"Horizon's intelligent computing technology, which combines software and hardware, will enable every vehicle to have environmental awareness, human-computer interaction and decision-making control capabilities. Horizon has laid out a full-stack technology that combines software and hardware, with software as the traction and hardware as the drive, to create a software and hardware combination intelligent driving system for all scenarios and all ecosystems, and implement the efficiency to the end to reach the end of intelligent driving. ”
Text丨Zhijia.com Wang Xin
Editor丨Langlang Mountain and Mingzhi Mountain
Is the battle of smart driving about to usher in the endgame?
Before the opening of the Beijing International Auto Show after a four-year absence, when will FSD enter China? What is a complete end-to-end? With or without a picture? Who is the right answer between pure vision and lidar?
Smart driving is going through the most lively and equally chaotic period in history.
On the afternoon of April 24, when the latest models of the world's automobile industry gathered in Beijing and were about to open the door, Yu Kai, CEO of Horizon Robotics, gave a speech in two hours at the National Speed Skating Hall next to the Bird's Nest in Beijing and released the industry's long-awaited Journey 6 series and SuperDrive equipped with Journey 6 series products, which is a high-end intelligent driving system that combines software and hardware.
It carries Yu Kai's ambition - to promote all smart cars to reach the endgame of intelligent driving.
He said: "Horizon's intelligent computing technology, which combines software and hardware, will enable every vehicle to have environmental awareness, human-computer interaction and decision-making control capabilities. Horizon has laid out a full-stack technology that combines software and hardware, with software as the traction and hardware as the drive, to create a software and hardware combination intelligent driving system for all scenarios and all ecosystems, and implement the efficiency to the end to reach the end of intelligent driving. ”
What is the endgame of intelligent driving?
It not only points to L3 and even L4 to adapt to the full scene of autonomous driving, but also not only to give every car on the road the ability to drive autonomously, but also intends to end the battle for the technical route of intelligent driving and the cost-effective entanglement of OEMs.
Can the Journey 6 Series and SuperDrive do it all?
Yu Kai, who bears the label of "scientist entrepreneurship", has been devoted to artificial intelligence for a long time, especially in deep learning, and is one of the earliest and most active research team members in the world.
With cutting-edge algorithms and architectures in hand, Yu Kai led Horizon to the door of the endgame of China's high-end intelligent driving.
Today, NOA in the urban area is ushering in the early stage of order reconstruction, bidding farewell to the era of pursuing the scale of Kaesong, and then entering the era of user sticky competition with the improvement of daily activity and practical experience as the core indicators.
A truly easy-to-use urban NOA solution not only pursues the acceleration of large-scale mass production, but also achieves rapid cost reduction in rounds of competition, and the equal rights of intelligent driving are in full swing.
Spring is the season of flowering.
01.
Why is end-to-end the best shortcut to urban NOA?
As high-speed NOA moves to urban NOA, the complexity of autonomous driving systems is increasing dramatically, and millions of lines of C++ code bring huge costs to the manual way of writing rules.
At this time, the perception module based entirely on artificial intelligence and neural networks will not have the confusion caused by the inefficiency caused by manually writing rules, so the current industry trend is towards end-to-end autonomous driving based on large models.
Although the market is full of expectations for urban NOAs, the current user experience has not yet reached the standard of "usable", let alone "easy to use" and "love to use", due to limitations such as poor driving experience and limited adaptation scenarios. Although technology is advancing, there is still a gap between the improvement of the user experience and the practical application.
Zhou Guang, CEO of Yuanrong Qixing, said at the Horizon Ecological Roundtable Forum during the Beijing Auto Show: "Last year, everyone was still talking about high-speed NOA, Transformer and BEV, but this year, the industry has become accustomed to urban NOA, and almost all of them are talking about mapless solutions and end-to-end. ”
The core of intelligent driving is returning to the essence of consumers. In Horizon's view, the intelligent driving system will experience usability (intelligent driving system 1.0 era), easy to use (intelligent driving system 2.0 era) to love to use (intelligent driving system 3.0 era).
The easy-to-use intelligent driving system is to provide an anthropomorphic driving experience, not only to ensure the physical safety of driving, but also to provide the driver with a sense of psychological security.
Yu Kai believes that in order to enter the intelligent driving system 3.0 and achieve the state of "love to use", while providing anthropomorphism, it is also necessary to realize the equal rights of science and technology, and achieve the inclusive popularization of high-end intelligent driving.
Horizon has also won a number of industry achievements in the end-to-end storm.
As early as 2016, Horizon Robotics took the lead in proposing the concept of end-to-end evolution of autonomous driving, and in 2022, it proposed the industry's leading end-to-end algorithm for autonomous driving perception, Sparse4D, and in 2023, the industry's first publicly published end-to-end autonomous driving UniAD with the participation of Horizon scholars won the best paper at CVPR 2023.
Sparse4D, an end-to-end algorithm for autonomous driving perception launched by Horizon, is a series of algorithms moving towards long-time sparsity 3D object detection, and is a completely end-to-end perception system. Due to the construction of a pure sparse fusion perception framework, Sparse4D can make the perception algorithm more efficient and accurate, and surpass the dense BEV algorithm in terms of perception effect and inference speed.
However, the ability to produce and research only in the laboratory is not the ultimate development destination of technology companies, and the vitality of technology lies in the support of a large number of market application needs, and research papers are not wasted.
02.
The combination of soft and hard can make the giant wheel rotate
For customer OEMs, it is important to be able to efficiently support the development and deployment of various intelligent driving applications with different positioning models, and the most important thing is to strike a balance between performance and cost-effectiveness.
Because of this, OEMs want to completely decouple from the perspective of efficiency and cost, and the more thorough, the better.
So what are the capabilities of the solution provider?
Achieving greater efficiency with dedicated chips specifically designed for autonomous driving scenarios is only the first step, and the more difficult thing is to build your own software system to support the continuous evolution and iteration of the user ecosystem.
At present, the integration of software and hardware is the "door of life and death" for autonomous driving companies.
Whether it is a software algorithm company or an intelligent driving computing solution company, the current focus of competition is how to achieve the ultimate system efficiency through software and hardware collaboration, and lead the mass production application. At present, some leading car companies, algorithm companies, and chip companies have begun to coordinate the layout of software and hardware collaboration.
In a way that combines breadth + depth, Horizon has built a rich and huge panorama of intelligent driving.
Starting from the breadth, the Journey 6 series has launched a total of six versions, including Journey 6B, Journey 6L, Journey 6E, Journey 6M, Journey 6H, and Journey 6P, with a computing power of 10+TOPS-560TOPS. From all-in-one machine to domain control, from active safety to full-scene intelligent driving, it can cover all levels of intelligent driving applications in an all-round way, and flexibly configure computing solutions for different intelligent driving scenarios.
Among them, the Journey 6P, as the flagship version, is oriented to the high-end solution of full-scene NOA, and the hardware system cost is less than 10,000 yuan according to the official introduction, which is 50% lower than that in the industry, and the Journey 6E supports the high-speed NOA hardware system cost of less than 5,000 yuan (based on the 7V3R system).
In the past, the core capabilities of Journey 2, Journey 3, and Journey 5 were all provided by Horizon's self-developed Bernoulli and Bayesian BPU architectures, as well as tool chains and software stacks.
The price war of automakers is full of gunsmoke, which is transmitted to all parties in the industrial chain, and they must jointly deal with it.
Horizon deliberately emphasized that in the generation of Journey 6, it has achieved "four cores in one", that is, a high degree of integration.
Journey 6 integrates BPU, MCU, CPU, GPU, and CPU, improving the cost performance of the system and reducing the difficulty of deployment.
The significance of high integration lies in the fact that it can help Tier1, OEMs, and algorithm partners to formulate a cost-effective solution more efficiently, and have more flexible choices for the configuration of models in different price ranges.
In addition to the joint appearance of 6 computing solutions, another highlight of the Journey 6 series is the recent announcement of its new generation of intelligent driving acceleration engine-BPU Nash architecture.
The BPU Nash architecture is designed for large-parameter transformers and large-scale interactive games, and is considered to be the best choice in the era of intelligent vehicles because of its innovative AI-assisted design, which greatly improves the programmability of the architecture by using AI-assisted design, and has a super-heterogeneous computing architecture, which can significantly enhance the diversity of computing power.
Horizon uses the Nash architecture as a foundational technology for the next 5-10 years.
Starting from the depth, SuperDrive, a full-scene intelligent driving solution launched by Horizon, is the answer to Yu Kai's thoughts on the endgame of intelligent driving.
SuperDrive can focus on the breakthrough of anthropomorphic experience, relying on the end-to-end perception architecture of dynamic, static, and OCC triple play.
Based on the end-to-end architecture of dynamic, static, and occupancy, the occlusion recall rate can be increased by 70%, the number of dynamic code lines can be reduced by 90%, and the network load can be reduced by 50%, effectively solving the problems of high latency, multiple rules, and heavy load in the current industry-aware architecture.
Horizon has also accumulated an end-to-end deep learning algorithm based on Monte Carlo tree search, which has greatly improved the perception and interactive game capabilities of the intelligent driving system in complex traffic environments.
Yu Kai believes that a more competitive intelligent driving system should be evaluated from three dimensions: scene traffic rate (coverage), traffic efficiency, and behavior anthropomorphism.
He said that the current intelligent driving scheme is more inclined to the scene traffic rate, and the performance in efficiency and behavior anthropomorphism is relatively poor. Progress on the latter two key capability indicators may require a fundamental breakthrough in the technology paradigm. The head player has realized this problem and has started to do a new round of layout.
To this end, based on the data-driven interactive game algorithm, SuperDrive can take into account the scene pass rate, traffic efficiency and behavior anthropomorphism in the road environment, and bring users a silky smooth experience in complex urban scenarios such as congestion confluence, intersection interaction - dynamic Driveline, courtesy to cyclists, congestion lane change, and urban roundabout.
At the same time, SuperDrive also has the ability to perceive light and high-precision maps, so as to quickly launch the NOA function of each city and achieve efficient city expansion.
What role does SuperDrive hope SuperDrive can play?
On the one hand, with software as the traction and hardware as the drive, we have created a software and hardware intelligent driving system for all scenarios and all ecosystems.
On the other hand, it provides high-end intelligent driving "model rooms" for partners' reference and provides more choices for the market.
Horizon's definition of SuperDrive is to focus on "human taste" and create an "easy-to-use" intelligent driving system 2.0. At present, Horizon Robotics has worked with JOYNEXT and ZeroBeam Technology to create a mass-produced SuperDrive domain controller solution.
On the eve of the large-scale outbreak of high-end intelligent driving, how did Horizon initiate "technological equality"?
With its large-scale mass production capacity, up to now, the Horizon Journey family has shipped more than 5 million pieces, reached pre-assembly mass production with more than 30 car companies around the world, and cooperated with more than 110 mass production models.
According to Horizon, in terms of cooperative car companies, Horizon's SuperDrive solution has reached cooperation with a number of car companies and top Tier 1, and it is expected that by the fourth quarter of this year, there will be a standard mass production plan launched, and by the third quarter of next year, the first model equipped with SuperDrive solution will also enter the delivery period.
Horizon's worldview and methodology have always pursued the concept of "combining software and hardware", adhered to Tier 2 positioning, and adhered to all-dimensional altruism.
Horizon's business model is like Intel + Windows in the PC era, empowered by computing solutions + tool chains combined with software model rooms. It doesn't make hardware or software — it's similar to the Wintel alliance between Windows and Intel, where hardware and OS are not separated, and the system operating system is built on its own hardware.
This conference can be said to be a milestone in Horizon's technological accumulation in the past 9 years.
The appearance of the full series of Journey 6 proves the acceleration effect of hardware architecture innovation on software, and SuperDrive not only accelerates the implementation of high-end intelligent driving, but also promotes the popularization of low-end intelligent driving.
At the end of the press conference, Wang Chuanfu, chairman of BYD, personally came to the scene to "stand up" for the horizon and affirmed Yu Kai's choice: "The horizon has done the right thing. ”
03.
A spark can start a prairie fire
The word "all-dimensional altruism" is a word that Yu Kai has repeatedly mentioned at many field horizon press conferences in recent years.
This is the sum of the experience extracted from the transformation of Horizon from building an industrial chain to an ecological chain.
Zhang Yufeng, Vice President of Horizon Robotics and President of the Intelligent Vehicle Division, said in the roundtable forum of the auto show: "We continue to do a good job in the computing foundation of the industry, and then we transform the ecological concept from full-dimensional altruism to public altruism, and strengthen our entire ecosystem from the perspective of user and consumer experience, and work together to create a cost-effective intelligent driving solution faster and more economically." ”
At the Horizon Intelligent Driving Product Launch Conference, Yu Kai announced a number of cooperative companies in the Journey 6 series, including: SAIC, Volkswagen Group, BYD, Li Auto, GAC Group, Deep Blue Automobile, BAIC Group, Chery Automobile, Xingtu Automobile, VOYAH Automobile, and a number of Tier 1, software and hardware partners.
It is reported that the Journey 6 series will start the delivery of the first front-mounted production model in 2024, and is expected to achieve mass production delivery of more than 10 models in 2025.
As one of the first batch of intelligent driving startups in China, Horizon has witnessed the rise and prosperity of China's intelligent driving industry.
According to the data of the Gaogong Intelligent Vehicle Research Institute, in 2023, Horizon will rank second in the market share of China's standard NOA domain control and standard forward-looking all-in-one computing solutions, with a market share of 35.5% and 23.7%, respectively. The first place in the two markets is Nvidia and Mobileye, with market shares of 48.9% and 29.4%.
Horizon's business continues to expand.
According to the prospectus, Horizon has obtained a large number of orders for non-mass-produced models, and as of the end of 2023, the models that have not yet achieved mass production account for more than 50% of all models that have been fixed. This indicates that Horizon is expected to further expand its market share in the future.
On the day that Horizon submitted the prospectus, Yu Kai wrote in the circle of friends:
“This is not the end. It is not even the beginning of the end. But it is,perhaps,the end of the beginning.” （这不是结束。 这甚至不是结束的开始。 但这也许是开始的结束）
The intelligent hardware industry has a long cycle, and the autonomous driving industry is changing rapidly, and to find certainty in uncertainty, we must first find the anchor that guides the emergence of certainty.
There will always be better products in the market, and for every company committed to driving the world's progress with technology, the best rhythm is not too backward, but not too advanced, just in sync with the times. "It's like Microsoft, not being led by competition. Yu Kai said.
The horizon has finally leaped over the horizon.