Horizon’s “Flip” Story: Scientists, Timing, and “Second Place”

Original link: https://www.latepost.com/news/dj_detail?id=1873

Scientist entrepreneurship is not a typical story in the Chinese business world. In different laboratories in the business field, the resources, leadership, and business acumen of the CEO (Chief Executive Officer) are usually more important than technology. The position set by large companies for technical talents is “CTO” (Chief Technology Officer).

In some industries, the weight of technology will increase, and there will be more entrepreneurs with scientist backgrounds, such as intelligent driving. New technological routes may subvert the market. But putting the technology into practice, earning a certain amount of income, and finding a long-term and reasonable business model are the common thresholds for all scientists to start a business.

Horizon founder Yu Kai is often asked whether scientists are suitable for entrepreneurship. He is a suitable candidate. In July 2015, Yu Kai founded the chip company Horizon. He was one of the first industry experts in China to come into contact with artificial intelligence.

Liu Qin, the founding partner of Wuyuan Capital, met Yu Kai for the first time, and the two chatted from 8 pm to 3 am. Hotel restaurants were closing one after another, and they were “rushed back and forth.” Liu Qin remembered that that night, Yu Kai kept “nagging” about his imagination of the future world, saying that “the whole world will be robots in the future.” There are huge opportunities here, which are rare in ten or twenty years. .

Liu Qin agreed, but he was more concerned about how to consolidate his wild imagination of the future world into a “valuable landing place.” This is the prerequisite for all startups to become “unicorns”.

Yu Kai’s “landing point” is chips. He believes that when the robot era comes, a new computing platform will be needed, just like the terminal products in the PC era have built-in Intel chips for computing. Horizon’s goal is to become Intel in the robot age.

At that time, Google’s AlphaGo had not yet defeated the world Go champion, artificial intelligence had not entered the public eye, and investors did not trust the story of startups making chips, not to mention that the founder had never made a chip. Many people wonder, “A bunch of people who work on algorithms make chips?”

It took Horizon four years and three months to sell its one millionth Journey series chip. Nine months later, it hit its second million. The third million took less than five months. Growth is still accelerating. In 2023, Yu Kai said that the shipment volume of the Journey series chips will reach 3 million pieces this year, exceeding the total of the past three years.

Yu Kai is also constantly adjusting between his identity as a scientist and entrepreneur. When the first chip was released, he once said in an interview that the more a scientist is, the more he likes to think carefully and prepare everything before doing something, but he realized that starting a business is about gambling. He is learning to transform from scientist to businessman. After crossing the first million pieces at the end of 2021, CEO Yu Kai has a new position of “chief scientist”.

Today Horizon has become a role that cannot be ignored in the intelligent driving industry. Its chips are used in BYD, Great Wall, Ideal, and Changan cars, and its cooperation list includes almost all domestic mainstream car companies. It currently has mass production designated cooperation with more than 25 car companies for a total of more than 150 models.

Just over three years ago, there was no Horizon chip for cars on the market. Yu Kai described those years as being like a tunnel, “Fumbling in the dark, thinking someone gave you a torch, but in fact there was no one, you illuminated yourself.”

The important reasons for getting out of the tunnel are the business vision of the founder, the judgment of management and the investment of the entire company, but not all – there is also the opportunity of the times that happens to come.

There is no oasis in sight. Find a nearby water point first.

When artificial intelligence first attracted attention from investors, algorithm companies received more attention.

From 2015 to 2017, more than 1,000 AI startups were established every year, most of which are software applications. They can help users predict stock trends, recommend a top, or match an age-appropriate story to their children. Investors and users alike understand. Instead of making applications, we made chips for software companies, which was not widely recognized at the time.

Chip investment is high and the return cycle is long. If you invest the money, you won’t even see whether the chip can be produced in three years, let alone be verified by the market. “It’s like throwing a stone out and hearing no sound after five years.” Moreover, “Does the market really need a new chip company? We already have NVIDIA. Even if it does, can a startup company make a chip?” Liu Qin said.

At that time, Horizon may not be the most suitable team to make chips. Before founding Horizon, Yu Kai worked in deep learning at Siemens and NEC Research Institute in the United States, and then took the lead in establishing Baidu Deep Learning Research Institute, which focused on speech and image recognition. He has never made hardware, let alone chips.

Yu Yinan, Horizon Vice President and President of the Software Platform Product Line, is Horizon’s employee No. 5. He came to Horizon together with Yu Kai. He is responsible for basic vision research and development at Baidu Deep Learning Research Institute. He described the early Horizon team as “a large group of people who seemed to understand but didn’t understand much”. “Others thought it was a group of people who made software to make chips.” Yu Yinan said.

Talents in the chip industry may not choose to start a business in this field with a long cycle and extremely high barriers to entry. “The more experienced people are in semiconductors, the more they feel that Yu Kai (them) can’t do it.” Liu Qin said.

When Horizon was founded, it was divided into two major businesses: automotive (AUTO) and Internet of Things (AIoT). Cars include computing chips and software solutions for smart driving and cockpits. The Internet of Things is oriented to various life scenarios and has been used in appliances, cameras, toys, etc., and later gradually covered urban transportation, smart retail, etc.

Compared with installing a chip into a car, the entry barrier for IoT devices is lower and easier to implement. Liu Qin once advised Yu Kai that any company that wants to build a platform or is ambitious will have to wait many years before it is launched. The more visionary a company has, the more it must do a good job of dismantling the business from 0 to 1. “While you are working on it, you must find a scene and do a simple closed loop first. You will understand how big the gap is between what you want to do and user needs.”

One of the suggestions he gave Yu Kai was: “As the CEO, you have to negotiate the first order yourself.” At that time, Horizon had dozens of people, and Yu Kai wanted to find someone with sales experience to join. Liu Qin said that this person must be found, but the CEO must personally experience the market, which may be “a chicken feather in one place”.

Later, at Liu Qin’s suggestion, Horizon made air conditioners, and Yu Kai personally went to talk to the first customer. At that time, Horizon had not yet produced its own chips, so it had to build software on third-party chip modules to optimize air conditioners. In the publicity of that year, this air conditioner could be controlled with gestures to change the air supply angle and size according to the person’s position.

Horizon made various early attempts, including toys and sweeping machines. Liu Qin believes that it is like looking for water in the desert. “The target must be an oasis, but you don’t know when the oasis will arrive. Can you find a nearby water point first?”

Yu Kai founded Horizon earlier than Google to make AI chips. It was the only startup company in China making intelligent computing chips at the time. In the first two years after its establishment, building a team, developing chips, and running the market were all new challenges, and there was no template to refer to.

In December 2017, Horizon released the first generation of self-developed chips, Rising Sun 1.0 for IoT scenarios, and Journey 1.0 for intelligent driving. Rising Sun 1.0 is mainly used for image recognition. Journey 1.0 is not a car-grade chip and is only used to develop and verify software solutions for intelligent driving systems.

What followed was a “rush from left to right” approach to commercialization. In the smart driving business, the largest customer is hundreds of vehicles, and each vehicle has a few chips, which cannot be spread widely. Making IoT chips depends on the accumulation of technology on the one hand and whether it can accurately position in large-scale images. On the other hand, it depends on the ability to obtain orders. The test is experience, reputation and customer relations. The three most popular ones are Hikvision, Dahua Technology and Uniview Technology. The “Four AI Tigers” are also rising, as well as Internet companies with strong funds and willingness to invest. Horizon, which is 3 years old, is not dominant.

The business is not progressing smoothly, which affects the confidence of investors. After the angel round, Horizon’s financing situation is not optimistic until mass production of automotive-grade chips in 2020.

Not long after starting the business, Horizon encountered the capital winter of 2016. During the A round of financing, Yu Kai met with sixty or seventy investors but didn’t get a penny. He recalled that a well-known investor told him that making chips was a “particularly bad” thing. He felt that the other person was right, because the challenges of making chips were greater than imagined.

Fortunately, it has received investment from Intel and SK Hynix, two global semiconductor giants who led the investment in 2017 and 2019 respectively. Liu Qin feels that this means that companies with real technical capabilities in the industry understand what Horizon is doing.

The capital market’s pursuit of trends sometimes puzzles Yu Kai. He came to Liu Qin and asked why investors love facial recognition so much.

“He was a little confused. He was doing something and almost no one liked it. He didn’t think it was a good thing, but a lot of people were very optimistic about it.” Liu Qin said.

Liu Qin believes that this is the “noise” encountered on the way Horizon grows. Yu Kai is occasionally upset by these noises, but he does not waver in the direction he has chosen. His suffering was that although he had a clear and independent judgment on the chip very early and kept moving forward, it took longer than expected to reach the end point.

There is only one bullet, which target should I shoot at?

A debate has been going on among Horizon management for nearly a year – should we abandon all other businesses and only focus on intelligent driving?

Working in the automotive and Internet of Things industries at the same time, developing two products, the team’s energy is greatly dispersed. To make products and services the best, more money and more people need to be invested in both areas. At that time, the funds in Horizon’s account were almost bottomed out.

“We held many rounds of strategic meetings, with more than a dozen or dozens of people. We pulled people to one place and locked them up for a day or two.” said Zhang Yufeng, vice president of Horizon and president of the smart car division.

Those who support shrinking the IoT business believe that the scenarios are too fragmented and the business potential is not big enough. “To put it bluntly, there is not enough money and no oil and water.” Opponents believe that “the direction that is thrown away is also a good direction.” Compared with smart driving, the IoT business iterates quickly and has a shorter payback cycle.

During a discussion, Yu Kai asked, even if a home service robot could clean tables, wash and fold clothes and put them in cabinets, how much would people be willing to spend? If it can be solved, someone may be willing to spend 50,000 yuan, but technically it is not possible at the moment.

After many discussions, Zhang Yufeng remembered that Yu Kai concluded that Horizon should ensure resources for saturation attacks. If the troops are too scattered and want to do everything, it will be difficult to do well.

“If there was only one bullet, which target would we shoot at?” he asked.

The automobile market is large enough and stable enough. What’s more, smart cars are seen as the harbinger of the robot age, a period of optimism about autonomous driving. According to IT Juzi statistics, in 2018, the financing amount of China’s autonomous driving parts and solution suppliers was three times that of the previous year, exceeding 16 billion yuan. Tencent demonstrated an autonomous driving test vehicle, and Baidu announced the launch of driverless taxis. plan, some car companies have stated that L3 autonomous driving will be achieved by 2020.

In 2019, the international environment is complex and ever-changing, and the policy risks of the AI ​​industry are also factors that entrepreneurs have to consider. In the end, Yu Kai decided to focus on the automobile business and shoot his only bullet at a bullseye.

After an internal consensus was reached, the largest business contraction since Horizon was founded soon came. An employee recalled that during that time, Yu Kai held several all-hands meetings, and what he said the most was that “the company must focus on some strategies and make certain choices.”

A month or two later, all businesses except automobiles were cut off, and there were only about 70 people left in the IoT department, most of whom were in sales and customer support. Some core executives from the founding team also left that year. Later, the Internet of Things department was renamed “Universal Robots” as a layout for the robot era.

Liu Qin commented that Yu Kai was very safe. He had never seen Yu Kai lose control of his emotions and become fragile. When his business partner left, he saw Yu Kai was under the most pressure, but at that time he only calmly analyzed how to handle the follow-up work.

Start-up companies do not have the qualifications or time to be discouraged. Once they have made a decision, they must overcome everything and implement it firmly. In August 2019, Horizon released its first car-grade chip, “Journey 2.0”, and began to use this product to discuss cooperation with car companies.

Yu Yinan remembers that when he led a team to a car company, the other party was only concerned about some basic issues: Putting aside the functionality, does it have car certification? Where is the chip taped out? Can it be shipped on time and in quantity?

They prepared a series of “evidence”, including certificates, plans, production processes, technical details, and piled up awards and data. The other party still didn’t believe it, “You bring 100 certificates, and there is no car running on the road that will convince you.” force”.

“This is a comprehensive hurdle.” Zhang Yufeng joined in 2017 and is responsible for the automotive business market. He understands this uneasiness, “If you choose an internationally renowned factory and it can’t make it, it’s definitely not the wrong choice. But if you choose a start-up company and the project fails, you may first ask the decision-maker why you chose this home?”

At that time, there were more mature manufacturers to choose from for smart driving chips. The products of Moblieye or Texas Instruments acquired by Intel can meet the requirements of L2-level assisted driving functions, while the higher-level assisted driving functions are mostly occupied by NVIDIA.

Car companies followed Tesla in developing self-developed intelligent driving software, which opened a hole in the horizon.

Tesla also adopted Moblieye’s solution in the early years, and the chips and algorithms were packaged together. But starting in 2016, Tesla replaced suppliers with self-developed algorithms, and domestic car companies began to follow suit. In 2020, Tesla only sold 500,000 vehicles, but its stock price rose 7 times in a year, and its market value exceeded US$1 trillion – although intelligence is not yet the decisive factor for people to choose a car, no one wants to Falling into the disadvantage of “everyone has what I don’t have”. For the first time, Chinese car companies have shown their enthusiasm for AI chips.

In a complex environment, choosing a local supplier to ensure supply has become a safe choice for car companies. Horizon released its first car-grade smart driving chip in 2019, one step ahead of other domestic chip manufacturers. It was the only local manufacturer that could provide smart driving chips at the time.

Both sides “bet” together

Challenges popped up in such dense numbers that it was “too many to count”, and if one had to pull out a list, “it would be estimated that there would be thousands of small problems.”

Before setting off for Chang’an, the Horizon team felt that they had made adequate preparations and “felt good about themselves.” When we arrive at the scene, there are always surprises. This algorithm has been tested many times internally, and the camera is used to determine the position of the lane line. There is no difference of 5 mm before and after, which is the size of a soybean. But the other party’s intelligent driving system has a rule that the lane line is centered. If it shakes slightly, the front of the car will swerve, twisting and turning.

At the test site of Changan Automobile’s Chongqing headquarters, everyone in the team was nervous for a moment. This is the first project for the mass production of Horizon’s first car-grade chip, Journey 2 ADAS, for pre-installation.


Employees from Horizon, Freetech and Changan Automobile are at the project site.

Horizon and Changan Automobile first cooperated in 2018. At that time, Changan was doing a preview of L4 level at an expo, and Horizon was the only supplier willing to provide free services. This cooperation has allowed Changan to initially establish trust in Horizon intelligent driving technology. Later, Changan UNI-T became the first mass-produced model of Journey 2, but the chip was used for cockpit intelligence.

In 2020, Changan decided to add assisted driving functions to its key model UNI-V. At that time, the Journey 2 chip had just been released a few months ago, and Horizon was already working on an intelligent driving solution based on this chip internally. Changan was often approached to “directly deliver a basically completed product.”

The communication took more than half a year. Before the formal cooperation, Changan also spent a month doing intensive demonstrations. Teams from both sides sent people to verify the entire process from indicator data to actual vehicle road tests, and then compared it with other companies. The results were good, and Horizon people were looking forward to it, but no one dared to rest assured until the project was approved.

On the evening of August 1, 2020, the cooperation was officially confirmed. At 10 o’clock that night, the people in charge of the project gathered online to determine which employees should go to the site, and the discussion lasted until nearly 1 o’clock in the morning. Seven hours later, more than 20 people named appeared at the airport, all wearing Horizon shirts – they had to catch the earliest flight to Changan Automobile’s Chongqing headquarters.

The entire project lasted from August 2020 to January 2022. The main team has been stationed in Chongqing for more than half a year. When there are many people, there are 40 people stationed in Chongqing, and when there are few people, there are more than 20 people. The company rented an apartment near the test site. There was no office at first. After testing every day, the two parties would huddle on the table at the front desk for a meeting, with a small fan set up.

The project leader is Yu Yinan. He stayed for nine months and only came back to the company once every two weeks. “Basically, he was in Chang’an full-time.” The personnel list was decided in advance. When they arrived at the site, Changan was pursuing full-stack development in the future. They also called the core personnel of the perception algorithm, and “the R&D director was also called.”

“At that time, friends in the industry were also surprised by our investment in the Chang’an project.” Liu Yongliang is the delivery manager of this project and is responsible for communicating with customers. He joined Horizon in 2018, has more than 15 years of working experience, and has participated in all early ADAS-related projects.

Typically, software is updated on a bi-weekly or monthly basis. But in the first two months of this project, Horizon, Freetech and Changan agreed that the pace was to release versions on a daily basis. “Except for Horizon, almost no one would do this,” Liu Yongliang said, and the entire team was “rotating” crazily.

They don’t have much experience in mass production, and they may encounter the most complex urban road conditions in China. Chongqing has many slopes, few straight roads, and no bicycle lanes. Motorcycles and bicycles all run in motor vehicle lanes. In August, Chongqing was “spicy and hot”. Test members worked three shifts a day, going out to drive cars in the morning, noon and evening. Problems were discovered in the morning and had to be solved that day. They were tested in the evening and had to be corrected in the next day’s version. Before handing it over to Changan, Horizon will conduct an internal self-test of 2 vehicles and more than 2 hours to ensure that no unexpected problems will occur.

Fifteen months later, Changan UNI-V equipped with Horizon Journey 2 was launched. In addition to basic ADAS functions, it also has highway driving assistance HWA, driver-triggered lane change assistance UDLC and other functions. In the fifth month of its launch, sales exceeded Ten thousand.

The cooperation between Horizon and Li Auto has attracted a wider range of attention, and also caused deeper anxiety – two start-up companies, two groups of people who others think “seem to understand but don’t understand much”.

Lideal ONE is Lideal’s first mass-produced car, and its assisted driving functions are provided by supplier Moblieye. Moblieye’s strategy is to bundle chips and algorithms and sell them to car manufacturers in packages. The solution is mature and easy to use, but car companies are also limited in updating the software.

In 2020, after Ideal went on the market, it began to increase investment in research and development, focusing on intelligent driving, and decided to use self-developed algorithms to replace supplier solutions for the 2021 Ideal ONE facelift. To this end, Li Auto needs a chip supplier that is willing to open up.

Both Xpeng and Weilai chose Nvidia’s Orin chip at the time, with a computing power of 254TOPS. Horizon just released the Journey 3 chip in September 2020, with a computing power of 5 TOPS. Although it is not as powerful as the Orin chip, it can meet basic assisted driving functions and is cost-effective.

The most important thing is that Horizon opens its chips, software algorithm technology and supporting tool chains to Ideal, supporting the rapid transplantation of Ideal algorithms. A team of dozens of people is resident and jointly developed with Ideal.

Before the official establishment of the project, Horizon conducted two rounds of presentations to its ideals.

A person involved in the project told “Late Auto” that the first step was to demonstrate the perception algorithm. Horizon modified a car, added standardized sensing hardware, connected it to a computer to view the data, and brought it to the ideal headquarters for demonstration. Several rounds. However, when it comes to actual cooperation, Horizon has not yet demonstrated the closed-loop operation of Journey 3, which means that the car can truly run stably.

“This is the part of the gamble between the two parties, wading across the river to find a way.” At the beginning of 2021, the project was officially launched. The two parties were developing at a test site in Donghai County, Lianyungang, and each had dozens of people. In his impression, Horizon selected the most elite team. In a small conference room, teams on both sides stayed up late and worked overtime together.

One team of people is running in the test field, and the other team of Horizon employees is guarding the ideal Changzhou factory. We all have no experience in mass production, so no one can predict what minor problems may arise on the production line. The assembly line moves forward day and night. The last step is to calibrate the sensors. There are 10 cameras in a car. Each camera has 6 parameters. One parameter may be 6 targets, and the time is only 30 seconds. When the car is delivered to the work station, it must be aligned with the “click”. If there is something wrong, it must be solved immediately, otherwise it will pile up behind.


The 2021 facelifted Ideal ONE is equipped with the Journey 2 project site.

This is just one of the procedures. A set of intelligent driving software is connected to chips on one side and sensors and various components on the other. It may be fine for 100 vehicles, but there will be problems for 1,000 vehicles. “This is a probabilistic event,” said the Ideal Project employee. During development, everyone was thinking about the ideal state, but there was no such large-scale verification in the office. “It’s like a long jumper. If he jumps 100,000 times, can he jump 100,000 times?” Perfect every time?”

The strings were tightened. Horizon sent five engineers to work in shifts 24 hours a day, standing by with computers. If there was a problem, they would rush forward and change the code on the spot.

Horizon’s product owner pivoted. One day at around 6 pm, he just drove back to the Donghai test site from the Changzhou factory. When he arrived in Donghai, he received a call saying that there was a problem with the production line. He drove back at 10 pm and arrived there at 2 midnight. After solving the problem in two hours, I drove for four hours and returned to Donghai at 10 o’clock. I slept for an hour, wiped my face and continued working.

In June 2021, Lili delivered the 2021 Lili ONE. Five months later, Lili launched the NOA (Navigation Assisted Driving) and AEB (Autonomous Emergency Braking, automatic emergency braking system) functions. Since then, the ideal self-research path has been on the right track. This cooperation also gave both parties trust. In July 2021, Horizon launched the Journey 5 chip, with a computing power of 128 TOPS, and mass production for the first time on the ideal L8.

The ideal intelligent driving hardware architecture is divided into two versions: Pro and Max. The models are divided into Max version, Pro version and Air version according to the high, medium and low configurations. The Max version is equipped with NVIDIA Orin chip, and both the Pro and Air versions are equipped with Horizon Journey 5 chip.

The chip company closest to the customer

Yu Kai will emphasize on various occasions that his company’s value is “customer success.” Horizon is a start-up company that wants car companies to abandon their original supply chain leaders and choose itself. As important as cost-effective products is good service.

“To put it bluntly, when your product is not so standardized at the beginning and does not even have clear instructions for use, you have no choice but to provide services yourself.” Yu Kai once said in an interview with the media.

Journey 2 When I got on the train in Changan, the number of on-site employees dispatched by Horizon exceeded 40 at most, but the industry generally defaulted that “a dozen people are considered too many.” The on-site employees are responsible for the front office, and the company also has team support, which is divided into the middle office and the back office. The middle office solves the technical problems returned by the front office, and the back office includes internal testing, administration and HR.

Members are virtually immersed in the project. An employee who participated in the Chang’an project remembered that the work lasted from 8 or 9 a.m. to 1 or 2 p.m. every day. There were no weekends. “Except for sleeping, I was basically working.” To prevent staff from becoming too exhausted, on-site staff are rotated every two weeks.

Horizon has cooperated with Ideal twice, both of which were the first mass-produced Alpha projects. Both projects mobilized “teams of hundreds of people”, with Yu Yinan personally leading the team.

Some employees are driven by passion, self-improvement or the pursuit of technology, especially in early volume production projects. “No one has ever complained about such a tiring project, and some students took the initiative to come to me and said they wanted to exercise.” Liu Yongliang said, “Everyone is holding back their energy and has an exciting mood to pursue from 0 to 1.” After the project ended, several employees also bought models developed by Horizon and Changan.

At the same time, “customer success” is also included in employee performance appraisals. It is broken down into 20 specific indicators, divided into levels of improvement, compliance or exemplary according to employee performance.

“Where are our opportunities? It’s not because we are smart, but because we have the determination to work more and sleep less than others.” Yu Kai said at the 2021 China Automobile Blue Book Forum.

This concept is implemented from top to bottom. Yu Kai also spends a lot of time on customer maintenance. One early morning two years ago, Liu Qin called him, and Yu Kai was still chatting with customers about cooperation.

Horizon uses this to shorten the project cycle. Journey 2 was launched in Changan, and it took half a year from cooperation to the function reaching the mass production level, which normally takes more than two years; Journey 3 launched the Ideal ONE, and achieved mass production and delivery in 8 months.

Last year, a car company had a shortage of chips and a large number of finished vehicles were unable to be launched after the production line. They were packed in the parking lot. The production line dare not stop, otherwise there will be no cars to sell. The other party chose Horizon, hoping to supply, switch and help car companies implement self-developed algorithms in a short period of time – one of the reasons for placing the order was that Horizon moved quickly enough. It only took Horizon four or five months to complete this project.

The competition for new energy vehicles has reached the knockout round, and if you fall behind, you may not be able to turn around. More agile, more flexible and faster, giving Horizon more opportunities to get on board. This concept has continued from the early days of Horizon’s entrepreneurship to the present.

A Horizon employee involved in Volkswagen’s cooperation said that in order to win the cooperation with Volkswagen, Horizon had 20 or 30 people with a car and rushed to the Volkswagen China headquarters to find a hotel for development. The senior management of the other party was pressed for time, so they demonstrated parking and other functions in the underground parking lot of the office.

At the 2021 China Automobile Blue Book Forum, Yu Kai gave a speech and told that he went to meet with executives from Great Wall and BYD and promised them that “we are the chip company closest to you in the world.” The distance is physical and the willingness to respond. This is something that Nvidia, the chip leader, cannot and is not willing to do. “You make appointments and send emails to the other side of the ocean. All the people who come to the meeting over there are a bunch of Indian engineers. You can’t even understand the English they speak to you, right?” Yu Kai teased.

In addition to good service attitude, Horizon also provides customers with more technical support. An intelligent driving executive from a car company who has had business dealings with Horizon once told Late Auto that Horizon’s products are in-vehicle intelligent computing chips, but its advantage lies in its software capabilities. In order to demonstrate to car companies that its own chips can realize intelligent driving functions, Horizon will make some of its own software, which is as capable as the intelligent driving teams of many car companies.

Unlike Moblieye’s black-box delivery model, which packages software and hardware together and sells them to car companies, Horizon adopts “white-box delivery” and the underlying software code is completely open to customers, allowing customers to directly upgrade or maintain the software.

Liu Yongliang said that in the Changan project, the Changan regulatory team would ask “very, very detailed questions”, “How do you do your sensing algorithm? What is the model? How to analyze the technology specifically?” Horizon will “break it into pieces” to The other party explained. During the entire development process, more than 100 technical reports were written.

“There are very few suppliers who are willing to tell customers all the core technologies. Basically none. International friends usually tell you when to solve it. You only need to know what interface to call after solving it. You don’t have to worry about the specific plan.” . He said.

A person close to Li Auto said that during the project of launching the Li Auto ONE in Journey 2, Li Auto’s intelligent driving team had only a small number of people at the time and had many technical exchanges with Horizon staff, which allowed the team to grow rapidly. In the early days, Horizon also had some Demo source code to demonstrate hardware effects. Ideal completed a complete code replacement before mass production.

break boundaries

After Changan and Ideal’s projects, Horizon’s ability to mass-produce smart driving chips was verified by the market. Since then, car companies have begun to come to the door. From 2020 to April 2023, the shipment volume of Horizon Journey series chips was 3 million. More than four months later, Yu Kai said at the Munich Auto Show that the number had increased to 4 million. Currently, Horizon has reached designated mass production cooperation with more than 25 car companies, totaling more than 150 models.

Multiple projects running in parallel pose a new test to Horizon’s delivery and service capabilities. In 2020, Horizon has about 1,000 employees and more than 20 pre-installation designated projects. Currently, Horizon has more than 2,000 employees and 150 pre-installed and mass-produced models.

The intelligent driving industry is in a period of technological change and has not stabilized. In the competition, car companies have different positions and progress, and their needs are also different. The leading car companies are eager to self-research and follow Tesla; some traditional manufacturers believe that the cost of self-developed intelligent driving systems is too high and not necessary, so they hope to directly purchase complete solutions; and some manufacturers, while While purchasing supplier solutions, we are also building a self-research team. If Horizon wants to serve every customer well, it needs to match corresponding solutions according to their different needs.

“We can provide whatever the market needs,” Yu Yinan said. Currently, Horizon offers three chips and three intelligent driving systems, covering a total of nine products. It can sell the chips individually or provide a complete set of intelligent driving solutions. There are basic solutions that use only one front-view camera and only perception functions, and there are also some that require An autonomous driving solution using 4 Journey 5 chips and 47 sensors.

As a supplier, it is easier to make money by providing standardized products. However, the diversification of customer needs has led Horizon to do a lot of customized development for customers, which is a challenge to Horizon’s business model.

“The ability to serve car companies is not a matter of strong marginal effects, but depends on strong operations. Everyone is willing to do things with higher marginal effects, such as chip production and software. Tier 1 needs to serve car companies, which requires a lot of manpower.” 1 A person in the intelligent driving business of a car company said.

In 2021, Horizon proposed an ecological strategy of “open, win-win, all-dimensional altruism”, with the aim of accelerating chip rollout and grabbing a greater market share. Beginning this year, Horizon established a joint venture with Continental and China Science and Technology Chuangda. It positions itself as a second-tier supplier, providing standardized chips and underlying software to “ecosystem partners” who provide delivery and services to car companies.

“We suddenly realized that there are many projects and many customers. In the case of large-scale mass production and multiple customers, we have to support Tier 1 and car manufacturers. How can we do it well?” Zhang Yufeng explained.

He said that for L2-level assisted driving, Horizon has rarely faced car companies directly. The current cooperation model is to leave the first-tier suppliers to be responsible for integrating other products and most of the delivery work. The boundary with ecological partners is “what they can do.” We will definitely not do whatever they can do, and give priority to them if they can do more.”

But for top customers or the first mass production project, Horizon invests a lot of manpower to support it “regardless of the cost.”

In April 2022, Horizon and BYD officially reached a cooperation. BYD will sell more than 1.8 million vehicles in 2022 and has sold 1.25 million vehicles in the first half of 2023. It is a major customer that all smart driving suppliers compete for. In this bidding, BYD negotiated with almost all suppliers. Zhang Yufeng saw with his own eyes the competition among suppliers. “Everyone worked very hard to find solutions for price, customer relations, etc.”

BYD’s primary criterion for selecting suppliers is the ability to ensure supply. Zhang Yufeng said, “Can it supply such a large quantity?” On the other hand, BYD will not rely on a single supplier, but will hand it over to multiple suppliers to spread risks. According to “Late Auto”, the smart driving chips currently used by BYD include Texas Instruments, Horizon, and NVIDIA.

Among the many suppliers that have contacted BYD, Horizon will replace the original Texas Instruments in some models starting in 2022. A BYD employee said that one of the reasons for choosing Horizon was “better connection” and a more positive attitude.

Not long ago, he experienced Horizon’s “extra” after-sales service and on-site training, which lasted for one week. This is not common. Usually formal training is only in projects, and they have participated in it. But after the cooperative model has been mass-produced, they hope to be more familiar with some underlying configurations. Once they communicate, several people rush to the horizon and the other party sends The engineer taught for a few days and opened some operating documents.

In 2022, Horizon officially established the “System Department” to specifically target a few strategic customer cooperation projects. Zhang Yufeng introduced that this is a full-time team, including business, pre-sales, technical experts, project delivery and other positions. There are about a dozen people. Their time and resources are dedicated to serving this customer, and they are granted greater decision-making power. When encountering You can decide on-site emergency issues yourself.

This was originally a virtual model established in 2020. After it was materialized, it became “like a combat unit on the front line. For example, if a technical problem requires engineers to sense and control it jointly to tackle the problem, don’t just report it to the original business leader through formalities, and hurry up.” Just mobilize people to solve it.”

Behind the dedicated team is the sales and service center, which is like a dispatch center that helps the system department mobilize company resources and form their experience into a set of reusable standard processes to save manpower on most regular projects.

In terms of products, Horizon also hopes to provide standardized solutions for routine projects to save manpower and avoid repeated development.但实际应用中,仍有问题待解决。

一名软件员工说,公司希望用几套软件产品服务多个项目,但其实各个项目间细节差异很大,这些差异点被吸收进方案中,致使软件方案越来越臃肿。“就像一个堆了太多东西的仓库”。他解释,要修改一个问题,从“仓库” 里翻检的难度也变大了。目前,他修改一个功能的版本是三年前的

3 年前,他修改一个功能的版本大概是20、30 兆,现在要到80 兆左右。背后都是工作量。



英伟达Orin 芯片的算力为254 TOPS,地平线2021 年发布的征程5 算力128 TOPS——两颗才能达到英伟达Orin 的算力。去年,英伟达发布了下一代SoC 芯片“Thor”,算力可达2000 TOPS,预计2025 年量产。

地平线需要追求2000 TOPS 算力吗?


他赞同第一名、第二名都要做好大算力,但同时也要符合市场需求。“你想住600 平方米的房子吗?肯定想,但是大部分人实际不需要,也买不起。如果用户都不需要,搞那么大干嘛?” 他说,英伟达的Orin 算力比地平线征程5 增加一倍,但售价增加了两至三倍。

张玉峰说,他今年最重要的任务是“做到足够的增长”,目标之一是争取拿到L2 辅助驾驶市场份额的第一。L2 辅助驾驶功能对应的芯片主要是征程2 ,目前地平线已出货的300 万颗芯片中,超过一半是征程2。

而在一些更高级别的辅助驾驶功能上,大部分车企选择用英伟达Orin-X,而不是地平线征程5。据弗若斯特沙利文统计,2022 年中国市场上的高算力(算力大于50TOPS)自动驾驶芯片中,超过80% 都来自英伟达,地平线排名第二,市场占比为6.7%。

地平线的重要客户理想也在最高版本MAX 使用2 颗英伟达芯片,地平线征程5 则搭载于售价相对便宜的Air 和Pro 版本。



地平线的下一代芯片征程6 预计将在2023 年底至2024 年上市。余轶南说,这会是一个“家族” 式产品,不止一款芯片,算力最高的版本会超过Orin。

在今年初的中国汽车百人会上,地平线总裁陈黎明也表示,地平线不会一味进行算力竞赛。“增加算力,但用户体验并没有随着算力增加而实现线性提升,这对算力是非常大的浪费。” 陈黎明说,“市场需要竞争,产品、能力在竞争中获得成长,但到最后还是要回归商业本质。”

前华为高管苏箐2022 年加入地平线,目前负责基于征程6 开发城市NOA 解决方案,《晚点Auto》了解到,该方案目前尚处于DEMO (特定有限场景下功能演示)阶段。今年8 月,地平线内部软件业务团队开了一场员工大会,正式宣布了这一消息。苏菁担任平台与技术部负责人,直接向余凯汇报。

城市NOA 即在城市中按照导航辅助驾驶直达目的地,是今天量产车型最高级别的辅助驾驶功能,特斯拉2020 年已开始测试,在国内,小鹏、理想、蔚来、问界、长城等都在跟进。


2022 年10 月,大众集团旗下软件公司CARIAD 与地平线宣布成立合资公司,大众汽车投入24 亿欧元,持有合资公司60% 的股权。《晚点Auto》了解到,目前合资公司仍在筹备阶段,尚未正式成立,地平线内部已启动“招募” 软件人员,近百名员工可能转去合资公司。

这场“招募” 已经在内部进行了几个月。一些员工接到业务领导的询问,了解其是否有意愿转岗去合资公司,也有员工了解到消息后主动报名。

据了解,地平线CTO 黄畅也将去合资公司,带着部分地平线高级别研发人员。2023 年8 月,这一消息已在公司内部小范围宣布。据接近地平线的人士说,这家合资公司将使用地平线的智能驾驶方案及算法训练平台,为大众汽车提供包括高速NOA 在内的辅助驾驶功能。

2022 年的周年庆(6 月1 日)上,余凯在员工大会上提出, 2025 年将会是智能驾驶行业的分水岭,市场格局至此将稳定下来。




2019 年,特斯拉放弃使用英伟达芯片,推出自研FSD,马斯克在发布会上说,“从来没有生产过芯片的特斯拉怎么能设计出世界上最好的芯片,但事实是,这发生了”。在国内,目前蔚来、小鹏、理想都组建了芯片团队,自研芯片都在进程中。





自研芯片当然要求巨大投入,但选择自研的车企相信,自研将更适配自身需求,销量规模足够大也能够降低成本。李想曾说,特斯拉FSD 硬件成本约1500 美元,国内厂商双Orin-X 的方案成本至少4000 美元。如果做好智能驾驶必须要自研芯片,全世界排在前几名的车企都不得不做这件事,那么留给地平线和其他芯片公司争夺的市场会越来越小。

车企们能不能做得出来是个未知数,挑战是肯定的。但2015 年夏天,初创的地平线也曾不被行业看好,幸运的是它把握住了时代机遇,成长为“独角兽”。




当燃油车变成新能源车,一辆车用到的零件数从30000 变成10000,最有潜力的汽车品牌大多是新面孔。时代机遇下,供应商们各显本领,老牌供应商转型、新型供应商崛起,一些其他行业的龙头企业也进入汽车零部件市场,共同重构整个产业链条。《晚点Auto》推出供应链新军系列文章,关注汽车供应链的重塑。

本文转自: https://www.latepost.com/news/dj_detail?id=1873
This site is only for collection, and the copyright belongs to the original author.