“If people abandon Horizon in the future, it means that Horizon, as a company, didn’t evolve and was too weak in the evolution process of this ecology, and it should be abandoned by the times.” (Yu Kaiyu)
During the Beijing Auto Show in late April 2018, I published an article on “Jianyue Auto Review” titled “This Start-up Might Break the Chip Monopoly of Nvidia and Mobileye”. Looking at the information released by Horizon during the Journey 5 press conference on July 29th this year, my prediction from three years ago is becoming a reality.
There were many highlights in this conference, which have been reported by the media almost over the past week. Here, I want to discuss one point: Mobileye, often compared and benchmarked against in the past, was “ignored” by Horizon at this press conference. It can be said that Horizon no longer regards Mobileye as a competitor.
Horizon has the confidence to “ignore” Mobileye in terms of product parameters, application scenarios, and customer relations. It can be said that Horizon has surpassed Mobileye in multiple dimensions.
In fact, unconsciously, the benchmarking target of Horizon has been “upgraded” from Mobileye to Nvidia. The Journey 5 launched in this press conference is benchmarked against Nvidia’s Orin.
Objectively speaking, due to the much later establishment time, Horizon still has a gap to catch up with Nvidia in terms of technology accumulation and developer ecology. However, compared to Nvidia, Horizon has many unique advantages in the commercial field.
Looking at a higher level, in the process of the automotive industry’s transformation toward intelligence, Horizon has the opportunity to become the “MediaTek” of the automobile industry.
Embrace deep neural networks to meet the high computing power and low power consumption requirements of autonomous driving
In mid-May, “Artificial Intelligence Opportunities” published an article titled “Why Mobileye’s Most Glorious Five Years Are Five Years Behind,” which mentioned that Mobileye almost perfectly missed the era of artificial intelligence based on deep neural networks.
Not only EyeQ1, EyeQ2, and EyeQ3 almost have no deep neural network acceleration unit, but also, due to heavy historical baggage, subsequent products such as EyeQ4 cannot completely abandon the old technology architecture, and therefore cannot embrace deep neural networks.
In contrast, Horizon decided to “put the neural network inside the chip” from the beginning when it was established in 2015, and continuously optimized the chip architecture to improve computing power and reduce power consumption.One reason for this difference is that Mobileye’s founder, Shashu, received funding from Toyota in 1998, which tied Mobileye to the traditional automotive industry from the beginning. In contrast, Horizon’s founder, Yu Kai, previously served as the head of Baidu’s Deep Learning Institute, which allowed Horizon to consider the demand for deep learning algorithms on chips from the beginning.
Due to being too closely tied to the traditional automotive industry, Mobileye’s approach inevitably carries some of the industry’s typical characteristics, such as being slow and conservative. The most typical example is that its fifth product, EyeQ 5, only has a computing power of 24 TOPS after being redesigned in 2017.
Currently, the cameras on mass-produced autonomous vehicles are upgrading from 1-2 million pixels to 8 million pixels, and the number of sensors such as cameras, millimeter-wave radar, lidar, and infrared is also increasing. Meanwhile, as the trend towards software-hardware decoupling continues, the processing of perception data is shifting from “post-fusion” to “pre-fusion.” All of this makes the demand for computing power for the perception system grow exponentially, and EyeQ 5 may soon face the embarrassing situation of “insufficient computing power.”
In comparison, Horizon is adjusting its product design rapidly according to the demand for computing power in the autonomous driving industry. For example, the initial computing power planned for Journey 5 was 96 TOPS, but the version released this time achieved a computing power of 128 TOPS.
From the information in the table below, the computing power of Journey 5 is just as strong as Mobileye’s EyeQ 6, but its energy efficiency is superior to that of EyeQ 6. In addition, the mass production time of Journey 5 will come earlier than that of EyeQ 6.
Attachment: Comparison of main autonomous driving chip energy efficiency
Originally, Journey 5 was benchmarking Tesla’s FSD chip, but on June 12th, Yu Kai mentioned at the 2021 13th China Automotive Blue Book Summit that “we benchmarked Nvidia’s Orin, but we’re ahead of them.” It is worth noting that Journey 5 is based on 16nm technology, but its energy efficiency is close to that of Orin, which is based on the 7nm process. As for Journey 5’s “disadvantage” in computing power compared to Orin, this can be compensated for by increasing the number of chips.
Not just for perception computing, but for central computing platforms.The perception process is the most computationally intensive for autonomous driving vehicles. Therefore, perception computing has become a key area of focus for autonomous driving chip manufacturers. One result of this is that many autonomous driving chips, including EyeQ 4 and Xavier, can only be used for perception computing and not for prediction and planning.
Currently, decision-making computing is done by traditional manufacturers such as Infineon and NXP. However, as autonomous driving evolves from ASAS to L2+, AI processors are needed for prediction and planning computing. Therefore, an AI chip that can support both perception and prediction/planning computing has become the ideal choice.
The first AI chip to have both capabilities is Tesla’s FSD chip. From the beginning of the design of the Xpeng P5, the goal was to benchmark the FSD chip. In addition to its 128 TOPS of AI computing power, it also includes eight A55 CPU cores, making it suitable for decision-making computing.
At the Xpeng P5 launch event, “full-scenario” was a key term repeatedly mentioned. The term “full-scenario” can be interpreted from at least two dimensions: supporting the expansion of autonomous driving functions in various scenarios and supporting application scenarios beyond autonomous driving.
The expansion of autonomous driving function scenarios has been extensively discussed in recent media reports, with the biggest highlight being “complex urban scenarios.” However, the dimension of “supporting application scenarios beyond autonomous driving” has not received enough attention. Here, we will briefly explain.
In the years before 2019, Horizon Robotics emphasized in its PR that the Xpeng series is an ASIC (application-specific integrated circuit). However, since 2020, Horizon Robotics has rarely highlighted the concept of ASIC and has instead emphasized more on the concept of “general-purpose chips.” So, are the two contradictory? No.
Actually, specialized or general-purpose is a relative concept rather than an absolute one.
The ASIC that Horizon Robotics emphasized early on is relative to “general-purpose chips” such as NVIDIA’s GPU and Intel’s FPGA. The ASIC is designed for different AI computing requirements in specific scenarios and can achieve higher energy efficiency (lower power consumption, higher computing power). The “general-purpose” emphasized after 2020 is mainly relative to Mobileye’s products. The Xpeng series can be used in both autonomous driving and intelligent cabin scenarios, supporting interactive functions, while Mobileye’s products cannot.
As for the Xpeng P2, it can be used for both ADAS functions and cabin perception computing such as DMS and multimodal voice interaction.During the 2021 Shanghai Auto Show in April, Horizon unveiled its “Horizon Halo™️ In-Car and Out-of-Car Linkage Solution” equipped with its Journey 2 software. The solution combines three abilities including exterior vehicle perception, in-cabin driver monitoring system (DMS), and in-cabin voice recognition to enable the system to identify user intent and understand human behavior and habits based on semantic understanding of scenes, allowing for proactive involvement in driving decisions.
A typical linkage scenario is: in a scene where the car is stuck in traffic or sandwiched between large trucks that can be overtaken, the system proactively asks the driver if they want to pass, achieving a natural and smooth human-machine driving experience, while improving driver’s psychological safety.
Whether it is Journey 2 or Journey 3, computing power is limited and insufficient to support both autonomous driving and intelligent cabin computing at the same time. However, with Journey 5’s 128 TOPS computing power, coupled with 8 A55 CPU cores, two DSPs, and ISPs, integrating autonomous driving and intelligent interaction into the same central computing platform is not difficult. After all, the Tesla FSD chip with 72 TOPS of computing power has already achieved it.
By integrating the cabin and autonomous driving computing together, Journey 5 solves the problem of high-bandwidth data transmission and communication. However, solving the problem of data transmission is not the only purpose of Horizon’s central computing chip, and it is not even the most important purpose.
On June 11th, at the second day of the 13th China Automotive Blue Book Forum, Yu Kai mentioned in his speech titled “Enabling the Next Generation of Intelligent Automotive Chips and Building an Open Software Ecosystem for Intelligent Cars” that “we believe that autonomous driving must be centered on people rather than machines. Our goal is not to make machines more powerful, but to make people the masters of the car. The real value creation here is that human-machine interaction is inseparable from autonomous driving and must be integrated.”
“For example, in the event of an emergency safety accident, the status of the occupants in the vehicle, the status of the seat belts, and the sitting posture are closely related to the external impact, so the external perception of autonomous driving must be closely integrated with the in-cabin perception.”
“Therefore, we and several international Tier 1s believe that the future direction is the computing platform for the entire vehicle and the intelligent vehicle for the whole scene, rather than separating the cabin from autonomous driving.”
From the messages released by some automotive companies, the so-called “navigation-assisted driving” function will become the standard configuration of new cars in the next one or two years. Regardless of whether the automotive companies define this function as so-called “L3,” “L2.5,” or NOA, according to the latest definitions of the Chinese Ministry of Industry and Information Technology (MIIT) and the American SAE, the current level of autonomous driving capabilities still remains at L2, with human drivers still being the main body responsible for driving.However, in practice, car manufacturers cannot resist constantly “tempting” drivers to use L2 as L3. Therefore, in this human-machine co-driving stage, it is necessary to use DMS and other interactive technologies to ensure that drivers can take over in emergency situations. It can be said that a reliable interaction experience is a necessary prerequisite for the mass application of navigation-assisted driving.
After the press conference, Li Xingyu, Vice President of Horizon’s Ecological Development and Strategic Planning, said in a media interview:
“If you believe that the autonomous driving system cannot be decoupled from people and must be fully interacted with people, then you must combine human-machine interaction with autonomous driving closely. The Journey 5 is based on the whole product coverage of the car body from the beginning, which results from the design concept of putting users first.”
However, the logic of the above paragraphs is that “interaction serves autonomous driving”. In fact, interaction is much more than that.
Currently, the aura of autonomous driving is much more dazzling than that of intelligent cabins and interactions, and the latter often plays a subordinate role. However, if we can take a longer-term view, will the cabin still be in a subordinate position after L4 implementation?
In fact, in the long run, autonomous driving is not the point of differentiation competition between car manufacturers or models. The autonomous driving system is just a driver. When consumers take a taxi, do they care about the name of the driver and which driving school they graduated from? No. After L4 implementation, the real point of differentiation competition is the entertainment and interaction experience in the intelligent cabin.
In May, Zhongjin issued a report titled “Ten-Year Outlook: Changes in the Valuation System of the Automotive Industry,” which expressed a similar view.
In addition, concentrating the calculation of in-cabin interaction and autonomous driving on the central computing chip also reduces the demand for chips in vehicles. This not only reduces the difficulty of supply chain management but also greatly improves development efficiency.
Respecting the existing order of the automotive industry has brought about “man and machine”.
In the first few years after its establishment, Horizon had a goal of Horizon Inside, targeting Intel Inside during the PC era. However, at the Shanghai Auto Show in April 2019, Horizon interpreted its commercial philosophy of AI on Horizon to the outside world (being the empowerer of the lowest level in the AI era). At that time, Yu Kai said:
“Horizon is positioned as Tier 2. We do not make systems or hardware products, do not directly serve the end customers, the host factory or travel service providers, but empower the Tier 1 through collaboration and provide product and technical services. Horizon only produces weapons, does not fight, does not touch data, and does not make products.”
Since then, Horizon Inside has not been mentioned.After the Horizon Journey 5 press conference, during the media Q&A session, when asked about the transition from “Horizon Inside” to “AI on Horizon,” Li Xingyu’s response was:
“AI on Horizon embodies Horizon’s values in the entire ecosystem. Horizon is positioned as the bottom ‘foundation,’ supporting our customers and ecosystem partners. This is our universal ecological belief.”
“The core essence of our ecology is altruism and openness. Altruism means that we support others with our bottom position, whether it is with OEMs or software partners, Tier 1s. Openness is reflected in every dimension, from the bottom hardware reference and design to the toolchain to the reference algorithm, the mass-produced algorithm, and the corresponding series of AI tools.”
This is completely consistent with Yu Kai’s explanation from two years ago.
In short, Horizon’s ecological values are mainly reflected in respecting the interests of car manufacturers and Tier 1 companies.
Respecting the Interests of Car Manufacturers
Why no longer mention Horizon Inside?
What can be “Inside”? It’s the soul. But car manufacturers cannot accept their souls being controlled by others, which is an indisputable fact.
A person who has been responsible for purchasing in car manufacturers for many years said:
“Car manufacturers with self-respect cannot tolerate another influential logo posting on their own product, especially the logo of a supplier, which would weaken the influence of the mother brand. Car manufacturers almost regard the brand as life.”
“Actually, it is rare to see a second logo on the exterior of a car for a hundred years, unless the second logo has completely overwhelming advantages for the car manufacturer or the car manufacturer has already completely ‘laid down’.”
“Even if the supplier is very powerful, the car manufacturer also hopes that the supplier can be modest enough in their cooperation, not to be too overbearing, and do well what a parts supplier should do.”
Compared to the software and hardware integrated solutions provided by companies such as Mobileye, car manufacturers prefer a collaboration model where they are responsible for system integration, and suppliers only need to be responsible for component development and technical support. What Horizon provides is only a chip and “a good way to use the chip.” The “non-binding” strategy just guarantees the car manufacturer’s independent choice and “system integration rights,” thereby ensuring the car manufacturer’s leading role in project development and “sense of security” for the future.
For car manufacturers, if they continue to use Mobileye’s software and hardware integrated solutions or other vendors’ full-stack solutions, regardless of how late they were established, because the algorithm cannot be OTA and the user experience cannot be self-directed, they will become a “traditional car manufacturer” in terms of ability, and eventually become “helpless mediocrity.” It is precisely because of this that in the past few years, Mobileye has almost offended every ambitious car manufacturer.However, if they cooperate with more open companies such as NVIDIA, Qualcomm, or Horizon Robotics and develop their own algorithms, then even if a car manufacturer has been established for several decades, they can still become a “new force in car-making” in terms of their capabilities.
Driven by the determination to develop their own technology, companies such as Li Auto and Great Wall Motors have all started to reach out to Horizon Robotics. At first, they were afraid to let Mobileye know, fearing that Mobileye would threaten “cutting off supply” in anger. But now, with the production of the third phase of Horizon Robotics technologies and the release of the fifth phase, these car manufacturers are confident and no longer concerned about Mobileye’s “jealousy”.
An interesting phenomenon is that the first batch of vehicles equipped with Horizon Robotics chips are often best-selling models. For example, Changan UNI-T, the first model equipped with the third phase of Horizon Robotics technology for intelligent cabins, sold over 10,000 units per month last year. The Ideal ONE, the first model to have the fifth phase of Horizon Robotics, is also the sales champion of new energy SUVs in China.
In addition to Li Auto, SAIC, and Great Wall Motors, the first batch of potential customers for the fifth phase of Horizon Robotics also includes NIO, whose monthly sales are already among the top four in the new energy vehicle industry. Some people say that NIO’s models are not high-end, but from the chip manufacturer’s perspective, this is not the point – even not-so-high-end models can bring installation volume, driving mileage, and data.
During the Shanghai Auto Show in April, Horizon Robotics showcased its NOA algorithm function. Perhaps, in cooperation with NIO, both parties have the opportunity to “popularize” the NOA function?
Respect the interests of Tier 1
Although Horizon Robotics has been in frequent direct contact with car manufacturers and is basically qualified to be a Tier 1 company, they still position themselves as Tier 2. The most typical example is that they do not build computing platforms, but only provide hardware reference designs to Tier 1.
When asked “Why do you position yourselves as Tier 2?” by the media during a group interview, Horizon Robotics Co-founder and CTO Huang Chang said, “The automotive industry has existed for more than a century, and the pattern is very stable. Therefore, it cannot be easily overturned, and we must respect this industry.”
When asked by the media, “What are the capabilities of Tier 1, how are their capabilities expected to improve, and at what stage will they form a more stable ecosystem?” Horizon Robotics Vice President and General Manager of Autonomous Driving Yu Yinan explained, “Tier 1 is responsible for the final product function. They have strong capabilities in system integration and user experience, which are not Horizon Robotics’ strong suits. Horizon Robotics’ strengths lie more in some lower-level capabilities, which happen to be relatively scarce at this stage, thus attracting high market attention.””But in the future, as the underlying capabilities evolve to a certain stage, such as increasing computing power, controllable power consumption and cost, and mature algorithms, I believe people’s focus will shift more towards user experience, such as speed, acceleration, and handling of the car. And these happen to be the core competitiveness of Tier 1 companies.
In the past few years, the entire industry has been in a period of great change, and there may not have been a clear understanding of what the industry will be like in the future. But after two years of development, the industry’s view of the endgame has become clearer and clearer.
As the industry matures, and as demands for responsibility, rights, and quality become further developed, today’s tumultuous industrial relations may calm down in a few years and the industry may return to the delivery relationships of Tier 1, Tier 2, and OEMs in the past.
Talking about the phenomenon of chip manufacturers directly contacting host plants, Li Xingyu said: “Today’s Tier 2 is no longer a traditional Tier 2. We will not be completely isolated from the host plants, but will actively cooperate with upstream and downstream partners to create overall solutions. Therefore, the Tier 2 we emphasize is not a level on the industrial chain, but rather adheres to the commercial boundaries.”
Because they adhere to commercial boundaries, Horizon also has many friends in the Tier 1 camp. For example, Foryou Corporation stated that they have won more than a dozen production projects for the front-loading of models based on Horizon’s IMS 3.0 solution.
Horizon’s organizational structure includes a “Chief Ecology Officer,” which shows how highly they value the ecosystem. According to the aforementioned analysis, not overturning and respecting the existing order of the automobile industry is the core of Horizon’s ecological view – which is officially expressed as “multi-dimensional altruism.”
The “multi-dimensional altruism” ecological view allows Horizon to enjoy “people-oriented” benefits on the basis of the two major external dividends, “timing” and “location.”
In mid-June, at the China Automotive Blue Book Forum, when talking about ecological planning, Yu Kai said: “Ecology should follow the laws of nature, which means less deliberate efforts and more going with the flow. An ecology is not planned or designed, it evolves.
“In the future, if everyone abandons Horizon, it means that in the evolution of this ecology, Horizon as a company has not evolved and is weak, and should be abandoned by the times. Therefore, I think we should have less deliberate effort and take things as they come.”
“NVIDIA is already quite cautious of Horizon.”
When car companies choose suppliers, they not only look at product quality, but also at their service attitude. Therefore, under the influence of the multi-dimensional altruism values, even if Horizon’s initial product performance is not particularly outstanding compared to its competitors, or even slightly inferior, it will be much easier to secure orders.”During the first half of this year, a friend in the industry told me that Mobileye was asking about Horizon’s progress. According to Mobileye, a major domestic automaker signed a strategic agreement with them and then gave a large portion of the order to Horizon.
This is quite interesting: Horizon no longer sees Mobileye as a competitor, but Mobileye remains wary of Horizon.
In addition, in June of this year, a person closely connected to automakers involved with NVIDIA stated, “NVIDIA no longer sees Mobileye as a competitor. Their biggest competitors, aside from Huawei, are now Horizons.”
Globally, NVIDIA’s Xavier only won contracts for XPeng P7 and Lexus LS. However, it should be noted that XPeng P7 does not come standard with Xavier, only the top-of-the-line version does. Meanwhile, the Horizon Journey 3 is standard on the Ideal One and has two installations per vehicle. Although the combined computations of two Journey 3s are only one third of the power of Xavier, according to Horizon’s MPAS standard measurement, their AI performance is similar.
As we mentioned earlier, Horizon Journey 5 is positioned to compete with NVIDIA Orin, and although it was released later, it is expected to go into production in approximately the same timeframe.
A person who was once responsible for purchasing at an automaker said: “NVIDIA may well have advantages in computation and brand, but only high-end brands like NIO or IM require such computational parameters to improve their brand image. To some extent, the demand is more for PR purposes than actual needs. Besides cost advantages, Horizon’s high compatibility is unmatched by NVIDIA. If the supplier’s compatibility is insufficient, it could affect the automaker’s production timeline and thus their product launch.”
“Furthermore, Horizon’s existence will not threaten the automaker’s brand recognition. For mainstream automakers, this is an ideal solution.”
Many automakers have weak accumulation in intelligent technology, and how to make good use of chips and maximize chip performance has become a major technological challenge. Both NVIDIA and Horizon will provide customers with open toolchains and can partner with customers to open up algorithms.
However, from the publicly available information, only Mercedes-Benz has established a deep cooperation relationship with NVIDIA, while multiple automakers such as Changan, SAIC, and Great Wall have established deep cooperation with Horizon.
In conclusion, “the MediaTek moment in the automotive industry” is still up for grabs.According to the author, Huawei, providing full-stack solutions including chips, is the “MediaTek” of the new energy vehicle industry. However, Mobileye was the first company to play the “MediaTek” role in the new energy vehicle industry. Now, in addition to Huawei, there are waymo (with self-developed chips) operating as a technology solution provider, and chip supplier NVIDIA and Horizon Robotics that can cooperate with car manufacturers to develop algorithms.
The transition of Waymo from a technology solution provider is not smooth. Currently, the main contenders for the “MediaTek” position are Huawei, NVIDIA, and Horizon Robotics.
Objectively speaking, Huawei and NVIDIA are much stronger than Horizon Robotics in terms of capability. However, the Chinese automotive market is large enough, and not one company can dominate the chip field. Moreover, from the current progress, Horizon Robotics’ market strategy is easily accepted by top automakers, so there is a great opportunity for them to become the “MediaTek” of the new energy industry.
From the perspective of car manufacturers, the advantages of first-movers such as NVIDIA and Huawei lie in the stability and reliability of their products. However, as a newcomer, Horizon Robotics has already won many front-end mass production orders with opportunities to continuously refine technology and products and iterate rapidly in cooperation with customers. For Horizon Robotics, the gap with the top players can be “smoothed” out through accumulated time.
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.