Author: Jaz

The ideal goal of retiring after success and the possibility of continuing the explosive myth of L8’s debut.

In less than two months, Ideal has successively released the full-size SUV model L9 and stopped production of Ideal ONE. On September 30th, L8 and L7 were officially released.

September of this year is a complete delivery month for Ideal L9. Monthly sales exceeded 12,000 vehicles. As a substitute model for Ideal ONE, L8 is also likely to reach a sales level of over ten thousand in the next month.

At the time when Ideal became a player in the new energy vehicle industry, Horizon also climbed aboard Ideal L8 with Journey 5, becoming a star manufacturer of domestic automotive chips and launching the first shot of the international mass production of high computing power self-driving chips.

According to Autohome, the Journey 5 chip from Horizon has received batch projects from multiple automakers, and the L8 Pro version from Ideal is the first mass-produced vehicle.

Getting on Journey 5 with L8 Pro is a “revolutionary” milestone for Ideal and Horizon.

This time, Ideal offers the Ideal L8 Pro and Max versions for the first time in intelligent driving and intelligent space.

Among them, the Ideal AD Pro equipped with Ideal L8 Pro is the first intelligent driving system in the world equipped with Horizon Journey 5 chips.

In addition, it is equipped with one 8 million pixel front-view camera, five 2 million pixel surround-view cameras, and 4 2 million pixel panoramic cameras, providing consumers with an intelligent driving experience like an experienced driver.

For Horizon, the computing power of a single Journey 5 chip can reach 128TOPS, and the real computing performance can reach 1531 FPS, marking the era of mass production of domestically-made high computing power self-driving chips.

Looking at the international arena of high computing power self-driving chips, the single-chip computing power of 25TOPS of Mobileye EyeQ5 was released in 2021, and it is not yet common to see actual mass production. Although NVIDIA OrinX has a high computing power and many batch projects, it is not cost-effective and not leading in mass production speed.

Journey 5 chip is the third-generation product of the self-driving chip Journey series developed by Horizon, which was officially released in July last year. From research and development to flowing chips to now entering mass production, Horizon’s Journey 5 took less than four years.

It can be said that Horizon is creating a “mass production sample of on-board high computing power self-driving chips” in China.

“Strong Alliance” Among Old Friends

This is not the first time that the Horizon Journey series chips have climbed aboard Ideal.In fact, Horizon and Li Auto reached a deep cooperation in May last year when the 2020 Li ONE was upgraded to the 2021 version, and Li Auto replaced Mobileye EyeQ4 chip with Horizon Journey 3.

Interestingly, the intelligent driving system of Li Auto was implemented first on the 2021 version of Li ONE.

In February 2021, Li Auto delivered complete AEB functionality and NOA navigation-assisted driving.

After the OTA3.0 update, the navigation-assisted driving function was standard on the 2021 Li ONE, whereas the 2020 model only added “visual parking” function.

At the time, there were many different interpretations and speculations about the “chip replacement” of Li ONE.

Some believe that the performance of the chips is closely related. The Journey 3 chip is superior to Mobileye EyeQ4 in terms of calculation power and energy efficiency. Others believe that this is the inevitable choice for Li Auto to choose more open manufacturers.

These views are among the influencing factors, but as Li Auto and Horizon Journey 5 embarked on their second cooperation, it gradually became clear that the fundamental reason for Li Auto to choose Horizon was that Horizon meets both the criteria of “good chips” and “user-friendly chips”, which opens up a “mass production flywheel” for chip research and development and integration.

How to define a “good chip”?

Horizon co-founder and CTO Huang Chang once mentioned the concept of “hexagonal warrior”, which means that a chip is a complex system that needs to achieve the overall optimal through coordinated software and hardware systems.

Therefore, in the arena of vehicle computing chips, a good chip is not just about one punch, but about every punch.

According to Huang Chang, a “good chip can continuously attack and hit hard”, that is to say, making a chip is not only about emphasizing its calculation power, but also about high performance, safety, openness, and ease of use to achieve overall optimization.

Journey 5 is a typical case of Horizon’s “hexagonal warrior”.

During the development process, Journey 5 achieved overall optimization of software, hardware, and architecture. According to Synced, a media company in the AI industry, Journey 5 not only has strong computational power but also low energy consumption.

It is reported that a single Journey 5 chip has a maximum calculation power of 128 TOPS, a super-low energy consumption of 30W, a super-low latency of 60 milliseconds, and an energy efficiency ratio of 4.3TOP/W.

In addition, Journey 5 is the first AI chip in China that follows the ISO 26262 functional safety certification process and has been certified with ASIL-B, which verifies the functional safety of the chip and software toolchain.

Practice proves the truth. While Horizon produces good chips, it also considers making “user-friendly chips” for automakers.An insider from an intelligent driving platform manufacturer disclosed to Autohome that the industry is highly sensitive to the phenomenon of futures trading, and training in the lab and demos have little significance. This implies that in addition to the chip itself, the second focus of the industry is the quality and speed of intelligent driving mass production. In other words, although a good chip may exist in the laboratory or on paper, the real test lies in the actual use of the chip in production cars. “Frankly speaking, making a chip to light up is less than half the work, and there is much more work to be done before it goes into mass production,” said Huang Chang.

With this awareness in mind, the Horizon Journey series of chips has been refreshing the “production speed” of popular car models. A simple review reveals that the Horizon Journey 2 was once equipped with the Changan UNI-T, which sold over ten thousand units per month, and the 2021 version of the Ideal ONE equipped with the Horizon Journey 3 was also very popular, with peak monthly sales exceeding 13,000 units.

Now, the Horizon Journey 5 with L8Pro is also likely to become a popular car model. At the same time, the production speed of the Horizon Journey series has basically maintained a record of “within one year.” According to Autohome, it took only 8 months for the Horizon Journey 3 to be equipped in the Ideal ONE, while the Horizon Journey 5 in the Ideal L8 only took 6-8 months.

Mass production is not simply about installing the chip into the car. The production and testing of cars test the coupling degree of the chip and the car, the coordination ability between the manufacturer and the car company, and the personalized problems caused by various defined configurations. The more complex the problem, the more cooperation is needed. Horizon believes that only when the three elements of product, technology, and mass production are closely linked, can chips be quickly mass-produced and put into production vehicles. This is like a mass production flywheel. At the beginning, the manufacturer often needs to push the product and technology flywheel with great effort to get it spinning, but once the flywheel effect is activated, the flywheel will turn faster and faster, and enter the mass production track.

So far, the shipping volume of the Horizon Journey series of chips has exceeded 1.5 million pieces, and it has reached cooperation on more than 70 mass production projects of front-loading vehicles with over 20 car companies.

In the post-Moore’s Law era, change makers are needed to transform the rules of intelligent driving chips. Good chips and good usable chips not only meet Horizon’s own high requirements but can also stand on the international stage of autonomous driving chips and compete for leading positions. Currently, the only chip that might rival the Horizon Journey 5 is the NVIDIA Orin X. They are both pillars of high computing power chips, and according to fixed-point on-board collaboration projects, the Horizon Journey 5 and Orin X are currently in a level position. This trend is also reflected in Ideal L8, with its PRO version equipped with a single Horizon Journey 5 chip, and its MAX version equipped with dual Orin X chips. It is rumored that the high-cost performance solution offered by the Horizon Journey 5 makes its cost only half that of NVIDIA.The car models announced to adopt Nvidia Orin X chip currently include, but are not limited to, the Ideal One (L9), Ideal One PRO version (L8), XPeng G9, NIO ET7, NIO ES7, Jidu Auto, WM Motor, IM Auto, etc.

As for the car makers that have announced to adopt Horizon Journey 5, they include Ideal, FAW Hongqi, BYD, Self-Driving Home, SAIC, and others.

Compared with Orin X in terms of computing power, Horizon Journey 5 is not as outstanding. Currently, a single Orin X chip has a computing power of 254TOPS, while Horizon Journey 5 is at 128TOPS.

However, the reason why Horizon Journey 5 can be on par with Orin X is that, in the post-Moore’s Law era, Horizon Journey 5 is a rule changer in the autonomous driving chip industry.

In the post-Moore’s Law era, where the number of transistors on an integrated circuit doubles approximately every 18 months, has become slow, even approaching failure.

On autonomous driving chips, the improvement of computing power is slow, and computing power is gradually no longer the “sole criterion” for measuring chip performance.

During this process, Horizon Journey found more “incremental space” from the software layer, such as operating systems and compilers, when the hardware level of chip design and process approaches saturation.

In 2016, Horizon Journey proposed a new Moore’s Law, using the frame per second (FPS) that processes images per unit time as a measure of the real AI performance of autonomous driving.

In the field of autonomous driving, the car perception device transmits images frame by frame to the processor, and whoever can process the images faster and more accurately within the same unit time, with higher FPS, also has stronger actual computing power of autonomous driving chips.

Therefore, FPS to some extent also represents the level of autonomous driving technology.

Currently, Horizon Journey 5’s real AI performance has reached 1531FPS.

Under the typical classification model, the AI performance of Horizon Journey 5 is not far behind that of Orin X, but, in processing millions of pixels per frame, it is also higher than Orin X at 3.3 times.

In Horizon’s view, the three main factors that improve real AI performance are:

• First, chip hardware architecture determines the theoretical peak computing power;
• Second, software architecture, including compilers that can compile and disassemble algorithms, and improve the efficiency of algorithm utilization;
• Third, improving efficiency through algorithm optimization.

Correspondingly, the strength of Horizon Journey 5 depends on the bottom-level chip architecture.

Huang Chang said: “Our architecture design is not for general computing, but for the computing needs of robots, while Nvidia caters to general computing as a whole.”According to the report from Autohome, the CPU of Journey 5 adopts an 8-core ARM Cortex A55 core. The most critical AI operation unit is the Double Bayesian architecture BPU dedicated to high-level autonomous driving that is designed by Horizon Robotics, which enhances the efficiency of data parallel operation on a proprietary platform.

At the same time, Horizon Robotics has proposed an innovative concept of “Starting from the End and Combining of Software and Hardware”.

The so-called “Starting from the End” means determining the data indicators that the chip needs to achieve based on its specific purpose, while “Combining of Software and Hardware” means using the power of software to cooperate with hardware to optimize the whole.

For example, Horizon Robotics spent one year iteratively compiling compilers and algorithm technology to improve the performance of Journey 5 chips by another 20 percent.

It is reported that when Journey 5 was initially released, its strongest computing performance was 1283 FPS, and later, Horizon Robotics upgraded software architecture and optimized compilers, etc., which increased its maximum computing performance to 1531 FPS.

“Software has the characteristics of continuity and upgradability. Journey 5 has not yet reached its limit and still has room for further improvement,” said Huang Chang at an industry conference.

Positioned as Tier 2, Building an Open Platform and Ecology

Looking at it from the perspective of chip manufacturers, having well-polished products is enough. However, excellent companies need to “adopt a different perspective” and think from the perspective of car manufacturers.

One trend is that car manufacturers are becoming more and more cautious in selecting autonomous driving chip suppliers, which means that chip manufacturers need to provide more than just high-quality products.

On the one hand, with the arrival of the intelligent car era, consumers’ demand for intelligence is becoming increasingly apparent, while on the other hand, it is closely related to the cost and investment of car manufacturers.

The high customization cost and difficulty in landing have become two obstacles to the development of autonomous driving for car manufacturers.

It can be noticed that in recent years, car manufacturers have successively established autonomous driving laboratories for software development and even launched plans to research and develop autonomous driving-related chips.

However, it needs to be emphasized that developing L2+ level autonomous driving platform requires a team of 200-300 people.

At the same time, in order to do good work, one must first sharpen one’s tools. Car manufacturers also need a mature development platform to support autonomous driving development, otherwise, the development risks would be uncontrollable.

The characteristics of long cycle and high investment require car manufacturers to choose an open and reliable development platform and commercial ecology.

“The reason why Horizon Robotics can quickly mass-produce on cars is that Horizon Robotics provides support for several links.” One of these links is the AI tool chain, which connects intelligent algorithms and AI chips to empower car manufacturers to release the ultimate performance of chips.Horizon has developed a mature, open, and easy-to-use intelligent vehicle development platform, which includes AI chips and software stacks that encompass algorithms, basic operating systems, and development kits, as well as Horizon’s open tools and infrastructure, including the TGOW AI toolchain and the Eddie AI development platform.

These underlying tool services optimize the efficiency and quality of autonomous driving development.

According to Autohome, although any AI team can collect data, it often takes several months, which cannot meet the efficiency requirements of automakers.

The Horizon TGOW toolchain includes the three core capabilities of model quantization, quantization training, and compilation optimization and deployment, providing developers with end-to-end support from model acquisition to application deployment.

By utilizing automated universal algorithms, the model can be quickly deployed on chips, with high operating efficiency and minimal accuracy loss. Ultimately, it helps automakers complete their projects with fewer human and financial resources.

At the same time, relying on its AI toolchain and development platform, Horizon has embarked on an “open path” in its volume production practice with automakers.

Mobileye, in the past, bundled autonomous driving chips with visual perception algorithms for sales purposes. Although it reduced the cost of use for automakers, its closed nature has become increasingly pronounced in the trend of autonomous driving end-to-end development, making it difficult to meet the differentiated needs of automakers, resulting in its being dubbed a black box mode.

Therefore, in recent years, automakers such as Tesla, NIO, and LiAuto have also sought other solutions, with stories of “escaping” from Mobileye.

In contrast, Horizon embodies an open attribute. This openness is reflected in two aspects. On the one hand, Horizon’s scope of cooperation and service is highly open. Horizon is positioned as a Tier 2 provider, empowering not only automakers but also Tier 1 companies.

At present, Horizon’s ecological partners mainly fall into three categories: OEMs, automotive supply chain enterprises, and MaaS/TaaS, etc.

It can be seen that Horizon’s partners are not a chain-like structure but rather an ecological circle of friends with complementary strengths.

“Horizon’s Tier 2 positioning continuously polishes our tools and chips to improve production development speed and reduce high-level development difficulty. Also, I think the entire industry needs to work together to optimize the cost of development and delivery,” said Lv Peng.

This year, according to Autonews, Horizon has already collaborated with companies such as Yingchi Technology, Jinmai, and Tianzhun on hardware IDH partnerships, where the three companies will work with Horizon on technology research and product development, jointly providing automakers with autonomous driving domain controllers.On the other hand, the openness of Horizon’s series of products and tools has turned closed black-box delivery into white-box delivery.

From the perspective of mass production, Lv Peng summarized the fast growth of Horizon: “Through large-scale production, we integrate our experience into products and technologies, finally depositing a set of perfect development platforms, which can accelerate the development process while ensuring quality and saving a lot of development costs.”

From a macroscopic perspective, the growth of Horizon not only benefits from the explosive demand brought by the “new four modernizations” of electrification, networking, intelligence and sharing in the automobile industry, but also from various innovative measures that meet efficiency standards, satisfying the pursuit of efficiency and quality in the industry.

In this context, the efficiency improvement in chip mass production brought by Horizon will not only benefit the enterprise, but also promote the automotive industry to move towards efficiency that complies with long-term social development.

In the era of intelligence, computing power is the “electricity, water and coal” of the digital economy.

Horizon has a strict sense of tension and fast pace in controlling chip research and development and on-board progress. According to Autohome, the journey to surpass 1000TOPS in computing power has already begun with the development of the 6-chip and is expected to be put into mass production by 2024.

Horizon has always upheld the corporate philosophy of “Innovation is not about big fish eating small fish, it’s about fast fish eating slow fish.” As chips enter the stage of high-level computing power, Horizon is empowering automakers to become the “fast fish” in the innovation deepwater zone.

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.