Yu Kai of Horizon Robotics stands at the historical crossroads of China's intelligent driving
Latest update time：2024-11-06
Reads：
“
The success of each generation of entrepreneurs is inseparable from the greatest pursuit of the young people of that era. Behind Horizon, there is not only Yu Kai’s ideal as a “
scientist
”, but also the common historical mission of a group of deep learning scholars and engineers.
”
Author | Li Yuchen
Editor
| Lin Jun
Among the entrepreneurs in the past decade, there are very few who have both "book smarts" and "street smarts". Yu Kai of Horizon Robotics is one of them.
Yu Kai has been a multi-dimensional person, with more than 100 academic papers, as the director of NEC Lab and the executive vice president of Baidu IDL. He told Leifeng.com that when he decided to start a business in 2015, he felt that computer vision was a transitional discipline that would eventually transform into Robotics.
"The reason for working on cars is that I want to work on robots. Cars are a form of robots. But now cars are our biggest base. Only with them can we talk about the future."
In the eyes of many people, Yu Kai is righteous, capable of taking on responsibilities, has high emotional intelligence, and a broad vision. He is a person with a clear ultimate goal - to be the brain of the era of intelligent robots, to enable every car and every appliance to have the ability to perceive the environment, and to make decisions and plan through human-computer interaction, so that machines can serve humans.
This is the underlying support for Yu Kai's personal long-termism, which enables him to make drastic measures and preserve his vitality when an entrepreneurial crisis comes, and to make bold bets and leave no way out when opportunities arise.
The success of each generation of entrepreneurs is inseparable from the greatest pursuit of the young people of that era. Behind Horizon, there is not only Yu Kai's ideal as a "scientist", but also the common historical mission of a group of deep learning scholars and engineers: to create a Chinese company with underlying technological innovation like Wintel with their own hands.
01
Yu Kai's Silicon Valley Past
Before deciding to start his own business, Yu Kai was best known for his academic achievements.
In academia, Yu Kai has served as a member of the program committee for many top international conferences such as NIPS, ICML, SIGIR, KDD, IJCAI, and as an expert judge for well-known journals such as JMLR, TPAMI, MLJ, TNN, TKDE, and the National Science Foundation of the United States.
As one of the few Chinese scholars, he is also the Area Chair of the two major international machine learning conferences, ICML and NIPS, which shows his status and influence in the international academic community.
Yu Kai is from Nanchang, Jiangxi Province. He completed his undergraduate and master's degrees at Nanjing University. In July 2004, he received his doctorate from the Department of Computer Science at the University of Munich. After graduation, he worked at Siemens Corporate Technology in Germany.
In 2006, Yu Kai came to NEC Lab, which at that time was equivalent to the base camp for neural network research.
When Yu Kai joined, deep learning had not become popular again, and support vector machines were the mainstream at the time. In addition to the support vector machine field, NEC Lab also has a group of experts in neural networks. Therefore, there is a strong accumulation of deep neural networks here.
Yann Lecun was also the head of the machine learning department at NEC Lab for a period of time. Vladimir Vapnik, who worked at NEC Lab from 2002 to 2014, is as important in the field of support vector machines as Geoffrey Hinton is today in deep learning.
In 2010, Yu Kai, who had become the director of NEC Lab, worked with Huang Xutao to bring together Lin Yuanqing, Zhu Shenghuo, Yang Ming, Lu Fengjun and others, UIUC interns Yang Jianchao, Zhou Xi, Cao Liangliang, Wang Xiaoyu and Rutgers University professor Zhang Tong to participate in ImageNet, the world's top image recognition competition.
During that time, at the recommendation of Trevor Darrell, a professor of computer vision at the University of California, Berkeley, Jia Yangqing also joined NEC Lab as an intern and began to engage in deep learning research under the guidance of Yu Kai and Huang Chang.
A few years later, Zhou Xi founded Yitu Technology, Wang Xiaoyu co-founded Yuntian Lifei, Cao Liangliang later became the head of machine learning for Apple Vision Pro, Yang Jianchao became the head of ByteDance Silicon Valley AI Lab, and Jia Yangqing became famous for the deep learning framework Caffe.
(2017, Jia Yangqing, Yu Kai and Li Mu, from left to right)
The ImageNet Challenge is jointly organized by Stanford University, Princeton University and the Visual Geometry Group (VGG) and is seen as a key symbol of progress in computer vision research and artificial intelligence.
In 2010, 35 programs from academic and industrial computer vision researchers around the world participated in the competition. Each contestant received 1.2 million labeled training images and a list of possible categories to which they belong, and the task of the participating program was to output the correct category for each image.
Later, Yu Kai commented that
the first great thing he did after entering the industry was to become the first champion of the ImageNet Challenge.
It was not easy to stand out from the 35 competing teams, because NEC Lab was at a critical moment with its mind wandering and the team in turmoil, and the impact of the 2009 financial crisis was still ongoing.
Even Huang Xutao, who was highly respected in the academic world, was worried about not being able to raise enough funds to train PhD students, so he could only urge students to graduate as early as possible. NEC found Yu Kai, who was quite influential in the team, hoping to find a successor to the laboratory.
Yu Kai is a natural leader and organizer. He often organizes parties at home, and people even suspect that all Chinese scientists in Silicon Valley gather at his home every weekend.
At the end of 2009, Yu Kai felt that he had a responsibility to reunite NEC's strength. He set two goals: to prove to the industry that he could attract the best talent and produce the best results in the world.
Yu Kai personally interviewed and recruited two great people. One was Huang Chang, who graduated from Tsinghua University with a bachelor's, master's and doctoral degree. He was very passionate about image recognition when he was an undergraduate, but until he graduated with a doctorate, he had not figured out how to apply image recognition on a large scale in the industry, so he went to the University of Southern California as a postdoctoral fellow. Later, Huang Chang rekindled his interest in the industry and found Yu Kai through a friend's introduction.
In Yu Kai's eyes, Huang Chang is a genius, and his research in the field of face recognition is world-class. Another is Indian Manmohan Chandraker, who has won many best paper awards. Yu Kai commented that he is "the best in 3D vision."
As the "parent" of NEC Lab, Yu Kai created very good research conditions for the lab. He took on most of the odd jobs, traveled frequently, and ran projects, maintaining a very good relationship between the lab and the headquarters, and the funding was sufficient. Yu Kai also established extensive cooperation with academia and industry, and was very active in external exchanges, such as cooperating with academic researchers such as Andrew Ng on papers and being invited to teach at Stanford University.
So, from this point of view, Yu Kai was not a typical scientist from the beginning. He was soft-hearted but determined, willing and good at handling interpersonal relationships. Without him, the glory of NEC Lab might have faded in 2009.
(A gathering between Yu Kai and his old friends from NEC Lab in 2016)
It was the eve of the 2012 Spring Festival.
At this time, Li Yanhong met with Yu Kai in Beijing and invited him to return to China. Yu Kai resolutely ended his six-year career in Silicon Valley, returned to China to join Baidu, established the Multimedia Department, and began to study speech recognition and image recognition.
Previously, Baidu had almost no research experience in AI. After Yu Kai joined Baidu, Baidu took the lead in injecting the gene of deep learning. In just a few months, Yu Kai preached the potential of deep learning to Robin Li.
Li Yanhong was shocked again and again, and even sent a letter to all product managers in the company, asking them to pay attention to the technology of deep learning. Another synonym for deep learning, neural network, also began to spread in China, and Baidu first supported the earliest batch of companies doing data labeling in China.
However, what really made Robin Li decide to invest heavily in deep learning was the failure of an auction in which Yu Kai participated on behalf of Baidu in December 2012.
In 2012, the neural network AlexNet developed by Geoffrey Hinton's team won the ImageNet championship, causing a sensation worldwide. Yu Kai told Leifeng.com, "No one in the world is more shocked than me, because I am the first champion, and I know what it means to increase the accuracy from 75% to 85%. So he immediately sent an email to Hinton, expressing his eagerness to cooperate.
This email inspired Hinton to launch a secret auction at the NIPS conference that year, attracting four of the world's leading AI companies: Google, Microsoft, Baidu, and DeepMind. In the auction, DeepMind and Microsoft were eliminated first; later Baidu and Google continued to bid until the final price of US$44 million. Hinton felt that if he took more money, he would not be able to live with his conscience, so he stopped the auction.
(Yu Kai
and Geoffrey Hint
on in the same frame
, source:
Wired)
Yu Kai judged that as long as Google and Baidu kept bidding the same, Hinton would definitely go to Google, because if Hinton came to China, he would only know himself.
However, this auction allowed Robin Li to personally see that a world-class company was willing to pay so much money for three people (Hinton's team, including Hinton and his two students), and realized the prospects of deep learning and the crisis of backwardness.
In January of the following year, Robin Li prepared to establish the world's first AI laboratory named after deep learning - Baidu Institute of Deep Learning (IDL). Robin Li served as the dean and Yu Kai served as the executive vice dean, which also opened the prelude to deep learning in China.
(Photo of Yu Kai in the early days of IDL)
Yu Kai was the actual operator of Baidu IDL in its early days. Under his influence, Baidu attracted a large number of outstanding AI scientists, such as Xu Wei, the founder of the PaddlePaddle platform, and Huang Chang, and later Lin Yuanqing, the director of Baidu Research Institute.
Baidu’s “Young Marshal Plan” was originally proposed by Robin Li with the aim of attracting young talents. Yu Kai mainly lobbied Li Mu (later Chief Scientist of Amazon AI and current founder of the big model startup Boson.ai), Li Lei (later Director of ByteDance AI Lab), and Gu Jiawei (current CEO of Wuling Technology) to join.
China's earliest autonomous driving project was born in IDL. Tao Ji, Ni Kai and other autonomous driving entrepreneurs were the first batch of employees recruited by Yu Kai for the project.
It was a coincidence that Yu Kai invited Andrew Ng to join Baidu.
In February 2014, Yu Kai went to the United States on a business trip and stayed at the Sheraton Hotel in Palo Alto, near Stanford. Andrew Ng happened to be at the hotel that day. The two hadn't seen each other for two years, so they chatted for a long time by the swimming pool. Not satisfied, they met again in the hotel restaurant in the evening.
At that time, Andrew Ng was working on Coursea, and Yu Kai felt that Andrew Ng was not doing well, so he told him that instead of working in education, he might as well return to what really touched him and where his passion was, which was AI. Andrew Ng took this to heart and started a conversation, and Yu Kai said that Baidu was willing to open its doors for him.
Later in March, Yu Kai arranged for Andrew Ng to meet Robin Li in Beijing, and Andrew Ng joined the company in May.
When Li Mu was at Baidu, Yu Kai discussed a project with Li Mu. Yu Kai believed that Baidu's voice, images, videos, etc. were all related as a whole, and a large model could be trained, which would definitely be very useful. This idea is actually a large model like ChatGPT today, but unfortunately they did not really do it at that time.
The first reason was that the resources invested were too large, and the second was that Yu Kai saw that Baidu had no intention to open source PaddlePaddle, so he suggested that Li Mu build an open source deep learning platform outside. This is one of the sources of MXNet.
One weekend, Yu Kai invited Li Mu to a cafe in Tsinghua Science Park and told Li Mu that an open source deep learning platform was a direction worth trying. Later, Li Mu began to cooperate with Chen Tianqi and others.
In fact, when it comes to key career choices, many colleagues are willing to listen to Yu Kai's opinions.
For example, Li Lei, the former director of ByteDance’s AI lab.
In 2015, after Yu Kai left Baidu, Li Lei called Yu Kai and asked him for advice. Li Lei is an outstanding scholar and won one of the best doctoral dissertations (second place) of the SIGKDD Committee on Data Mining of the American Computer Association. In addition, Li Lei is also one of the co-authors of the best paper of ACL 2021.
Yu Kai said to Li Lei, "If you want to choose a recommendation algorithm, you should go to ByteDance."
In 2015, ByteDance was not yet an important company, but Yu Kai keenly predicted the potential of ByteDance in the future. At that time, Yu Kai was on his way to a meeting with investors, and he spent two hours persuading Li Lei to join ByteDance.
Before Li Lei, Yu Kai also introduced Chen Yuqiang to Zhang Yiming. Chen Yuqiang studied under Professor Yang Qiang, a top scholar in the field of artificial intelligence, and focused on research in the field of transfer learning. At Baidu, Chen Yuqiang is the person who works most closely with Yu Kai.
When it was rumored that Sun Jian was leaving Microsoft Asia Research, Yu Kai had a meal with Sun Jian and asked him if he was willing to join Horizon Robotics. Sun Jian thought that Horizon Robotics was working on hardware, so he didn't want to go.
Sun Jian joked, "When I was an undergraduate at Xi'an Jiaotong University, there was a guy in my dormitory who worked on hardware. We would sleep soundly every night, but he would come back in the middle of the night."
From his experience at NEC to Baidu IDL, we can see that Yu Kai already has the ability to predict the future evolution of technology. Of course, he already has sufficient psychological expectations for the "hard work" of working in hardware.
(An interesting phenomenon is that scholars born in 1974-1976 and around 1984 are two important driving forces of China's deep learning wave. Yu Kai's two "leading" research and development experiences at NEC and Baidu have led to many historical anecdotes with his successors. Interested readers can add the author's WeChat Gru1993 to communicate)
02
From J1 to J6, the horizon is bright
Starting a business was a watershed decision for Yu Kai’s career in academia and as a worker.
In March 2015, Yu Kai resigned and began planning to start his own business.
One version of the origin of the name Horizon comes from the poem "Love Life" by poet Wang Guozhen: "Since the goal is the horizon, all that can be left to the world is the back."
Yu Kai once told Leifeng.com about the ideas and logic behind founding Horizon Robotics and developing smart driving chips:
First, Baidu was the first company in the world to use GPUs on a large scale since 2012. However, "normal software runs very fast on CPUs, so why doesn't deep learning run very fast on CPUs? This shows that the separation of software and hardware is just an ideal."
Yu Kai believed that using GPU for deep learning was an accidental achievement, and that it would be a better direction to make a dedicated chip for deep learning. Yu Kai went to Li Yanhong and wanted to design (hardware) specifically for deep neural networks, but Li Yanhong thought that the hardware business was difficult and the search engine was at its peak of profitability at that time.
Second, after Baidu launched the autonomous driving project in 2013, Yu Kai discovered that autonomous driving could not be mass-produced without making chips.
Yu Kai comes from an algorithm background, but he is not superstitious about algorithms. Instead, he is determined to start from chips.
He realized the difficulty of automotive-grade chips very early. An industry insider once joked to Leifeng.com: "Yu Kai's hair has become thinner and thinner over the years since he started his business, from a side part to a flat top."
When Horizon Robotics was first established, it needed to find application scenarios. Horizon Robotics listed four major business lines: toy robots, smart homes, smart retail, and smart cars. To be honest, the least promising one at the time was cars. The other three either had sexy concepts or could get projects right away. At the time, the four AI unicorns were in full swing and could generate hundreds of millions of dollars in revenue.
On December 20, 2017, Horizon Robotics held a press conference at the China World Hotel in Beijing, launching two smart chips for the first time: Journey 1.0 (J1 for short) and Rising Sun 1.0 (X1 for short). The two chips provide solutions for application scenarios such as smart driving and smart business.
Yu Kai is one of the few scientists who is particularly good at marketing. The product names of contemporary chip companies are mostly similar to AK000/MHI587, which are not easy for customers to remember. Yu Kai named Horizon's chips Journey and Rising Sun, which greatly simplified the cognitive burden of users.
Just like the naming of the horizon, there is a beautiful vision behind the naming of the rising sun: the rising sun marks the beginning of a journey - perhaps this is a kind of romanticism of Yu Kai as a scientist.
As a "non-automotive-grade" automotive chip, J1 was limited by the R&D capabilities at the time, and its overall performance was average. The OEMs were also unclear about the application scenarios of this chip in automobiles.
At that time, Yu Kai did not have too much revenue expectation for the intelligent driving chip. His understanding was that the automotive business was relatively slow and the OEM's fixed-point process was longer. The AIoT market would be relatively more mature, and the focus should be on generating revenue around the X1 chip.
Therefore, the significance of J1 and X lies in proving to the outside world that Horizon Robotics, which originated from software algorithms, can successfully develop hardware by combining software and hardware.
In August 2019, Horizon released its first automotive-grade chip J2.
J2 is equipped with Horizon Robotics' independently developed high-performance computing architecture BPU2.0 (Brain Processing Unit), which can provide 4 TOPS of computing power. In Yu Kai's words at the time, "Horizon Robotics took the lead in launching J2, which not only achieved a breakthrough in the mass production of automotive-grade smart chips in China, but also completed the key link in the construction of the domestic autonomous driving industry ecosystem."
However, at the time of its release, domestic OEMs did not show enough attention to Horizon. The new energy industry was in the depths of winter, and Tesla's stock price was at its lowest level in the past five years.
In 2018, Changan Automobile ranked third among domestic automakers. At that time, FAW and SAIC both had strong joint venture brands. As the third largest automaker in the market, where would Changan’s future growth point be? Changan Automobile approached McKinsey for strategic consulting. McKinsey’s advice to Changan was to make smart cars and rely on smart cars to enhance Changan Automobile’s overall brand image.
In this context, Changan Automobile paid attention to Horizon Robotics, which was a newcomer in the automotive market at the time, and hoped to use automotive chips to enhance the intelligent experience of the entire vehicle. The J2 chip has a computing power of 4 TOPS and is installed on the Changan UNI-T to officially achieve mass production.
An insider told Leifeng.com, "The reason why J2 has a small computing power is that Horizon Robotics was very tight on money at the time, so it could only use extreme software and hardware optimization to achieve a 10-fold cost-effectiveness. In order to make customers accept it, the cost-effectiveness must be so great that they can't resist it. When the J2 chip demo was given to the Germans, they didn't believe it and didn't think such a small chip could handle forward-looking perception tasks."
In fact, the launch of J2 was the darkest time for Horizon, and the most difficult period was between October and December. In other words, if there was no Changan, there might not be Horizon today.
(Photos of the Changan project during the breakthrough)
At that time, Horizon's AIoT had revenue, but there was no way to make further breakthroughs, and there was no new financing. Funding, strategy, and personnel issues were like a tightening hoop that hung over Yu Kai's head.
During the most difficult period, the number of people at Horizon dropped by one-third from more than 1,000 to only about 800. Recalling the situation at that time, Yu Kai said: "
Every day was like falling into the abyss, with no handholds.
"
According to Horizon's cash flow at the time, if AIoT was not cut, the cash flow could only support three years. If AIoT was cut, the cash flow could support two more years. For the founder, it was difficult and painful to make such a decision.
But investors and customers are not here to listen to stories. Where is Horizon's influence? Where is its certainty? What is the long-term competitive situation in the business? The automotive scene is big enough, but Horizon's smart driving chips have not yet been developed. When will the demand in this track turn?
......
In the second half of 2019, Yu Kai was repeatedly tortured by similar questions.
During that half year, he held rounds of meetings with the executive team, trying to answer these questions that had been intentionally or unintentionally ignored. In the end, Horizon decided to abandon the AIoT business and focus on the automotive track.
People close to Yu Kai admitted that "the revenue of the AIoT business is not healthy enough. It is particularly deceptive. For example, a project may have revenue of tens of millions or even hundreds of millions, but there is no growth potential, which distorts the team."
Therefore, Yu Kai thought clearly about one thing:
the scenarios in the AIoT field are too fragmented. If he continues to bet on it, Horizon will only be dragged down.
Yu Kai's decision this time gave Horizon a chance to survive. It was also during these six months that Yu Kai, in pain, truly completed the initial transformation from a scientist to a businessman and began to think about the essence of business.
He transformed himself into a super salesman, and whether he was drinking and banging the table or trying to persuade customers, he would spend every day at the car factory, concentrating on trying to convince customers and talking business with them.
A consultant close to Yu Kai told Leifeng.com, "Entrepreneurs should not spend too much time on strategy. Because in many cases, the strategy is right, but the problems all arise in the execution stage. Humans will always overestimate others' ability to understand them, but in fact, only when you take the lead and set an example, will they understand the logic of decision-making."
Yu Kai has a unique understanding of how to deal with the relationship with the OEMs. Compared with his peers who only focus on hardware, Yu Kai is open-minded and sincere to the OEMs. For example, BYD sent people to Horizon Robotics to learn. Horizon Robotics not only opened up its code and development tools, but also provided internal training to more than 100 visitors, and even jointly developed projects with them.
In the eyes of many people, Yu Kai is a person with high emotional intelligence and good German. When a leader of Volkswagen visited China, he personally received him and arranged his itinerary.
In addition, Yu Kai is very good at speaking and telling jokes. Whenever he attends a dinner party, there is no possibility of the whole party being silent. He is also well-known for his rich experience and extensive network of contacts.
After a dinner between Leifeng.com and Yu Kai in August this year, he left in an Ideal Mega worth more than 500,000 yuan. According to people around him, he would support his clients with real money out of his own pocket.
People close to Yu Kai have never seen him cry, and he rarely gets angry, but the one time they have seen him get emotional is when he gets recognition from a client. In the final analysis, Yu Kai has a sense of the underworld, values human relationships, and has a clear understanding of how the business world works.
These traits are rare in scientists. When most scientists change their identities to become entrepreneurs, they often have shortcomings such as arrogance, low emotional intelligence, and lack of understanding of customers and marketing. But Yu Kai has never left such an impression on the outside world.
If the cooperation with Changan was the key battle for Horizon to open the door to the automobile industry and build confidence, then Ideal made Horizon completely famous in the industry.
In the industry, ideals are not easy to fool. But Horizon, led by Yu Kai, chose to take the initiative to plunge into ideals. With the active matchmaking of Zhang Lei of Hillhouse Capital, Yu Kai and Li Xiang had a meeting. Yu Kai's high emotional intelligence and social attributes made Li Xiang very happy, and the meeting was a great success for both parties.
In 2020, Ideal was caught in the vortex of chip shortage along with the industry, and is actively looking for ways to ensure supply. But despite Yu Kai's eloquent speech, Li Xiang, who had no idea about Horizon Robotics, finally responded to Yu Kai's kindness with a vague "give it a try".
Li Xiang's so-called "give it a try" is to agree to try to use Horizon's chip solution on the new car, but at the same time he also put forward harsh additional conditions: Ideal will only provide limited personnel support and will not place a heavy bet on Horizon.
As a newcomer, Yu Kai had to convince the outside world that Horizon Robotics had the ability and determination to make smart driving chips. Therefore, the first customer had to be established. Horizon Robotics's tough tactics and flexible posture became the key variables.
"When you only have one bullet in your hand, things can be accomplished." This became what Yu Kai often said in Horizon.
In order to respond to Ideal's needs in a timely manner, Yu Kai took the lead in mobilizing all the elites of Horizon and formed a 300-person team to be stationed at Ideal to provide real-time assistance. At that time, Horizon, which had already undergone a major downsizing, accounted for almost a quarter of the total force.
In order to cooperate with the countdown of Ideal's new car, the Horizon team even worked overtime day and night with the Ideal Auto team to tune the chip. Even Li Xiang himself has repeatedly praised the Horizon team's desperate work attitude.
In September 2020, Horizon released J3, and it only took 8 months for J3 to be mass-produced on Ideal ONE.
This achievement made Horizon Robotics famous in the industry. As for the J5 chip used in the Ideal AD Pro system solution, Horizon Robotics even set a record of 7 months of mass production and delivery.
At that time, the entire delivery process was short and the task was heavy. But Yu Kai was still able to arrange hundreds of people to develop on site. You know, Ideal is notoriously strict with suppliers, including Li Xiang himself, who would constantly make demands and "scold people".
Under such conditions, Yu Kai and his team of Horizon engineers persevered.
After rigorous training within Ideal, the Horizon team has also developed the ability to respond to OEMs in a timely manner. After differentiation and combination, this team has begun to become the embodiment of Horizon's core technical and engineering capabilities. Ideal has also rewarded Horizon for its ultimate service. In 2021, Horizon's J3 chip will be launched for the first time on models such as Ideal One and L7.
A former mid-level manager of Horizon Robotics told Leifeng.com, "The essence of business is to build credibility, not to exchange money, but to establish a trusting relationship with your customers. Only when that is established can everything else come true."
Subsequently, SAIC, BYD, Geely, Ideal and Changan all used J3.
In the one year from September 2020 to September 2021, Horizon Robotics received eight rounds of financing with a total of more than US$2.5 billion. It took only one year for Horizon Robotics to get a fixed position, gain the favor of capital, and go from falling to the bottom to attracting everyone's attention.
03
Atypical chip company
After J2 and J3, Horizon’s next generation product is J5, but this is a product that makes Horizon "full of pain".
At that time, Wu Zheng, the general manager of J5 chip, brought 6 directors to Horizon Robotics. There were not enough people, and competitors were trying to poach people. Secondly, the IP left over from J2 and J3 was originally planned to be reused, but later it was discovered that functional safety had to be supported, so the CPU and bus had to be replaced, the IP had to be replaced with a new one, and all modules had to be replaced, which was equivalent to starting from scratch.
A core member of Horizon J5 told Leifeng.com, "The pressure in the later stage of J5 was very high. We worked overtime during the National Day and were busy until 11-12 o'clock every night. There were no weekends or New Year's Day. During the Spring Festival of 2021, on the third day of the Lunar New Year, we worked overtime together until tape-out in mid-February. It was really hard at that time."
Another core member recalled, "At that time, the team encouraged each other and said, 'Brothers, don't forget your original intention. We are here to get things done. No matter how hard it is, we must persevere.'"
The launch of J5 has given Horizon Robotics a weapon to impact the high-end intelligent driving market, and it is also a key watershed for Horizon Robotics' in-vehicle intelligent chips. It can be said that the J5 R&D team has made an indelible contribution to Horizon Robotics.
(One-time lighting for Journey 5)
However, now it seems that there is still some controversy about the evaluation of J5 internally and externally. J5 was defined in 2018, and at that time it was impossible to predict the changes in market demand five years later, so it is inevitable that there will be defects.
Since Horizon's previous J2 and J3 were both low-computing chips with limited computing power, it was difficult to support the rich chip ecosystem that Yu Kai wanted. It was the launch of J5 that enabled Horizon to move from closed to open and have its own chip ecosystem.
There is no need to explain how important ecosystem building is to a chip startup like Horizon.
Each generation of chips is like a pond dug by Horizon itself, but if the chip is to have real vitality, it also needs a rich operator library, developers, and a large number of users, just like the long-term prosperity of a pond is inseparable from a variety of aquatic plants, animals, insects, and plankton.
"The advantage of J5 is that Horizon has many partners working with it to optimize algorithms. Other domestic chips currently have few customers and are a bit closed-door. Car companies will think about how to use their chips."
An industry insider analyzed to Leifeng.com that from the perspective of customer acquisition, the chip ecosystem is based on division of labor and cooperation, and has a certain degree of stickiness. At this time, as long as one of the many ecological partners has an order, everyone will have work to do.
Yu Kai has also mentioned on many occasions when evaluating J5 that ecology is the key in future competition, because technology can easily be overturned by latecomers, but a well-developed ecology can attract customers, which is the same logic as WeChat.
"If we have so many customer resources, then even if we don't do well in one generation of chips, they can still hold on until the next generation of chips comes out."
More importantly, J5 also snatched away some customer resources from the international chip giant Mobileye, which also made Nvidia feel a sense of crisis.
People familiar with the matter revealed that when Nvidia was initially planning the Orin chip, it actually defined a series of products for all-in-one machines, medium, high, and low computing power. After releasing the high-computing power Orin-X, it was discovered that the cost was too high and that the performance was a bit excessive for car manufacturers. In order to cope with the competition from Horizon J5, it launched the medium-computing power chip Orin-N.
However, it cannot be denied that Horizon J5 also has problems.
The J5 was originally designed for CNN (convolutional neural network) design, which does not require a high CPU. Its benchmark was also Mobileye's EyeQ5 chip with a computing power of only 30TOPS. In addition, Horizon's budget was insufficient in 2019 and it could not afford Arm's more expensive CPU IP core Cortex-A78. It had to settle for the second best and bought Arm's Cortex-A55 IP with slightly weaker performance and efficiency.
However, with the rapid update of autonomous driving algorithms, Transformer has gradually replaced the traditional CNN. More sensors such as lidar and multi-input cameras are widely used, and the intelligent driving system's requirements for CPU computing power are also rising.
In fact, it takes about three years for each chip to go from definition to mass production. Three years ago, the R&D team predicted market demand and defined the product, but three years later the market changed, resulting in chip sales falling short of expectations. This situation is common.
In the opinion of industry insiders,
with the help of J5, Horizon's automotive-grade intelligent driving chip research and development and mass production can take root, which is a milestone work for Horizon.
A viable chip company will always have to go through one or two generations of product iterations before it can truly mature. This is an industry with a long cycle and huge tuition fees.
Moreover, the development trajectory of Horizon J1, J2, J3, and J5 is largely a reflection of Yu Kai's product definition and business understanding. From the idealism of the first generation, the entanglement and self-denial in the middle period, to the ecological awareness and altruism in the later period, Yu Kai hopes that Chinese companies can have a place in the tide of changing the process of human civilization, and that bottom-level technology innovation companies like Microsoft and Intel will emerge.
In April this year, Horizon released Horizon SuperDrive, a new high-end full-scenario intelligent driving solution developed based on the Journey 6 flagship chip. The chip computing power reaches 560TOPS. A single Journey 6P can support full-stack computing tasks such as perception, planning and decision-making, and control. Its functions can cover mainstream intelligent driving scenarios such as high-speed NOA, urban NOA, and autonomous parking.
(Currently, Horizon's major ecosystem partners such as Qingzhou, Jianzhi, and Yihang are developing high-end solutions based on J6, which will be mass-produced and delivered as early as next year. For stories about the mass production and delivery of Horizon's ecosystem and car manufacturers, interested readers can add the author's WeChat Gru1993 to communicate)
In October 2023, Horizon Robotics launched the SuperDrive project, investing more than a thousand engineers. After about 10 months of development, SuperDrive conducted a public media exchange in August this year.
Yu Kai said, "Horizon Robotics has been building full-stack model houses on J2, J3, and J5. But this time, we not only want to create a benchmark, but also a "breakthrough" smart driving benchmark worldwide. Horizon Robotics is doing just this one thing, and is putting all its ammunition and firepower into this one thing (intelligent driving solution)."
Yu Kai has never given up his original intention of building China's "Wintel".
He believes that Horizon is an atypical chip company. It is the chip company that understands software algorithms best, the company that understands chips best among software algorithms, and the company that understands automotive regulations best among companies that do software algorithms + chips.
A former mid-level person at Horizon Robotics told Leifeng.com, "Once, after an important leader visited Horizon Robotics, he said that Yu Kai was a bit like Musk. Yu Kai was very pure and he really hoped to change the world like Musk."
Yu Kai also clearly knows that intelligent driving is Horizon's current base, but it is far from the end.
"If my coming to this world is no different from not coming to this world, then my life has no meaning. People are born to do things that change the world."
//
Recent Hot Articles
京公网安备 11010802033920号