. They can devote more effort to polishing algorithms and investigating scenario functions before the product goes on sale.

Since 2019, Tang Libin has been visiting customers constantly. Many banks and he repeatedly mentioned the complaint rate. Sometimes customers feel that the business process is complex and they are prone to get angry when waiting for a long call. At this time, appeasement is the first.

They pilot an application in Agricultural Bank of Guangdong, the background can see how many people online shop, each person will show a state overhead, the general waiting time limit is 15 minutes, if there is no call, you need to go to greet, or pour a glass of water to let the customer's mood downgrade.

These things that go directly to customers are all as an artificial intelligence manufacturer that does not make chips, algorithms.

Saying No to Zha's Ten-Year Stepping

However, the computer vision camp is still unable to sit still, almost at the same time found problems, a choice is for people, a choice to practice.

In May this year, Zhu long, founder, announced in Shanghai Bund, taking advantage of the preheating of the news broadcast.


Display according to the diagram

They plan to build chips in February 2017 and then invest in ThinkForce, a veteran of the chip industry, to make Eto's core-building easier. According to the diagram to provide visual algorithm, ThinkForce is responsible for hardware research and development. This is printed in Zhu long's mouth.

It's a complete story. There's Hess on the terminal side. It may not be easy to follow the picture, but on the cloud side, they have their own power needs.

But the chip industry has been infiltrating experts for many years, Zhang Bing's question is whether the demand of Yitu can support enough, if not, Yitu is doing both business and chip, the situation is still slightly embarrassing. Compared with the Cambrian period when the third party is the sole underneath, it has more chances in terms of chips.

Hu Xiaojun, a resigned Ito engineer, and Netease Intelligent's Insight, said that in fact, Zhu Long often stressed that he should take a ten-year step in his early years. Why is it that he is in a hurry now?

Because in 2017, when we went to raise funds, the investors were the first to disagree with Zamabu. And now Yitu's revenue pressure is soaring. This may be the best choice.

Anyway, every road may lead to Rome. This is their story about chips and the evidence of the rise of China's core. No matter the front is the mine array or the abyss, everyone seems to be in high spirits.

Universal chips are finally over.

(At the request of interviewees, Wu Lei, Zhang Bing and Hu Xiaojun are aliased.)



China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments